# ANGER AND INTERPERSONAL AGGRESSION

EDITED BY : Nelly Alia-Klein, Annegret L. Falkner, Gabriela Gan, Klaus A. Miczek, Aki Takahashi and Rosa Maria Martins De Almeida PUBLISHED IN : Frontiers in Behavioral Neuroscience

#### Frontiers eBook Copyright Statement

The copyright in the text of individual articles in this eBook is the property of their respective authors or their respective institutions or funders. The copyright in graphics and images within each article may be subject to copyright of other parties. In both cases this is subject to a license granted to Frontiers. The compilation of articles constituting this eBook is the property of Frontiers.

Each article within this eBook, and the eBook itself, are published under the most recent version of the Creative Commons CC-BY licence. The version current at the date of publication of this eBook is CC-BY 4.0. If the CC-BY licence is updated, the licence granted by Frontiers is automatically updated to the new version.

When exercising any right under the CC-BY licence, Frontiers must be attributed as the original publisher of the article or eBook, as applicable.

Authors have the responsibility of ensuring that any graphics or other materials which are the property of others may be included in the CC-BY licence, but this should be checked before relying on the CC-BY licence to reproduce those materials. Any copyright notices relating to those materials must be complied with.

Copyright and source acknowledgement notices may not be removed and must be displayed in any copy, derivative work or partial copy which includes the elements in question.

All copyright, and all rights therein, are protected by national and international copyright laws. The above represents a summary only. For further information please read Frontiers' Conditions for Website Use and Copyright Statement, and the applicable CC-BY licence.

ISSN 1664-8714 ISBN 978-2-88963-904-5 DOI 10.3389/978-2-88963-904-5

#### About Frontiers

Frontiers is more than just an open-access publisher of scholarly articles: it is a pioneering approach to the world of academia, radically improving the way scholarly research is managed. The grand vision of Frontiers is a world where all people have an equal opportunity to seek, share and generate knowledge. Frontiers provides immediate and permanent online open access to all its publications, but this alone is not enough to realize our grand goals.

#### Frontiers Journal Series

The Frontiers Journal Series is a multi-tier and interdisciplinary set of open-access, online journals, promising a paradigm shift from the current review, selection and dissemination processes in academic publishing. All Frontiers journals are driven by researchers for researchers; therefore, they constitute a service to the scholarly community. At the same time, the Frontiers Journal Series operates on a revolutionary invention, the tiered publishing system, initially addressing specific communities of scholars, and gradually climbing up to broader public understanding, thus serving the interests of the lay society, too.

### Dedication to Quality

Each Frontiers article is a landmark of the highest quality, thanks to genuinely collaborative interactions between authors and review editors, who include some of the world's best academicians. Research must be certified by peers before entering a stream of knowledge that may eventually reach the public - and shape society; therefore, Frontiers only applies the most rigorous and unbiased reviews.

Frontiers revolutionizes research publishing by freely delivering the most outstanding research, evaluated with no bias from both the academic and social point of view. By applying the most advanced information technologies, Frontiers is catapulting scholarly publishing into a new generation.

#### What are Frontiers Research Topics?

Frontiers Research Topics are very popular trademarks of the Frontiers Journals Series: they are collections of at least ten articles, all centered on a particular subject. With their unique mix of varied contributions from Original Research to Review Articles, Frontiers Research Topics unify the most influential researchers, the latest key findings and historical advances in a hot research area! Find out more on how to host your own Frontiers Research Topic or contribute to one as an author by contacting the Frontiers Editorial Office: researchtopics@frontiersin.org

# ANGER AND INTERPERSONAL AGGRESSION

Topic Editors:

Nelly Alia-Klein, Icahn School of Medicine at Mount Sinai, United States Annegret L. Falkner, Princeton University, United States Gabriela Gan, University of Heidelberg, Germany Klaus A. Miczek, Tufts University, United States Aki Takahashi, University of Tsukuba, Japan Rosa Maria Martins De Almeida, Federal University of Rio Grande do Sul, Brazil

Citation: Alia-Klein, N., Falkner, A. L., Gan, G., Miczek, K. A., Takahashi, A., De Almeida, R. M. M., eds. (2020). Anger and Interpersonal Aggression. Lausanne: Frontiers Media SA. doi: 10.3389/978-2-88963-904-5

# Table of Contents

*06 Brains in Competition: Improved Cognitive Performance and Inter-Brain Coupling by Hyperscanning Paradigm With Functional Near-Infrared Spectroscopy*

Michela Balconi and Maria E. Vanutelli


Gadi Gilam, Adi Maron-Katz, Efrat Kliper, Tamar Lin, Eyal Fruchter, Ron Shamir and Talma Hendler


José Manuel García-Fernández


Helena L. Rohlf, Anna K. Holl, Fabian Kirsch, Barbara Krahé and Birgit Elsner

*102 Exogenous Testosterone Enhances the Reactivity to Social Provocation in Males*

Lisa Wagels, Mikhail Votinov, Thilo Kellermann, Albrecht Eisert, Cordian Beyer and Ute Habel

*113 Tantrums, Emotion Reactions and Their EEG Correlates in Childhood Benign Rolandic Epilepsy vs. Complex Partial Seizures: Exploratory Observations*

Michael Potegal, Elena H. Drewel and John T. MacDonald

*123 Aggression, Social Stress, and the Immune System in Humans and Animal Models*

Aki Takahashi, Meghan E. Flanigan, Bruce S. McEwen and Scott J. Russo


Marjolein M. J. van Donkelaar, Martine Hoogman, Irene Pappa, Henning Tiemeier, Jan K. Buitelaar, Barbara Franke and Janita Bralten

*167 Anger Modulates Influence Hierarchies Within and Between Emotional Reactivity and Regulation Networks*

Yael Jacob, Gadi Gilam, Tamar Lin, Gal Raz and Talma Hendler


*228 Differential Roles of the Two Raphe Nuclei in Amiable Social Behavior and Aggression – An Optogenetic Study*

Diána Balázsfi, Dóra Zelena, Kornél Demeter, Christina Miskolczi, Zoltán K. Varga, Ádám Nagyváradi, Gábor Nyíri, Csaba Cserép, Mária Baranyi, Beáta Sperlágh and József Haller

*244 The COMT Val158Met Polymorphism and Reaction to a Transgression: Findings of Genetic Associations in Both Chinese and German Samples*

Cornelia Sindermann, Ruixue Luo, Yingying Zhang, Keith M. Kendrick, Benjamin Becker and Christian Montag

### *253 The Urge to Fight: Persistent Escalation by Alcohol and Role of NMDA Receptors in Mice*

Herbert E. Covington III, Emily L. Newman, Steven Tran, Lena Walton, Walae Hayek, Michael Z. Leonard, Joseph F. DeBold and Klaus A. Miczek

*267 The Role of Estrogen Receptor* b *(ER*b*) in the Establishment of Hierarchical Social Relationships in Male Mice* Mariko Nakata, Anders Ågmo, Shoko Sagoshi and Sonoko Ogawa

# Brains in Competition: Improved Cognitive Performance and Inter-Brain Coupling by Hyperscanning Paradigm with Functional Near-Infrared Spectroscopy

#### Michela Balconi\* and Maria E. Vanutelli

Research Unit in Affective and Social Neuroscience, Department of Psychology, Catholic University of Milan, Milan, Italy

Hyperscanning brain paradigm was applied to competitive task for couples of subjects. Functional Near-Infrared Spectroscopy (fNIRS) and cognitive performance were considered to test inter-brain and cognitive strategy similarities between subjects (14 couples) during a joint-action. We supposed increased brain-to-brain coupling and improved cognitive outcomes due to joint-action and the competition. As supposed, the direct interaction between the subjects and the observed external feedback of their performance (an experimentally induced fictitious feedback) affected the cognitive performance with decreased Error Rates (ERs), and Response Times (RTs). In addition, fNIRS measure (oxyhemoglobin, O2Hb) revealed an increased brain activity in the prefrontal cortex (PFC) in post-feedback more than pre-feedback condition. Moreover, a higher inter-brain similarity was found for the couples during the task, with higher matched brain response in post-feedback condition than pre-feedback. Finally, a significant increased prefrontal brain lateralization effect was observed for the right hemisphere. Indeed the right PFC was more responsive with similar modalities within the couple during the post-feedback condition. The joined-task and competitive context was adduced to explain these cognitive performance improving, synergic brain responsiveness within the couples and lateralization effects (negative emotions).

### Edited by:

Nelly Alia-Klein, Icahn School of Medicine at Mount Sinai, United States

#### Reviewed by:

Christos Frantzidis, Aristotle University of Thessaloniki, Greece Anna Song Huang, Icahn School of Medicine at Mount Sinai, United States

#### \*Correspondence:

Michela Balconi michela.balconi@unicatt.it

Received: 05 April 2017 Accepted: 17 August 2017 Published: 31 August 2017

#### Citation:

Balconi M and Vanutelli ME (2017) Brains in Competition: Improved Cognitive Performance and Inter-Brain Coupling by Hyperscanning Paradigm with Functional Near-Infrared Spectroscopy. Front. Behav. Neurosci. 11:163. doi: 10.3389/fnbeh.2017.00163 Keywords: hyperscanning, competition, emotion, cognition, fNIRS

### INTRODUCTION

Competition essentially implies a social dynamic that requires a comparison between two or more subjects during an interpersonal performance. Previous research suggested a crucial role of competitive social interactions in achieving accurate self-representation of our social position. Conversely, it was found that social perception affects performance during situations that require to compare our own behavior with that of others (Munafò et al., 2005). That is, the analysis of our social role in competition may influence the cognitive performance by improving or decreasing the actual outcomes (Munafò et al., 2005).

About the brain contribution, it was observed that an extended neural network, including limbic areas, the prefrontal cortex (PFC) and striatal structures, may represent the behavioral cognitive and emotional and correlates of social interactions, respectively (Levitan et al., 2000). Preliminary evidence in support of this neural mechanism of the social brain comes from previous studies exploring the structures and functions of brain areas associated with social representation, social ranking and self-efficacy. Specifically, both dorsal (DLPFC) and ventral (VLPFC) portions of the lateral PFC are generally involved in response to social status inference and interpersonal tasks (Chiao et al., 2009b; Balconi and Pagani, 2014, 2015). The activation of DLPFC and VLPFC during social interactions probably represents the recruitment of brain areas that apply top-down control over some processes, such as emotional behavior in response to social demands (Marsh et al., 2009; Balconi and Vanutelli, 2016). Indeed these brain areas are generally associated with socio-emotional regulation and behavioral inhibitory mechanisms.

Recent research on cooperation/competition also showed enhanced cortico-cortical communication and interconnections between these prefrontal areas. For example, the effects of competitive tasks in more than one brain was recently explored (Decety et al., 2004; Liu et al., 2015; Cui et al., 2016). In addition, some studies confirmed that the social context jointly affects subjects' reactions to their environment and consequently their brain activity. For example it was noted that one's own action planning is facilitated during cooperation since others' actions are joined with our actions, in opposition to competitive conditions (Knoblich and Jordan, 2003; Sebanz et al., 2003).

Therefore the cortical activity is also modulated based on the different forms of social interaction, since competitive or cooperative situations are qualitatively distinct contexts. Indeed, it should be noted that cooperation and competition are two basic types of interpersonal interaction (Decety et al., 2004). That is, based on the interactive condition (cooperation vs. competition) people may either facilitate or hinder the goals of others. Specifically competition, as a social-evaluative phenomenon, can increase the amount of cognitive resources beyond what is needed to simply execute the task demands. In particular, the cognitive effort could be increased in competition when subjects have potentially contrasting goals (De Cremer and Stouten, 2003; Decety et al., 2004). We previously focused on cooperation with specific measures (EEG and neuroimaging near-infrared spectroscopy (NIRS); Balconi and Vanutelli, 2017a,b) and then we applied EEG and Functional NIRS (fNIRS) to study competition.

It was also found that competition may improve the effective cognitive performance and the self-perception of higher social position (Goldman et al., 1977). The higher demand may explain how competition affects performance (with an ''improving effect'') and brain responsiveness due to the attendant modifications of the subject's mental condition and underlying neural activities (Rietschel et al., 2011). More specifically, the self-perception during competition may affect the subjective internal judgment and manifest as an increased cerebral responsiveness in those areas related to competitive conditions, and positively affect the performance outcomes.

However, whereas the available previous results indicate that social exchanges involve a specific network of cortical areas, further analysis is required to clarify the specific contribution of the brain structures in different social conditions, i.e., when subjects compete toward a personal goal during a joint action. Second, the presence of a real interlocutor may affect the interbrain responsiveness and cognitive outcomes, as suggested by hyperscanning research (Konvalinka and Roepstorff, 2012). In a recent study Cui et al. (2016) have measured the prefrontal activation during cooperative and competitive tasks by using NIRS. Dyads of participants were asked to press two keys either simultaneously (to obtain synchronized action in cooperative condition), or as fast as possible to obtain a better result than their partner during competitive condition. The participants showed increased inter-brain synchronization in the right superior frontal areas during cooperation, but not competition: such result emerged because of the necessity to model others' behavior during a cooperative task. It should also be considered that the increase in cortico-cortical communication was high and significant, and involved heightened responses between all non-motor areas with strategy planning regions (such as prefrontal areas).

Third, it should be considered whether and how an increase in brain activity and cognitive performance is specifically promoted by an external feedback which is able to manipulate the cognitive performance though self-other evaluation. In fact no previous research has considered the social environment and the cognitive outcomes by using a direct competitive task. Generally, previous studies implied only single subjects and their isolated performance in abstract social tasks, since they did not include paired joint actions and interactive tasks. In other cases research explored the response in asynchronous conditions (subsequential response by the participants) by two or more subjects interacting each other (Boone et al., 1999; Decety et al., 2004). In this regard, the hyperscanning approach introduces an innovative perspective to explore two interacting brains (Holper et al., 2012; Konvalinka and Roepstorff, 2012). However, when an hyperscanning paradigm was used, it was applied only in response to cognitive performance without a specific interactive feedback (Saito et al., 2010; Dommer et al., 2012; Cui et al., 2016).

Therefore, to summarize, compared to previous research, two relevant aspects were underestimated and deserve to be considered to evaluate inter-subjective brain activity and the cognitive performance during competition: the presence of a dual interaction, and the feedback furnished by the social context to (fictitiously) represent the effectiveness of the joint action. That is, the effect of an external feedback (positive or negative feedback about the competitive performance) on the interbrain responses and cognitive performance was not adequately considered. An external feedback is supposed to modify the selfrepresentation, the effective cognitive outcomes and the brain responsiveness to social contexts (Montague et al., 2002). In the present study the performance was manipulated in a visà-vis competitive situation which stressed the subjects' ability to win and to perform better than the partner. Compared to other studies (Zink et al., 2008), we planned a more ecological and realistic scenario where subjects were asked to directly compare their outcomes with the other partner by monitoring their performance. Specifically this request underlined the necessity to increase subject's effectiveness during the task (''your performance is better than. . .''). In this regard we

formulated a clear and reinforced social condition based on cognitive skill during a dyadic interaction. Second, brain-tobrain coupling effect induced by the competitive task has to be explored, by using an adequate hyperscanning paradigm, which is able to reveal the common strategies applied by the participants to obtain a better performance and the effects of this synergic planning. To test these double effects, fNIRS was applied to acquire subjects' brain response during a task performed simultaneously in paired subjects. Classical neuroimaging approach (i.e., functional Magnetic Resonance, fMRI) was not able to exhaustively show the social nature of the inter-personal processes since the temporal course of such activation was scarcely addressed. fNIRS measure has a resolution which is considered high enough for monitoring event-related fNIRS responses (Elwell et al., 1993; Montague et al., 2002; Decety et al., 2004; Dötsch and Schubö, 2015). More importantly, fNIRS proved to be much more suitable for ecological hyperscanning applications since it imposes significantly milder physical burdens than other techniques such as fMRI, it is not noisy or uncomfortable, and is robust to exogenous noise thus permitting interactive contexts (Balconi and Molteni, 2015).

Therefore, based on our hypotheses, the artificially increased performance during competition may effectively modulate the behavioral performance in social contexts, with improved outcomes mainly after receiving the feedback. Therefore, a consistent better performance should be found for the post-feedback condition (that is in the case of perception of improved outcomes), as a result of a higher reinforcing situation. Second, the cortical effect of these social and cognitive representation processes are hypothesized to be supported by the PFC (Hall et al., 2005; Chiao et al., 2009a; Balconi and Pagani, 2014, 2015), with significant higher responsiveness of the PFC mainly after the feedback. Third, we intended to study inter-brain activity in competitive conditions and in relationship with the positive (improved) feedback. Indeed we expected a higher brainto-brain coupling induced by the feedback, which may induce a more synergic activity between the subjects.

### MATERIALS AND METHODS

### Subjects

Fourteen couples of subjects (28 subjects, all undergraduate students: M = 23.78, SD = 1.98 years old) were recruited for the present research. Each couple was composed by two players of the same gender matched for age. The subjects were all right-handed, with normal or corrected-to-normal visual acuity. To exclude history of psychopathology Beck Depression Inventory (BDI-II, Beck et al., 1996) was administered to the participants or immediate family. Moreover State-Trait-Anxiety-Inventory (STAI, Spielberger et al., 1970) was submitted in the post-experimental session. No neurological or psychiatric pathologies were revealed. The research was approved by the local ethics committee of the Department of Psychology, Catholic University of Milan. The subjects gave informed written consent to participate in the study in accordance with the 1964 Helsinki Declaration and its later amendments or comparable ethical standards. No payment was provided for subjects' performance.

### Procedure

Subjects were seated with monitor in front of them (positioned approximately 60 cm) in a moderately darkened room. Participants were seated side-by-side and separated by a black screen in order to not seeing each other. They performed a cognitive task for sustained selective attention (modified version of Balconi and Pagani, 2014).

### Task

Subjects were taught that specific attentional measures were registered to evaluate the subjective skills. They were required to recognize target stimuli from non-targets, based on four different combinations of shape and color: circles or triangles, green or blue. The target remained on the video until subjects were able to memorize it. Then, stimuli were displayed one after another. Size and color features changed every experimental block, each composed by 25 trials. Subjects were asked to press a left/right button after each stimulus to decide for target/non target. Each stimulus was displayed on the screen for 500 ms (300 ms inter-stimulus interval, ISI), and each trial was constituted by three stimuli. After each trial a feedback appeared on the screen in the form of two up-arrows (better performance than the competitor); a dash (comparable performance); or two down-arrows (worse performance). The feedback lasted 5000 ms, followed by an inter-trial interval (ITI) of 5000 ms. The task was subdivided in two sub-sessions: the first without a specific feedback to subject's performance (four blocks of stimuli before the feedback, for a total of 100 trials); the second preceded by the feedback about the performance (four blocks of stimuli with the feedback, for a total of 100 trials; **Figure 1**).

To increase subjects' intrinsic motivation, they were told that accuracy, number of errors (Error Rate, ER) and response times (RTs) were usually used to screen future professional career success in term of teamwork abilities. Moreover, participants were told that the final cognitive outcome was based on the ability to produce a better performance than the competitor, in order to highly stress the competitive nature of the task.

Halfway, participants received also a synthetic feedback of their cognitive performance. Both feedbacks (both trial-related and general feedback) were prearranged (without the awareness of the participants), and participants were told that their outcome was ''well above (score with 91% in terms of speed, and 92% in terms of accuracy)''. In addition they were pressed to maintain their higher performance level during the course of the task (''The measures recorded till now reveal that your performance is very good. Your response profile is well superior to your competitor's one. If you want to win, keep going like this in the following part''). Trial feedbacks constantly reinforced them about their high and competitive performance by presenting the up-arrows (70% of the cases) and the dash or the down-arrows (30% of the cases, and they were mainly positioned at the beginning of the task), to make the outcome more credible and plausible.

Finally, after each block, subjects were required to rank their outcome and perceived self-efficacy on a 7-point Likert scale (1 = most decreased performance; 7 = most improved performance). As reported in a post-experimental phase, the participants were strongly engaged in the social and ranking process. Participants were also requested to self-report their degree of trust based on the external feedback. They showed very

emitters were placed on positions FC3-FC4 (purple, number 1 and 2) and F1-F2 (number 3 and 4), while detectors were placed on FC1-FC2 (blue, number 3 and 4) and F3-F4 (number 1 and 2). Resulting channels are displayed in pink color.

high trust (92%) and a self-represented relevance of the task for their social position (94%).

#### Performance Scoring

The RTs (ms) measures were registered from the stimulus onset, and ERs were calculated as the total number of incorrect detections out of the total trial, for each experimental category (therefore higher values represented higher number of incorrect responses).

#### fNIRS

fNIRS recordings were conducted with NIRScout System (NIRx Medical Technologies, LLC. Los Angeles, CA, USA) with an 8-channel array of optodes placed on the prefrontal area (four light sources/emitters and four detectors). Emitters were placed over FC3-FC4 and F1-F2 positions, while detectors were placed on FC1-FC2 and F3-F4 positions (**Figure 2**). Emitter-detector distance was maintained at 30 mm for contiguous optodes and near-infrared light of two wavelengths (760 and 850 nm) were considered. According to the international 10/5 system, NIRS optodes were placed on the subject's head using a NIRS-EEG compatible cup. Changes in the concentration of oxygenated (O2Hb) and deoxygenated hemoglobin (HHb) from a 120 s resting baseline were acquired. By using NIRStar Acquisition Software, Signals obtained from the eight NIRS channels were acquired (sampling rate of 6.25 Hz), then transformed in values for the changes in the concentration of oxygenated and deoxygenated HHb in each channel (scaled in mmol∗mm).

The raw O2Hb and HHb data from each channel were digitally band-pass filtered at 0.01–0.3 Hz. Then, the mean concentration of each channel was calculated (from the trial onset for the following 5 s) as the average across trials. According to the mean concentrations in the time series, the effect size was calculated for each channel and participant in every experimental condition. The Cohen's z effect sizes were obtained as the difference of the means of the baseline and trial divided by the standard deviation (SD) of the baseline, as reported in the formula: d = (m1 − m2)/s. m1 and m2 were the mean concentration values during baseline and trial, respectively, and s the SD of the baseline. The baseline was calculated considering the 5 s period immediately before the trial beginning. Then, in order to increase the signal-to-noise ratio, the effect sizes obtained from the eight channels were averaged. Although NIRS raw data were originally relative values and for his reason they could not be directly averaged across experimental conditions (subjects or channels), effect sizes normalized data could be averaged regardless of the unit since the effect size is not affected by differential pathlength factor (DPF; Schroeter et al., 2003; Matsuda and Hiraki, 2006; Shimada and Hiraki, 2006).

### Data Analysis

Three levels of analyses were performed for behavioral (ER; RTs) and neurophysiological (fNIRS, O2Hb measures) measures.

For the first level of analysis, a repeated measure ANOVA with one factor (Condition, Cond: pre vs. post feedback) was applied to ERs and RTs data. A second ANOVA was applied to O2Hb dependent measure, with repeated factors Cond and Lateralization (Lat: left vs. right). This preliminary set of ANOVAs was finalized to test the general effect of Condition (for ERs and RTs) and Condition and Lateralization (for O2Hb) in the whole sample.

For the second level, a similarity measure for continuous data was applied to each couple of subjects in pre- and post-feedback condition (for the 100 trials). These similarity measures for interval data, i.e., Pearson correlation as a measure of distance between vectors, finds the ratio between the covariance and the SD of both subjects (Sheldon, 2014). By using this measure specific similarities between each couple of subjects was monitored for cognitive (ERs and RTs) and O2Hb dependent measures.

For the third level, to analyze the systematic effect of the independent within subjects factors Cond on the similarities

coefficients, repeated measure ANOVAs were applied to the coefficients calculated for ERs and RTs as dependent variables. In addition to Cond, Lateralization factor was added in the case of O2Hb coefficients as dependent measure.

For all the ANOVA, the Greenhouse–Geisser epsilon was used for degrees of freedom correction where appropriate. Post hoc comparisons (contrast analyses) were used when necessary and Bonferroni test was used in the case of multiple comparisons.

To exclude a possible learning effect due to pre-/post feedback condition, a preliminary analysis was conducted, comparing distinctly the first groups of intervals (four pre-feedback intervals, for each 25 trials) and the second group of intervals (four post feedback intervals, for each 25 trials) for all the dependent variables (RTs, ERs, O2Hb). Since no significant differences among the four intervals, respectively for before and

as a function of right (C) and left (D) hemisphere.

after feedback condition, were found, we did not include this factor in the successive phases of the analysis.

### RESULTS

### ANOVA (Raw Data)

#### ER and RTs

For ER measure, ANOVA indicated significant effect for Cond (F(1,27) = 8.90, p ≤ 0.001, η <sup>2</sup> = 0.37). Indeed ER decreased in post-feedback (M = 0.03; SD = 0.009) compared to pre-feedback (M = 0.05; SD = 0.01; **Figure 3A**).

For RTs, ANOVA revealed significant main effect for Cond (F(1,27) = 9.05, p ≤ 0.001, η <sup>2</sup> = 0.39), showing reduced RTs for post-feedback (M = 223; SD = 0.25) compared to pre-feedback (M = 258; SD = 0.31; **Figure 3B**).

### O2Hb

The successive ANOVAs were applied to d measure for both O2Hb and HHb-values. Since the analysis on HHb did not show any significant results only statistical results for O2Hb were reported. The data over left (Ch1: FC3-F3; Ch2: FC3-FC1; Ch5: F1-F3; Ch6: F1-FC1) and right (Ch3: FC4-F4; Ch4: FC4-FC2; Ch7: F2-F4; Ch8: F2-FC2) channels were averaged.

ANOVA showed Lat × Cond significant interaction effect (F(1,27) = 11.32, p ≤ 0.001, η <sup>2</sup> = 0.40) with increased right brain responsiveness for post-feedback (M = 0.72; SD = 0.02) compared to pre-feedback condition (M = 0.38; SD = 0.01; **Figure 3C**).

## Similarity Measures

#### ER and RTs

The Pearson similarity coefficients (Fisher's z transform) were reported in the following **Figures 4A–B** for each couple of subjects in pre- and post-feedback. As indicated in **Figure 4A**, for ER, five couples showed significant coefficients for the pre-feedback condition, whereas 10 couples showed significant coefficients for the post-feedback condition. **Figure 4B** indicates the coefficients for RTs measures. As reported, nine couples revealed significant joined RTs modulation for the pre-feedback condition, whereas 13 couples showed significant coefficients in post-feedback.

### O2Hb

Significant Pearson coefficients were found in pre-feedback condition: six couples in the right and four in the left hemisphere revealed significant coefficients, whereas 12 couples were matched in post-feedback in the right side and five couples in the left side (**Figures 4C,D**).

### ANOVA on Similarity Measures

The third level of analysis considered the Pearson coefficients derived for ER, RTs and O2Hb as dependent measure in the repeated measures ANOVAs.

### ER and RTs Coefficients

Significant differences in ER were found for Cond (F(1,27) = 9.06, p ≤ 0.001, η <sup>2</sup> = 0.39), with increased Pearson coefficients values in post-feedback (M = 0.68; SD = 0.01) than pre-feedback

(M = 0.54; SD = 0.01) condition (**Figure 5A**). For RTs, a significant result was found for Cond (F(1,27) = 7.76, p ≤ 0.001, η <sup>2</sup> = 0.34), with increased Pearson values in post-feedback (M = 0.67; SD = 0.03) than pre-feedback (M = 0.57; SD = 0.02) condition (**Figure 5B**).

#### O2Hb Coefficients

Significant effect was found for Cond (F(1,27) = 8.79, p ≤ 0.001, η <sup>2</sup> = 0.36) and Cond × Lat (F(1,27) = 7.52, p ≤ 0.001, η <sup>2</sup> = 0.34). Indeed, increased coefficients were revealed in post-feedback (M = 0.60; SD = 0.02) than pre-feedback condition (M = 0.49; SD = 0.01). Second, about the interaction effect, during post-feedback the right hemisphere showed higher coefficient values (M = 0.67; SD = 0.02) compared to the left hemisphere (M = 0.57; SD = 0.01; F(1,27) = 7.12, p ≤ 0.001, η <sup>2</sup> = 0.34). In addition the right hemisphere registered increased Pearson values in post-feedback (M = 0.67; SD = 0.01) than in pre-feedback (M = 0.47; SD = 0.03; F(1,27) = 7.43, p ≤ 0.001, η <sup>2</sup> = 0.35; **Figure 5C**).

### GENERAL DISCUSSION

The present research explored the effects of a competitive joint-action on cognitive performance and brain activity by using a hyperscanning paradigm. Specifically, inter-brain similarities measures were acquired in couples of subjects during a competitive task, by using fNIRS. Based on our results, the following effects were observed. A first main effect was the systematic prefrontal (PFC) increased activity when a positive reinforce (post-feedback) was furnished to the participants about their performance. Indeed significant increased PFC activity in response to a positively reinforced joint action was found for all participants when compared to pre-feedback condition. Second, a better performance for both RTs and ER measures was revealed after the reinforcing feedback. Third, a higher inter-brain similarity was found for the couples after the feedback. Specifically when participants perceived (experimentally induced) to have performed better, a homologous and similar brain response was produced, with higher coherent PFC activity within the couple. Finally, it should be noted that a significant prefrontal brain lateralization effect was present, with the right hemisphere being more engaged in post-feedback condition.

About the first result, previous evidence revealed that prefrontal areas are crucial in social status monitoring and joint actions (Karafin et al., 2004; Haruno and Kawato, 2009; Suzuki et al., 2011). Also, using EEG-based hyperscanning technique, specific DLPFC activation emerged during reciprocal interaction in iterated Prisoner's Dilemma paradigm (De Vico Fallani et al., 2010). In the present research we observed a similar effect, with significant increased PFC activity in response to positively reinforced joint action during the cognitive task. This prefrontal brain area was hypothesized to have an evolutionary role in social perception mainly when hierarchy in social groups is crucial (Chiao et al., 2009a). Therefore we may suggest that this area has dedicated mechanisms to perceive social position and interaction significance during an interpersonal task.

More interestingly, the post-feedback condition induced an increased PFC responsiveness than pre-feedback condition. In fact, we observed that the PFC was mainly implicated when subjects were informed on their efficient interaction. This fact may indicate a central role of this prefrontal area in the case of a positive self-perception (to be a good performer) within a social situation where the competition is relevant and stressed. It is interesting to note that this ''improved brain effect'' was also accompanied by a significant increased cognitive performance (decreased ER and RTs). Indeed it was found that subjects highly improved their cognitive outcomes in response to the external reinforce. Due to these results we may suppose that the improved self-representation in term of social ranking and social position may have enhanced the real subjective performance. It should also be noted that the cortical and behavioral data showed to be matched, with a similar trend of higher activity for both behavior and cortical activity, which underlined the main effect of the (artificially) induced positive social reinforce on the intersubjective joint performance.

About the inter-brain relation, we observed a consistent and relevant increased brain-to-brain coupling for the dyads, mainly in concomitance with the positive social feedback. That is, this homologous inter-brain activity emerged in post-feedback condition for most of the couples. Therefore we may state that the externally induced reinforcing condition influenced the joint cortical responsiveness. That is, it could be suggested that, although the task was competitive, the self-perceived efficacy produced a sort of ''glue'' between the two brains, orienting the subjects on the same direction. Therefore, the present results provides initial evidence for the hypothesis of a significant interbrain effect during competitive tasks and offer suggestions for future studies examining the extent to which the competition in two brains is selectively related to a better cognitive joint performance for the two inter-agents.

This fact was further underlined by the significant effect of positive feedback on the cognitive joint-performance. Indeed the common strategy was evident also for the cognitive measures (RTs and ER) in addition to the brain measures. Higher similarity coefficients were found for the cognitive variables, thus underlining the impact of the external feedback on both the hemodynamic and cognitive level. In other terms, we may suppose that the external reinforce may have modulated the effective joint-behavior inside the couple, with relevant convergence of the increased performance by the two interagents.

About inter-subjective joint neural activity it should be noted that this prominent effect was mainly observed for the right hemisphere with respect to the left one. This result may be understood taking into account the social role of PFC and the lateralized effect observed in previous studies (Balconi et al., 2012). At this regard, we may consider the increased responsiveness in the right hemisphere as a possible marker of the competitive goal, oriented toward the maximization of the personal profit. Indeed, as previously demonstrated, the prefrontal asymmetry in favor of the right hemisphere may represent the withdrawal

### REFERENCES


motivation in opposition to approach motivation (Davidson, 1993; Jackson et al., 2003; Urry et al., 2004; Balconi and Mazza, 2010; Harmon-Jones et al., 2010; Koslow et al., 2013).

An alternative second explanation of the present result may relate the increased right hemisphere responsiveness to a significant increasing of more negative and avoidance emotions toward the competitor, linked to the competitive condition. As previously shown, the right hemisphere is supporting the aversive situations where the subjects are required to manage the conflictual and potential divergent goals (Balconi et al., 2012). Therefore, a sort of a ''negative echo'' may be induced by the individualistic and competitive aims of the task for each subject, with a significant increasing of more withdrawal attitudes. Consequently, prefrontal brain activity can be regarded to be highly involved in the processing of emotional behavior which affects the competitive context (Adolphs, 2002; Chiao et al., 2009a). However actually few studies have tried to study the emotional effects of competition on brain activity, taking into consideration the role of emotions on the cortical response (and on inter-brains responsiveness) when it responds to social situations as competition of cooperation. For this reason future research should better explore the distinct effect of emotions and competition on the cortical responsiveness to disambiguate their reciprocal relation.

However some limitations could be suggested for the present study. First, some adjunctive analyses could be used, to elucidate the gender effect in interactions, since the couples were composed by males or females. Second, a more accurate analysis for the dynamical changes of the inter-subjective strategy during the task should be conducted, in order to verify the progression of the learning mechanisms related to the inter-brain and cognitive processes. Finally, future research should better explore the effect of competitive in comparison with cooperative task, to verify the significant differences in brain-to-brain coupling and cognitive performance in response to these two different experimental conditions. This comparison should also allow to comprehend the significance of the positive feedback per se, separated by the competitive/cooperative task effect.

### AUTHOR CONTRIBUTIONS

MB designed the research, supervised the experiment, analyzed the data and wrote the text. MEV realized the experiment, analyzed the data and wrote the text.

and behavioural inhibition (BIS) and activation (BAS) systems. Laterality 15, 361–384. doi: 10.1080/13576500902886056


dynamic interaction. Soc. Neurosci. 10, 166–178. doi: 10.1080/17470919.2014. 977403


**Conflict of Interest Statement**: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

The reviewer AH and handling Editor declared their shared affiliation, and the handling Editor states that the process met the standards of a fair and objective review.

Copyright © 2017 Balconi and Vanutelli. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Hit or Run: Exploring Aggressive and Avoidant Reactions to Interpersonal Provocation Using a Novel Fight-or-Escape Paradigm (FOE)

Frederike Beyer 1† , Macià Buades-Rotger 2,3 \* † , Marie Claes <sup>2</sup> and Ulrike M. Krämer 2,3

1 Institute of Cognitive Neuroscience, University College London, London, United Kingdom, <sup>2</sup>Department of Neurology, University of Lübeck, Lübeck, Germany, <sup>3</sup> Institute of Psychology II, University of Lübeck, Lübeck, Germany

Interpersonal provocation presents an approach-avoidance conflict to the provoked person: responding aggressively might yield the joy of retribution, whereas withdrawal can provide safety. Experimental aggression studies typically measure only retaliation intensity, neglecting whether individuals want to confront the provocateur at all. To overcome this shortcoming of previous measures, we developed and validated the Fight-or-Escape paradigm (FOE). The FOE is a competitive reaction time (RT) task in which the winner can choose the volume of a sound blast to be directed at his/her opponent. Participants face two ostensible opponents who consistently select either high or low punishments. At the beginning of each trial, subjects are given the chance to avoid the encounter for a limited number of times. In a first experiment (n = 27, all women), we found that fear potentiation (FP) of the startle response was related to lower scores in a composite measure of aggression and avoidance against the provoking opponent. In a second experiment (n = 34, 13 men), we altered the paradigm such that participants faced the opponents in alternating rather than in random order. Participants completed the FOE as well as the Dot-Probe Task (DPT) and the Approach-Avoidance Task (AAT). Subjects with higher approach bias scores in the AAT avoided the provoking opponent less frequently. Hence, individuals with high threat reactivity and low approach motivation displayed more avoidant responses to provocation, whereas participants high in approach motivation were more likely to engage in aggressive interactions when provoked. The FOE is thus a promising laboratory measure of avoidance and aggression.

#### Keywords: approach, avoidance, aggression, provocation, retaliation

### INTRODUCTION

Aggressive behavior is a great challenge to individuals and to society. It is thus not surprising that research on aggression has been conducted for decades and continues to be an important problem addressed by scientific experiments today (Anderson and Bushman, 1997). Laboratory experiments on human aggression usually employ one of several well-established paradigms, which are derived from different theories of human aggression. One central aspect is interpersonal provocation: two of the most widely used aggression paradigms, the point-subtraction-aggression-paradigm (PSAP; Cherek, 1981) and the Taylor Aggression Paradigm (TAP; Taylor, 1967) confront the participant with an ostensible opponent, who in some way inflicts harm upon the participant. The participant then has the option to retaliate, which is directly measured within the paradigms.

#### Edited by:

Nelly Alia-Klein, Icahn School of Medicine at Mount Sinai, United States

#### Reviewed by:

Nathalie Holz, Zentralinstitut für Seelische Gesundheit, Germany Gennady Knyazev, Institute of Physiology and Basic Medicine, Russia

\*Correspondence:

Macià Buades-Rotger macia.rotger@neuro.uni-luebeck.de

†These authors have contributed equally to this work.

Received: 16 June 2017 Accepted: 28 September 2017 Published: 17 October 2017

#### Citation:

Beyer F, Buades-Rotger M, Claes M and Krämer UM (2017) Hit or Run: Exploring Aggressive and Avoidant Reactions to Interpersonal Provocation Using a Novel Fight-or-Escape Paradigm (FOE). Front. Behav. Neurosci. 11:190. doi: 10.3389/fnbeh.2017.00190 This approach is based on the well-established theory that reactive aggression is usually elicited as a response to some form of provocation or frustration (Anderson and Bushman, 1997; Lawrence, 2006). However, one shortcoming of these paradigms is that they largely limit the participant's behavioral options to showing lower or higher levels of aggression. In real-life hostile situations, one usually has the option to avoid the aggressive interaction altogether and withdraw from the situation. Limiting the range of behavioral options possibly limits the applicability of laboratory findings to real-life aggression. Thus, implementing an escape option in laboratory aggression paradigms is an important step towards improving research on aggression (Tedeschi and Quigley, 1996).

In the PSAP, participants are led to believe that they will interact with another player in another room while trying to earn points, which can later be exchanged for money. The participant can earn points by pressing a button as quickly as possible. With another button, the participant can subtract points from his co-player. Provocation is implemented as the co-player's subtraction of the participant's points, whereas aggressive behavior is measured as the number of times the participant uses the point subtraction button to inflict cost upon his co-player. In some versions of the PSAP, a protective button is implemented, which protects the participant from point subtraction for a certain time (Cherek et al., 1997).

In the TAP, participants are also led to believe that they are competing against another player. Subjects are told they will engage in a reaction time (RT) competition with the opponent. They are required to respond quickly to a stimulus, and are led to believe that the faster player in each run wins. The loser gets punished with an aversive stimulus (e.g., a mild electric shock or a sound blast), which can be adjusted in intensity. In each trial, the winner determines the intensity of punishment for the loser. Provocation in this paradigm is manipulated as the punishment level assigned to the participant, whereas aggression is measured as the punishment level selected by the participant for the opponent.

These paradigms have been widely used in behavioral research on aggression and, more recently, also in research on the neural basis of aggressive behavior (Krämer et al., 2007; Lotze et al., 2007; Kose et al., 2015). In both paradigms, aggressive behavior is non-instrumental insofar as the main outcome of the task (i.e., winning the RT task; earning points) is not improved by aggressive behavior but may actually (in case of the PSAP) be hindered by it. One conceptual advantage of the TAP is that while in the PSAP aggression is costly (the participant cannot simultaneously subtract and earn points), in the TAP aggressive behavior can be measured independently of cost-benefit considerations. Similarly, while avoidance behavior is assumed to be related to fear (Carver, 2004), in the PSAP the protective button may be used based on considerations of monetary tradeoff (weighing the points missed while pressing the protective button against the points saved by avoiding subtraction). As such, it may constitute an imperfect model of real-life avoidance behavior.

This latter point, however, also poses an important limitation for the extrinsic validity of the TAP. We have previously argued (Beyer et al., 2014) that based on theories about the role of emotional reactivity in aggressive behavior, one would expect a negative relationship between fear reactivity and aggressive behavior in the TAP. Specifically, anger, as an approachrelated affect (Carver and Harmon-Jones, 2009) should be most reliably elicited by provocation in participants low in fear reactivity and high in approach motivation. On the other hand, participants high in fear reactivity should react to a provocative confrontation with increased avoidance tendencies, rather than aggressive approach tendencies. In a functional magnetic resonance imaging (fMRI) study using the TAP, in which we measured threat reactivity as fear potentiation (FP) of the startle response, we found support for this hypothesis on a neural level (Beyer et al., 2014). The startle response in humans can be measured as the eye-blink amplitude in response to a sudden burst of white noise. FP is defined as the amplification of this amplitude when the participant is watching threatening rather than neutral pictures, and has been shown to be a good measure of emotional reactivity to threat (Vaidyanathan et al., 2009). In participants low in FP, we observed increased activity in areas of the so-called mentalizing network when they were confronted with a provocative opponent in the TAP. For participants high in FP, we observed the opposite effect, a reduction of activity in the mentalizing network due to provocation. The mentalizing network consists of cortical structures recruited when people take the perspective of another person in order to infer his/her thoughts, wishes or intentions (Lieberman, 2007). We therefore interpreted this effect as cognitive avoidance of the aggressive interaction in highly fearful participants (Beyer et al., 2014). However, we observed no effects on a behavioral level. One potential reason for this is the lack of an avoidance option in the TAP. Participants high in fear reactivity had no option of escaping the confrontation with the aggressive opponent and consequently may have adopted a tit-for-tat-like strategy. In many everyday incidents of provocation, however, avoidance is a realistic and valid behavioral option. Thus, we expect that the ecological validity of aggression paradigms should be increased by including a true avoidance option, producing the proposed relationship between personality traits (namely fear reactivity) and aggressive behavior.

In this study, we present a novel interactive aggression paradigm with an avoidance option: the ''Fight-Or-Escape'' (FOE) paradigm. In a first experiment, we implemented the task in a female student sample, also measuring FP of the startle response in a similar setup as we previously used for the TAP. This experiment is designed to test our previous interpretation of the non-existing relationship between FP and aggression in the TAP. To further validate our new paradigm, in a second experiment, we combined the task with two other well-established tests of social avoidance tendencies, the approach-avoidance-task (AAT; Roelofs et al., 2005) and a dot-probe-task (DPT; MacLeod et al., 1986) using emotional facial stimuli. We expected that fear reactivity and social avoidance should be associated with avoidant behavior and less aggression towards a provocative opponent.

### EXPERIMENT 1

### Materials and Methods

#### Participants and Procedure

Forty-three healthy female volunteers (Mage = 22 ± 2 years) participated in this study. We recruited only women in order to keep comparability with the previous study (Beyer et al., 2014). Participants were invited to the lab in groups of three. They were informed that two different experiments would be carried out: one EMG-measurement (the startle measure) and one group-task. The order of the two tasks was randomized across participants.

For the aggression task, the three participants received written instructions together. Prior to this task, participants filled out questionnaires assessing approach/avoidance tendencies and empathy (see below). After the aggression task, participants filled out a questionnaire probing for suspicion concerning the task, as well as a questionnaire on trait aggression. Finally, all participants were fully debriefed and reimbursed for their participation with 8 Euro per hour. Importantly, participants always received the same endowment regardless of their performance, and this was made clear to them before the measurement. This study was carried out in accordance with the recommendations of the University of Lübeck Ethics Commission (Ethikkommission der Universität zu Lübeck) with written informed consent from all subjects. All subjects gave written informed consent in accordance with the Declaration of Helsinki. The protocol was approved by the University of Lübeck Ethics Commission (Ethikkommission der Universität zu Lübeck).

### Fight-or-Escape (FOE) Paradigm

The FOE-paradigm was set up as a competitive RT game for three people. Participants were instructed that each would randomly be assigned one of three characters from the ''Lord of the Rings'' trilogy (Tolkien, 1954). As a background story, participants were told that Sauron and Saruman were sending out orcs to obtain the Ring of Power from Frodo, and that during the game it would be decided whether Frodo succeeds in destroying the ring or whether it is taken by his opponents. The three available characters were Frodo, Sauron and Saruman. In fact, each participant was assigned the role of Frodo. Participants were instructed that Sauron and Saruman would compete together against Frodo.

In each trial, Frodo would be playing against one of his two opponents and the winner of each trial would receive one point, with the points of Sauron and Saruman being summed together. At the end of the game, the party with the highest score (Frodo or Sauron and Saruman) would win. Participants were informed that the outcome would not affect their endowment. The winner of each trial was decided in a simple RT task: an exclamation mark was presented, followed by the picture of an orc. Frodo and his opponent had to respond to the orc by button press as quickly as possible. The faster player received one point, whereas the loser was punished with an aversive sound. At the beginning of each round, participants selected a punishment level (i.e., the noise level, ranging from 1 to 8) for their opponent, in case that the participant would win. Additionally, at the beginning of each trial, Frodo had the option of putting on the ring and thus becoming invisible, to avoid the confrontation. In that case, nobody received a point and nobody got punished; this choice constituted the ''escape'' option, i.e., avoiding potential punishment while foregoing the chance of earning a point. Thus, the sequence for one trial was as follows (**Figure 1**): (1) information on which opponent the participant was playing against; (2) choice of putting on the ring; in case Frodo put on the ring, a short message (''You escape the orc'') was displayed and the trial ended; (3) if Frodo did not put on the ring: punishment selection; (4) RT task; and (5) outcome phase: information on who won and which punishment level the opponent selected.

During screen number 2 (option of putting on the ring), Frodo was presented with the remaining number of times he could use the ring. Throughout the game, he could put on the ring a maximum number of 10 times. If a participant selected the ring after the 10 permitted escape options had been used, the message ''You cannot use the ring anymore'' was displayed and the trial then automatically proceeded to the punishment selection.

The game was programed such that subjects would win about 50% of trials against each opponent. Winning and losing was not related to RT, unless participants were slower than 500 ms, in which case they would lose. The behavior of Sauron and Saruman was programed such that one opponent was non-provocative, selecting low punishments (range 1–4, mean = 2.3), whereas the other was highly provocative, selecting high punishments (range 4–8, mean = 6.0). Punishments were randomly distributed across trials for each opponent. Frodo played 20 trials against each opponent in randomized order. Thus, Frodo had the option of avoiding up to 50% of the confrontations with the provocative opponent. Frodo could distribute using the ring between the two opponents in any ratio, and he was not obliged to use all 10 escape options. Since nobody received a point in escape trials, putting on the ring did not affect the ultimate outcome of the game. The identity of the provocative opponent (Sauron or Saruman) was randomized across participants. The intensity of the punishment (a Styrofoam scratching sound) was adapted to each participant's tolerance before starting the task. After the task, participants rated the perceived unpleasantness of the loudest and lowest tone in a scale from 1 to 8.

### Behavioral Measures

A range of behavioral measures can be derived from this task: pure aggression measures were obtained by calculating mean punishment selection for each opponent. However, this score would be identical for a participant who avoided the aggressive opponent in half the trials but otherwise retaliated with high punishment selections, and a participant who never avoided and behaved aggressively. Avoidance measures were obtained by counting the number of times the ring was used to avoid each opponent. Similar to the problem mentioned for the mean aggression score, this avoidance measure would be identical for participants high and low in punishment selections, if both avoid the same number of trials. To address these issues, we additionally calculated a combined aggression-avoidance score by summing punishment selections for each opponent across all trials. For this measure, avoidance trials are scored as zero. Consequently, this score is affected both by the number of times a participant chose not to play against an opponent, and by the punishment she selected when she did. Accordingly, this measure reflects the absolute amount of aggression shown towards the opponent. A medium score for the provocative opponent could be reached by a participant who frequently avoided him, but behaved aggressively in the remaining trials, or a participant who did not avoid him, but showed moderate levels of aggression.

### Personality Questionnaires

A German version (Herzberg, 2003) of the Buss and Perry Aggression Questionnaire (AQ; Buss and Perry, 1992) was used to assess trait aggressiveness. The AQ consists of four subscales: physical aggression, verbal aggression, hostility and anger. To assess approach and avoidance tendencies, a German version (Strobel et al., 2001) of the behavioral inhibition and activation scales (BIS/BAS; Carver and White, 1994) was used. We also used the Interpersonal Reactivity Index (IRI; Davis, 1983) in our own translation to measure empathy and perspective taking tendencies.

### Measurement of Fear Potentiation

To measure FP, we used a setup which was adapted from previous studies (Caseras et al., 2006; Conzelmann et al., 2009). Participants were presented with 51 pictures from the International Affective Pictures System (IAPS; Lang et al., 1999). Half of these pictures were threatening (e.g., a gun pointed at the viewer, an attacking dog), the other half were neutral (e.g., a secretary on the phone; household objects). Pictures were presented in a fixed order which was set up randomly with the constraint that no more than two pictures of the same valence were presented consecutively. Each picture was presented for 6 s with a 12 s inter trial interval (ITI), during which a white central cross was presented on a black background. During 18 threatening and 21 neutral pictures, a short burst of white noise (50 ms, 95 dB), was presented over speakers 1.5, 2.8 or 4.0 s after picture onset. For the remaining 12 pictures, the startle probe was presented during the ITI and these trials were not analyzed. To account for initial habituation of the startle response, four startle probes were presented while participants watched the fixation cross. Additionally, the first three picture trials (all neutral), were discarded.

### EMG Measurement and Analysis

Two Ag-AgCl electrodes were placed below the left lower eyelid, one in line with the pupil and the other 1–2 cm to the left of the first. A ground electrode was positioned centrally on the forehead. Prior to electrode placement, the skin was treated with a peeling paste and alcohol. The EMG signal was amplified and recorded at a sampling rate of 250 Hz using an EEG amplifier (32-channel Brainamp; Brain Products).

We analyzed EMG recordings with EEGLAB, a MATLABbased open-source toolbox (Delorme and Makeig, 2004). EMG signals were high-pass filtered at 10 Hz, low-pass filtered at 500 Hz and baseline-corrected using the 50 ms prior to onset of the startle probe as baseline. We then visually inspected each startle trial for artifacts. Trials with excessive noise or eyeblinks in the 50 ms baseline period were excluded. Blink magnitude was measured as the maximum absolute amplitude in an interval of 20–160 ms following the startle probe. Blink scores were z-transformed within each participant across all trials. We then subtracted the mean standardized blink amplitude for neutral pictures from the respective value for threatening trials, to get individual FP scores.

#### Statistical Analyses

We first conducted paired t-tests to compare mean aggression and mean avoidance scores for the provoking vs. non-provoking opponent, as well as mean blink amplitude for threatening vs. neutral pictures.

To investigate the relationship between FP and aggressive and avoidant behavior, we correlated FP scores with the respective behavioral measures (mean aggression, number of avoidance choices, sum of punishment selections across trials) for each opponent. We hypothesized that FP should be negatively related to aggressive behavior and positively to avoidant behavior, resulting in a negative correlation between FP and the summed punishment score. In the presence of a significant relationship, we post hoc compared correlation coefficients between opponents in R v1.3.1 with the r.test() function, available in the psych package v1.5.6 (Revelle, 2017). On an exploratory level, we also correlated personality questionnaire scores with FP and behavioral measures from the task. Significance was set in all cases at p < 0.05. **Table 1** shows descriptive statistics for all measures in Experiment 1.

### Results

Of the 43 participants, 16 had to be excluded due to the following reasons: technical problems during startle measurements and/or bad EMG data quality (13); suspicion concerning the aggression task (3). Participants rated the highest tone (M = 5.8, SE = 1.8) as significantly more unpleasant than the lowest tone (M = 1.8, SE = 0.2), t(26) = 11.1, p < 0.001.

On average, participants selected higher punishment levels for the provoking (M = 4.0, SE = 0.3) than the unprovocative opponent (M = 3.1, SE = 0.2), t(26) = 3.2, p < 0.01. Of the 10 avoidance options, participants used on average 5.9 (SE = 0.7; range 0–10). There was no significant difference in the number of times participants avoided the provocative (M = 3.3; SE = 0.5) and non-provocative opponent (M = 2.5; SE = 0.4), p = 0.33. There was also no significant difference (p = 0.61) between startle responses to neutral (M = −0.015, SE = 0.13) and threatening pictures

TABLE 1 | Descriptive statistics for variables in Experiment 1 (n = 27).


AQ, Aggression Questionnaire; BIS, Behavioral Inhibition System; BAS, Behavioral Activation System; IRI, Interpersonal Reactivity Index; Z, Z scores; HP, High Provocation; LP, Low Provocation.

(M = 0.015, SE = 0.12). Across participants, however, there was great variability in FP (range −0.59 to 0.40; mean = 0.03; SE = 0.05).

We found a negative correlation between FP and the summed punishment score for the provocative opponent (r = −0.38, p < 0.05; **Figure 2A**). This relationship was not significant for the non-provocative opponent (r = −0.28, p = 0.14). These correlations did not significantly differ from each other (p = 0.58). Concordantly, the difference in aggression sum scores between the provoking and non-provoking opponent was not correlated with FP (p = 0.44). We observed similar effects for mean aggression scores, with a negative correlation between FP and aggression against the provocative opponent (r = −0.44, p < 0.05; **Figure 2B**), but no effect for the non-provocative opponent (r = −0.17, p = 0.39). These two effects only differed marginally from each other, t(26) = 1.8, p = 0.08. The difference in mean aggression between opponents was negatively correlated at trend level with FP (r = −0.35, p = 0.07). We found no relationship between FP and the number of times the avoidance option was chosen for either opponent or the absolute number of trials avoided (all p > 0.2). Mean aggression towards the provocative opponent was positively correlated with trait anger (r = 0.44, p < 0.05). There was a negative correlation between FP and trait anger as assessed with the AQ (r = −0.46, p < 0.05). Thus, FP was negatively related to trait anger, aggressive behavior towards the provocative opponent, as well as to a conglomerate measure of aggression and avoidance for this opponent. We observed no other relationships with self-report data (all p > 0.1).

### Discussion

In this experiment, we found a negative relationship between fear reactivity, as measured using a fear potentiated startle paradigm, and aggressive behavior towards a provocative opponent. In contrast to our previous study, where we found no such effect, participants here had the option of avoiding the aggressive interaction. Interestingly, we found no direct relationship between avoidance behavior and FP. The intrinsically low variability in avoidance (i.e., a maximum 10 times) might have curtailed the possibility to find direct relationships between avoidance and other parameters. Nonetheless, there was a negative relationship between FP and a conglomerate measure of aggression and avoidance against the provoking opponent. Thus, in situations where avoiding a confrontation is explicitly possible, participants high in fear reactivity behave less aggressively towards an aggressive opponent than participants low in fear reactivity. Note anyway that the relationship between aggression scores and FP did not significantly differ between opponents. Hence, FP might reflect general avoidant tendencies rather than reactivity to provocation specifically. Such tendencies seem to be slightly more pronounced when facing a provoking rival, but apply to interpersonal confrontation in general.

The negative correlation we found between FP and trait anger supports previous findings using startle paradigms which suggest that anger is an emotion related to behavioral approach tendencies (Amodio and Harmon-Jones, 2011). FP, on the

other hand, is a defensive reflex associated with behavioral avoidance, like freezing behavior in animals (Davis et al., 1993). The positive correlation we found between trait anger and aggression towards the provocative opponent further underlines the role of emotional reactivity in aggressive retaliation. Thus, people high in fear reactivity are overall less prone to feelings of anger and are less likely to retaliate. People low in fear reactivity, on the other hand, report more angry impulses, which are related to more aggressive behavior. One must note, however, that we conducted correlations with personality questionnaires in an exploratory approach and the observed effects would not survive correction for multiple testing.

Our results might seem partly at odds with fMRI studies linking aggression with increased amygdala reactivity to threat (Coccaro et al., 2007; Carré et al., 2013; McCloskey et al., 2016). However, in a previous study from our group we implemented the FOE in the scanner and observed that the amygdala was recruited when avoiding a highly provoking opponent (Buades-Rotger et al., 2017). Thus, emotional reactivity to threat appears to favor either avoidance or aggression depending on threat escapability. Indeed, recent accounts of amygdala function posit that this structure generally codes for biologically relevant events, regardless of their aversive or appetitive nature (Pessoa and Adolphs, 2010; Weymar and Schwabe, 2016).

It is also intriguing that we did not find a clear FP effect, or a difference in avoidance between opponents. The lack of FP might be due to a failure in eliciting strong emotional reactions and/or to high interindividual variability in these affective responses (Bradley et al., 2001). We deem it unlikely that the white noise was itself too aversive, since we applied the same regime as in our previous study, where we did observe FP (Beyer et al., 2014). The absence of a provocation effect on avoidance may be attributable to the fact that the avoidance option only provided a momentary respite, as subjects could face the opponent they just avoided. Hence, in Experiment 2 we programed the task so that participants faced the opponents in alternating order, which should render the avoidance option more meaningful.

## EXPERIMENT 2

### Introduction

In our first experiment, we used FP as a measure of fear reactivity in order to compare our findings to our previous study using the TAP. This showed the proposed negative relationship between FP and aggression and supports our previous argument that the lack of an avoidance option in the TAP reduces interpersonal variability in aggression, as it imposes an unnatural limitation of behavioral options upon the participant. To further validate the paradigm and test the proposed negative relationship between avoidance and aggression, in a second experiment we combined the FOE-paradigm with two well-established tasks designed to measure implicit approach and avoidance tendencies towards social stimuli. Furthermore, we set task so subjects confronted each opponent alternatingly. We did so in order to potentiate the meaningfulness and salience of the avoidance option.

In the AAT (Roelofs et al., 2005), participants are asked to perform approach movements (i.e., pulling a joystick towards themselves) or avoidance movements (pushing a joystick away from themselves) in response to visual stimuli. Using stimuli that typically elicit approach or avoidance tendencies, congruency effects can be observed: participants are faster in pushing away stimuli that elicit avoidance tendencies than they are at pulling them close, whereas the opposite effect is observed for approach-related stimuli. This has been found for social stimuli as angry and happy faces (Roelofs et al., 2005), as well as other affective stimuli, as pictures of spiders in participants high in fear of spiders (Rinck and Becker, 2007). The AAT thus constitutes a measure of behavioral avoidance.

The DPT (MacLeod et al., 1986) assesses automatic orientation of attention towards one of two visual stimuli. In each trial, two stimuli are shortly presented at opposing sides of the screen center. After stimulus offset, a dot is presented on one side, and the participant is asked to press a corresponding response button (i.e., left or right). As individuals tend to initially allocate attention to threatening stimuli and then avoid them (Cooper and Langton, 2006; Rinck and Becker, 2006), they should be slower to respond to a dot presented on the side of the threatening stimulus at long exposition times (Mogg et al., 2010). The DPT can thus be used as a measure of attentional avoidance.

In this experiment, we used the AAT with pictures of angry and happy facial expressions and the DPT with pictures of angry and neutral facial expressions. We hypothesized that participants who showed high behavioral and attentional avoidance of angry faces would be less aggressive towards an aggressive opponent in the FOE-paradigm and would more frequently use the avoidance option for that opponent compared to participants showing less avoidance of angry faces.

### Materials and Methods

### Participants and Procedure

Forty-two healthy volunteers (18 men; Mage = 22 ± 3 years) participated in this study. Participants were invited to the lab in same-sex triads. They were told that they would perform three computer tasks, the first of which would be interactive. The order of the two individual tasks was randomized across participants. After the computerized tasks, participants filled out the same questionnaires as in Experiment 1 (i.e., German versions of AQ, BIS/BAS and IRI, and the deception check). All participants were debriefed regarding the true study goals and reimbursed for their participation with 8 Euro per hour. This study was carried out in accordance with the recommendations of the University of Lübeck Ethics Commission (Ethikkommission der Universität zu Lübeck) with written informed consent from all subjects. All subjects gave written informed consent in accordance with the Declaration of Helsinki. The protocol was approved by the University of Lübeck Ethics Commission (Ethikkommission der Universität zu Lübeck).

### FOE-Paradigm

The setup of the FOE-paradigm was identical to that used in Experiment 1, with one exception: instead of randomizing the trials against the two opponents, we now implemented an alternating order. Thus, participants knew that they would be playing against the two opponents alternatingly. We modified this in order to increase the salience of the avoidance option, as reasoned above. Whereas in Experiment 1, subjects could only minimize the absolute number of trials they played against the aggressive opponent, a trial where they chose to avoid this opponent could still be followed by a trial against the same opponent. As such, the avoidance was realized on a global level rather than immediately. To make the avoidance more prominent, the alternating schedule implemented here ensured that if a participant avoided the aggressive opponent, in the following trial they would always be playing against the non-aggressive one.

### Approach-Avoidance-Task

For this task, we used photographs of angry and happy facial expressions of the Radboud Faces Database (Langner et al., 2010). We used pictures of 30 individuals (15 females), each showing an angry facial expression in one picture, and a happy expression in the other. Pictures of nine different individuals (four females) were used for practice blocks. The pictures were cropped into an oval shape, removing hair, ears and neck.

Participants were given a standard joystick (Speedlink<sup>r</sup> Dark Tornado) as response device. At the beginning of each trial, a fixation cross was presented centrally on a white background. The participant started the trial by pressing the ''shoot'' button on the joystick. Following this, one picture was presented centrally on the screen. Participants could reduce picture size by pushing the joystick away from them. By pulling the joystick towards themselves, they could increase picture size. Picture size was varied gradually in seven steps.

In one block, participants were instructed to ''push away'' angry faces and ''pull close'' happy faces as quickly as possible; in the other block, the reverse instruction was given. Block order was randomized between subjects. Each block consisted of 30 happy and 30 angry trials, and the same pictures were used in both blocks. Each block was preceded by a practice run. During practice runs, each trial was followed by feedback (a green check-mark for correct reactions, a red cross for errors). The practice run for the first block consisted of 20 trials. The second practice run consisted of 28 trials, since participants had to reverse their response patterns from the first block.

For RT, we analyzed the interval between stimulus presentation and movement onset. Incorrect trials (including trials in which the initial movement was performed in the incorrect direction, followed by a correction), and trials with response latencies shorter than 150 ms or longer than 3 SD from the own mean were excluded from analysis. We calculated the pull minus push difference in RT (higher scores meaning higher avoidance) for angry and happy faces separately. We compared both biases with a paired t-test.

### Dot-Probe-Task

For the DPT, we used 40 pictures from a set of previously validated videos (Kircher et al., 2013) showing angry and neutral facial expressions of 20 different professional actors (nine women). In each trial, the neutral and angry pictures of one person were presented together, to the left and right of the screen center.

Each trial began with a centrally located fixation cross being presented for a random interval between 500 ms and 1000 ms. Then, the neutral and angry pictures of one person were presented for 1000 ms. At picture offset, a dot was presented located to the left or right of the screen center, at the coordinate where the center of the respective picture had been. Participants were instructed to react as quickly as possible to the dot, by pressing a left button (A) on the computer keyboard if the dot was presented on the left, and a right button (L) if it was presented at the right side. We chose a presentation time of 1000 ms to allow sufficient time for eliciting avoidance, based on previous results showing that an avoidant bias can already be observed at 500 ms (Cooper and Langton, 2006; Pintzinger et al., 2017).

The task consisted of two blocks. In each block, 80 trials were presented, with each picture pair presented four times, once in each of the four conditions: angry picture on the left, dot probe on the left; angry picture on the left, dot probe on the right; angry picture on the right, dot probe on the right; and angry picture on the right, dot probe on the left.

We analyzed attentional approach vs. avoidance tendencies for angry pictures by subtracting mean RT for the neutral condition (dot probe in location of neutral picture) from angry conditions (dot probe in location of angry picture). The higher this score was for a given participant, the greater was this participant's RT cost for reacting to a probe in the location of an angry picture. Thus, higher scores for this task represent greater attentional avoidance of angry facial expressions.

#### Statistical Analyses

We compared mean punishment selections and mean avoidance against the provoking vs. non-provoking opponent with paired t-tests, as in Experiment 1. We also compared approach vs. avoidance scores in the AAT, as well RTs for angry vs. neutral trials in the DPT. Additionally, we inspected whether men and women differed in their aggression and avoidance scores against each opponent with an analysis of variance (ANOVA) with provocation as within-subject factor and gender as betweensubject factor.

To explore the relationship between avoidance of threatening social stimuli and behavior in the FOE-paradigm, we correlated behavioral scores of the aggression task (mean aggression scores for each opponent; summed avoidance scores for each opponent; conglomerate avoidance-aggression score for each opponent) with the RT scores of the AAT and the DPT. As in Experiment 1, significant correlations were compared between opponents with the r.test() function (Revelle, 2017). We also correlated self-report scores with behavioral measures from the task on an exploratory basis. Significance was again set at p < 0.05. Descriptive statistics for all measures in Experiment 2 are provided in **Table 2**.

TABLE 2 | Descriptive statistics for questionnaires and computerized measures in Experiment 2 (n = 34).


AQ, Aggression Questionnaire; BIS, Behavioral Inhibition System; BAS, Behavioral Activation System; IRI, Interpersonal Reactivity Index; AAT, Approach-Avoidance Task; DPT, Dot-Probe Task; HP, High Provocation; LP, Low Provocation.

FIGURE 3 | (A) Avoidant behavior against each opponent in the FOE. Values are mean ± standard error. (B) Results of the Approach-Avoidance Task (AAT). (C) Correlation between the approach bias in the AAT (pull minus push for happy faces) and avoidant behavior against the highly provoking (HP) opponent in the FOE.

### Results

Of the 42 participants, eight had to be excluded (five due to nondeception, one due to incomplete task data, two due to extreme bias scores of ±3 SD in AAT and DPT). Hence, analyses for this experiment were performed on 34 participants (13 men). Subjects reported the highest tone (M = 5.5, SE = 0.2) to be significantly more unpleasant than the lowest tone (M = 1.5, SE = 0.1), t(33) = 13.9, p < 0.001.

Regarding behavior in the FOE paradigm, participants tended to select marginally higher punishments against the provoking opponent (M = 4.1, SE = 0.2) relative to the non-provoking one (M = 3.7, SE = 0.2), t(33) = 1.9, p = 0.06. Crucially, they avoided the provoking opponent (M = 2.6, SE = 0.3) more often than the non-provoking one (M = 1.8, SE = 0.3), t(33) = 2.1, p < 0.05 (**Figure 3A**). Participants used the avoidance option 4.3 times on average (SE = 0.5; range 0–10). The interaction between gender and provocation was non-significant for both aggression (p = 0.18) and avoidance (p = 0.55), indicating that behavior in the FOE did not differ between men and women.

We found the expected effect in the AAT, t(33) = 2.8; p < 0.01 (**Figure 3B**), such that participants showed an avoidant bias for angry faces (M = 47 ms, SE = 14 ms) and an approach bias towards happy faces (M = −44 ms, SE = 20 ms). We did not observe the hypothesized avoidant bias in the DPT (p > 0.2), since RT were similar in neutral (M = 365 ms, SE = 8 ms) and angry trials (M = 364 ms, SE = 8 ms). DPT and AAT scores were uncorrelated (all p > 0.1).

The DPT or AAT biases for angry faces were not related to avoidance or aggression in the FOE (all p > 0.1). There was, however, a significant correlation between AAT scores for happy faces and avoidance against the provoking opponent (r = 0.43, p < 0.05). Namely, participants who were quicker to pull happy faces towards them, but were slower to push them away, avoided the provoking opponent less often (**Figure 3C**). The approach bias for happy faces was unrelated to avoidance of the non-provoking opponent (r = 0.01, p = 0.97). These two correlation coefficients differed significantly from each other, t(33) = 2.23, p < 0.05, and, concordantly, the difference in avoidance between opponents was also associated with AAT approach scores (r = 0.39, p < 0.5). There were no correlations between self-report data and avoidance or aggression (all p > 0.1).

### Discussion

In this second experiment, we modified the task to make the avoidance option more meaningful, so that participants faced each opponent alternatingly, i.e., they could not face the same opponent in two consecutive trials. As intended, this caused participants to avoid the provoking opponent more than the non-provoking one. Participants selected slightly higher punishments against the provoking opponent than against the non-provoking one, but less so than in Experiment 1, indicating that subjects retaliated more evenly against both rivals. There were no gender differences in avoidance or aggression, suggesting that the provocation effect for avoidance observed in Experiment 2 is not due to the inclusion of male participants. Given that in Experiment 2 avoidance was a more attractive alternative strategy and the task was more predictable, participants might have experienced an increased sense of safety and confidence. This should favor the activation of appetitive, rather than defensive, motivational systems (Lang and Bradley, 2013). Hence, perhaps aggression in Experiment 2 reflected an appetitive drive, and not so much a defensive reflex. The results of correlational analyses, which are subsequently commented, support this account.

Instead of the expected relationship between an avoidant bias for angry faces and avoidance in the FOE, we found that participants with high approach scores towards happy faces engaged in more aggressive encounters in high provocation trials. Since happy faces constitute a social reward signal (Rademacher et al., 2010; Ruff and Fehr, 2014), our results indicate that individuals who are more strongly motivated by positive stimuli will tend to be less avoidant when provoked. This is consistent with the notion of aggression as an approach-related behavior (Carver and Harmon-Jones, 2009; Berkowitz, 2012). On the other hand, the fact that avoidance was related to AAT, but not to DPT scores, suggests that fight-vs.-flight decisions as implemented in the FOE are driven by general behavioral tendencies rather than implicit attentional biases.

The finding that AAT scores for happy rather than angry faces were related to avoidance deserves however further discussion. Some authors have argued that individuals showing an approach bias towards angry faces should be more aggressive, as they should be more prone to interpret anger expressions as a challenge rather than a threat, i.e., as an appetitive stimulus (van Honk et al., 2001; Beaver et al., 2008). However, happy faces are generally less ambiguous than angry faces (Coupland et al., 2004; Becker et al., 2011; Parmley and Zhang, 2015) and they more clearly convey reward and positive valence (Averbeck and Duchaine, 2009; Furl et al., 2012). Hence, happy facial expressions should more consistently elicit approach motivation than angry ones. In line with this formulation, and dovetailing our findings, a recent study using the AAT in veterans found that anxious symptomatology was related to avoidance of happy faces, but not to biases towards or away from angry expressions (Clausen et al., 2016). Similarly, another study with the AAT found that approach scores towards positive stimuli predicted reactive aggression, but no effect was found for angry faces (Lobbestael et al., 2016).

### General Discussion

Most established laboratory measures of aggression do not allow participants to avoid confrontation. We addressed this issue by developing and validating a version of the TAP that included an avoidance option: the FOE paradigm. In two separate experiments, we showed that reactivity to threat as measured by FP relates to reduced aggression and avoidance, and that participants with stronger approach tendencies towards positive stimuli more frequently chose to engage in an aggressive interaction than participants who tended to avoid positive stimuli.

In Experiment 1, participants with stronger FP responses were less aggressive on average in response to provocation. FP was also negatively related to aggression sum scores against the provoking opponent, which can be understood as a composite measure of avoidance and aggression. In our previous fMRI study, we found no relationship between threat reactivity and aggression in inescapable encounters (Beyer et al., 2014). Here, by giving participants the possibility to avoid confrontation, we observed the previously hypothesized negative correlation between threat reactivity and aggression. Nevertheless, we found no direct relationship between FP and avoidance in the FOE (i.e., number of avoidance options). This might be due to the fact that the avoidance option was not salient enough, as participants could face the highly provoking opponent in the trial after avoiding her.

In Experiment 2, we set the task so participants played against each opponent alternatingly instead of pseudo-randomly. In so doing, the avoidance option became more meaningful, and participants avoided the provoking opponent more often than the non-provoking opponent. Crucially, subjects showing a stronger avoidant bias for happy faces used the avoidance option against the provoking opponent more frequently. We found no relationship between the AAT avoidant bias for angry faces and avoidance in the FOE. DPT scores, which represent attentional avoidance of angry faces, were also uncorrelated with behavior in the task.

Taken together, our results suggest that aggressive behavior as implemented in the FOE is driven by approach motivation to engage in aggressive interactions. Participants who tended to react to threatening stimuli with strong defensive reflexes behaved overall less aggressively towards a provoking rival. This supports our initial theory that participants high in fear reactivity should behave less aggressively if given the opportunity to escape. On the other hand, participants with strong approach tendencies towards positive stimuli chose to confront the aggressive opponent more often. Likely, fight-vs.-flight choices in these participants were mainly driven by the prospect of potentially being able to retaliate against the aggressor, whereas participants low in behavioral approach preferred to withdraw from the aggressive interaction. These findings are in line with prominent theories of anger and aggression as driven by appetitive approach (Harmon-Jones, 2003; Carver and Harmon-Jones, 2009; Berkowitz, 2012).

### Limitations and Future Directions

A few limitations should be mentioned. First, the avoidance option is available only during the game and can be used several times. In real provocation situations one can only retreat once and usually before the proper physical confrontation ensues. Moreover, there are often alternative strategies to curb the provocateur. In future studies, one could implement friendly choices (e.g., praising the opponent or being punished in her stead), and/or an option to aggress proactively (e.g., preemptive punishments). This would allow for a more direct measurement of aggressive relative to prosocial tendencies than existing paradigms such the Help/Hurt Task (Saleem et al., 2015), or the Inequality Game (Klimecki et al., 2016), and would also better cover the range of possible responses to —or safeguards from interpersonal provocation. Related to this point, our sample might have suffered from range restriction due to situation selection (Dijkman and Devries, 1987), as highly avoidant subjects are unlikely to volunteer for such a study in the first place, and our participants were all healthy young students. Our task should thus be further validated on samples preselected based on extreme approach and avoidance. Future work would also benefit from cross-validating the FOE with other tasks such as the PSAP. In that case, one could relate avoidance in the FOE with the number of times one uses the defensive strategy, and aggression with the number of times one uses the offensive

### REFERENCES


strategy. Nevertheless, as stated in the ''Introduction'' section, both the defensive and aggressive options in the PSAP are costly, and therefore can be assumed to entail a conflict between approach and avoidance tendencies. An additional possibility would entail using more realistic measures of approach and avoidance, such as those based on interpersonal distance, which can be implemented physically (Perry et al., 2016) and/or in virtual reality environments (McCall and Singer, 2015; McCall et al., 2016).

### CONCLUSION

In the present article, we have provided validation of the FOE paradigm. The FOE is a competitive RT task in which participants can avoid the confrontation against either of two purported opponents: a provoking and a non-provoking one. We showed that subjects with higher reactivity to threat and lower approach motivation towards positive stimuli are less aggressive and more avoidant when taunted by a provocateur. Thus, the FOE can be readily implemented to measure not only retaliation intensity, but also subjects' proneness to avoid or engage in aggressive encounters.

### AUTHOR CONTRIBUTIONS

FB, MB-R, and UMK conceived the study and coordinated the experiments. FB, MB-R and MC performed the experiments. FB and MB-R analyzed the data and wrote the manuscript. MC and UMK revised the manuscript. All authors read and approved the manuscript.

### FUNDING

This study was funded by the (Deutsche Forschungsgemeinschaft) German Science Foundation (Grant number KR3691/5-1 awarded to UMK). We acknowledge financial support by Land Schleswig-Holstein within the funding program Open Access Publikationsfonds.


**Conflict of Interest Statement**: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Beyer, Buades-Rotger, Claes and Krämer. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Community Violence Exposure and Conduct Problems in Children and Adolescents with Conduct Disorder and Healthy Controls

Linda Kersten<sup>1</sup> \*, Noortje Vriends <sup>1</sup> , Martin Steppan<sup>1</sup> , Nora M. Raschle<sup>1</sup> , Martin Praetzlich<sup>1</sup> , Helena Oldenhof <sup>2</sup> , Robert Vermeiren<sup>2</sup> , Lucres Jansen<sup>2</sup> , Katharina Ackermann<sup>3</sup> , Anka Bernhard<sup>3</sup> , Anne Martinelli <sup>3</sup> , Karen Gonzalez-Madruga<sup>4</sup> , Ignazio Puzzo<sup>5</sup> , Amy Wells <sup>4</sup> , Jack C. Rogers <sup>6</sup> , Roberta Clanton<sup>6</sup> , Rosalind H. Baker <sup>6</sup> , Liam Grisley <sup>6</sup> , Sarah Baumann<sup>7</sup> , Malou Gundlach<sup>7</sup> , Gregor Kohls <sup>7</sup> , Miguel A. Gonzalez-Torres <sup>8</sup> , Eva Sesma-Pardo<sup>8</sup> , Roberta Dochnal <sup>9</sup> , Helen Lazaratou<sup>10</sup> , Zacharias Kalogerakis <sup>10</sup> , Aitana Bigorra Gualba<sup>11</sup> , Areti Smaragdi <sup>12</sup> , Réka Siklósi <sup>9</sup> , Dimitris Dikeos <sup>13</sup> , Amaia Hervás <sup>11</sup> , Aranzazu Fernández-Rivas <sup>8</sup> , Stephane A. De Brito<sup>6</sup> , Kerstin Konrad<sup>7</sup> , Beate Herpertz-Dahlmann<sup>7</sup> , Graeme Fairchild<sup>14</sup> , Christine M. Freitag<sup>3</sup> , Arne Popma<sup>2</sup> , Meinhard Kieser <sup>15</sup> and Christina Stadler <sup>1</sup>

<sup>1</sup>Department of Child and Adolescent Psychiatry, Psychiatric University Hospital, University of Basel, Basel, Switzerland, <sup>2</sup>Department of Child and Adolescent Psychiatry, VU University Medical Center, Amsterdam, Netherlands, <sup>3</sup>Department of Child and Adolescent Psychiatry, Psychosomatics and Psychotherapy, University Hospital Frankfurt, Goethe University, Frankfurt am Main, Germany, <sup>4</sup>Department of Psychology, University of Southampton, Southampton, United Kingdom, <sup>5</sup>Broadmoor High Secure Hospital, West London Mental Health NHS Trust, Crowthorne, United Kingdom, <sup>6</sup>School of Psychology, University of Birmingham, Birmingham, United Kingdom, <sup>7</sup>Department of Child and Adolescent Psychiatry, Psychosomatics and Psychotherapy, University Hospital, RWTH Aachen, Aachen, Germany, <sup>8</sup>Basurto University Hospital, Bilbao, Spain, <sup>9</sup>Faculty of Medicine, Child and Adolescent Psychiatry, Department of the Child Health Center, Szeged University, Szeged, Hungary, <sup>10</sup>Department of Child and Adolescent Psychiatry, National and Kapodistrian University of Athens, Athens, Greece, <sup>11</sup>University Hospital Mutua Terrassa, Barcelona, Spain, <sup>12</sup>Center of Addiction and Mental Health, Toronto, ON, Canada, <sup>13</sup>Department of Psychiatry, Medical School, National and Kapodistrian University of Athens, Athens, Greece, <sup>14</sup>Department of Psychology, University of Bath, Bath, United Kingdom, <sup>15</sup>Institute of Medical Biometry and Informatics, University of Heidelberg, Heidelberg, Germany

Exposure to community violence through witnessing or being directly victimized has been associated with conduct problems in a range of studies. However, the relationship between community violence exposure (CVE) and conduct problems has never been studied separately in healthy individuals and individuals with conduct disorder (CD). Therefore, it is not clear whether the association between CVE and conduct problems is due to confounding factors, because those with high conduct problems also tend to live in more violent neighborhoods, i.e., an ecological fallacy. Hence, the aim of the present study was: (1) to investigate whether the association between recent CVE and current conduct problems holds true for healthy controls as well as adolescents with a diagnosis of CD; (2) to examine whether the association is stable in both groups when including effects of aggression subtypes (proactive/reactive aggression), age, gender, site and socioeconomic status (SES); and (3) to test whether proactive or reactive aggression mediate the link between CVE and conduct problems. Data from 1178 children and adolescents (62% female; 44% CD) aged between 9 years and 18 years from seven European countries were analyzed. Conduct problems were assessed using the Kiddie-Schedule of Affective Disorders and Schizophrenia diagnostic interview.

#### Edited by:

Nelly Alia-Klein, Icahn School of Medicine at Mount Sinai, United States

#### Reviewed by:

Monique LeBlanc, Southeastern Louisiana University, United States Kathleen Monahan, Stony Brook University, United States

#### \*Correspondence:

Linda Kersten linda.kersten@upkbs.ch

Received: 30 August 2017 Accepted: 20 October 2017 Published: 06 November 2017

#### Citation:

Kersten L, Vriends N, Steppan M, Raschle NM, Praetzlich M, Oldenhof H, Vermeiren R, Jansen L, Ackermann K, Bernhard A, Martinelli A, Gonzalez-Madruga K, Puzzo I, Wells A, Rogers JC, Clanton R, Baker RH, Grisley L, Baumann S, Gundlach M, Kohls G, Gonzalez-Torres MA, Sesma-Pardo E, Dochnal R, Lazaratou H, Kalogerakis Z, Bigorra Gualba A, Smaragdi A, Siklósi R, Dikeos D, Hervás A, Fernández-Rivas A, De Brito SA, Konrad K, Herpertz-Dahlmann B, Fairchild G, Freitag CM, Popma A, Kieser M and Stadler C (2017) Community Violence Exposure and Conduct Problems in Children and Adolescents with Conduct Disorder and Healthy Controls. Front. Behav. Neurosci. 11:219. doi: 10.3389/fnbeh.2017.00219

**28**

Information about CVE and aggression subtypes was obtained using self-report questionnaires (Social and Health Assessment and Reactive-Proactive aggression Questionnaire (RPQ), respectively). The association between witnessing community violence and conduct problems was significant in both groups (adolescents with CD and healthy controls). The association was also stable after examining the mediating effects of aggression subtypes while including moderating effects of age, gender and SES and controlling for effects of site in both groups. There were no clear differences between the groups in the strength of the association between witnessing violence and conduct problems. However, we found evidence for a ceiling effect, i.e., individuals with very high levels of conduct problems could not show a further increase if exposed to CVE and vice versa. Results indicate that there was no evidence for an ecological fallacy being the primary cause of the association, i.e., CVE must be considered a valid risk factor in the etiology of CD.

Keywords: community violence exposure, conduct disorder, reactive aggression, proactive aggression, adolescence, antisocial behavior

## INTRODUCTION

Community violence exposure (CVE) is defined as the witnessing of violence within a community, falling victim to violent acts oneself, or being subjected to a combination of both experiences (Schwab-Stone et al., 1995). CVE is a common and persistent public health issue in many inner city neighborhoods (Buka et al., 2001; Stein et al., 2003). Although the prevalence of CVE has been reported to be lower in many European countries compared to the US (Mercy et al., 2003; Hillis et al., 2016), it has nevertheless been recognized as a global public health problem by the World Health Organization (2002). A comprehensive meta-analysis of 114 studies on the effects of CVE on adolescent mental health by Fowler et al. (2009) found that the effects of CVE were strongest on the subsequent development of post-traumatic stress disorder, followed closely by externalizing problems. Specifically, effect sizes for the relationship between CVE and externalizing problems were 0.72–0.78 for witnessing violence and victimization, respectively. Fowler et al. (2009) found that relationship between exposure to CVE and externalizing behaviors was stronger in adolescents compared to children. Further factors that have been shown to influence the effects of CVE on mental health outcomes are gender and socioeconomic status (SES). Namely, being male (Javdani et al., 2014), and coming from lower SES strata (Anderson et al., 2001) have been found to increase the strength of the association between CVE and conduct problems.

Overall, research evidence to date suggests that the association of CVE and conduct problems is reciprocal, enhancing the chances of a negative spiral of increasing conduct problems and greater violence exposure. For instance, a bi-directional relationship between CVE and externalizing problems has been reported by Mrug and Windle (2009). The authors found that CVE was linked to the development of later conduct problems and delinquency. Likewise, baseline delinquency predicted higher rates of later CVE. Although there is much evidence indicating that violence exposure in early childhood is a major risk factor for the development of Conduct Disorder (CD; for a review see Burke et al., 2002), to date it is not known whether there are similarly strong associations between recent CVE and current conduct problems in adolescents with CD. If a reciprocal relationship between CVE and conduct problems exists, strong associations between recent CVE and current conduct problems would be expected.

Children and adolescents with CD constitute a group that is particularly prone to experiencing violence exposure due to the nature of their diagnosis. CD is defined as a repetitive and persistent pattern of violent and antisocial behavior (American Psychiatric Association, 2013). For children and adolescents with CD, it is difficult to separate CVE as a form of unintended environmental exposure from self-provoked situations that reflect part of the adolescent's symptomatology (Halliday-Boykins and Graham, 2001; Lynch, 2003). Children and adolescents with CD may encounter violent situations in ways other than as innocent bystanders, e.g., as a result of being present when a friend initiates a fight or robs a person or indeed as the perpetrator themselves. Much of the CVE literature focuses on community samples derived from urban, low socio-economic backgrounds, representing ethnic minorities and living in neighborhoods with high crime rates (Dempsey et al., 2000; Gorman-Smith et al., 2004; Frey et al., 2009; Copeland-Linder et al., 2010; Goldner et al., 2015). As such, these studies have likely included a mixture of healthy and clinically impaired youth. According to epidemiological research around 22.2% of adolescents within a national representative US sample reported a history of a psychiatric disorder that was accompanied by severe impairment or distress, of which 9.6% comprise behavioral disorders, such as CD or attention-deficit/hyperactivity disorder (Merikangas et al., 2010). For European countries specifically, it is estimated that around 38.2% of the general population of the European Union (EU) exhibit a mental disorder each year, with 5% of that proportion relating to externalizing behavior (Wittchen et al., 2011). Generally, adolescents from low-income neighborhoods exhibit greater mental health problems than those living in higher-income neighborhoods (Aneshensel and Sucoff, 1996). Although past CVE studies have offered unique insights into the debilitating effects of CVE on adolescents' psychosocial adjustment, the effects of CVE remain to be disentangled among a sample in which healthy and clinically-impaired individuals can be distinguished. Investigating these two groups separately allows precluding the presence of an ecological fallacy, i.e., the finding of stronger associations between CVE and conduct problems than is actually the case, due to the aggregation of healthy and clinically impaired adolescents. Specifically, insight into the associations between recent CVE and current conduct problems in an adolescent sample with CD and a healthy control sample will answer the following questions: does recent CVE continue to be of relevance in terms of determining current conduct problems in healthy adolescents as well as in those who have developed diagnosable levels of conduct problems, i.e., those with a CD diagnosis? Can we exclude an ecological fallacy that may have developed due to a lack of studies investigating the effects of CVE and conduct problems in an exclusive group of healthy adolescents vs. adolescents with CD?

Studies have shown that effects of CVE on later conduct problems persisted even when controlling for an individual's initial aggression level (e.g., Schwab-Stone et al., 1995; Farrell and Bruce, 1997; Miller et al., 1999; McCabe et al., 2005; Weaver et al., 2008). However, aggression is heterogeneous and may take different forms. Two key forms of aggression that are commonly distinguished are reactive and proactive aggression (for overview, see Kempes et al., 2005). Reactive aggression refers to impulsive forms of aggression, usually evoked by high arousal levels and strong emotions such as anger or fear. In contrast, proactive aggression is an instrumental, often pre-meditated form of aggression, characterized as callous and goal-oriented behavior and thought to be associated with low levels of arousal. Reactive aggression may be explained through the frustration-anger model (Dollard et al., 1939), explaining why this form of aggression is commonly linked to provocations or threats. Proactive aggression, on the other hand, is better understood through social learning theory (Bandura, 1973). This theory outlines why proactive aggression is often motivated by reward-orientation and is reinforced by positive outcomes following aggressive behavior. There have been more recent theories proposed since then which have set out hypotheses regarding the distinct neurobiological bases of these two aggression subtypes. For instance, reactive aggression has been linked to orbitofrontal cortex dysfunction and impaired emotion regulation (Bechara et al., 2000; Blair and Cipolotti, 2000) while proactive aggression is thought to be associated with amygdala dysfunction and a diminished response to distress cues (Blair, 1995, 2005). Research has shown that proactive but not reactive aggression may be predictive of later delinquency, conduct problems and violent offending in mid-adolescence as well as criminal behavior later in life (Pulkkinen, 1996; Vitaro et al., 1998; Raine et al., 2006). Conversely, reactive aggression was found to predict impulsivity and hostility (Raine et al., 2006).

Thus, when investigating the association between CVE and conduct problems, it is not only necessary to parse out effects of aggression but also to examine the role of each of these aggression subtypes separately. To date, it remains to be investigated how the relationship between recent CVE and current conduct problems may differ according to aggression subtypes. One possibility is that greater violence exposure is associated with more proactive aggression, perhaps because such exposure normalizes violence or leads to a desensitization to the effects of violence. More proactively aggressive children and adolescents, in turn, may intentionally choose to enter violent situations. Another possibility is that greater violence exposure is associated with more reactive aggression, possibly due to its effects on sensitivity to threat or even the neural circuits implicated in reactive aggression. Individuals with high levels of reactive aggression may, in turn, act out aggressively in response to CVE.

In summary, many studies have shown that there is a strong association between CVE and conduct problems. However, to date no study has investigated this association separately in children and adolescents with a diagnosis of CD and a sample exclusively made up of healthy controls to examine the deleterious effects of CVE separately in high and low-risk groups. Finally, the literature has not differentiated between reactive and proactive forms of aggression in terms of possible mediators of the association between CVE and conduct problems. We know that the relationship between CVE and aggression still holds when controlling for levels of prior aggression. Understanding how different types of aggression (i.e., reactive vs. proactive) may explain the link between CVE and conduct problems within healthy controls vs. children and adolescents with CD might be important for further specifying etiological models.

We had the following hypotheses:


### MATERIALS AND METHODS

This study was conducted within the framework of the ongoing European multi-disciplinary FP7 (i.e., European Commission's 7th Framework Health program) project ''Neurobiology and Treatment of Adolescent Female Conduct Disorder: The Central Role of Emotion Processing'' (FemNAT-CD). A detailed outline of the methodological aspects of the project is available on the official website<sup>1</sup> . Assessments were conducted at clinical sites from seven European countries: Germany, Greece, Hungary, Netherlands, Spain, Switzerland and the UK.

### Participants

Child and adolescent participants between the ages of 9 and 18 years were recruited through various means, including distribution of study information in schools, sports and leisure clubs, through street promotion and contacts with psychiatric clinics, youth offending services, or youth welfare institutions.

The inclusion criterion for the CD sample was a current diagnosis of CD according to DSM-IV-TR criteria (American Psychiatric Association, 2000). Exclusion criteria for both CD and control groups were a history and/or current diagnosis of autism spectrum disorder, schizophrenia, bipolar disorder or mania, fetal alcohol syndrome (all according to DSM-IV-TR), any known monogenetic disorders, chronic or acute neurological disorders, severe medical conditions or valid indications of an IQ < 70 (measured with the vocabulary and block design subtests of the Wechsler Intelligence Scale for Children (WISC) or vocabulary and matrix reasoning subtests of the Wechsler Adult Intelligence Scale (WAIS; Wechsler, 2008) depending on the participant's age; at UK sites, the Wechsler Abbreviated Scale of Intelligence was used for all ages (Wechsler, 1999). Additional exclusion criteria for healthy controls included any other current disorder according to DSM-IV-TR criteria as well as a past history of Attention-Deficit/Hyperactivity Disorder, Oppositional Defiant Disorder or CD.

From the current study sample, 1178 children and adolescents had complete data on all key measures and thus were included in the present analysis. Consistent with the aim of the study to over-recruit female participants, there were more females than males (62 vs. 38%) and slightly more than half of the overall sample was healthy controls (56%). The number of male children and adolescents was spread evenly across CD and control subjects (50% each). With regard to females, there were slightly more controls than CD subjects (60 vs. 40%). Comparison of the CD and control groups suggested that children and adolescents with CD were significantly older than controls (M age CD = 14.4, SD = 2.3 vs. M age controls = 13.9, SD = 2.6, t = 3.63, p < 0.001) and were characterized by significantly lower SES (M SES CD = −0.3, SD = 0.9 vs. M SES controls = 0.3, SD = 1.0, t = −9.36, p < 0.001).

### Procedure

Participants and their legal guardians received detailed study information via telephone, mail or email prior to the day of assessment. On the first assessment day, participants were given the opportunity to ask questions and it was assured that both parent/legal guardian and children knew that participation could be declined or stopped at any point during the course of the study. Written informed consent was obtained from the participants and their legal guardians. If consent of the legal guardian was unavailable, participants were included only if considered old enough according to the

<sup>1</sup>www.femnat-cd.eu

ethical requirements of the respective country (i.e., ≥16 in Switzerland and UK, ≥18 all else). Research was carried out in compliance with the fundamental ethical principles as stated by the Declaration of Helsinki and its later amendments as well as with the ethical standards of the institutional and/or national research committee. All subjects and legal guardians gave written informed consent in accordance with the Declaration of Helsinki. If it was not possible to obtain consent from the legal guardian, participants were included only if considered old enough to provide informed consent according to the ethical requirements of the respective country (i.e., <=16 years in Switzerland and the UK, <=18 years in all other countries). Ethical approval was obtained from all local ethics committees (Basel—Ethikkommission Nordwest- und Zentralschweiz, Frankfurt—Ethik-Kommission Fachbereich Medizin Klinikum der Johann Wolfgang Goethe-Universität, Aachen—Ethik-Kommission an der medizinischen Fakultät der rheinisch-westfälischen technischen Hochschule Aachen, Amsterdam—Medisch Ethische Toetsingscommissie Vrije Universiteit Medisch Centrum, Birmingham and Southampton—NHS Research Ethics Committee, Bilbao—Comite Etico de Investigacion Hospital Universitario Basurto, Barcelona—Comite Etico de Investigacion Clinica Parc de Salut Mar, Szeged—Human Reproduction Committee, Athens—Ethics Committee of Aiginiteio University Hospital of Athens).

In order to obtain information on mental health problems, semi-structured interviews were conducted with the children/adolescents and, if available, with their legal guardian in separate rooms/consecutively by trained, postgraduate-level investigators. Information obtained from both interviews was then combined to obtain a final summary judgment. Questionnaires assessing CVE were handed to the participant subsequent to the interview. Investigators were available to provide help to participants and clarify the meaning of items if requested. In line with the ethics committees' decision for the respective universities, participants were compensated with a gift card or a small monetary payment.

### Measures

### Community Violence Exposure

Initially developed by Richters and Saltzman (1990), and modified by Schwab-Stone et al. (1995, 1999) and Ruchkin et al. (2004), two scales of the Social And Health Assessment (Weissberg et al., 1991) assessing direct victimization as well as the witnessing of violence in the community, served as the measure of CVE. For the victimization scale, seven items assessed how often in the past year participants had been: (1) beaten up or mugged; (2) threatened with serious physical harm by someone; (3) threatened because of their race/ethnicity; (4) shot or shot at with a gun; (5) attacked or stabbed with a knife; (6) chased by gangs or individuals; or (7) seriously wounded in an incident of violence. Participants reported their answer on a 5-point Likert scale from 0 (never), 1 (1–2 times), 2 (3–5 times), 3 (6–9 times) to 4 (10 times or more). The witnessing scale included seven items asking the respondents how frequently they had seen someone else being exposed to the same violent acts in their community within the past year on the same 5-point Likert scale as described above. The two scales were found to have good psychometric properties in a sample of American inner-city youth (Richters and Martinez, 1993). With respect to our sample, the witnessing subscale produced a Cronbach's alpha of 0.89 for cases and 0.77 for controls, i.e., good/satisfactory internal consistency respectively. The victimization subscale resulted in an alpha of 0.81 for cases and 0.67 for controls indicating good/questionable internal validity (Bland and Altman, 1997).

### Conduct Problems

The Kiddie Schedule for Affective Disorders and Schizophrenia—Present and Lifetime (K-SADS-PL; Kaufman et al., 1997) is a semi-structured diagnostic interview used to screen for the current presence or lifetime history of a broad range of disorders ranging from affective disorders (i.e., depression, bipolar disorder), schizophrenia and substance use disorders through to externalizing disorders (i.e., CD, ADHD). The K-SADS-PL is administered independently to the adolescent as well as their caretaker to assess the presence of DSM-IV-TR psychiatric disorders (for this study, the CD present section was used). Summary ratings are derived from clinical judgment using both interview sources as well as other information available on file. The items of the instrument are scored on a scale from 0 to 3. A rating of 0 indicates no (insufficient) information, a score of 1 indicates a given symptom is not present, 2 indicates a subclinical expression, while a score of 3 is given when a symptom is present and clinically significant. Scores were recoded, so that a clinical rating of ''not present'' is represented by 0, a subclinical rating by a score of 1, and a clinically significant rating by a score of 2. Unknown ratings were recoded into missing data. For the purpose of the current study, mean item scores were calculated for CD based on the current summary ratings. That is, a mean score was calculated across all CD symptoms for each individual. This procedure allowed inspection of current conduct problems at the symptom-level and therefore represented a more comprehensive estimation of problematic behavior symptoms with regard to the healthy control group. In other words, conduct problems were assessed on a dimensional level including subclinical expressions to assess the association between CVE and conduct problems. Inter-rater reliability for the K-SADS-PL section used was based on 75 CD individuals and found to be almost perfect with a percentage agreement of 94.67 and Cohen's κ of 0.907 (95% CI: 0.819–0.995; Landis and Koch, 1977).

### Aggression Subtypes

Developed by Raine et al. (2006), the Reactive-Proactive aggression Questionnaire (RPQ) measures self-reported reactive (11 items, e.g., ''I have damaged things because I felt mad'', ''I have gotten angry when frustrated'', ''I have had temper tantrums'') and proactive (12 items, e.g., ''I have had fights to show that I was on top'', ''I have vandalized something for fun'', ''I have gotten others to gang up on someone'') aggression. Each item is answered on a 3-point Likert scale (0 = never, 1 = sometimes, 2 = often). The two scales were found to have good internal validity (0.84 for reactive, 0.86 for proactive aggression) in its original evaluation study comprising a sample of American school boys (Raine et al., 2006). Further validation studies in several countries have subsequently confirmed reliability and validity across the genders and various populations (e.g., different age groups, non-offender vs. criminal samples; Fossati et al., 2009; Fung et al., 2009; Cima et al., 2013). With regard to the present sample, Cronbach's alpha was 0.84 for adolescents with CD and 0.79 for controls for the reactive aggression subscale indicating good to satisfactory internal consistency. The proactive subscale yielded estimates of 0.85 for adolescents with CD and 0.67 for controls indicating good/questionable internal consistency (Bland and Altman, 1997).

### Socioeconomic Status

SES was calculated based on parental income, education as well as occupational status. Classifications were made using the International Classification of Education (UNESCO Institute for Statistics, 2015) and the International Standard Classification of Occupations (International Labour Organization, 2012). Human rater and computer-based ratings were combined into a standardized factor (M = 0, SD = 1) score using Principal Component Analysis. Internal consistency of the composite SES score was acceptable (α = 0.74). Due to potential economic variation on the country level, SES was centered and scaled within each country, in order to obtain an indicator of relative socioeconomic position.

### Data Analysis

The data were analyzed using the Statistical Package for Social Sciences (SPSS-23; IBM Corp, 2016, Armonk, NY, USA), Analysis of Moment Structures (AMOS-23; Arbuckle, 2014) and R (R Core Team, 2013) with the packages ''plot3D'', ''ggplot2'' and ''localgauss''. For descriptive results sample mean scores were calculated for witnessing, victimization, conduct problems, reactive and proactive aggression measures to characterize the two groups. Furthermore, Mann-Whitney-U tests and two-sample sample t-tests were calculated to gain more insight into group differences. Local Gaussian correlations and 2-D plots were computed to approximate density functions and obtain further insight into the distribution of CVE and conduct problems within the two groups. Structural equation modeling (SEM) was used for analyzing the primary models since it allowed us to compare the model fit of successively nested models with each other. In all SEM models age, gender, site and SES were used as control variables. For the final model age, gender and SES were inspected as moderating variables, while site served as a control variable.

Analyses were conducted on two different CVE constructs: (1) witnessing violence was examined as a latent variable by parceling the seven items comprising this scale into three indicators; (2) using the same procedure, victimization was inspected separately as well. Through the use of latent variables, we were able to reduce measurement errors and improve the


TABLE 1 | Parcel composition and standardized loadings of parceled indicators by group.

accuracy of the findings (Little et al., 1999). The method of parceling was chosen in order to overcome low communality and reliability frequently encountered with the use of individual items and to decrease the likelihood for distributional violations (Little et al., 1999, 2002). Items were grouped into parcels based on item-total correlations (Little et al., 1999, 2002). Items with highest and lowest item-total correlations were grouped together, resulting in two groups with two items and one group with three items (see **Table 1** for detailed list of parcel composition). The parceled indicators loaded well on their respective factors with loadings ranging between 0.59 and 0.88 (see **Table 1**).

Chi-square, the Root Mean Square Error of Approximation (RMSEA; Browne and Cudeck, 1993) and the Comparative Fit Index (CFI; Bentler, 1990) were used as indicators of goodness of fit. While for the commonly employed Chi-Square test greater (insignificant) p-values generally indicate better fit, the RMSEA requires values of 0.05 or less and CFI values of 0.95 or greater to consider a model to be of acceptable fit (Bentler, 1990; Browne and Cudeck, 1993).

To test the mediation hypotheses, change in model fit when direct paths from CVE to conduct problems were removed was assessed controlling for site and moderating effects of age, SES and gender (Holmbeck, 1997). Additionally, the magnitude of the indirect effects of CVE on conduct problems via reactive or proactive aggression was estimated (Holmbeck, 1997). For the group comparison, a series of up to four models were examined for cases and controls separately, including comparisons between an exploratory model in which all paths between the variables were free to vary for each group. This model served to pinpoint the variables of interest for each group. Then, a fully constrained model was examined, in which all primary paths were set as equal for both groups. Consequently, the second model hypothesized no group differences for all associations/paths. If this model was true, constraining the paths to the same value should not significantly decrease the overall model fit (as compared to the first model). If indicated (i.e., if second model significantly decreased model fit) a third model, in which some selected paths were non-constrained, was inspected. These selected paths were identified by re-examining the paths of the first model and selecting potentially different associations between patients and controls. The selected paths were unconstrained and allowed to differ by group. If the model fit significantly improved compared to the second model (and was not worse than that for the first model), it would suggest the presence of group differences in the model. A final select model would then be produced in which all insignificant paths are deleted and again compared against the model fit of the previous model.

### RESULTS

### Descriptive Analyses

Children and adolescents with CD reported significantly greater CVE within the past year than healthy controls for both witnessing violence, U = 88840, p < 0.001 (M witnessing CD = 0.62, SD = 0.75 vs. M witnessing controls = 0.13, SD = 0.29) and victimization, U = 97557, p < 0.001 (M victimization CD = 0.28, SD = 0.46 vs. M victimization controls = 0.03, SD = 0.11; see **Figure 1**). As healthy controls rarely reported victimization events within the past year, only the witnessing violence subscale of the SAHA was included in all further analyses. In both groups the distribution of CVE was skewed as many individuals reported zero to low frequency of exposure within the past year. **Table 2** presents the means and prevalence rates (i.e., the percentage of individuals having experienced the respective item at least once within the past year) of each witnessing item by group and shows that children and adolescents with CD experienced all of the listed events to a much greater extent than their healthy counterparts. Supplementary Table S1 (presented in Supplementary Material) shows the exact percentage of endorsed frequencies within the past year by group. In both groups, ''threats with physical harm'' was the most frequently endorsed form of violence exposure (49.5% vs. 16.0%), while ''getting shot'' was the least frequently encountered event (12.2% vs. 0.9%) by children and adolescents with CD and healthy controls, respectively. In addition, means and standard deviations for reactive and proactive aggression are presented in **Table 2**.

**Figures 2A,C** shows a broader range of witnessed violence for adolescents with CD than for healthy controls, although high values for witnessing are rare amongst both groups. For controls, recent witnessing and current conduct problems have a highly left-skewed distribution, since most individuals have low current conduct problems and low frequency of recent exposure to witnessed violence. For adolescents with CD, current conduct problems are more normally distributed (**Figure 2C**). **Figure 2B** shows that for both adolescents with CD and healthy controls a positive linear trend (green line) between recent witnessing and current conduct problems can be observed. Testing the significance of the association of CVE and conduct problems between groups in a SEM model, while controlling for site effects as well as the moderating effects of age, gender and SES indeed revealed significant associations for both groups (CD: beta = 0.36, p < 0.001; and controls: beta = 0.20, p < 0.001). A Loess fitting function (red line) shows that the fit line flattens or even becomes

slightly negative in the higher range of current conduct problems or witnessing violence. This finding can be corroborated assessing local Gaussian parameters (**Figure 2D**): Local Gaussian parameters show a positive trend across the whole spectrum of current conduct problems/witnessing violence, whereas in the higher range of both variables, the association becomes neutral (white) or even negative (purple). These findings indicate a ''ceiling'' effect, i.e., that beyond a high level of current conduct problems, witnessing violence is not able to increase the level of symptoms, and vice versa.

### Mediation Analysis of Reactive and Proactive Aggression in the Overall Sample

Reactive and proactive aggression partially mediated the relationship between witnessing community violence and conduct problems. Mediation was tested by a 1 df Chi-Square change test of the model with and without estimating direct paths from witnessing community violence to conduct problems (also refer to **Table 3**, ''Introduction'' section). The model fit changed significantly after deleting the respective path: χ 2 (19) = 152.21, p < 0.001; RMSEA = 0.075, CI = 0.064–0.086; CFI = 0.969; change in χ 2 (1) = 64.46, p < 0.001, suggesting that reactive and proactive aggression were not acting as full mediators. In order to assess partial mediation further, the path between witnessing community violence and conduct problems was constrained to the regression weight of a direct effects model. A significant change in model fit (χ 2 (19) = 241.75; RMSEA = 0.097, CI = 0.086–0.108; CFI = 0.948; change in χ 2 (1) = 154.00; p < 0.001) supported partial mediation. The association between witnessing community violence and conduct problems (i.e., the path coefficient) still remained highly significant (beta = 0.25, p < 0.001) even when accounting for the indirect effects of proactive and reactive aggression.

### Proactive and Reactive Aggression as Mediators between CVE and Conduct Problems in Adolescents with Conduct Disorder and Healthy Controls

A multilevel analysis examining the mediating effects of reactive and proactive aggression on the association between witnessing community violence and conduct problems separately within children and adolescents with CD and healthy controls and controlling for site and moderating effects of SES, age and gender revealed partly differential effects of aggression subtypes between groups.

When examining the impact of witnessing violence, we followed a series of models to determine the best explanatory model (see ''Materials and Methods'' section above). A first unconstrained model fit the data well (see **Table 3**, ''Materials and Methods'' section). A constrained model significantly decreased model fit, indicating that there were at least some differences in the model paths between adolescents with CD and controls (χ 2 (41) = 188.07, p < 0.001; RMSEA = 0.054, CI = 0.046–0.061; CFI = 0.942; change in χ 2 (5) = 58.19, p < 0.001). Re-examination of the first model and chi-square change tests confirmed significant increases in model fit when compared to the constrained model for all primary paths except for the path between reactive aggression and witnessing violence (1. path between witnessing violence and conduct problems: χ 2 (40) = 174.27, p < 0.001; RMSEA = 0.052, CI = 0.044–0.060; CFI = 0.947; change in χ 2 (1) = 13.80, p < 0.001; 2. path between reactive aggression and conduct problems: X 2 (40) = 165.06, p < 0.001; RMSEA = 0.050,



<sup>1</sup>Values represent corresponding standard deviations.

CI = 0.042–0.058; CFI = 0.951; change in χ 2 (1) = 23.01, p < 0.001; 3. path between proactive aggression and conduct problems: χ 2 (40) = 150.99, p < 0.001; RMSEA = 0.047, CI = 0.039–0.055; CFI = 0.956; change in χ 2 (1) = 37.08, p < 0.001; 4. path between proactive aggression and witnessing violence: χ 2 (40) = 184.18, p < 0.001; RMSEA = 0.054, CI = 0.046–0.062; CFI = 0.943; change in χ 2 (1) = 3.89, p < 0.05). Therefore, all other paths were unconstrained in a selected-paths-free model. This select model fit the data significantly better than the constrained model (χ 2 (37) = 130.78, p < 0.001; RMSEA = 0.045, CI = 0.037–0.053; CFI = 0.963; change in χ 2 (4) = 57.29, p < 0.001). Further, the select model did not significantly decrease model fit compared to the unconstrained model (change in χ 2 (1) = 0.9, n.s.). In a final step, all non-significant paths were deleted. The fit of this final model was not significantly worse than that of the select model (change in χ 2 (14) = 15.47, n.s.). Therefore, **Figure 3** contains the standardized path coefficients from the final, most parsimonious model.

Across groups, the coefficients demonstrate a partial mediation effect. Further, in both groups both aggression subtypes had a significant impact on the association between witnessing violence and conduct problems, however, proactive aggression showed a stronger effect. Between groups, children and adolescents with CD showed a significantly stronger link between witnessing violence, proactive aggression and conduct problems compared to controls.

Furthermore, moderating effects of age, gender and SES were considered. Across groups, age played a significantly moderating role with regard to witnessing violence, such that older individuals tended to witness more violent events (CD: beta = 0.22, p < 0.001; controls: beta = 0.14, p < 0.001). In addition, older children and adolescents with CD showed significantly more conduct problems compared to their younger counterparts (beta = 0.08, p < 0.05). With regard to gender, male controls witnessed significantly more violence than female controls (beta = 0.08, p < 0.05). Finally, controls with a lower SES tended to

#### TABLE 3 | Fit statistics for all models.


<sup>1</sup>Refers to the two-group analysis, i.e., CD and controls analyzed separately; <sup>2</sup>assumption of no group differences; <sup>3</sup> full mediation model; <sup>4</sup>partial mediation model.

show significantly more conduct problems than control subjects with a higher SES background (beta = −0.09, p < 0.05).

### DISCUSSION

The present study demonstrated that recent witnessing of community violence is strongly positively associated with levels of current conduct problems in both children and adolescents with CD as well as healthy controls. Furthermore, the association between witnessing community violence and conduct problems remained significant in both groups even when including aggression subtypes (i.e., reactive/proactive aggression) as mediators in the model while controlling for the moderating effects of SES, gender and age and accounting

Note: results for CD group are displayed in bold, with those for the control group presented in normal font; coefficients in brackets represent the direct path model (analyses were controlled for site); only significant coefficients are displayed. ∗∗∗p < 0.001; ∗∗p < 0.01; <sup>∗</sup>p < 0.05.

for site effects. Proactive aggression had a stronger impact on the association between witnessing community violence and conduct problems than reactive aggression in both children and adolescents with CD and controls. When comparing the two groups, proactive aggression accounted for a greater proportion of the relationship between witnessing violence and conduct problems in children and adolescents with CD. Increased age was associated with greater rates of witnessing violence in both groups, while it was additionally associated with greater conduct problems in adolescents with CD. For controls, a lower SES was associated with greater conduct problems and being male was associated with greater exposure to witnessed violence. As such, the results of the present study are in line with findings on the well-established link between CVE and conduct problems, and extend the existing literature by demonstrating that: (1) the association between recent witnessing and current conduct problems is strongly detectable even in a group with a formal diagnosis of CD as well as a group with no clinical impairments; (2) the associations between recent witnessing of violence and current conduct problems persist even when accounting for reactive/proactive aggression across adolescents with CD and healthy controls; (3) the association between recent witnessing of violence and current conduct problems is primarily explained by proactive aggression across the two groups. Thus, the present study highlights the importance of not only taking into account early childhood risk factors known to predict the development of conduct problems and CD (Loeber et al., 2009; Bernhard et al., 2016), but also focusing on current factors in the young person's life, such as witnessing community violence, that are likely to maintain or exacerbate conduct problems.

The present study indicates that a strong association between recent witnessing of violence and current conduct problems exists even in groups characterized by the absence of any clinically significant impairment or the presence of CD. Past studies have never investigated the relationship between CVE and conduct problems in an exclusively healthy population nor in an adolescent sample with CD. The fact that current findings indicate a robust positive association in both groups allows us to reject the possibility of an ecological fallacy due to potential confounding effects within community samples comprising a mixture of healthy and clinically impaired adolescents.

Furthermore, the present finding that recent witnessing of community violence has a strong impact on current conduct problems in children and adolescents with CD as well extends the results of studies suggesting that greater levels of violence exposure and high levels of conduct problems tend to co-occur (Sanchez et al., 2013; Cecil et al., 2014; Voisin et al., 2016). Results of a longitudinal study indicated bi-directional effects between CVE and conduct problems, suggesting a downward spiral (Mrug and Windle, 2009). The present finding of a strong association for children and adolescents with CD might point to a similar pattern. However, findings also underline that in the presence of a high rate of CVE as well as severe levels of conduct problems, eventually a ceiling effect sets in, where the strength of association between the two variables is reduced or even becomes negative. Gaylord-Harden et al. (2017) specifically investigated the cumulative impact of CVE on psychopathology in a male adolescent community sample. CVE was found to show a curvilinear relationship with internalizing problems and a positive linear relationship with violent behavior. Present findings suggest that in the case of an extremely violent group with high levels of CVE at baseline, effects of CVE on subsequent violent behavior may be much smaller due to ceiling effects.

Self-reported proactive and reactive aggression subtypes did not fully explain the link between recent witnessing of community violence and current conduct problems for children and adolescents with CD as well as for healthy controls. As such, the present findings are in line with studies indicating persisting effects of CVE on conduct problems when controlling for baseline levels of aggression (Schwab-Stone et al., 1995; Farrell and Bruce, 1997; Miller et al., 1999; McCabe et al., 2005; Weaver et al., 2008). Specifically, the finding that the association between witnessing violence and conduct problems remains when controlling for aggression subtypes suggests that it is not just the CD individual's own level of reactive/proactive aggression that explains the link between witnessing community violence and conduct problems.

Furthermore, results of the present study showed that, for both groups, proactive aggression had a stronger mediational effect on the link between witnessing violence and conduct problems than reactive aggression. This finding aligns with studies on chronic CVE and associated desensitization processes (i.e., emotional numbing and use of aggression increasingly being seen as acceptable) resulting in higher levels of externalizing behavior (Ng-Mak et al., 2004; Boxer et al., 2008; Mrug et al., 2016; Gaylord-Harden et al., 2017). Translating these findings to the present study, one might expect that those children and adolescents witnessing more violence in their neighborhoods may have also undergone desensitization processes that have led them to become more proactively aggressive. In turn, these desensitized children and adolescents might also more readily seek out situations where violence is likely to occur or commit acts of violence which they also ''witness''.

Reactive aggression, on the other hand, explained less of the association between witnessing violence and conduct problems than proactive aggression. In addition, the mediating effects of reactive aggression on the relationship between witnessing violence and conduct problems were similar in children and adolescents with CD and healthy controls. This result is surprising in the context of studies that have identified impulsivity as a correlate and predictor of CVE (Lambert et al., 2010) as well as of studies that have identified impulsivity as a relevant moderator of the effects of CVE on adolescent deviant behavior (Low and Espelage, 2014). However, Monahan et al. (2015) previously suggested that declines in impulsecontrol happen only in response to increases in CVE rates independent of an individual's baseline rate (Monahan et al., 2015). While we have not directly investigated this, the impact of differences in CVE exposure rates over time (e.g., a change that might occur if one moves from a high to a low violence neighborhood) is something that should be considered in future studies. Furthermore, the current findings could partly be explained by the type of CVE assessed. Studies have shown that as the proximity of the exposure increases, the effects on emotional distress and internalizing symptoms increase as well (Fowler et al., 2009). Compared to direct victimization, witnessing community violence might evoke less emotional arousal and thus less of a response due to the more distal nature of the exposure. Individuals might feel less personally involved and are thereby less likely to act out in response.

Finally, consistent with the literature (Fowler et al., 2009), older adolescents in both groups tended to witness more violent events compared to their younger counterparts. In addition, older adolescents with CD showed more conduct problems than children. This finding aligns with studies identifying adolescence as the peak time for the majority of referrals to child and adolescent psychiatric clinics (Loeber et al., 2000) and studies that show that self-reported rates of violent offending are highest at age 16–17 years (Elliott, 1994). For controls, being male was linked to greater exposure to violence, while coming from a lower SES background was associated with greater conduct problems. Both findings have been demonstrated previously (Anderson et al., 2001; Buka et al., 2001; Javdani et al., 2014). The fact that gender and SES exerted no influence on levels of CVE and conduct problems for children and adolescents with CD might be based on the fact that the group represents a homogenous group. That is, children and adolescents with CD came from lower than average SES strata and the females within this group exhibited clinically significant levels of conduct problems. Furthermore, all children and adolescents with CD were exposed to some level of CVE and displayed some level of conduct problems by definition. As such, gender and SES would be expected to have less impact in this population compared to a more representative sample.

### Strengths and Limitations

While this study adds to our understanding of the specific associations and mediation effects regarding the link between CVE and conduct problems, it remains a snapshot of the sample's situation at the time of assessment, i.e., a crosssectional investigation. Hence, it is limited when it comes to exploring pathways of CVE to conduct problems as well as the long-term or cumulative effects of CVE. Consequently, it cannot shed light on whether CVE and/or reactive/proactive aggression precede the development of conduct problems or emerge as a consequence of conduct problems. Future studies using longitudinal designs will be able to shed more light on the consequences and persistence of such effects. Furthermore, it has been shown that cross-sectional approaches to mediation may result in misleading estimates (Maxwell et al., 2011). Again, it would be more valid to apply the present mediation models to data collected as part of a longitudinal study with repeated measurements of CVE, conduct problems and aggression subtypes.

Due to very low victimization rates in healthy children and adolescents, here we only focused on witnessing community violence. Fowler et al. (2009) concluded that witnessing and victimization were equally predictive of externalizing behaviors. Nevertheless, it would have been interesting to see whether reactive aggression might play a greater role in mediating the association between direct victimization and conduct problems.

In the present study, only age, gender and SES were considered as additional moderators. Aside from these variables, some additional important factors identified in previous research have been family structure, school characteristics and peer relationships (Buka et al., 2001; Chen et al., 2016). These variables present further risk or protective factors in the relationship between CVE and psychopathology but were unavailable to the authors. Future studies should therefore take these moderators into account when analyzing differences between children and adolescents with CD and healthy controls.

Finally, information obtained with regard to CVE and aggression subtypes relied on self-report data and may have been subject to social desirability effects (e.g., respondents exaggerating the extent of community violence in their neighborhoods to sound tough). Particularly with regard to CVE, past studies have found differences between informants, with parents reporting lower CVE rates for their offspring than the adolescents themselves (Kuo et al., 2000). As the present sample was mostly comprised of adolescents, it seemed safer to rely on self-report data in order to avoid potential under-reporting.

Despite these limitations, the present study is the first to systematically investigate and disentangle the effects of witnessing community violence and conduct problems in a clinical population (i.e., children and adolescents with CD) and a healthy population. As such, it is the first study to demonstrate that recent witnessing is related to current conduct problems in an exclusively healthy sample as well as in children and adolescents with CD even when taking their levels of reactive/proactive aggression into account. Further, this study is the first to illustrate that recent CVE is associated with the level of current conduct problems and therefore may play a role in the development or maintenance of conduct problems even for those with pre-existing externalizing behavior.

### Implications

An important implication for etiological models of the development of conduct problems is that neighborhood violence might be an important contributing factor, for healthy youths and particularly for children and adolescents with CD. From a clinical perspective, the strong association between witnessing community violence and conduct problems highlights the need for prevention and/or intervention strategies, as the relationship between the two variables is likely to intensify over time. The present results emphasize the need to consider recent CVE in addition to early risk factors. Correspondingly, it has been demonstrated that multimodal intervention programs with an additional focus on the adolescent's environment (e.g., Multisystemic Therapy, Multidimensional Family Therapy) are more effective in reducing conduct problems as opposed to programs that do not take an individual's environment into account (Weisz and Kazdin, 2010; National Collaborating Centre for Mental Health, Social Care Institute for Excellence, 2013).

An important direction for future research will be to test for relationships between CVE and conduct problems in young people with CD and controls using a longitudinal design. Based on the findings of Fowler et al. (2009), we know that the strongest relationships hold between lifetime measures of externalizing behavior and lifetime CVE illustrating the cumulative impact of CVE on conduct problems over time (Fowler et al., 2009).

### CONCLUSION

Witnessing community violence is a highly prevalent experience for children and adolescents in Europe, and is strongly associated with the individual's level of current conduct problems, in healthy controls as well as in children and adolescents with CD. As such, the present study is able to show a robust relationship between recent CVE and conduct problems not only in a clinically impaired sample but also in a healthy group, thereby reducing the possibility that previously reported associations between these variables were explained by an ecological fallacy. Furthermore, the present study demonstrates a strong association between recent CVE and current conduct problems, which is primarily mediated by proactive aggression. The challenge for the future lies in breaking the dangerous cycle of young people being exposed to community violence, and going on to perpetrate violence against others as a result.

## AUTHOR CONTRIBUTIONS

CMF coordinates the FemNAT-CD FP7 research project. DD, AH, AF-R, SADB, KK, BH-D, GF, CMF, AP, MK and CS designed the study and took over site-specific coordination of the FemNAT-CD FP7 research project. LK, MP, HO, RV, LJ, KA, AB, AM, KG-M, IP, AW, JCR, RC, RHB, LG, SB, MG, GK, MAG-T, ES-P, RD, HL, ZK, ABG, AS and RS recruited subjects and collected data. LK drafted the manuscript together with CS, NV and MS. NMR, MP, SADB, KK, GF, CMF, AP and MK helped in manuscript preparation and critically reviewed the article.

### FUNDING

The present study is part of the FemNAT-CD consortium (Neurobiology and Treatment of Adolescent Female Conduct Disorder: The Central Role of Emotion Processing). This collaborative project is funded by the European Commission under the 7th Framework Health Program with Grant Agreement no. 602407.

### REFERENCES


### ACKNOWLEDGMENTS

We would like to thank all children and adolescents, families and institutions participating in the project, as well as members of the FemNAT-CD consortium.

### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fnbeh. 2017.00219/full#supplementary-material


from the National Comorbidity Survey Replication-Adolescent Supplement (NCS-A). J. Am. Acad. Child Adolesc. Psychiatry 49, 980–989. doi: 10.1016/j. jaac.2010.05.017


World Health Organization. (2002). The World Health Report 2002: Reducing Risks, Promoting Healthy life. Geneva, Switzerland: World Health Organization.

**Conflict of Interest Statement**: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Kersten, Vriends, Steppan, Raschle, Praetzlich, Oldenhof, Vermeiren, Jansen, Ackermann, Bernhard, Martinelli, Gonzalez-Madruga, Puzzo, Wells, Rogers, Clanton, Baker, Grisley, Baumann, Gundlach, Kohls, Gonzalez-Torres, Sesma-Pardo, Dochnal, Lazaratou, Kalogerakis, Bigorra Gualba, Smaragdi, Siklósi, Dikeos, Hervás, Fernández-Rivas, De Brito, Konrad, Herpertz-Dahlmann, Fairchild, Freitag, Popma, Kieser and Stadler. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Tracing the Neural Carryover Effects of Interpersonal Anger on Resting-State fMRI in Men and Their Relation to Traumatic Stress Symptoms in a Subsample of Soldiers

#### Edited by:

Nelly Alia-Klein, Icahn School of Medicine at Mount Sinai, United States

#### Reviewed by:

Gabriela Gan, Zentralinstitut für Seelische Gesundheit, Germany Dardo G. Tomasi, National Institutes of Health (NIH), United States Birgit Derntl, Universität Tübingen, Germany

\*Correspondence:

Gadi Gilam gadi.gilam@gmail.com Talma Hendler talma@tlvmc.gov.il

†These authors have contributed equally to this work.

‡Present address:

Gadi Gilam, Division of Pain Medicine, Systems Neuroscience and Pain Laboratory, Department of Anesthesiology, Perioperative and Pain Medicine, Stanford University School of Medicine, Palo Alto, CA, United States

Received: 05 June 2017 Accepted: 11 December 2017 Published: 20 December 2017

#### Citation:

Gilam G, Maron-Katz A, Kliper E, Lin T, Fruchter E, Shamir R and Hendler T (2017) Tracing the Neural Carryover Effects of Interpersonal Anger on Resting-State fMRI in Men and Their Relation to Traumatic Stress Symptoms in a Subsample of Soldiers. Front. Behav. Neurosci. 11:252. doi: 10.3389/fnbeh.2017.00252 Gadi Gilam1,2 \* †‡ , Adi Maron-Katz 1,3,4† , Efrat Kliper <sup>1</sup> , Tamar Lin1,2 , Eyal Fruchter <sup>5</sup> , Ron Shamir 4,6 and Talma Hendler 1,2,3,6 \*

<sup>1</sup>The Tel Aviv Center for Brain Function, Wohl Institute for Advanced Imaging, Tel Aviv Sourasky Medical Center, Tel Aviv, Israel, <sup>2</sup>School of Psychological Sciences, Tel-Aviv University, Tel Aviv, Israel, <sup>3</sup>Sackler Faculty of Medicine, Tel-Aviv University, Tel Aviv, Israel, <sup>4</sup>Blavatnik School of Computer Science, Tel-Aviv University, Tel Aviv, Israel, <sup>5</sup>Division of Mental Health, Israeli Defense Force Medical Corp, Haifa, Israel, <sup>6</sup>Sagol School of Neuroscience, Tel-Aviv University, Tel Aviv, Israel

Uncontrolled anger may lead to aggression and is common in various clinical conditions, including post traumatic stress disorder. Emotion regulation strategies may vary with some more adaptive and efficient than others in reducing angry feelings. However, such feelings tend to linger after anger provocation, extending the challenge of coping with anger beyond provocation. Task-independent resting-state (rs) fMRI may be a particularly useful paradigm to reveal neural processes of spontaneous recovery from a preceding negative emotional experience. We aimed to trace the carryover effects of anger on endogenous neural dynamics by applying a data-driven examination of changes in functional connectivity (FC) during rs-fMRI between before and after an interpersonal anger induction (N = 44 men). Anger was induced based on unfair monetary offers in a previously validated decision-making task. We calculated a common measure of global FC (gFC) which captures the level of FC between each region and all other regions in the brain, and examined which brain regions manifested changes in this measure following anger. We next examined the changes in all functional connections of each individuated brain region with all other brain regions to reveal which connections underlie the differences found in the gFC analysis of the previous step. We subsequently examined the relation of the identified neural modulations in the aftermath of anger with state- and trait- like measures associated with anger, including brain structure, and in a subsample of designated infantry soldiers (N = 21), with levels of traumatic stress symptoms (TSS) measured 1 year later following combat-training. The analysis pipeline revealed an increase in right amygdala gFC in the aftermath of anger and specifically with the right inferior frontal gyrus (IFG).We found that the increase in FC between the right amygdala and right IFG following anger was positively associated with smaller right IFG volume, higher trait-anger level and among soldiers with more TSS. Moreover, higher levels of right amygdala gFC at baseline predicted less reported anger during the subsequent anger provocation. The results suggest that increased amygdala-IFG connectivity following anger is associated with maladaptive recovery, and relates to long-term development of stress symptomatology in a subsample of soldiers.

Keywords: anger, rumination, recovery, stress, PTSD, amygdala, IFG, fMRI

### INTRODUCTION

''Do not mix anger with profusion and set them before your guests. Profusion makes its way through the body, and is quickly gone: but anger, when it hath penetrated the soul, abides for a long time.'' -Epictatus

Anger is experienced on a daily basis and mostly during or following social interactions (Averill, 1983; Baumeister et al., 1990), possibly leading to aggression and violence towards the environment (Anderson and Bushman, 2002; Rosell and Siever, 2015). Excessive anger may also have negative consequences on one's health, wellbeing and social rapport (Johnson, 1990; Williams, 2010). Moreover, uncontrolled anger is prevalent in numerous psychopathological conditions (Novaco, 2010), such as in post-traumatic stress disorder (PTSD; Olatunji et al., 2010). The importance of regulating anger and adapting it to socially accepted norms is thus unequivocal (Davidson, 2000; Gilam and Hendler, 2015). Notably, as Epictatus's quotation illustrates, the challenge of coping with and recovering from anger extends beyond the termination of the angerinducing provocation since feelings of anger tend to linger and outlast the provocation itself (Potegal, 2010). In fact, individuals with a chronic tendency to be angry (i.e., high in trait-anger) have difficulties in disengaging from such lingering anger, paralleled by impaired recruitment of regulatory resources (Sukhodolsky et al., 2001; Wilkowski and Robinson, 2010).

There are various emotion regulation strategies one may recruit to cope with anger, some more adaptive than others. For example, cognitive reappraisal, which refers to the reinterpretation of an emotional event, was shown to be effective in reducing angry feelings (e.g., Ray et al., 2008; Szasz et al., 2011). On the other hand, angry rumination, which refers to recurrent thought patterns on causes and consequences of the angering episode, was shown to intensify and prolong the experience of anger and increase subsequent aggression (e.g., Bushman et al., 2005; Pedersen et al., 2011). Indeed, rumination is considered a maladaptive regulatory response, common also in PTSD patients (Michael et al., 2007).

The neural bases of emotion regulation generally engages prefrontal cortex (PFC) regions such as the ventro-medial PFC (vmPFC) and inferior frontal gyrus (IFG) which exert control over emotion reactivity regions such as the amygdala and insula (Diekhof et al., 2011; Frank et al., 2014; Etkin et al., 2015). In the context of an anger experience, it was previously demonstrated that while reappraisal and rumination similarly engaged the activation of such regions, differences between these strategies emerged only in connectivity (Fabiansson et al., 2012). Specifically, there was a positive correlation between the IFG and both amygdala and thalamus during rumination but not reappraisal. However, emotion regulation was explicitly instructed in that study, limiting our understanding of naturalistic anger regulation. We recently developed a modified Ultimatum Game in which participants faced unfair angerinducing monetary offers infused with verbal provocations by a competitor (Gilam et al., 2015). We demonstrated that vmPFC activation and posterior insula-medial thalamus connectivity modulated angry feelings leading to increased acceptance of unfair offers and thus gaining more money throughout the game, providing a neural model for spontaneous anger regulation. In the current study, we aimed to trace the changes in spontaneous neural processing in the aftermath of an angering episode, assuming this may shed light on endogenous regulatory processes enabling recovery from such turmoil. This may also reveal if recovering from anger in its aftermath engages similar or different neural processes as regulating anger during on-going anger provocation and potentially inform efforts to mitigate the negative implications of anger on people's lives.

The resting-state (rs) fMRI paradigm in which participants' brain is scanned while they let their thoughts wander without any instruction (Gruberger et al., 2011) may be particularly relevant to trace changes in spontaneous neural dynamics related to an immediately preceding emotional experience. A few previous studies adopted such an approach in regards to an induced emotional experience, mostly revealing increased neural coupling between prefrontal, limbic and paralimbic brain regions including the medial PFC, amygdala, cingulate and insular cortices, some of which were associated with sustainment of the emotional experience (e.g., Harrison et al., 2008; van Marle et al., 2010; Veer et al., 2011; Schultz et al., 2012; Vaisvaser et al., 2013; Maron-Katz et al., 2016; Clemens et al., 2017). In some of these studies a data-driven analysis was used instead of a seed-based analysis, having the benefit of being independent of any prior hypothesis and of an unbiased identification of brain regions in the entire brain (e.g., Harrison et al., 2008; Maron-Katz et al., 2016).

To reveal neural dynamics of an angering experience beyond its immediate occurrence, we used a data-driven analysis to examine changes in rs-functional connectivity (FC) from before to after an interpersonal angering experience. At first we performed a whole-brain analysis using a measure of global FC (gFC) which captures the level of FC between each region and all other regions in the brain, and examined which brain regions manifested changes in gFC following anger. This allowed assessing regional changes in gFC. GFC has been suggested to reflect the level of neural integration of a certain brain region and thus potentially having a role in coordinating cognition and behavior (Cole et al., 2010, 2012). Therefore, changes in gFC following anger might reflect processes associated with recovery from the emotional experience. We next used the identified brain regions in a secondary analysis in which we examined the changes in all functional connections of each individuated brain region with all other brain regions to reveal which connections underlie the differences found in the gFC analysis of the previous step.

Participants were those who underwent the anger-infused Ultimatum Game (UG) mentioned above. Thus, to examine if the identified rs-FC modulations associated with participants' anger experience and reaction during provocation, we tested whether the identified modulations, either in gFC or in specific connections, corresponded to participants' self-reported angry feelings and their total monetary gain accumulated throughout the game. While these represent state-like measures of anger, we also questioned whether trait-like measures of anger corresponded to any of the identified modulations. In fact it was previously demonstrated that trait-rumination levels corresponded to gray matter volume in the IFG and cingulate (Kühn et al., 2012). We therefore tested the correspondence of both trait-anger and gray matter volume in the same brain regions in which the rs-FC modulations were identified.

Lastly, our sample consisted of newly recruited infantry soldiers from a unit in the Israeli Defense Forces (IDF) and civil-service volunteers. Since the specific infantry unit recruits only male soldiers, participants in this study were only male. The soldiers were at the beginning of an intense combat-training period of 1 year which was subsequently shown to increase traumatic stress symptoms (TSS) levels (Gilam et al., 2017). The civilians had no change in TSS along a similar period of civil-service. Since anger dysregulation and rumination are characteristic of PTSD patients (Michael et al., 2007; Olatunji et al., 2010), we aimed to examined whether the identified rs-FC modulations identified in the entire sample at the beginning of their respective programs could be predictive of TSS levels in soldiers towards the end of combat-training, possibly linking neural processes of recovery from anger to later development of stress symptomatology.

### MATERIALS AND METHODS

### Participants

Our sample consisted of 60 male participants (age = 18.62 ± 0.88, mean ± SD) that underwent the anger-infused UG task as previously reported (Gilam et al., 2015). All participants completed secondary education, had no reported history of psychiatric or neurological disorders, no current use of psychoactive drugs and normal or corrected-to-normal vision. Due to MRI malfunctions rs-fMRI data was unavailable for 11 participants, and an additional five were discarded due to excessive head movements (>2 mm/2◦ ). Therefore the final sample for rs-fMRI analyses consisted of 44 participants. Twenty-nine of these participants were newly recruited soldiers designated to an infantry unit in the IDF, while the other 15 were pre-army civil-service volunteers. Sample size was based on the number of soldiers and civilians willing to volunteer as participants. All participants were at the beginning of their respective programs and since there were no differences between soldiers and civilians in any measures of the anger induction task they were collapsed as one group. For example, no differences were found in behavior in the modified UG (p = 0.22), emotion ratings (p = 0.61), physiological arousal (p = 0.58), neural activations (e.g., vmPFC p = 0.38), trait anger (p = 0.15) and stress symptoms (p = 0.51; Gilam et al., 2015, 2017). All participants provided written informed consent and the study was approved by the Institutional Ethics Committee of the Tel-Aviv Sourasky Medical Center and of the IDF Medical Corps in accordance with the Helsinki Declaration.

### Procedure

Two 6-min rs-scans with eyes open on a fixation cross were recorded immediately before (rest1) and after (rest2) an interpersonal anger-induction task previously extensively reported (Gilam et al., 2015). Briefly, participants played an anger-infused version of the Ultimatum Game in which they had to agree on how to split a sum of money between themselves and another putative player. The game consisted of 10 such rounds and after each round players verbally negotiated between them, while participants were led to believe that monetary offers they received were decided in real time and influenced by their bargaining. Anger was induced by a predefined sequence of mostly unfair monetary divisions offered by the putative player, who was actually a professional actor trained with scripted improvisations to further intensify the angry experience during negotiations by incorporating personal insults, violating norms of conduct and direct confrontations regarding the game. This manipulation reflected the importance of embedding social interactions when investigating emotional experiences (Gilam and Hendler, 2016). Before entering the scanner participants filled out a trait-anger questionnaire and post-scan they reported on their emotional experience in relation to the game.

### Behavioral Measures

Trait-Anger was assessed using the gold-standard State-Trait Anger Expression Inventory (Spielberger and Sydeman, 1994) and comprised 10 items rated on a 4-point frequency scale from 1 (not at all) to 4 (very much) related to the frequency of angry feelings experienced over time. Trait-Anger was calculated as the sum score of these items and showed good internal consistency (Cronbach's α = 0.74).

Total-Gain was calculated as the sum of money accumulated throughout the entire game by accepting offers and used as an objective measure of individual differences reflecting the final outcome of the anger-infused Ultimatum Game. Rejecting more offers and gaining less money is associated with aggressive reactions while accepting more offers and gaining more money is associated with anger-coping conciliatory reactions.

Angry feelings were assessed based on an iterated version of the Geneva Emotion Wheel (GEW; Scherer, 2005) scheme which was used to obtain subjective reports of the emotional experience during the anger-infused game, on a round-by-round basis and in accordance with participants' actual decisions. Rating was performed on a 7-point intensity scale from 0 (none) to 6 (very high), in relation to how they felt in that exact period during the actual game. An anger cluster of emotions, which includes Anger, Hostility, Contempt and Disgust, was the highest reported cluster of emotions, compared to all other GEW clusters of emotions (positive clusters including Pride, Elation, Happiness, Satisfaction, Relief, Hope, Interest and Surprise, and an additional negative cluster including Shame/Guilt, Boredom, Sadness and Anxiety; Gilam et al., 2015). Indeed as previously noted, anger was the highest reported emotion from all emotion categories used in the GEW (p < 0.05 with Tukey's correction; uncorrected only for hostility and contempt). Importantly, the other negative affect related emotion categories were unaffected by the manipulation nor explained any variance in behavior. We could also show differences in this anger-cluster measure by types of provocations in the task (fair/unfair offers) as well as showing an escalation along the task. This supported that our anger induction was not a mere general negative mood induction. Here we used the average reported emotions for all periods and all rounds of the game in the anger cluster as the measure of angry feelings of each individual during the game.

Traumatic stress-symptoms (TSS) were quantified using the Post-Traumatic Stress Disorder Check-List—Military (PCL-M; Weathers et al., 1991; Spoont et al., 2015) questionnaire which assess stress-symptoms experienced specifically in relation to military experiences. Respondents rate each of 17 stresssymptoms items on a 5-point frequency scale from 1 (not at all) to 5 (extremely), indicating the extent to which they have experienced a specific symptom during the past month of military service. This measure was evaluated towards the end of combat-training and was available for 21 of the 29 soldiers initially recruited to the experiment. As previously indicated in the literature (Novaco and Chemtob, 2002; Jakupcak et al., 2007), a potential methodological confound may exist in correlations between anger related measures and TSS measures since arousal, anger and physical reactions are all anger concomitants as well as being symptoms of post traumatic stress. To avoid circularity between measures and refute this possible confound we removed these symptoms' items (#5, #14, #16 and #17) from the PCL-M score (Gilam et al., 2017).

### MRI Data Acquisition

Brain imaging was performed by a GE 3T Signa Excite scanner using an 8-channel head coil at the Wohl Institute for Advanced Imaging, Tel-Aviv Sourasky Medical Center. Functional whole-brain scans were performed with gradient echo-planar imaging (EPI) sequence of functional T2<sup>∗</sup> -weighted images (TR/TE = 3000/35 ms; flip angle = 90◦ ; FOV = 200 × 200 mm; slice thickness = 3 mm; no gap; 39 interleaved top-to-bottom axial slices per volume; in-plane resolution = 1.5625 × 1.5625 mm<sup>2</sup> ). Anatomical T1-weighted 3D axial spoiled gradient (SPGR) echo sequences (TR/TE = 7.92/2.98 ms; flip angle = 15◦ ; FOV = 256 × 256 mm; slice thickness = 1 mm; in-plane resolution = 1 × 1 mm<sup>2</sup> ) were acquired to provide high-resolution structural images.

### fMRI Data Preprocessing

fMRI data preprocessing was performed with SPM5 (Wellcome Department of Imaging Neuroscience, London, UK). It included correction for head movements via realignment of all images to the mean image of the scan using rigid body transformation with six degrees of freedom, normalization of the images to Montreal Neurological Institute (MNI) space by co-registration to the EPI MNI template via affine transformation (re-sliced voxel size was 3 × 3 × 3 mm<sup>3</sup> ), and spatial smoothing of the data with a 6 mm FWHM. Participants' head motion displayed a framewise displacement (FD; Power et al., 2012) averaging at FD = 0.085 ± 0.047, with three participants having 0.20 < FD < 0.26. There was no difference in FD between the two rs-sessions (FDrest1 = 0.085 ± 0.057, FDrest2 = 0.085 ± 0.044, p = 0.91). The first six images of each functional resting scan were excluded to allow for T2<sup>∗</sup> equilibration effects. Before further analysis, blood oxygenation level-dependent (BOLD) signals were filtered to low frequency fluctuations (0.01–0.08 Hz) using DPARSF toolbox (Chao-Gan and Yu-Feng, 2010).

### fMRI Data Parcellation

We used a previously reported whole brain functional parcellation which was based on the application of correlationbased clustering procedure on rs-fMRI data of healthy subjects that partitions the brain volume into 517 regions or parcels (Craddock et al., 2012). To note, using parcels diminishes the spatial resolution compared to using voxels, yet using information at the voxel level entails a large redundancy in information, while information at the level of parcel has been demonstrated as reliable. Moreover, this procedure enabled us to focus subsequent analyses on functionally defined distinct brain regions which are less sensitive to noise and are interpreted more easily. Parcels were masked to include gray matter voxels only using the WFU Pick Atlas Tool (Maldjian et al., 2003; Stamatakis et al., 2010) and 54 parcels that had less than five voxels in common with the gray matter mask were excluded, leaving 463 parcels. This ensured that only gray matter voxels were used in the analysis. For each subject, average BOLD value across all gray matter voxels was calculated within each parcel at each time point of the two rest periods. These time series were used as the parcel's signal. To reduce the effect of physiological artifacts and nuisance variables, six motion parameters, cerebrospinal fluid, and white matter signals were regressed out of these parcel signals.

### fMRI Data Analysis

To examine experimental effects of anger on rs-fMRI, we conducted a parcel-based analysis (**Figure 1**) inspired by previous such efforts (Cole et al., 2010; Maron-Katz et al., 2016), in which the relationship between each two brain regions was estimated by calculating the Pearson correlation coefficient between their corresponding signals. This was done for each subject and each rs-session separately. Coefficient values were next Fisher Z transformed to fit a normal distribution. We initially tested these coefficients for changes in rs-FC between rest1 (before) and rest2 (after) but no changes in these pairwise rs-FC maps were significant following correction for multiple comparisons

using the false discovery rate (FDR) procedure (Benjamini and Yekutieli, 2001). To note, while FDR is a more lenient procedure compared to Bonferroni, it has considerable power advantages especially compared to conservative procedures such as Bonferroni which assume complete independency between tests (an assumption which is clearly violated in the case of fMRI data). Therefore we computed for each parcel and rest period the sum of all the correlation coefficient values with all other parcels. This measure of gFC captures the level of FC a certain brain region has with all other brain regions in the entire brain. Since the sum over all coefficient values holds the risk of positive and negative values cancelling each other out, we also computed the sum of only positive and only negative values separately. Finally, we calculated the change in rs-FC by subtracting gFC level estimates of rest1 from the corresponding estimates in rest2, resulting in three gFC change values (denoted ∆gFC, ∆gFC<sup>+</sup> and ∆gFC−) for each brain region and for each subject. To identify brain regions that demonstrated significant change following the anger induction, we applied a one-sample t-test on the three change values of each parcel across all subjects and applied the FDR procedure to account for multiple comparisons. We next performed a secondary analysis in which we examined the changes in all functional connections of the brain regions identified in the previous step with all other brain regions. This step aimed to reveal which specific and more localized connections underlie the differences found in the gFC analysis. FDR procedure was again applied to account for multiple comparisons.

### Volumetric Data Preprocessing and Analysis

The volumetric analysis was performed using the FreeSurfer V5.3 image analysis suite<sup>1</sup> which is an automated software for brain segmentation based on probabilistic atlas and intensity values. Briefly, the automated procedure includes skullstripping, intensity normalization, Talairach transformation (the MNI305 template is also used in this processing step; Collins et al., 1994), tissue segmentation, and surface tessellation (Dale et al., 1999; Fischl et al., 1999a,b). The complete FreeSurfer analysis pipeline was performed with manual intervention and quality assurance of the data. Based on the automated segmentation and the fMRI data-driven results we extracted for each subject the right amygdala, right IFG (pars orbitalis) and intra-cranial volumes (mm<sup>3</sup> ). We subsequently calculated the adjusted volume of amygdala and IFG by dividing each subjects' volume by his intra-cranial volume.

### RESULTS

### Data-driven Analysis

To obtain a data driven account of rs-FC modulations following anger induction, we employed the parcel-based univariate gFC analysis. The results revealed a single significant brain region located in a medial region of the right amygdala

<sup>1</sup>http://surfer.nmr.mgh.harvard.edu/

right medial Amygdala (rAmy; to the left; MNI coordinates: x = 18, y = −3, z = −18) for which global positive FC (gFC+) significantly increased between rest1 and rest2 (the extent of change is shown on the right). (B) The scatter plot illustrates all 462 amygdala connections per rest1 (x-axis) and rest2 (y-axis) as t-values of the across participants FC calculated in comparison to zero. All dots above the diagonal (311 in number) reflect connections that increased between rs-sessions. All dots beyond the red square have significant t-values (t(43) = ±2.017, p < 0.05), 145 of which had positive FC in both rs-sessions. (C) Examining all pairwise FC changes involving the amygdala parcel revealed a single significant change characterized by an increase in FC with a parcel located in the right Inferior Frontal Gyrus (rIFG; x = 26, y = 23, z = −18). The orange dot in (B) represents the rAmy-rIFG connection. Error bars indicate standard error of mean, <sup>∗</sup> false discovery rate (FDR) q < 0.05, n = 44.

(31 voxels centered at MNI coordinate: x = 18, y = −3, z = −18) for which positive gFC increased between rest1 (53.70 ± 15.86) and rest2 (62.91 ± 18.32; ∆gFC+=9.21 ± 14.51; t(df <sup>=</sup> <sup>43</sup>) = 4.21, p = 0.0001; FDR q < 0.05; **Figure 2A**). In order to account for the possibility that the sum of all positive FCs may have involved different parcel pairs for each subject, we validated the finding using, for each session, a fixed set of 145 parcels that demonstrated a significant positive FC with the right amygdala in both sessions (**Figure 2B**). The amygdala ∆gFC<sup>+</sup> based only on these 145 connections per subject per rs-session similarly showed a significant increase between rest1 (27.37 ± 12.29) and rest2 (33.11 ± 10.94; ∆gFC<sup>+</sup> <sup>145</sup> = 5.73 ± 9.79; t(43) = 3.88, p = 0.0003). To note, there was no group differences between civilians and soldiers in the ∆gFC<sup>+</sup> of the amygdala between before and after anger (t = 1.51, p = 0.14).

### Amygdala Parcel Analysis

We subsequently examined all 462 functional connections of the right amygdala parcel using a similar univariate analysis as implemented in the initial pairwise FC analysis (see ''fMRI Data Analysis'' section in ''Materials and Methods''), in order to individuate the connections that significantly contributed to the increased ∆gFC<sup>+</sup> of the amygdala between rs-sessions. We found a single connection with a parcel located in the right IFG pars orbitalis (24 voxels centered at x = 26, y = 23, z = −18) that showed a significant increase between rest1 (0.20 ± 0.28) and rest2 (0.37 ± 0.23; ∆FC = 0.17; t(43) = 4.29, p = 0.0001; FDR q < 0.05; **Figure 2C**). This increase in connectivity was found also when applying global signal removal by extracting the mean time course over all white matter and CSF voxels and adding this mean to the set of confound variables regressed out of the time course of each parcel (p = 0.0135). To note, there was no group differences between civilians and soldiers in the ∆FC of the amygdala-IFG connectivity between before and after anger (t = 0.27, p = 0.79).

### Correlation Analyses with State Anger Measures

We next tested the relation between the rs-FC changes, namely right amygdala ∆gFC<sup>+</sup> and right amygdala-right IFG ∆FC, and individual differences in state measures associated with anger induction, namely total-gain and angry feelings during the game, using Spearman's correlation coefficient with a two-tailed significance test. No significant relationships were found (p-values > 0.319).

### Correlation Analyses with Trait Anger Measures

We next tested the relation between the identified angerinduced rs-FC modulations and trait-like measures, namely traitanger, adjusted right amygdala volume and adjusted right IFG volume. We found significant relationships such that higher trait-anger levels and smaller adjusted right IFG volume were associated with a greater increase in right amygdala-right IFG FC between rs-sessions (ρ = 0.469, p = 0.001, FDR q < 0.05; **Figure 3A** and ρ = −0.304, p = 0.045, uncorrected; **Figure 3B**, respectively).

### Correlation Analyses with TSS

Finally, and in view of results thus far, we examined whether right amygdala gFC+ at baseline or right amygdala-right IFG ∆FC between rs-sessions would predict TSS levels among soldiers at the end of a 1 year period of intense combattraining. We found a positive relationship such that a greater

n = 21). Finally, for state measures of anger we found that higher gFC<sup>+</sup> of the right Amygdala before (rest1) playing an anger-inducing Ultimatum Game predicted (D) lower reported feelings of anger experienced during the game (ρ = −0.332, p = 0.027, uncorrected; n = 44) and (E) higher total-gain accumulated throughout the game (ρ = 0.353, p = 0.019, uncorrected; n = 44).

increase in right amygdala-right IFG FC following anger induction was associated with higher TSS levels as measured by the PCL-M (ρ = 0.459, p = 0.036, uncorrected; n = 21; **Figure 3C**).

### Post hoc Correlation Analysis

We further explored whether baseline (i.e., rest1) levels of right amygdala gFC<sup>+</sup> or right amygdala-right IFG FC were associated with state or trait anger measures. We found significant relationships such that more right amygdala gFC<sup>+</sup> at baseline was associated with less angry feelings (ρ = −0.332, p = 0.027, uncorrected; **Figure 3D**) and with more total-gain (ρ = 0.353, p = 0.019, uncorrected; **Figure 3E**). To note, we also found a significant relationship such that smaller adjusted right IFG volume was associated with more reported angry feelings during the game (ρ = −0.278, p = 0.032, uncorrected; n = 60).

### DISCUSSION

The current study implemented a data-driven whole-brain analysis to investigate the change in endogenous neural dynamics in the aftermath of an interpersonal anger experience by examining modulations in FC of rs-fMRI. We revealed an increase in positive gFC of the right amygdala, and specifically an increase in rs-FC between the right amygdala and right IFG, following an anger-infused Ultimatum Game. Higher levels of right amygdala positive gFC before the angering game predicted less reported anger and more monetary gain during anger provocation. Moreover, a greater increase in the amygdala-IFG connection in the aftermath of anger was associated with smaller volumes of the right IFG, higher trait-anger levels and with higher TSS among soldiers, measured 1 year later at the end of combat-training. Together, though some correlations did not survive correction for multiple comparisons, these findings potentially link neural dynamics related to the traces of anger with state and trait like characteristics of anger, and with later development of pathological symptomatology.

The amygdala and IFG were recently associated with a neural model of anger (Gilam and Hendler, 2015), whereby the amygdala was suggested to be involved in threat detection and in mediating negative affect and thus subsequently contributing to reactive aggression, while the IFG was attributed a role in regulating anger experience and inhibiting aggressive impulses. We previously demonstrated that adaptive coping with anger during provocation was associated with increased vmPFC activation and increased posterior Insula-medial Thalamus connectivity (Gilam et al., 2015). The increased amygdala-IFG connectivity following anger demonstrated here suggests that different neural processes are engaged when coping with anger in its aftermath compared to during provocation. A recent model postulated a reverse relationship between a brain network involved in salience processing, which includes the amygdala and insula, and a brain network involved in executive control, which includes regions of the PFC, in the dynamics of exposure to and recovery from acute stress (Hermans et al., 2014). Congruent with current findings, an increase in resource allocation to the salience network and the reverse for the executive control was suggested to mediate maladaptive recovery following stress. However, while anger involves stress-related neuro-physiological systems such as the Locus Coeruleus-Noradrenergic system (Gilam et al., 2015), it is not strictly a stress response. The present findings may thus be informative for a broader account on the interaction between emotion generation and regulation processes and their unfolding beyond the immediate emotional experience.

In agreement with a previous study that explicitly instructed participants to ruminate about anger (Fabiansson et al., 2012), we revealed an increase in amygdala-IFG neural coupling in the aftermath of anger. We extend these results by detecting this increase during non-instructed rs-fMRI using a wholebrain data-driven analysis. In fact, our results are concordant with a broader involvement of amygdala and IFG activity and connectivity in rumination, not necessarily specific to anger (Ray et al., 2005; Kross et al., 2009; Hooker et al., 2010; Kühn et al., 2012; Milazzo et al., 2016). Of particular interest, it was demonstrated that high trait-rumination was associated with smaller gray matter volume and overlapping lower neural activations in the right IFG (Kühn et al., 2012). Congruently, we showed that a larger increase in amygdala-IFG rs-FC in the aftermath of an angering experience was associated with lower gray matter volume of the IFG. The increase in amygdala-IFG connectivity also positively correlated with trait-anger which was shown to have a positive association with rumination (Sukhodolsky et al., 2001; Wilkowski and Robinson, 2010). In contrast to this result, it was elsewhere shown that higher traitanger was associated with lower rs-FC between amygdala and a region in the orbito-frontal cortex, just anterior to the IFG (Fulwiler et al., 2012). However, in that study only one rs-fMRI scan was acquired irrespective of an emotional experience, while here we probed two rs-fMRI scans immediately before and after an interpersonal experience of anger, which might explain this discrepancy. In this respect, we also found that smaller gray matter volume in the IFG, which was previously associated with more trait-rumination (Kühn et al., 2012), was associated with higher anger reported to have been experienced during the anger-infused game, alluding to the general positive association between anger and rumination (Bushman et al., 2005; Pedersen et al., 2011).

Nevertheless, it should be noted as a limitation that while we examined the volume of specific brain regions in anatomical overlap with those regions found in the rs-FC analysis, namely right amygdala and right IFG pars orbitalis, the anatomical labeling of brain regions in the volumetric analysis is approximated to Talairach space (Collins et al., 1994; Dale et al., 1999). This was not used in the rs-FC analysis. Thus said, it was previously demonstrated that the volume calculated by FreeSurfer is comparable to that of other software which calculate brain volume and do not use such an approximation procedure (Tae et al., 2008; Klauschen et al., 2009; Grimm et al., 2015). An additional limitation should be noted regarding the gFC approach that was used. We choose this approach due to the low statistical power of the rs-FC pairwise approach, which did not yield any significant results. Unlike the pairwise approach, gFC is a summary metric which does not capture the whole brain FC at the level of each functional connection. Thus said, to minimize information loss, we calculated positive and negative gFC separately. Future studies able to reveal large-scale network changes following anger could potentially implement graph-theoretical measures to further characterize the underlying properties of these network changes.

While findings of increased amygdala connectivity and specifically with the IFG seem to indicate maladaptive recovery from anger, possibly related to ruminative processing, three main limitations should be cautioned. First, we did not have a direct measure of rumination. Second, we were only able to show an indirect relation between state measures of anger and the identified rs-FC modulations. Namely, while only the amygdala connectivity patterns changed between before and after anger, the relation to angry feelings and monetary gain accumulated in the anger-infused UG was found to amygdala connectivity at baseline, before anger. Therefore, and third, it is possible that a latent variable such as habituation or fatigue contributed to the increase in amygdala-IFG connectivity in the aftermath of anger. Nevertheless, it should also be noted that studies investigating neural dynamics during rs-fMRI following acute stress have similarly shown increased amygdala connectivity (van Marle et al., 2010; Veer et al., 2011), which was also associated with a sustained negative experience and thus with maladaptive coping (Vaisvaser et al., 2013). Elsewhere, positive amygdala-IFG neural coupling was demonstrated to associate with unsuccessful or enhanced efforts to exert cognitive control via reappraisal over an evoked negative emotion (Wager et al., 2008). In contrast, a different study reported that greater amygdala-IFG neural coupling during rs-fMRI was associated with subsequent success in using reappraisal to down regulate negative affect associated with angry faces (Morawetz et al., 2016). We therefore do not know what the amygdala-IFG connectivity reflects per se. Ultimately, future studies will hopefully disentangle between the different regulation strategies (implicit/explicit), conditions (rest/task) and timings (before/after emotional experiences) in order to better understand and interpret these findings.

However, we further support that the increase in amygdala-IFG neural coupling following anger associates with maladaptive or dysregulated emotional processing by demonstrating that among soldiers undergoing combattraining, greater increase in amygdala-IFG connectivity at the beginning of combat-training predicted higher levels of TSS at the end of training. Though hyper-reactive amygdala and hypo-reactive IFG are of the most robust and consistent findings in studies comparing PTSD patients with healthy and trauma exposed controls (Hayes et al., 2012; Patel et al., 2012), amygdala-IFG connectivity has yet been demonstrated as a potential predisposing risk-factor on TSS development (Admon et al., 2013). There is a growing number of studies demonstrating altered rs-FC in PTSD patients (Peterson et al., 2014), several of which have shown weaker positive amygdala-IFG connectivity compared to control groups (Sripada et al., 2012; Brown et al., 2014; Zhang et al., 2016). Though seemingly incongruent with current results, unlike those previous studies, we examined rs-FC changes in the aftermath of anger, emphasizing neural processing associated with emotional recovery and possibly related to rumination. Indeed, rumination was found as a powerful predictor of persistent stress symptomatology (Michael et al., 2007) and in fact also as a mediator between TSS and anger (Orth et al., 2008). Though previous findings provided a neural link between anger and TSS (Gilam et al., 2017; Lin et al., 2017), the results found here provide a first such link between emotional coping in the aftermath of anger and stress related symptomatology. Nevertheless, our findings necessitate additional inquiry since they were limited by the relatively small sample size which consisted of males only, and that soldiers did not experience actual traumatic events and evidenced moderate levels of TSS. In this regard it should be noted as a limitation that this result as well as several other significant correlations did not survive for correction of multiple comparisons.

In conclusion, the increase of amygdala-IFG FC found here seems to be associated with maladaptive processing in the aftermath of anger and thus may serve as a potential target for understanding and treating psycho-pathological conditions characterized by excessive anger and/or emotion dysregulation such as anxiety, depression, and personality disorders.

### AUTHOR CONTRIBUTIONS

GG designed the experiment, collected the data, performed behavioral and integration analyses, wrote the original draft and reviewed and edited the manuscript. AM-K preprocessed and

### REFERENCES


analyzed resting-state fMRI data and reviewed and edited the manuscript. EK preprocessed and analyzed anatomical data. TL designed the experiment and collected the data. EF provided resources and collected the data. RS supervised resting-state analysis and reviewed and edited the manuscript. TH designed the experiment, funded and supervised the project and reviewed and edited the manuscript.

### ACKNOWLEDGMENTS

We thank Ofir Shany and Itamar Jalon for providing comments on an earlier draft of the manuscript and the Sagol Network for Neuroscience. This work was financially supported by the University of Chicago's Arete Initiative—A New Science of Virtues Program (39174-07; awarded to TH, Rakefet Sela-Sheffy and Judd Ne'eman); the U.S. Department of Defense award (W81XWH-11-2-0008 awarded to TH); the I-CORE Program of the Planning and Budgeting Committee (51/11 awarded to TH); and the Israeli Ministry of Science, Technology and Space (3-11170 awarded to TH).


amygdala-hippocampal functional connectivity. Front. Hum. Neurosci. 7:313. doi: 10.3389/fnhum.2013.00313


**Conflict of Interest Statement**: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2017 Gilam, Maron-Katz, Kliper, Lin, Fruchter, Shamir and Hendler. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Biological Clocks and Rhythms of Anger and Aggression

Suzanne Hood<sup>1</sup> and Shimon Amir <sup>2</sup> \*

<sup>1</sup>Department of Psychology, Bishop's University, Sherbrooke, QC, Canada, <sup>2</sup>Department of Psychology, Concordia University, Montreal, QC, Canada

The body's internal timekeeping system is an under-recognized but highly influential force in behaviors and emotions including anger and reactive aggression. Predictable cycles or rhythms in behavior are expressed on several different time scales such as circadian (circa diem, or approximately 24-h rhythms) and infradian (exceeding 24 h, such as monthly or seasonal cycles). The circadian timekeeping system underlying rhythmic behaviors in mammals is constituted by a network of clocks distributed throughout the brain and body, the activity of which synchronizes to a central pacemaker, or master clock. Our daily experiences with the external environment including social activity strongly influence the exact timing of this network. In the present review, we examine evidence from a number of species and propose that anger and reactive aggression interact in multiple ways with circadian clocks. Specifically, we argue that: (i) there are predictable rhythms in the expression of aggression and anger; (ii) disruptions of the normal functioning of the circadian system increase the likelihood of aggressive behaviors; and (iii) conversely, chronic expression of anger can disrupt normal rhythmic cycles of physiological activities and create conditions for pathologies such as cardiovascular disease to develop. Taken together, these observations suggest that a comprehensive perspective on anger and reactive aggression must incorporate an understanding of the role of the circadian timing system in these intense affective states.

#### Edited by:

Nelly Alia-Klein, Icahn School of Medicine at Mount Sinai, United States

#### Reviewed by:

Jorge Mendoza, UPR3212 Institut des Neurosciences Cellulaires et Intégratives (INCI), France Etienne Challet, Centre National de la Recherche Scientifique (CNRS), France

> \*Correspondence: Shimon Amir shimon.amir@concordia.ca

Received: 22 September 2017 Accepted: 09 January 2018 Published: 23 January 2018

#### Citation:

Hood S and Amir S (2018) Biological Clocks and Rhythms of Anger and Aggression. Front. Behav. Neurosci. 12:4. doi: 10.3389/fnbeh.2018.00004 Keywords: anger, aggression, circadian rhythm, infradian rhythm, clock genes

### INTRODUCTION

In the writings of Galen and Aristotle, changes in human tempers were associated with the passage of time, where summer was the season of yellow bile, a humor responsible for a ''nature that is angry, insolent, or fierce'' (Grant, 2000, p.17). Although humorism has long since been abandoned as a medical perspective, the notion that states of anger and aggressive behavior nonetheless exhibit predictable cycles of waxing and waning across time holds some merit. Essentially all species on Earth possess internal timekeeping mechanisms that govern a multitude of cellular, physiological, and behavioral processes. An abundance of evidence demonstrates that these timekeeping mechanisms form a vital part of our physical and mental health, and that disruptions to their normal functioning can severely compromise emotional state and well-being.

Anger and aggressive behaviors are normal parts of the human behavioral repertoire, and an absence of these can be highly disadvantageous for survival (Green and Phillips, 2004; Waltes et al., 2016). At the same time, there is a maladaptive relationship between excessive anger and health, and this maladaptive relationship implicates the operation of the body's internal timekeeping mechanisms. Here, we review evidence indicating that there are predictable cycles of anger and aggression in humans and non-human species, and offer critical insight into the mechanisms by which normal operation of the body's circadian system influences these patterns of aggressive behavior across time. We also examine evidence that demonstrates a complex relationship between excessive anger in humans and disruption of circadian clocks: specifically, that disturbances of the body's time keeping system increase the likelihood of aggression and irritability; and, reciprocally, that the physiological symptoms of anger and aggression perturb the normal functioning of this system.

### DEFINING ANGER AND AGGRESSION

The variety of operational definitions of anger and aggression presents a challenge for establishing whether there are predictable rhythms of these behaviors in humans and non-human species (Kempes et al., 2005; Mathias et al., 2007; Fung et al., 2009). It is important to consider these differences in terminology carefully, because some behaviors and emotional states appear to exhibit predictable rhythms whereas others do not. For example, reactive aggression in humans is argued to differ from proactive aggression, in that the former represents a response to a potential threat and is associated with high arousal and impulsivity, whereas the latter is a low arousal, calculated behavior intended to obtain instrumental ends such as a reward (Hubbard et al., 2002; Kempes et al., 2005). As discussed below, some evidence suggests that humans exhibit predictable cycles in displaying reactive aggression (e.g., Leggett et al., 2015; Hwang et al., 2016), but not proactive aggression. Distinguishing among types of aggressive behaviors is also important for identifying the environmental and physiological mechanisms that underlie each, and for determining how these mechanisms are linked to the internal timekeeping system. For example, in song sparrows, cyclic increases in androgen activity strongly influence territorial aggression in breeding season, but these hormone rhythms do not play the same role in regulating aggressive behavior outside of the breeding season (Wingfield, 2012). For these reasons, we have tried in the following sections to specify the type of aggression under analysis; state the context in which the behavior is displayed; and describe the features of the behavior, where possible.

### THE CIRCADIAN OSCILLATORY NETWORK AND BIOLOGICAL RHYTHMS

Endogenous biological clocks regulate patterns of physiological activity and behavior on several time scales. Cycles of change that complete within 24 h are known as circadian rhythms and include examples such as the sleep/wake cycle, body temperature change, and release of hormones such as melatonin and cortisol. Circadian rhythms provide an adaptive mechanism for organisms to coordinate physiological functions and behaviors with the predictable 24-h cycle of light and dark on Earth. In mammals, the suprachiasmatic nucleus (SCN) in the hypothalamus contains the master circadian clock, and exposure to daylight provides the dominant cue to synchronize this master clock to the external environment (RW.ERROR—Unable to find reference:249). Other powerful synchronizing cues, or zeitgebers, include food consumption and social interaction (Stephan, 2002; Mistlberger and Skene, 2004). Subordinate clocks, or oscillators, also exist throughout the body in tissues such as brain, heart, lungs, liver and endocrine glands (Schibler et al., 2015). By receiving time-of-day information from the SCN via synaptic and diffusible signals, these subordinate clocks coordinate the timing of rhythmic activities throughout the body to the external environment (Mohawk et al., 2012; Dibner and Schibler, 2015).

Although zeitgebers are vital for keeping the body's network of oscillators in time with the 24-h day, they are not sufficient in and of themselves for circadian rhythms to occur. Rather, true clock-controlled functions persist even in the absence of environmental cues (known as ''free running'' rhythms), and will typically exhibit a period that deviates slightly from 24 h. Furthermore, certain zeitgebers such as daylight may be rendered ineffective if the SCN is destroyed. Among the criteria used to establish that a physiological process or behavior is truly under control of an endogenous clock, it must persist under constant (i.e., zeitgeber-free) conditions, and be able to synchronize anew (or re-entrain) to the re-introduction of an appropriate zeitgeber.

At a molecular level, biological circadian clocks are driven by a core group of genes that regulate their own transcription and translation over 24 h via a series of interacting negative feedback loops (**Figure 1**; for reviews, see Huang et al., 2011; Mohawk et al., 2012). ''Clock'' genes regulate their own levels of expression in a predictable cycle that completes in approximately 24 h. Beyond their self-regulation of expression across the day, clock genes play a vital role as transcription factors and control the timing of expression of a wide variety of other genes (referred to as ''clockcontrolled genes, CCGs'').

Rhythms that complete over time periods exceeding 24 h, such as monthly, seasonally, or annually, are known collectively as infradian rhythms. Familiar examples of such rhythms include the menstrual cycle, seasonal breeding and migration behaviors. Although the circadian system is implicated in the expression of infradian rhythms (Oster et al., 2002), the precise mechanisms that regulate these long oscillations remain less clear compared to those underpinning circadian rhythms. A variety of environmental cues regulate the timing of infradian rhythms, depending on the specific rhythm and species under consideration. For example, day length (or photoperiod), ambient temperature and food availability play important roles in the regulation of seasonal breeding rhythms in mammalian and non-mammalian species (for a review, see Paul et al., 2008). These cues exert an impact at a cellular level by initiating changes in melatonin release from the pineal gland, and regulating other endocrine factors such as pituitary hormone signaling (e.g., thyroid, prolactin) and hypothalamic peptides. In turn, these factors ultimately trigger downstream physiological and behavioral changes over time, likely through epigenetic mechanisms (Dawson et al., 2001; Lincoln et al., 2006; Stevenson and Prendergast, 2015; Lynch et al., 2017). In contrast to the presence of a master circadian clock in the SCN, no single tissue appears to house a master infradian timekeeper: rather, evidence to date suggests

that several tissues are involved, and that different infradian rhythms engage different networks of tissues (Paul et al., 2008).

### RHYTHMS OF ANGER AND AGGRESSIVE BEHAVIOR

### Infradian Cycles

Studies of non-human species provide the strongest evidence for seasonal rhythmicity in the expression of certain types of anger and aggression (for a visual summary of these rhythms, see **Figure 2**). Aggressive behaviors in many different species of mammals, birds, reptiles, fish and insects exhibit predictable peaks and valleys across the year, and the timing of these behavioral patterns typically exhibits a stable phase relationship with the expression of other seasonal behaviors, such as mating, territory selection, or challenges to social hierarchies (Wilson and Boelkins, 1970; Michael and Zumpe, 1978; Ruby, 1978; Klukowski and Nelson, 1998; Soma, 2006; Perrone et al., 2009; Muschett et al., 2017). For example, some species of non-human primates and ungulates that breed seasonally display more aggression towards others within the social group and sustain more physical injuries from intra-group member attacks (males vs. male) during the few weeks before and at the beginning of sexual partnering. This seasonal rise in aggression has been observed in animals living in laboratory conditions as well as in the wild (Wilson and Boelkins, 1970; Blank et al., 2015). Seasonality in aggressive behaviors is also well-documented

in many species of birds and rodents, although the phase relationship of an aggression rhythm with other seasonally fluctuating behaviors such as mating varies. In contrast to the coincident rise of aggression with mating in primates, several species of hamsters and beach mice exhibit more vigorous attack responses towards territory intruders when reproductive activity is low during short photoperiods (Jasnow et al., 2000; Trainor et al., 2007).

The physiological mechanisms that drive seasonal changes in aggression remain only partially understood; however, substantial research in mammals and birds species has identified complex relationships between photoperiod, thyroid hormone levels, and sex hormone levels. In rodents and songbirds, sex hormones play a fundamental role in the regulation of aggression (Ogawa et al., 2000; Sato et al., 2004; Sperry et al., 2010), yet the impact of androgens and estrogens on aggressive behaviors does not remain constant across the seasons. For example, in rodents, estrogens promote aggressive responses to territory intruders when animals are housed in short photoperiod conditions (i.e., winter-like, non-breeding season), but diminish aggression under long photoperiod conditions (Trainor et al., 2007; Laredo et al., 2013). In song sparrows, the role of androgens in modulating aggression also exhibits a seasonal shift, whereby circulating levels of luteinizing hormone and testosterone are necessary for territorial defense behaviors during long photoperiod conditions (the breeding season; Wingfield, 1994, 2012; Sato et al., 2004). Outside of the breeding season, however, displays of territorial aggression are not dependent on these hormones. The triggering of these seasonal differences in behavior appears to rely in part on a photoperiod-mediated endocrine cascade including melatonin and thyroid hormones that drive seasonal fluctuations in sex hormones, although there are significant inter-species differences in this process. For example, the daily period of melatonin release is an important intermediary driver between daylight length, hormone profile and seasonal aggression in rodents (Ono et al., 2008; Rendon et al., 2015), whereas deep brain photoreceptors directly mediate the impact of photoperiod changes, hypothalamic and pituitary hormones, and behavior in songbird species (Nakao et al., 2008; Mukai et al., 2009). Significant work remains to be done to further elucidate these mechanisms. Furthermore, much remains unknown regarding the mechanisms by which nonphotoperiod–based seasonal zeitgebers, such as food availability, influence circannual cycles of aggressive behavior (Bailey et al., 2016).

In humans, epidemiological evidence indicates that rates of physically aggressive crime fluctuate in phase with seasonal changes in temperature and photoperiod. Analysis of violent crime statistics suggest that events involving personal physical attack (simple and aggravated assault, sexual assault, intimate partner violence) are more likely to occur during the summer season and less likely during the winter, whereas violent crimes not involving direct physical contact (e.g., robbery) do not exhibit seasonal trends (Michael and Zumpe, 1983, 1986; Lauritsen and White, 2014). Other multinational studies of criminally aggressive behavior suggest a similar pattern, whereby rates of physically violent crime (assault, sexual assault) increase during the geographic summer season and decline during the winter (e.g., an increase during July and August in northern nations such as Denmark, and during December and January in southern nations such as Australia); in contrast, non-physical, or propertybased crime such as theft showed no seasonal variation (Schreiber et al., 1997). Perhaps surprisingly, however, these studies have not identified similar seasonal cycles in murder rates. Furthermore, global rates of suicide (conceptualized by some as a self-aggressing behavior) also appear to follow a seasonal rhythm, with completed suicides using either violent (e.g., use of firearms, drowning, hanging) or non-violent (e.g., overdose) means increasing slightly in the spring compared to other times of year (Hakko et al., 1998; Woo et al., 2012).

Seasonal trends in physically violent criminal acts have led some to propose a causal relationship between aggressive behavior, temperature and photoperiod (Anderson et al., 2000; Hsiang et al., 2013; discussed in Michael and Zumpe, 1983). Such proposals have not gone without challenge, and alternative interpretations of the relationship include the increased likelihood of social interaction during milder temperatures as compared to periods of cold or extreme heat (Rotton and Cohn, 1999). To date, there is no evidence to suggest that endogenous timekeeping mechanisms are responsible for seasonal changes in violent crime; for example, we are not aware of any demonstration that these patterns free-run in the absence of these environmental cues. Furthermore, there is limited evidence for seasonal variation in human hormone levels or receptor function that is comparable to the mechanisms in non-human species discussed above (Huhtaniemi et al., 1982; Smith et al., 2013).

### Circadian Cycles

Rhythmic patterns in anger and aggressive behavior also have been documented on a circadian time scale (**Figure 2**). In humans, some evidence suggests that an individual's chronotype (i.e., if one is a morning person or an evening person) is associated with expressions of anger and hostility. Specifically, young to middle-aged adults identifying as evening types tend to score higher on self-report scales of impulsivity, state and trait expressions of anger, and irritability (Park et al., 2015; Hwang et al., 2016; however, see Chrobak et al., 2017), and children and adolescents who are evening types are more likely to be rated by parents or teachers as displaying rule-breaking and externalizing behaviors such as conflict with others, lying, screaming, or swearing compared with individuals who identify as being morning types (reviewed in Schlarb et al., 2014). Interesting experiments also suggest that chronotype may interact with time of day to influence the likelihood of displaying socially transgressive behavior such as cheating: for example, young adults are more likely to lie about their success on monetarily rewarded tasks, like self-reported puzzle solving or dice throwing scores, if they are tested at a time of day that does not coincide with their chronotype (i.e., if morning-type people are tested in the late evening; Gunia et al., 2014; Kouchaki and Smith, 2014; Ingram et al., 2016).

Additionally, there is some evidence for a daily pattern in physically aggressive or agitated motor behaviors and verbal outbursts in individuals suffering from dementia-related disorders such as Alzheimer's disease. An increase in these behaviors in the late afternoon and early evening has been described as ''sundowning'' (reviewed in Bachman and Rabins, 2006). Although it is debated in the literature whether sundowning is a real behavioral phenomenon or is instead a methodological artifact (e.g., due to reporting bias in caregivers; Yesavage et al., 2003), some studies indicate that an increase in late-afternoon agitation is likely not attributable to sleep disturbances, which are common in this population (Volicer et al., 2001).

The foregoing examples of daily fluctuations in aggressive behaviors are suggestive of behavioral rhythms; however, these patterns have not yet been shown to meet the criteria of being controlled by an endogenous timekeeper. Nevertheless, several interesting hypotheses could explain the physiological mechanisms by which the circadian system could regulate 24-h susceptibility to anger and aggression. Studies of both human and non-human species indicate that there is a significant genetic component to aggression and anger (for reviews see Takahashi and Miczek, 2014; Waltes et al., 2016), and some of the candidate genes identified to date appear to be clock controlled (Duffield, 2003). For example, individual differences in the activity of several monoamine neurotransmitter systems such as dopamine (DA), norepinephrine (NE) and serotonin (5-HT) are clearly associated with the likelihood of aggressive behaviors in both human and non-human species (**Figure 1**; Marino et al., 2005; Miczek et al., 2007; Alia-Klein et al., 2008), and some regulatory elements of each of these systems such as catalytic and metabolic enzymes, and receptor proteins are indeed clock controlled (Aston-Jones et al., 2001; Ueda et al., 2002; Weber et al., 2004; Malek et al., 2005). The metabolizing enzyme monoamine oxidase A (MAO-A), which degrades DA, NE and 5-HT, provides a compelling illustration of this. Brain levels of MAO-A are inversely associated with self-reported trait anger in humans (Alia-Klein et al., 2008), and pharmacological blockade of MAO-A activity in rodents increases aggressive responses to intruders (Shih, 2004). The clock proteins BMAL and PER2 positively regulate mao-a gene expression in regions of the rodent brain including the striatum and ventral tegmental area, and daily fluctuations in levels of MAO-A in these tissues could presumably influence the likelihood of responding in a more or less aggressive fashion to events happening at different times of day (Hampp et al., 2008). A role for brain glutamatergic and gamma aminobutyric acid (GABA) signaling in aggression has also been demonstrated in non-human species. PER2 positively regulates the expression of the excitatory amino acid transporter (eaat) in astrocytes, which contributes to the re-uptake of glutamate from the extracellular space (Abarca et al., 2002; Spanagel et al., 2005; Hampp and Albrecht, 2008; Takahashi and Miczek, 2014).

Clock-controlled regulation of these genes may be particularly important for influencing the activity of brain structures known to mediate the expression of aggression and hostility in mammals. For example, activity in the amygdala is closely linked with aggressive behavior and trait anger, and blunted serotonin activity in this region is associated with elevated aggression and impulsivity in both non-human species and humans (Rosell and Siever, 2015; Suzuki and Lucas, 2015; da Cunha-Bang et al., 2018). Notably, circadian oscillators have been identified within the central and basolateral nuclei of the amygdala (Lamont et al., 2005), and the timing of these oscillators can be shifted by events such as acute glucocorticoid hormone release and exposure to psychological stressors (Segall and Amir, 2010; Al-Safadi et al., 2015). Such events are also known to increase displays of reactive aggression (Mikics et al., 2007).

The genetic circadian clock also plays an important role in mediating the activity of signaling systems such as sex hormones in the ventromedial hypothalamus (VMH), a region strongly implicated in the expression of aggressive behaviors (Cai et al., 2008; Falkner and Lin, 2014). A number of studies demonstrate that changes in the expression of sex hormone receptors in the VMH and in the firing rates of neurons in this region contribute to the display of aggressive behaviors by rodents and songbirds (Spiteri et al., 2010; Lee et al., 2014; Falkner et al., 2016; Wacker et al., 2016). Evidence suggests that the VMH houses a circadian oscillator that is under control of signals from the SCN master clock (Inouye, 1983; Egawa et al., 1993; Ono et al., 1987).

More recent evidence has implicated the role of the nicotinamide adenine dinucleotide-dependent sirtuin proteins (SIRT)—specifically SIRT1—in both the regulation of the genetic circadian clock and the risk for diagnosis of antisocial personality disorder in humans (Chang et al., 2017). SIRT1 regulates the expression of bmal1 and per2 through deacetylation (Asher et al., 2008; Nakahata et al., 2008), and overexpression of sirt1 in rodents leads to elevated levels of bmal1 and per2, and shortens the period of the circadian clocks in the brain controlling locomotor activity (Chang and Guarente, 2013). In a sample of young men, some of whom were juvenile offenders and had received a diagnosis of antisocial personality disorder, a particular single nucleotide polymorphism (SNP) in the sirt1 gene was associated with a lower risk of antisocial personality disorder diagnosis, whereas a second SNP was found to be more frequent among youth who had received a diagnosis. Although these associations were modest, these observations, when combined with additional evidence linking SIRT1 to the development of midbrain dopaminergic activity, raise interesting questions regarding a link between SIRT1, the circadian clock, monoamine neurotransmitter systems and aggressive behaviors (Lee et al., 2009; Kishi et al., 2011).

Individual genetic differences leading to chronotype may also underlie the association of daily fluctuations in anger and aggression and the circadian timekeeping system. Recent genome-wide association studies in humans have identified multiple loci associated with a morningness type, and many of these loci are near clock genes or genes implicated in the phototransduction process mediating the transfer of daylight information to neurons in the SCN. Interestingly, loci were also identified near genes implicated in the regulation of serotonin activity (5htr6) as well as GABAergic activity in brain (plcl1; nol4), suggesting that the genetic profile that influences morningness or eveningness preference could also involve differences in the activation or sensitivity of these neurotransmitter systems (Hu et al., 2016; Jones et al., 2016). At a behavioral level, chronotype influences changes in cognitive alertness across the day, whereby performances on tests of memory function, reaction time and decision making are worse if one is tested at a time of day that does not align with self-reported morningness/eveningness preference or phase of the genetic circadian clock (May, 1999; Ingram et al., 2016). This worsening of cognitive performance outside of one's self-identified ''optimal'' time may increase the likelihood of poorer decision making and impulsiveness, and in turn predispose the expression of anger or hostility in the face of challenging or frustrating circumstances.

### DISRUPTION OF THE CIRCADIAN SYSTEM AND AGGRESSION

In addition to the forgoing evidence suggesting a role for biological clocks in regulating anger and aggressive behaviors, a significant body of research suggests that disruptions of normal biological rhythms also influence these behaviors. The exact nature of this relation remains unclear, as much of the evidence collected on this topic to date in human populations is correlational. As such, it is an open question as to whether one particular direction of relationship is more influential than the other—that is, whether disruptions of rhythmic behaviors promote aggression, or whether heightened arousal and anger disrupt biological clocks, or both (Kamphuis et al., 2012). Below, we examine the evidence that disruptions of circadian rhythms such as the sleep/wake cycle do in fact modulate the expression of anger and hostility in several species. The inverse relation—whether intense or chronic expressions of anger and aggression can impact the functioning of the circadian system itself—will be discussed in the final section of this review.

A common example of a perturbation of the circadian system is the disruption of the normal sleep/wake cycle. Although sleep is not exclusively governed by biological clocks, even short-term periods of sleep deprivation can negatively affect a number of other physiological and behavioral rhythms, and transiently alter the genetic clock in a variety of tissues (for review, see Archer and Oster, 2015; Cedernaes et al., 2015; Gil-Lozano et al., 2016). A number of studies have explored the link between the disruption of sleep and the expression of aggressive behavior. To date, this literature has revealed a complex relationship, in that the impact of sleep impairment on aggression varies depending on the species and the nature of the aggressive behavior under consideration (for review see Kamphuis et al., 2012). In humans, poor sleep quality and loss of sleep correlate with higher self-reported ratings of irritability and experience of anger, and, in the case of studies of children and adolescents, greater instance of aggressive physical activity and hyperactivity based on observer ratings (Waters et al., 1993; Gregory et al., 2004; Grano et al., 2008; Coulombe et al., 2011). Interventions to improve sleep quality appear to lessen these behaviors in some cases (Haynes et al., 2006; Mitchell and Kelly, 2006). It has been noted that many of the studies investigating the relation between sleep loss and aggression in adults have used self-report measures of irritability or responses to hypothetical social scenarios as a measure of aggression, rather than objective assessments of external behavior. However, an interesting exception to this includes observations of on-the-job behavior in shift workers. For example, a large-scale study of police officers in North America revealed that approximately 40% experienced some form of chronic sleep disorder such as apnea or insomnia, and that affected individuals were more likely to have demonstrated adverse work-related behavior such as displaying uncontrolled anger towards suspects or citizens (Rajaratnam et al., 2011). Studies in animal models, particularly rodents, are consistent with the view that sleep deprivation is associated with an increased likelihood of physically aggressive behaviors (Licklider and Bunch, 1946; Webb, 1962; Hicks et al., 1979).

Complicating this picture of the relation between sleep and aggression are several studies that have failed to demonstrate an impact of sleep disruption on aggression, and still others that suggest sleep deprivation may actually diminish aggressive behaviors. For example, in humans, an acute period of sleep deprivation (33 h) decreased the likelihood in male participants of displaying retaliatory behavior towards an opponent in a computer game (Cote et al., 2013). In Drosophila, a 12-h period of sleep deprivation similarly reduced physically aggressive behaviors towards other males, and following a sleep recovery period these behaviors were restored (Kayser et al., 2015). Given these differences in the literature, it is difficult to determine conclusively how sleep interacts with aggression, although there is clearly some kind of modulatory influence.

Several mechanisms by which sleep disruption might drive the likelihood of aggressive behavior have been proposed. Sleep loss is known to negatively impact both simple and more complex aspects of cognitive performance including executive function skills and inhibitory regulation (as reviewed in Killgore, 2010). Sleep loss also worsens mood state; increases sensitivity to negatively valenced emotional stimuli; impairs the accuracy of judgment of emotion expressed by human faces; and is associated with poorer inhibitory emotion regulation and decreased empathy towards others (Pilcher and Huffcutt, 1996; Yoo et al., 2007; Killgore et al., 2008; Hisler and Krizan, 2017). This pattern of behavioral change has been linked to impaired functioning of prefrontal cortex, and indeed metabolic activity in this brain region has been observed to decrease following a brief period of sleep deprivation (Harrison and Horne, 2000; Drummond and Brown, 2001). Together, a reduction in executive control, emotion regulation, and weakened perception of emotional states in others could create conditions whereby sleepdeprived individuals are at greater risk of displaying poorly controlled emotional reactions in situations of interpersonal conflict.

Mutations of the genetic clock itself in animal models have also been documented to increase the likelihood of aggression and associated behaviors such as hyperactivity and impulsivity. For example, mice with a knockout of the clock gene reverb alpha display more aggression towards a territory intruder compared to wild type mice, as well as exhibit greater locomotor activity and exploration behavior (Chung et al., 2014; Jager et al., 2014). The clock delta 19 mutant mouse, which lacks a CLOCK protein capable of transcription regulation, also exhibits hyperactivity and impulsivity (Roybal et al., 2007; Coque et al., 2011). Associated with the behavioral features of both of these models appears to be the dysregulation of midbrain DA systems, whereby the corresponding gene mutation has been found to create a hyperdopaminergic state through the disinhibition of DA synthesis in midbrain cell groups (in the case of the rev-erb alpha KO model) or hyperexcitability of midbrain DA neuron firing (in the case of the clock delta 19 mutant). Correspondingly, manipulations that diminish this hyperdopaminergic activity appear to attenuate these aggressive phenotypes. Catecholamine dysregulation also appears to be implicated in the effects of sleep deprivation on aggression in drosophila, whereby the administration of an octopamine agonist in sleep deprived flies restores the display of physically aggressive behaviors towards another male (octopamine in insects is akin to NE in mammals; Kayser et al., 2015). Taken together, these findings suggest that the dysregulation of midbrain dopaminergic systems by an atypically functioning clock may predispose aggressive and hyperactive behaviors in these models.

### IMPACT OF ANGER AND AGGRESSION ON ENDOGENOUS CLOCKS

Because the circadian network is a physiological system that is exquisitely sensitive to feedback signals arising from both within the body and outside it, it is perhaps not surprising that intense emotional states themselves have been found to influence the pattern of several rhythms acutely. For example, during the night after an episode of work-related or interpersonal conflict, individuals report poorer sleep quality and more sleep disruptions. The strength of this effect appears to be mediated by the extent to which individuals exhibit cynical hostility (a personality construct characterized by negative attitudes towards others, mistrust and defensiveness) and their tendency to ruminate on distressing events (Brissette and Cohen, 2002; Radstaak et al., 2014). In chronically stressed individuals, previous day experiences of anger have been found to predict a blunted amplitude of the cortisol rhythm on the following day (Leggett et al., 2015; Oberle et al., 2017; although see Adam et al., 2006). Acute, anger-provoking stressors have also been observed to blunt the diurnal rhythm of reactive oxygen species production, often used as a marker of immune system activity, in young adults completing a computerized visual judgment task (Atanackovic et al., 2003). Additional evidence suggests that personality disposition towards anger and hostility may perturb biological rhythms over longer time scales. Higher trait anger is predictive of greater sleep fragmentation, poorer self-reported sleep quality, and increased risk for sleep disorders such as insomnia (Waters et al., 1993; Grano et al., 2008; Taylor et al., 2013; Hisler and Krizan, 2017).

The experience of anger is often characterized by physiological symptoms of intense autonomic nervous system activation including transient elevations in blood pressure, heart rate, and sympathetic neurochemical tone, and changes in each of these activities have been found to influence the timing of biological rhythms and clock gene expression in several respects (for a review, see Buijs et al., 2016; Sasaki et al., 2016). For example, substantial research in cell cultures and animal models has demonstrated convincingly that acute surges in corticosteroids can act as a zeitgeber for tissue-specific clocks such as within the limbic system of the brain, liver, kidney and heart (Balsalobre et al., 2000; Al-Safadi et al., 2015). As such, it is possible that frequent and/or intense spikes in sympathetic activation in highly emotional or aggressive individuals may create ''noisy'' internal feedback conditions that impair the daily entrainment of biological clocks to other important signals.

It should be noted that the effect sizes reported in several of the studies of humans discussed above are modest; nevertheless, even small changes in rhythms such as sleep, when accrued across time, may exert a negative effect on physical health and cognitive performance (Vgontzas et al., 2004; Wang et al., 2016). Given the delicate relationship between biological rhythms and vital physiological functions, anger-induced disruptions of the biological clock may be a key mechanism underlying the increased risk of high trait-anger individuals to physical illness such as cardiovascular disease and inflammatory disorders (Gouin et al., 2008; Haukkala et al., 2010; Boylan and Ryff, 2013). Individuals predisposed towards trait anger or aggression exhibit larger, more prolonged changes in heart rate, galvanic skin responses and blood pressure to stressful events (for review, see Waters et al., 1993; Smith et al., 2004), and this exaggerated reactivity may underlie the attenuated nighttime declines in heart rate and blood pressure observed in these individuals (Linden et al., 2008; Beatty and Matthews, 2009; Pavek and Taube, 2009). Given the evidence that chronic variability in blood pressure and heart rate are risk factors for cardiovascular

### REFERENCES


disease (Kikuya et al., 2000), it is possible that anger-induced perturbation in the biological rhythms of vascular activity may be an important contributor towards exacerbating this risk. If so, the circadian system could represent an important target for therapeutic interventions intended to reduce the risk of illness in individuals who have poorly controlled anger or hostility.

## CONCLUSION

Taken together, the findings reviewed here demonstrate a meaningful role for biological clocks in anger and aggression. Seasonal changes in the behaviors of several non-human species provide some of the most convincing demonstrations of predictable cycles in the expression of certain types of aggression. It is less clear that true, clock-controlled seasonal cycles exist in human aggression; however, stronger evidence suggests that circadian patterns are present, and that clockrelated individual differences such as chronotype are associated with the propensity for anger and aggression. There are a number of hypothesized mechanisms that could account for how the biological timekeeping system at a genetic level influences the cellular and physiological factors that regulate aggressive behaviors, and further research into the role of clock genes in controlling catecholaminergic midbrain systems may be particularly useful in understanding these mechanisms more deeply. Further studies at this level of analysis may also provide more definitive insight into how disruptions of endogenous clocks increase the likelihood of hostile behaviors. Advancements in our understanding of the circadian system's role in aggression and anger may perhaps be most valuable for improving our ability to prevent and treat the health complications that individuals high in trait anger are at higher risk of developing.

### AUTHOR CONTRIBUTIONS

SH and SA wrote the manuscript.

### ACKNOWLEDGMENTS

This work was supported by the Natural Sciences and Engineering Research Council of Canada, les Fonds de la Recherché en Santé Québec, and the Canadian Institutes for Health Research (grant no. #MOP142458) to SA.


emotional intelligence and constructive thinking skills. Sleep Med. 9, 517–526. doi: 10.1016/j.sleep.2007.07.003


**Conflict of Interest Statement**: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Hood and Amir. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Implicit and Explicit Motivational Tendencies to Faces Varying in Trustworthiness and Dominance in Men

Sina Radke1,2\*, Theresa Kalt <sup>1</sup> , Lisa Wagels 1,2,3 and Birgit Derntl 4,5,6

<sup>1</sup>Department of Psychiatry, Psychotherapy and Psychosomatics, Medical Faculty, RWTH Aachen University, Aachen, Germany, <sup>2</sup>JARA—BRAIN Institute Brain Structure-Function Relationships: Decoding the Human Brain at Systemic Levels, RWTH Aachen University, Aachen, Germany, <sup>3</sup> Institute of Neuroscience and Medicine (INM-10), Research Center Jülich, Jülich, Germany, <sup>4</sup>Department of Psychiatry and Psychotherapy, Medical School, University of Tübingen, Tübingen, Germany, <sup>5</sup>Werner Reichardt Center for Integrative Neuroscience, University of Tübingen, Tübingen, Germany, <sup>6</sup>LEAD Graduate School, University of Tübingen, Tübingen, Germany

Motivational tendencies to happy and angry faces are well-established, e.g., in the form of aggression. Approach-avoidance reactions are not only elicited by emotional expressions, but also linked to the evaluation of stable, social characteristics of faces. Grounded in the two fundamental dimensions of face-based evaluations proposed by Oosterhof and Todorov (2008), the current study tested whether emotionally neutral faces varying in trustworthiness and dominance potentiate approach-avoidance in 50 healthy male participants. Given that evaluations of social traits are influenced by testosterone, we further tested for associations of approach-avoidance tendencies with endogenous and prenatal indicators of testosterone. Computer-generated faces signaling high and low trustworthiness and dominance were used to elicit motivational reactions in three approach-avoidance tasks, i.e., one implicit and one explicit joystickbased paradigm, and an additional rating task. When participants rated their behavioral tendencies, highly trustworthy faces evoked approach, and highly dominant faces evoked avoidance. This pattern, however, did not translate to faster initiation times of corresponding approach-avoidance movements. Instead, the joystick tasks revealed general effects, such as faster reactions to faces signaling high trustworthiness or high dominance. These findings partially support the framework of Oosterhof and Todorov (2008) in guiding approach-avoidance decisions, but not behavioral tendencies. Contrary to our expectations, neither endogenous nor prenatal indicators of testosterone were associated with motivational tendencies. Future studies should investigate the contexts in which testosterone influences social motivation.

Keywords: approach-avoidance, trustworthiness, dominance, testosterone, 2D:4D

## INTRODUCTION

Approach and avoidance motivation are fundamental in regulating behavior. While avoiding potentially harmful stimuli ensures survival, approaching potentially rewarding stimuli facilitates thriving (Elliot, 2008). In research with human participants, these motivational tendencies can be reliably quantified by approach-avoidance tasks, in which participants pull a joystick toward

#### Edited by:

Gabriela Gan, Zentralinstitut für Seelische Gesundheit, Germany

#### Reviewed by:

Macià Buades-Rotger, University of Lübeck, Germany Frederike Beyer, University College London, United Kingdom

\*Correspondence:

Sina Radke sradke@ukaachen.de

Received: 24 October 2017 Accepted: 11 January 2018 Published: 23 January 2018

#### Citation:

Radke S, Kalt T, Wagels L and Derntl B (2018) Implicit and Explicit Motivational Tendencies to Faces Varying in Trustworthiness and Dominance in Men. Front. Behav. Neurosci. 12:8. doi: 10.3389/fnbeh.2018.00008 (approach) or push it away from their body (avoidance). For emotional facial expressions, healthy individuals are typically faster to initiate approach movements to happy faces and to initiate avoidance movements to angry faces than vice versa (Rotteveel and Phaf, 2004).

Not only emotional expressions, but also expressions conveying socially relevant information may serve as signals whether to approach or avoid a person (Oosterhof and Todorov, 2008). Structural facial features, such as the distance between the eyes and the eyebrows or the facial width-to-height ratio, bias holistic inferences on social traits and behavioral patterns, e.g., aggression (Carré et al., 2009a; Shasteen et al., 2015). Thus, it is not surprising that common mechanisms underlie the evaluation of emotional expressions, e.g., anger, and of social traits, e.g., trustworthiness (Todorov and Engell, 2008; Engell et al., 2010). In fact, untrustworthy faces are likely to be perceived as angry or threatening, signaling avoidance behavior (Oosterhof and Todorov, 2008). The data-driven model by Oosterhof and Todorov (2008) has identified two underlying orthogonal dimensions of face evaluation which account for more than 80% of the variance: The first dimension can be approximated via trustworthiness judgments, i.e., the intention of a person to cause harm. The second dimension can be interpreted as dominance, i.e., the ability of a person to inflict harm. This model has been extensively validated (Oosterhof and Todorov, 2008; Dotsch and Todorov, 2012; Todorov et al., 2013), and provides the opportunity to manipulate the stable facial cues associated with trustworthiness and dominance, respectively, while keeping other visual features constant. Although faces with mild variations in the two dimensions are rated as emotionally neutral, they have been proposed to elicit motivational tendencies (Oosterhof and Todorov, 2008).

Along these lines, slower approach movements toward untrustworthy, compared to trustworthy faces, were obtained in a joystick-based approach-avoidance task (Slepian et al., 2012). Here, participants had to react to faces and houses, without being aware of the faces varying in trustworthiness. As approach-avoidance tendencies are most consistently found when expressions are explicitly evaluated (e.g., ''if you see a happy face, pull the joystick toward yourself'') and to a lesser extent with implicit task instructions (Phaf et al., 2014), the current study set out to test in how far implicit and explicit evaluations influence behavioral tendencies towards faces varying in social traits. In light of the mixed findings, we investigated several ''levels'' of approach-avoidance tendencies to stimuli signaling varying degrees of trustworthiness and dominance. Although mainly natural faces have been investigated in approach-avoidance tasks, we used the computerized faces from Oosterhof and Todorov (2008) to adhere to the proposed framework, and to enable control of other features of the face (e.g., symmetry, hair) that influence trustworthiness judgments (e.g., Bakmazian, 2014). Rinck et al. (2010) also emphasized the advantage of immersive virtual environments in perfect experimental control over facial expressions and observed similar approach-avoidance behavior as in real-life settings. Inferences about trustworthiness from facial appearance are elicited by both natural and computerized faces (Klapper et al., 2016), so that, putting the proposition of Oosterhof and Todorov (2008) to test, we expected highly trustworthy faces to elicit approach and highly dominant faces to elicit avoidance.

Interestingly, evaluations of these social traits are influenced by the steroid hormone testosterone. Administration of testosterone decreased ratings of facial trustworthiness in females (Bos et al., 2010), but not in males (Bird et al., 2017). For endogenous testosterone, sex- and context-dependent effects were observed as a rise in testosterone levels after a competitive interaction predicted decreased trust ratings in men (but not in women; Carré et al., 2014). Exogenous testosterone did not alter males' perceptions of dominance in emotionally neutral faces (Bird et al., 2017), but increased their self-perceived dominance (Welling et al., 2016).

Changes in the perception of social signals may influence engagement in motivational behavior such as aggression. For social threat conveyed by emotional valence, testosterone has been shown to bias behavior toward the approach of social threat, i.e., angry faces (Enter et al., 2014, 2016), subserved by increased amygdala activation (Radke et al., 2015b). As testosterone also increased amygdala reactivity to untrustworthy faces (Bos et al., 2012), behavioral approach toward faces signaling threat via the intention or the ability to cause harm, i.e., untrustworthy or dominant faces, may be similarly influenced by endogenous testosterone. In addition, these activational effects may be facilitated by testosterone's impact on brain organization during prenatal development (Sisk and Zehr, 2005). The secondto-fourth digit ratio (2D:4D), a putative index of prenatal testosterone exposure (Zheng and Cohn, 2011), has been related to aggression and dominance (Turanovic et al., 2017). Therefore, we also tested for associations of approach-avoidance tendencies with endogenous and prenatal indicators of testosterone. In order to rely on a homogenous population, only males were included in this initial study.

### MATERIALS AND METHODS

### Ethics Statement

This study was carried out in accordance with the recommendations of the World Medical Association Declaration of Helsinki with written informed consent from all subjects. The protocol was approved by the local ethics committee at the Medical Faculty of RWTH Aachen University.

### Participants and Procedure

The sample consisted of 50 healthy young men (Mage = 25.1 years, range = 18–33), recruited from the local university and via personal networks. Using an in-house checklist, participants were screened for severe somatic, endocrine and psychiatric disorders. Additional exclusion criteria were participation in a pharmacological study within the last month, use of medication, hormones or illegal substances, and smoking more than five cigarettes per day. Participants were asked to abstain from eating and drinking (except water) 3 h before the experimental session as saliva samples were obtained to determine testosterone concentration. All sessions took place between 2 pm and 5 pm, to control for diurnal hormonal variation. Sample size was based on previous research in this field (Rotteveel and Phaf, 2004; Slepian et al., 2012; Radke et al., 2013, 2014).

### Stimulus and Response Materials

Stimuli consisted of computer-generated faces, which were originally developed with FaceGen (Singular Inversions, Toronto, ON Canada) by Oosterhof and Todorov (2008), applying a 2D statistical model of face evaluation. This model was developed in a multi-step procedure, with, in short, first obtaining trait ratings of faces, then applying a principal component analysis to these ratings which reduced them to the dimensions of trustworthiness and dominance, and cross-validating the model with ratings on 300 computergenerated faces (Oosterhof and Todorov, 2008; Todorov et al., 2013). This database was chosen to enable a highly controlled manipulation of variations in facial features. In particular, specific structural facial features have been associated with the two dimensions, such as a larger distance between eyes and eyebrows with trustworthiness, and masculine features, e.g., a larger facial width-to-height ratio, with dominance. Keeping other features constant, and controlling for non-facial attributes, e.g., hair, which may influence evaluations of social traits (Bakmazian, 2014; Kalogiannidou and Peters, 2015), any differences in response should be caused by the manipulation in facial characteristics along the specific dimension.

Pictures from 25 different models were selected from the ''fi'' face set, all depicting Caucasian males without hair or facial hair in a front view. For each model, four pictures were used, i.e., a high and low expression of each feature of interest (trustworthiness, dominance), corresponding to the −2SD and +2SD version of each feature (out of seven available versions, ranging from −3SD to +3SD). This procedure resulted in a set of 100 experimental stimuli (during practice trials, different pictures, i.e., from the ''nexus'' set, were used).

For the explicit joystick task and the rating task, stimuli were presented in their natural coloring, whereas for the implicit joystick task, pictures were presented with a blue or yellow filter (see **Figure 1**). The computer screen had a resolution of 1024 × 768 pixel. In the joystick tasks, initial picture size was 304 × 363 pixel, with full joystick displacement resulting in a picture size of 556 × 663 pixel for pulling, and a picture size of 112 × 134 pixel for pushing, respectively. For the rating, pictures were presented with a size of 332 × 396 pixel, via Presentation Software (Neurobehavioral Systems, Albany, CA, USA). A Logitech Attack 3 (Logitech, Newark, CA, USA) was used for responding in the joystick tasks.

### Tasks

### Implicit Joystick Task

Each stimulus was presented twice, once with yellow and once with blue coloring, entailing a total number of 200 experimental trials, distributed across two blocks. Each trial was self-paced and started with a blank screen and the joystick in the resting (upward) position. To initiate stimulus presentation in the center of the screen, participants pressed the fire button. Participants were instructed to respond as fast as possible to the color of the face by pushing the joystick away from or pulling it toward their body. The joystick movement implied that participants' arm moved likewise and caused the stimulus to shrink or grow in size. It disappeared when the minimum, respectively the maximum size was reached. During the eight practice trials, the stimulus remained visible after an erroneous response (allowing participants to practice until the response was correct). The stimulus-response mapping remained constant

per participant, but was counterbalanced between participants, i.e., half of the participants reacted to yellow with pull movements and to blue with push movements, and the other half vice versa.

### Explicit Joystick Task

The general setup was identical to the implicit joystick task, but instructions differed. Stimuli were presented in six blocks with 50 experimental trials each, preceded by eight practice trials each. In the first four blocks, stimuli were presented separately per feature, i.e., two blocks with faces varying in trustworthiness, and two blocks with faces varying in dominance. In these blocks, participants had to judge whether that particular feature was present, e.g., react to trustworthy faces with pull movements and to non-trustworthy faces with push movements. After the first block of each feature, the stimulus-response mapping was reversed (for this example, pull non-trustworthy faces and push trustworthy faces). Half of the participants started with evaluating trustworthiness, the other half started with evaluating dominance. The last two blocks combined both features and consisted of pictures corresponding to +2SD trustworthiness and +2SD dominance. Participants were instructed to react to one feature with pull movements, and to the other feature with push movements. This instruction was kept constant with the instruction in block #4, e.g., if participants were to pull trustworthy faces in block #4, they had to pull trustworthy faces and push dominant faces in block #5, which then reversed from block #5 to block #6. Taken together, this implies that participants had to differentiate either between high and low expressions of the same feature, or between high-trustworthy and high-dominant faces, but not between low-trustworthy and low-dominance faces within the same block.

### Approach-Avoidance Rating

The same 100 stimuli as in the previous paradigms were presented each in the center of the screen with the rating scale below until a response was given. Participants should imagine standing face to face with the person depicted and explicitly rate their tendency to approach or avoid him as the number of steps they would make towards (+) or away (−) from him on a scale from −4 to +4. This rating scale was intended to capture conscious behavioral tendencies, in contrast to motor reactions as in the joystick tasks. As in prior studies (e.g., Radke et al., 2015a), the rating was always presented after the joystick paradigms to prevent carry-over effects from conscious evaluation of the pictures to behavioral tendencies.

### Endogenous Testosterone Concentrations

Saliva samples were taken using SaliCaps (IBL International, Hamburg, Germany) at the beginning of the experimental session. Samples were stored at −30◦C until assessment by a commercial laboratory (ISD, Malente, Germany). Free active testosterone was determined via enzyme-linked immunosorbent assay (ELISA, Demeditec Diagnostics GmbH), based on the principle of competitive binding, and with a sensitivity of 2.2 pg/ml. Samples were analyzed in duplicate and the average was used in subsequent analyses. The intra-assay CV was 6.8%, and the inter-assay CV from this laboratory was below 10%. Testosterone levels could not be determined for one participant, and another participant showed high levels of >400 pg/ml. These two participants were excluded from the correlational analyses involving salivary testosterone.

### Digit Ratio (2D:4D)

Testosterone's early organizational effects were indexed by participants' digit ratio (Zheng and Cohn, 2011). For that purpose, participants' left and right hands were photocopied. On these photocopies, the length of the second (2D) and the fourth (4D) fingers were measured from tip to basal crease later by two independent investigators. As the reliability between the two judgments was high (α > 0.98 for all measures), the mean of the two measurements was used for subsequent analyses, i.e., calculating the 2D:4D. The 2D:4D of the left and the right hand were significantly correlated, r = 0.546, p < 0.001.

### Approach-Avoidance-Related Traits

All participants completed the German version of the BIS-BAS scale (Action Regulating Emotion Systems Scale; ARES; Hartig and Moosbrugger, 2003), which assesses BIS-sensitivity with the subscales anxiety and frustration and BAS-sensitivity with the subscales drive and gratification. Additional demographic data was surveyed, e.g., education and relationship status, and is presented in **Table 1**.

### Statistical Analyses

For the joystick tasks, reaction time (RT) was defined from stimulus onset until movement onset. Trials with incorrect and extreme responses (>3SDs of the subject-specific mean per condition) were excluded [implicit: 5.5%, explicit: 20%]. For each joystick task, mean RTs were calculated for each level of the three experimental factors (Feature, Level, Movement). Participants with <5 operational trials per condition were excluded, yielding data of 49 participants for the implicit joystick task, and data of 44 participants for the explicit joystick task. Mean RTs were subjected to two repeated-measures ANOVAs with the within-subject factors Feature (trustworthiness, dominance), Level (high, i.e., +2SD, low, i.e., −2SD) and Movement (approach, avoid). Analogously, error rates were analyzed by using two repeated-measures ANOVAs with the within-


Note: BIS, Behavioral Inhibition System; BAS, Behavioral Approach System.


TABLE 2 | Performance of study participants in the two joystick tasks in ms (presented as Mean [SD]).

subject factors Feature (trustworthiness, dominance), Level (high, i.e., +2SD, low, i.e., −2SD) and Movement (approach, avoid).

In addition, means of the rating scores per Feature and Level were analyzed using a repeated-measures ANOVA with the within-subject factors Feature (trustworthiness, dominance) and Level (high, i.e., +2SD, low, i.e., −2SD).

Following previous studies with emotional expressions (e.g., Radke et al., 2013, 2014), individual behavioral tendencies were calculated by subtracting mean RTs for pull movements from individual mean RTs for push movements. Here, positive scores reflect a relative approach tendency, while negative scores denote an avoidance tendency. These tendencies were used for computing Pearson's correlations: (i) between tasks; and (ii) with indicators of endogenous and prenatal testosterone as well as trait approach-avoidance motivation. Statistical testing was performed with IBM SPSS 22.0 with an α-level of p < 0.05 and partial eta squared as estimate of effect size.

### RESULTS

### Implicit Joystick Task

The Feature × Level × Movement ANOVA on the RTs showed a significant main effect of Movement, F(1,48) = 7.08, p = 0.011, partial η <sup>2</sup> = 0.13, a significant main effect of Level, F(1,48) = 10.88, p = 0.002, partial η <sup>2</sup> = 0.19, and a significant Feature × Level interaction, F(1,49) = 4.29, p = 0.044, partial η <sup>2</sup> = 0.08. No other effects were significant, Fs < 2.49, ps > 0.12.

The main effect of Movement was due to faster avoidance (M = 485.91 ms, SD = 67.36) than approach (M = 496.44 ms, SD = 66.70) reactions. The main effect of Level was due to faster reactions for the faces indicating the presence of a particular feature, i.e., high-trustworthy and high-dominant faces (Mhigh = 487.67 ms, SD = 65.59, compared to low-trustworthy and low-dominant, Mlow = 494.68 ms, SD = 66.41). This effect was driven by the faces varying in trustworthiness as decomposing the Feature × Level interaction revealed faster reactions for faces signaling high, compared to low, trustworthiness, F(1,48) = 18.84, p < 0.001, partial η <sup>2</sup> = 0.28, without significant differences in RTs to faces varying in dominance, F(1,48) = 1.73, p = 0.19, partial η <sup>2</sup> = 0.04 (see **Table 2** for means and **Figure 1**).

The Feature × Level × Movement ANOVA on the error rates showed no significant effects, Fs < 3.82, ps > 0.06 (see **Table 3** for means).

### Explicit Joystick Task

The Feature × Level × Movement ANOVA on the RTs showed significant main effects of Movement, F(1,43) = 4.41, p = 0.042, partial η <sup>2</sup> = 0.09, of Level, F(1,43) = 4.76, p = 0.035, partial η <sup>2</sup> = 0.10, and of Feature, F(1,43) = 9.81, p = 0.003, partial η <sup>2</sup> = 0.19. There was also a significant Feature × Movement interaction, F(1,43) = 4.15, p = 0.048, partial η <sup>2</sup> = 0.09. No other effects were significant, Fs < 0.41, ps > 0.53.

As in the implicit task, the main effect of Movement was due to faster avoidance (M = 982.08 ms, SD = 231.85) than approach (M = 1006.29 ms, SD = 235.28) reactions. Similarly, the main effect of Level was due to faster reactions for the faces indicating a high level of trustworthiness or dominance (Mhigh = 970.25 ms, SD = 227.23, compared to low-trustworthy and low-dominant faces, Mlow = 1018.12 ms, SD = 255.22). The main effect of Feature was evident in faster reactions to faces varying in dominance (M = 966.45 ms, SD = 216.54) than to faces varying in trustworthiness (M = 1021.93 ms, SD = 257.30). Follow-up analyses of the Feature × Movement interaction indicated that the effect of Movement, i.e., faster avoidance than approach RTs, was only present for faces varying in trustworthiness, F(1,43) = 7.17, p = 0.01, partial η <sup>2</sup> = 0.14, but not for those varying in dominance, F(1,43) < 0.01, p = 0.97, partial η <sup>2</sup> < 0.01 (see **Table 2** for means and **Figure 2**).


high dominance (upper face) and high trustworthiness (lower face) from Oosterhof and Todorov (2008). Means were significantly different at p < 0.05 for avoidance vs. approach responses, for high vs. low features, for faces varying in dominance vs. faces varying in trustworthiness, and for avoidance vs. approach to faces varying in trustworthiness. dom, dominance; tw, trustworthiness.

For the error rates, the Feature × Level × Movement ANOVA showed a significant main effect of Movement, F(1,43) = 8.46, p = 0.006, partial η <sup>2</sup> = 0.16. No other effects were significant, Fs < 3.43, ps > 0.07. More errors were committed when approach movements were required (M = 19.40%, SD = 15.75) than when avoidance movements were required (M = 17.18%, SD = 14.23; see also **Table 3**).

### Approach-Avoidance Rating

The Feature × Level ANOVA on the rating showed a significant main effect of Feature, F(1,49) = 144.95, p < 0.001, partial η <sup>2</sup> = 0.75, and a significant Feature × Level interaction, F(1,49) = 181.89, p < 0.001, partial η <sup>2</sup> = 0.79. The main effect of Level was not significant, F = 0.33, p = 0.57.

The main effect of Feature was due to higher ratings for the faces varying in trustworthiness than for the faces varying in dominance. This effect, however, needs to be viewed in the context of the Feature × Level interaction, which revealed differences between ''high'' and ''low'' versions for both features, yet in opposite directions. In other words, for faces varying in trustworthiness, highly trustworthy faces were rated as more approachable (M = 1.30, SD = 0.90) than faces signaling low trustworthiness (M = −0.66, SD = 0.80). For faces varying in dominance, low-dominant faces received higher approach ratings (M = 0.32, SD = 0.85) than highly dominant faces (M = −1.73, SD = 1.06; see also **Figure 3**).

### Correlations

First, there were no significant correlations between behavioral tendencies in the joystick tasks and the rating (all ps > 0.13). Secondly, there were no significant correlations between indicators of testosterone and approach-avoidance tendencies in any of the three tasks. The lowest p-value was p = 0.052, obtained for the (negative) association between salivary testosterone and the approach tendency toward low-trustworthy faces in the explicit joystick task (r = −0.28). Endogenous and prenatal indicators of testosterone were not significantly correlated with one another, r = −0.23 (p = 0.12), and r = −0.17 (p = 0.26) for left and right hand respectively, nor with trait approach-avoidance motivation (all ps > 0.18). Trait approach-avoidance motivation was not significantly correlated with motivational tendencies (all ps > 0.08).

### DISCUSSION

The current study investigated three ''levels'' of approachavoidance tendencies to faces varying in the social traits of trustworthiness and dominance. When participants explicitly rated their behavioral tendencies, highly trustworthy faces elicited approach, while untrustworthy faces elicited avoidance. Conversely, highly dominant faces yielded behavioral avoidance, whereas faces signaling low dominance yielded approach. This pattern, however, did not manifest in the joystick tasks where action tendencies were assessed via the initiation of approachavoidance movements. Instead, more general effects of faster avoidance movements and faster reactions to faces signaling high trustworthiness or high dominance emerged. Neither endogenous nor prenatal indicators of testosterone were related to motivational tendencies.

The two dimensions proposed by Oosterhof and Todorov (2008) can provide a useful framework for investigating motivational tendencies to social traits. As derived from their model of face evaluation, our rating data underpinned that facial variations in trustworthiness and dominance influence approach-avoidance decisions. High trustworthiness and low dominance prompted approach, and low trustworthiness and high dominance prompted avoidance. Of these, faces indicating the presence of a particular feature, i.e., high-trustworthy and high-dominant, evoked the strongest tendencies. As for

All means differ significantly from zero and from each other at p < 0.05. dom, dominance; tw, trustworthiness.

emotional expressions, recognition of these ''present'' features may be easier because they lie closer to the prototypical expressions (Young et al., 1997), thereby reducing ambiguity. Moreover, the mechanisms and neural circuits involved in evaluating emotional expressions also take effect in the judgment of social traits, and judgments are often highly correlated (Todorov and Engell, 2008; Engell et al., 2010). However, we did not assess judgments on these social traits, but probed participants for the behavioral consequences.

In contrast, behavior in the joystick tasks did not manifest as approach-avoidance tendencies, i.e., faster approach toward one type of stimuli and faster avoidance toward another, as commonly observed for emotional faces (Rotteveel and Phaf, 2004; Phaf et al., 2014). Comparing not the two movement directions, but trustworthy and untrustworthy faces, Slepian et al. (2012) reported an asymmetry with trustworthy faces eliciting faster approach than untrustworthy faces, without differences in avoidance. In fact, our data also reveal faster approach movements for high than for low trustworthy faces. However, this effect was neither limited to approach nor reversed for avoidance movements, i.e., there was also faster avoidance of high vs. low trustworthy faces. Therefore, this partial replication of Slepian et al. (2012) needs to be interpreted with caution as it is likely overshadowed by general effects, e.g., of movement. Accordingly, the overall faster initiation of avoidance than approach movements in the current study was unexpected based on previous studies using the same response measures and setup with emotional faces (e.g., Radke et al., 2013, 2014). Taken together, this limits the interpretation of the findings in terms of behavioral tendencies, i.e., RT differences, instead of general responses to faces varying in social traits.

The findings from the joystick tasks may also indicate that the judgment of social traits does not map as readily on approachavoidance tendencies as evaluations of emotional valence. The stimulus set of the current study was drawn from the range of faces rated as emotionally neutral (Oosterhof and Todorov, 2008) in order to exclude this potential affective confound. One may speculate whether more extreme facial displays of social traits would evoke distinct behavioral tendencies in the joystick tasks, but disentangling valence judgments from judgments of trustworthiness and dominance is inherently difficult at the extremes of the two dimensions. Moreover, it seems plausible that emotional expressions, being dynamic, demand faster action and more distinct behavioral tendencies than less transient social signals. In light of the unexpected overall faster avoidance responses, using context-deprived computerized faces may also warrant further investigation. The computerized faces from this database have been used in a variety of studies investigating social perception, such as when selecting allies for a team (Kret and De Dreu, 2013) and in automatic threat processing (Shasteen et al., 2015). Still, they might be perceived as artificial, especially when combined with a colored filter, and they might be less familiar than natural faces, possibly leading to negative evaluations due to difficulties in extracting the relevant diagnostic information (Winkielman et al., 2003).

Ascending ''levels'' of approach-avoidance tendencies were to be assessed by instructions that varied in the explicitness of the evaluation, i.e., from reacting to the irrelevant characteristic of color, to judging the facial traits, to indicating one's behavioral tendency. In line with previous research (Phaf et al., 2014; Radke et al., 2015a), approach-avoidance tendencies were evident for the most explicit instructions yielding conscious evaluation of one's behavior. Yet, no tendencies emerged for either variant of the joystick task. Instead, faster reactions to high-trustworthy faces were evident in the implicit task version, whereas in the explicit task, it was faces varying in dominance that elicited the fastest responses. Facial trustworthiness is neurally evaluated without perceptual awareness (Freeman et al., 2014; Marzi et al., 2014), whereas intentional trustworthiness judgments are driven by a different neural circuitry (Winston et al., 2002). As there are no analogous studies on dominance perception which would point to a similar precedence in neural processing, trustworthiness might more easily influence behavior particularly when attention is directed toward other stimulus features (i.e., color in our task). Interestingly, however, dominance perception is influenced by contextual cues, with masculinized distractor faces decreasing perceived dominance of the target face (Re et al., 2014). Such a relative evaluation of dominance may facilitate more deliberative judgments, as required in our explicit joystick task. Yet, these dominance evaluations might not only be affected by the faces shown within the same block, but also by participants' perception of their own dominance and their interpretation of ''dominance'', which we did not assess or specify in terms of physical or social dominance (see also Re et al., 2014). Taken together, these results may suggest that trustworthiness and dominance might be differentially processed in implicit and explicit contexts.

Unlike prior studies that relied on affect and gender evaluations as explicit and implicit evaluations, respectively (Roelofs et al., 2009), the two currently used instructions turned out not to be matched for difficulty. Of these two, the explicit joystick task prompted longer RTs and higher error rates, which was underlined by participants' spontaneous accounts of increased task difficulty and confusion. Moreover, handling a more stringent cutoff (e.g., at least 70% valid trials per condition) would have drastically diminished the sample size for analyses in the explicit joystick task. We acknowledge that this unexpectedly high task difficulty may limit the interpretation of the results from this task. Still, the pattern of error rates matches that of RTs, with overall more errors and more hesitation when approach movements were required, suggesting that approach movements were particularly difficult in this task. In part, this may reflect the increased processing demands associated with evaluating faces in terms of trustworthiness or dominance compared to following an arbitrary response mapping. Adding difficulty, the stimulusresponse mapping changed after each block of the explicit joystick task, but remained constant throughout the implicit version. To move toward matching task demands, future studies should consider using a block-wise changing of the mapping in both tasks.

Contrary to our hypotheses, testosterone was not associated with approach tendencies to faces signaling threat via the intention or the ability to cause harm, i.e., untrustworthy or dominant faces. Previous research indicates that testosterone can alter the perception of trustworthiness (Bos et al., 2010; Carré et al., 2014) and dominance (Welling et al., 2016; but see Bird et al., 2017) as well as induce threat approach (Enter et al., 2014, 2016). Despite the null effect in the current study, the possibility still exists that heightened testosterone went hand in hand with a decreased sensitivity to cues of trustworthiness or dominance, without translating into behavioral changes. Given the absence of approach-avoidance tendencies in the joystick tasks, the lack of associations with testosterone may not be surprising. However, testosterone did not influence explicit approachavoidance ratings either, although distinct motivational patterns were evident on this level. Together, these findings could point to the fact that variations in endogenous testosterone, on their own, may not be sufficient to explain variations in motivational behavior, and may need stronger contextual triggering or further regard of individual differences, e.g., trait dominance or cortisol (Mehta and Josephs, 2010; Carré and Mehta, 2011; Carré et al., 2017).

Another consideration is that many of the above mentioned findings were obtained in the context of testosterone administration leading to supraphysiological levels (in females: Bos et al., 2010; Enter et al., 2014, 2016), while our design was correlational and included only one measure of salivary testosterone in its normal range. Fluctuations in testosterone may be more relevant than baseline levels for modulating social motivational behavior such as aggression (Carré et al., 2009b, 2013, 2014). In a similar vein, prenatal testosterone seems to play a stronger role in activational effects of the hormone for higher order social cognition than for basic reactions to threat (Terburg and van Honk, 2013; Terburg et al., 2016). However, including only male participants certainly prevents generalization across sexes.

Nevertheless, the current study helps clarify the hypothesized link between social traits and approach-avoidance responses by testing the proposition of Oosterhof and Todorov (2008). By showing that trustworthiness prompted approach ratings and dominance prompted avoidance ratings, without similar behavioral responses, our findings add to a body of research on approach-avoidance tendencies to socially relevant stimuli. They also contribute to our understanding of the impact of hormones on social motivation. The current null-relation between testosterone and motivational tendencies may point to specific contextual boundaries under which effects are (not) to be expected that need to be investigated further. Together with the inconsistencies of past literature and the limitations of the present study, future research is needed in order to replicate and expand upon the current findings.

### AUTHOR CONTRIBUTIONS

SR and BD designed the study. TK collected and processed the data as part of her dissertation. SR performed data analyses of the joystick tasks with support from LW. TK performed data analyses of the rating task as part of her dissertation. SR drafted the manuscript. All authors revised the manuscript and gave final approval of the version to be published.

### FUNDING

This work was supported by the International Research Training Group (IRTG 2150) of the German Research Foundation (Deutsche Forschungsgemeinschaft, DFG). We further acknowledge the DFG and the Open Access Publishing Fund of the University of Tübingen for covering publication costs.

### ACKNOWLEDGMENTS

The authors acknowledge Frank Leonhardt and Mike Rinck as the authors of the approach avoidance task software.

REFERENCES

#### Bakmazian, A. (2014). The man behind the beard: perception of men's trustworthiness as a function of facial hair. Psychology 5, 185–191. doi: 10.4236/psych.2014.53029

Bird, B. M., Geniole, S. N., Little, A. C., Moreau, B. J., Ortiz, T. L., Goldfarb, B., et al. (2017). Does exogenous testosterone modulate men's ratings of facial dominance or trustworthiness? Adapt. Hum. Behav. Physiol. 3, 365–385. doi: 10.1007/s40750-017-0079-7


dominance. Metab. Brain Dis. 64, 136–142. doi: 10.1016/j.psyneuen.2015. 11.016


**Conflict of Interest Statement**: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Radke, Kalt, Wagels and Derntl. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Aggression Profiles in the Spanish Child Population: Differences in Perfectionism, School Refusal and Affect

María Vicent <sup>1</sup> \*, Cándido J. Inglés <sup>2</sup> , Ricardo Sanmartín<sup>1</sup> , Carolina Gonzálvez <sup>1</sup> and José Manuel García-Fernández <sup>1</sup>

<sup>1</sup>Department of Developmental Psychology and Didactics, Faculty of Education, University of Alicante, Alicante, Spain, <sup>2</sup>Department of Clinical Psychology, Faculty of Social-Health Sciences, Miguel Hernández University of Elche, Elche, Spain

The aim of this study was to identify the existence of combinations of aggression components (Anger, Hostility, Physical Aggression and Verbal Aggression) that result in different profiles of aggressive behavior in children, as well as to test the differences between these profiles in scores of perfectionism, school refusal and affect. It is interesting to analyze these variables given: (a) their clinical relevance due to their close relationship with the overall psychopathology; and (b) the need for further evidence regarding how they are associated with aggressive behavior. The sample consisted of 1202 Spanish primary education students between the ages of 8 and 12. Three aggressive behavior profiles for children were identified using Latent Class Analysis (LCA): High Aggression (Z scores between 0.69 and 0.7), Moderate Aggression (Z scores between −0.39 and −0.47) and Low Aggression (Z scores between −1.36 and −1.58). These profiles were found for 49.08%, 38.46% and 12.48% of the sample, respectively. High Aggression scored significantly higher than Moderate Aggression and Low Aggression on Socially Prescribed Perfectionism (SPP), Self-Oriented Perfectionism (SOP), the first three factors of school refusal (i.e., FI. Negative Affective, FII. Social Aversion and/or Evaluation, FIII. To Pursue Attention), and Negative Affect (NA). In addition, Moderate Aggression also reported significantly higher scores than Low Aggression for the three first factors of school refusal and NA. Conversely, Low Aggression had significantly higher mean scores than High Aggression and Moderate Aggression on Positive Affect (PA). Results demonstrate that High Aggression was the most maladaptive profile having a high risk of psychological vulnerability. Aggression prevention programs should be sure to include strategies to overcome psychological problems that characterize children manifesting high levels of aggressive behavior.

Keywords: aggressive behavior, profiles, childhood, socially prescribed perfectionism, self-oriented perfectionism, school refusal, positive affect, negative affect

### INTRODUCTION

Aggression has been typically defined as a ''behavior directed toward harming or injuring another living being who is motivated to avoid such treatment'' (Blair, 2016, p. 4). Although the study of aggression was originally limited to its direct physical and verbal forms (Archer, 2009), it is currently considered to be a complex construct involving multiple components, forms and functions.

#### Edited by:

Nelly Alia-Klein, Icahn School of Medicine at Mount Sinai, United States

#### Reviewed by:

Inti Brazil, Radboud University Nijmegen, Netherlands Gianluca Serafini, Department of Neuroscience, San Martino Hospital, University of Genoa, Italy

> \*Correspondence: María Vicent maria.vicent@ua.es

Received: 14 October 2017 Accepted: 16 January 2018 Published: 30 January 2018

#### Citation:

Vicent M, Inglés CJ, Sanmartín R, Gonzálvez C and García-Fernández JM (2018) Aggression Profiles in the Spanish Child Population: Differences in Perfectionism, School Refusal and Affect. Front. Behav. Neurosci. 12:12. doi: 10.3389/fnbeh.2018.00012 This study is based on the conceptualization of aggression as proposed by Buss and Perry (1992) who consider aggressive behavior to be the combination of three components: emotional (Anger), cognitive (Hostility) and motor (Physical and Verbal Aggression). Anger is an emotion that involves feelings of variable intensity ranging from mild irritation to intense fury (Lubke et al., 2015). Hostility refers to a cognitive state consisting of attitudes and feelings of negative evaluation toward others, such as cynicism, mistrust and suspicion (Fabiansson and Denson, 2016). Lastly, the motor component of aggression implies any physical or verbal action that is carried out in order to injure others (Leary et al., 2006).

Far from being an exclusive manifestation of adolescence or adulthood, aggressive behavior may develop from a very early age (Hay, 2017) and is one of the most common causes for child therapy referrals (Shachar et al., 2016). Over recent years, research on aggressive behavior during childhood has become particularly relevant. This is due in part to the high prevalence of aggressive manifestations in the child population, such us bullying (Modecki et al., 2014), as well as the adverse consequences derived from each of these three components (cognitive, emotional and motor), including a tendency to develop physical and mental health problems, drug use, delinquency, etc. (e.g., Harachi et al., 2006; Kerr and Schneider, 2008; Hampson et al., 2010; Garaigordobil et al., 2013). Furthermore, aggressive behavior has been also identified as having a putative modulating role between genetic factors and the emergence of suicidal behavior in psychosis (Serafini et al., 2012).

Research on the stability of aggression tends to reveal more patterns of continuity than discontinuity (Piquero et al., 2012). Advances in statistical methods for data modeling have evidenced a nonlinear continuity of aggression, also revealing the existence of certain risk factors that explain variance in aggression that is above and beyond its continuity from childhood to adulthood (Petersen et al., 2015). This continuity is particularly strong over time in individuals manifesting early highly aggressive behavior (Piquero et al., 2012). Therefore, identifying groups of aggressive behavior is of particular importance during the school age, allowing for the examination of the psychological profile of children exhibiting high levels of aggression. Furthermore, individual and contextual differences between high and low aggression groups may provide clues as to both protective and predisposing variables that should be considered in intervention/prevention actions for highly aggressive individuals. More specifically, this study analyzed the role of perfectionism, school refusal and affect as possible mitigating or enabling factors of such tendencies.

The interest in studying these three constructs is mainly due to their close link with overall psychopathology. Thus, perfectionism has even been considered as a transdiagnostic process (Egan et al., 2011, 2012). Specifically, with respect to child perfectionism, it has been conceptualized according to two dimensions (Flett et al., 2016): Socially Prescribed Perfectionism (SPP), understood as the tendency to consider the environment as highly demanding of perfectionism; and Self-Oriented Perfectionism (SOP), which captures the tendency to be sharply self-critical and to impose excessive high performance goals on oneself.

Second, although there is a declining trend in overall school dropout rates (Freeman and Simonsen, 2015), high levels of this problematic prevail. Moreover, advances in the reduction of school dropout rates has not taken place equally in terms of cultural and socioeconomic aspects. This situation justifies the relevance of addressing cases of school refusal which are, in turn, associated with multiple internalizing and externalizing problems (e.g., Maynard et al., 2012; Ingul and Nordahl, 2013). School refusal refers to the ''avoidance of a child attending school and/or persistent difficulty staying in the classroom throughout the school day'' (Inglés et al., 2015, p. 37). According to the functional model of Kearney and Diliberto (2014), the school refusal behavior may be explained based upon four reasons or factors which are not mutually exclusive: FI. To Avoid Negative Affectivity (associated with younger students who refuse to attend school and have difficulties identifying the cause of their discomfort); FII. To Avoid Social Aversion and/or Evaluation (linked with those students who present social difficulties and suffer in assessment situations, such as exams or oral presentations); FIII. To Pursue Attention (related to those students who prefer staying at home or with their parents or loved ones instead of going to school); and FIV. To Pursue Tangible Reinforcement (characterized by truancy based on the desire to engage in leisure activities outside of the educational center, such us staying at home watching the TV, playing computer games or spending time with friends).

Third, Positive affect (PA) is a positive, energetic, emotional, affiliation and dominion dimension of an individual, whereas Negative Affect (NA) is characterized by mood states such as sadness, aversion, anger, contempt, disgust, guilt, fear and nervousness (Watson et al., 1988; Clark et al., 1994). Both dimensions have been widely used by research as indicators of adjustment and maladjustment, respectively (e.g., Schütz et al., 2013; Liu et al., 2014).

However, despite the importance of these three variables, previous studies examining the relationship between perfectionism, school refusal, affect and aggressive behavior are limited or non-existent. Regarding perfectionism (i.e., SPP and SOP), García-Fernández et al. (2017) were the first to analyze the association between SPP and aggressive behavior in accordance with the model of Buss and Perry, using a Spanish sample of students aged 8–11. Results revealed that all components of aggressive behavior were significant and positive predictors of high levels of SPP. Likewise, students with high levels of SPP scored significantly higher on all components of aggressive behavior than their peers with low levels of SPP. Also, regarding Spanish child population, Vicent et al. (2017) found that a cluster defined by high SPP and SOP obtained significantly higher levels of Anger, Hostility and Physical and Verbal Aggression than any other combination of SPP and SOP. Lastly, Stoeber et al. (2017) used bivariate and partial correlations (controlling for the effect of the other perfectionist facets) in three samples of English university students. Thus, while the results indicated a positive and significant relationship between SPP and all components of aggressive behavior, except for Verbal Aggression; SOP obtained positive and significant bivariate correlations with Hostility and Verbal Aggression, as well as negative and significant partial correlations with Physical Aggression.

On the other hand, no study to date has examined the relationship between the functional model of school refusal and the aggression model of Buss and Perry (1992). However, research has revealed how this type of aversion to the school may influence the development of the aggressive behavior. For instance, Wallinius et al. (2016), in a sample of Swedish prisoners, found that school absenteeism at early ages was one of the major predictors of antisocial behaviors during adulthood. Similarly, other studies with American clinical child and/or adolescent populations have agreed to positively link FIII. To Pursue Attention and FIV. To Pursue Tangible Reinforcement with externalizing behaviors (Kearney and Silverman, 1993; Higa et al., 2002; Kearney, 2002). Specifically, Kearney and Silverman (1993) found positive and significant correlations between FIII and FIV and externalizing behavioral problems. Nevertheless, in the study of Higa et al. (2002), these correlations were only significant for FIV. Similarly, Kearney (2002) also concluded that both externalizing and internalizing problems jointly prevailed in FIII, whereas only externalizing problems prevailed in FIV.

Lastly, of the two affective dimensions, NA is the most closely linked with the aggressive behavior (Donahue et al., 2014). Based on the model of Buss and Perry, several studies have examined how both affective dimensions have been linked with aggression. Specifically, Verona et al. (2002), in a sample of American undergraduates, found that participants with high negative emotion scored significantly higher on all components of aggressive behavior as compared to their peers with low negative emotion. Considering only the Anger dimension, Harmon-Jones (2003) found positive and significant correlations between Anger and NA, as well as non-significant correlations with PA, in a sample of American undergraduate students. In contrast, Hewig et al. (2004), using German undergraduates, observed that PA was significantly and negatively associated only with Hostility, whereas NA was not significantly linked with Anger, Hostility and Physical and Verbal Aggression. Similarly, Dufey and Fernández (2012), in a sample of Chilean undergraduates, found that PA was negatively and significantly linked with all of the components of aggressive behavior, whereas NA was linked in a positive sense. More recently, Shachar et al. (2016), in a sample of Israeli students (grades 3–6) with observed aggressive behavior, obtained positive and significant correlations between NA and Anger, Hostility and Physical Aggression, with no data provided regarding Verbal Aggression. In contrast, such correlations were negative and significant between PA and Hostility and Physical Aggression, whereas the relationship with Anger was not statistically significant.

This study has a two-fold goal. First, it aims to verify whether there are different profiles of students with aggressive behavior, considering the cognitive, emotional and motor components established by Buss and Perry (1992). On the other hand, some important questions about the relationship between aggressive behavior and perfectionism, school refusal and affect remain unanswered. Thus, no study to date has examined the link between SOP, Positive and NA and the three components of aggressive behavior during childhood. Regarding school refusal, there is no previous empirical evidence about how the four functions of school refusal are associated with aggressive behavior. So, in order to overcome this limitations, after the identification of the profiles of aggressive behavior, the significant differences between the profiles identified on perfectionism (i.e., SPP and SOP), the four factors of school refusal, and Positive and NA are determined.

Thus, it is expected that profiles with more levels of aggressive behavior shall obtain: (a) Hypothesis 1. Significantly higher levels in SPP and SOP, in accordance with those studies that suggest a positive relationship between both perfectionist dimensions and all or some of the components of the aggressive behavior (García-Fernández et al., 2017; Stoeber et al., 2017; Vicent et al., 2017); (b) Hypothesis 2. Significantly higher scores in FIII and FIV on school refusal, in line with previous research that has identified a positive relationship between these factors and externalizing problems (Kearney and Silverman, 1993; Higa et al., 2002; Kearney, 2002); and (c) Hypothesis 3. Significantly lower scores on PA and higher on NA, in accordance with previous works that observed a negative and significant association between PA and some components of the aggressive behavior (Hewig et al., 2004; Dufey and Fernández, 2012; Shachar et al., 2016), as well as positive in the case of NA (Verona et al., 2002; Harmon-Jones, 2003; Dufey and Fernández, 2012; Shachar et al., 2016).

### MATERIALS AND METHODS

### Participants

The sample was recruited using multi-stage random cluster sampling, which means that all clusters were randomly selected in each stage, in the geographic areas: central, north, south, east and west of the Alicante province (Spain). Between one and three centers were randomly and proportionally selected from each geographical zone. Thus, a total of 16 schools were selected. From each of these schools, one group per academic grade was randomly selected from 3rd to 6th grade of primary education. Following this procedure, an initial sample of 1397 students was obtained, of which 195 were excluded because: (a) their parents and/or legal guardians did not give written contest to participate in the study (N = 74); (b) they did not have the minimum reading level required; (N = 68); or because; of (c) errors or omissions in the questionnaire completion (N = 53). Thus, the final sample consisted of 1202 Primary Education students aged 8–12 (Mage = 10.25, SD = 1.28). 48.6% of participants were males and 51.4% were females. Sample distribution across age was: 12.6%, 17.3%, 20.3%, 32.3% and 17.5%, respectively for students from 8 to 12 years. As for ethnic composition, 88.1% were Spanish, 5.9% South American, 4.7% Arab and 1.3% were of other origins.

### Instruments

### The Aggression Questionnaire (AQ; Buss and Perry, 1992)

The Aggression Questionnaire (AQ) is a 29-item self-report measure with a 5-point Likert scale of four dimensions of aggressive behavior: Anger, Hostility, Physical Aggression and Verbal Aggression. Specifically, the Spanish version of the scale validated by Santisteban and Alvarado (2009), whose levels of reliability range from α = 0.65 (Anger) to 0.80 (Physical Aggression), was used. Acceptable internal consistency indices were obtained in this study: α = 0.71, 0.73, 0.80 and 0.75, respectively, for Anger, Hostility, Physical Aggression and Verbal Aggression.

### The Child-Adolescent Perfectionism Scale (CAPS;

### Flett et al., 2016)

The Child-Adolescent Perfectionism Scale (CAPS) is a 22-item self-report measure with a 5-point Likert scale of SPP and SOP. It is the most widely employed measure to assess child perfectionism (García-Fernández et al., 2016). Specifically, the Spanish translation of the scale provided by Castro et al. (2004), whose levels of reliability were α = 0.82/92 (SPP) and α = 0.75/92 (SOP), was used. Acceptable internal consistency indices were obtained in this study: α = 0.74 for SPP and 0.76 for SOP.

### The School Refusal Assessment Scale Revised for Children (SRAS-R; Kearney, 2002)

The School Refusal Assessment Scale Revised for Children (SRAS-R) is a 24-item self-report measure with a 7-point Likert scale of four functions of school refusal: FI. To Avoid Negative Affectivity, FII. To Avoid Social Aversion and/or Evaluation, FIII. To Pursue Attention, and FIV. To Pursue Tangible Reinforcement. In this study, the Spanish version developed by Gonzálvez et al. (2016), whose levels of reliability range from α = 0.70 (FI) to 0.87 (FIII), was employed. Acceptable internal consistency indices were obtained in this study: 0.72, 0.74, 0.76 and 0.71, respectively, for the four factors of the SRAS-R.

## The Positive and Negative Affect Schedule for

Children (10-Item PANAS-C; Ebesutani et al., 2012) The 10-Item Positive and Negative Affect Schedule for Children (10-Item PANAS-C) is a self-report measure with a 5-point Likert scale of PA and NA. Internal consistency levels of 0.86 and 0.82 were obtained for PA and NA (Ebesutani et al., 2012). The 10-Item PANAS-C was translated to Spanish using a back-translation method, in accordance with the recommendations of Hambleton and Lee (2015). Acceptable internal consistency indices were obtained in this study: 0.78 and 0.82, respectively, for PA and NA.

### Procedure

A meeting was held with the head teachers of the selected educational centers in order to inform them of the aims of the study and to request their collaboration. All centers agreed to cooperate in the investigation. Subsequently, the written informed parental consent was requested from all participants. Since the study participants were minors, a letter describing the aims of the study was provided to parents and/or legal guardians of the students selected to participate. The letter included a section that parents and/or legal guardians were to sign and return to the school in case they give their consent to participate in the study. Only those minors who provided the parental consent participated in the study. The four tests were administrated in a single 60-min session, in which a researcher was present. During test administration, the researcher highlighted the strictly voluntary and anonymous character of the activity. Participants did not receive compensation for their contribution to this study.

This research was carried out in accordance with the recommendations of the ethical standards of the 1964 Helsinki Declaration and its subsequent amendments. The protocol was approved by the Ethics Committee of the University of Alicante (Spain) (UA-2017-09-05). Written parental informed consent was obtained from all parents or legal guardians of minors participating.

### Data Analysis

The profiles of child aggressive behavior were defined based on the different combinations of Anger, Hostility, Physical Aggression and Verbal Aggression. To determine the number of profiles, a Latent Class Analysis (LCA) was performed. LCA is a model-based technique that is currently considered to be the best method of identifying homogeneous classes of subjects given that it overcomes all of the problems related to K-means clustering (Schreiber, 2017). Data-driven calculations begin with one class. Individuals are then successively allocated to an ascending number of classes. The fit indices and the criteria taken into account when choosing the most adequate class solution were the lowest Bayesian Information Criteria (BIC) and Entropy values closer to one (Nyland et al., 2007; Schreiber, 2017; Smeets et al., 2017).

Once the aggressive behavior profiles were established, the inter-class differences in the obtained scores on perfectionism, school refusal and affective dimensions were analyzed using the analysis of variance (ANOVA). Moreover, gender was included

as a covariate to analyze its moderator effect. The Scheffé method was used to analyze the post hoc tests, as well as the Cohen's d index to calculate the effect size of the observed differences. Specifically, d levels between 0.20 and 0.49 indicate a small effect magnitude; between 0.50 and 0.79 indicate a moderate magnitude; and ≥0.80, a large one (Cohen, 1988).

### RESULTS

### Aggressive Behavior Profiles

The LCA allowed for the identification of three profiles characterized by different levels of aggressive behavior (see **Figure 1**). As shown in **Table 1**, this three-class model was the best-fitting, having the lowest BIC and the highest entropy. The first class, High Aggression, included 590 students (49.08%) having high levels in all components of aggressive behavior. The second class, Moderate Aggression, consisted of 462 participants (38.46%) characterized by moderate levels in all components of aggressive behavior, whereas the third class, Low Aggression, classified 150 students (12.48%) with low levels of aggressive behavior.

### Inter-class Differences in Perfectionism, School Refusal and Affect

The results of the ANOVA indicate the existence of statistically significant differences between the three profiles of aggressive behavior for all of the variables considered in this study, with the exception of FIV of the SRAS-R (see **Table 2**). These differences were maintained when gender was included as a covariate, i.e., still a significant main effect of class.

Specifically, post hoc comparisons revealed that High Aggression obtained significantly higher mean scores in SPP,

TABLE 1 | Fit indices of the latent class analysis (LCA) values in bold show the best model fit.


TABLE 3 | Cohen's d index to post hoc contrasts between the mean scores obtained and the three classes in the factors of perfectionism, school refusal and affect.


Note: FI, Factor I. To Avoid Negative Affectivity, FII, Factor II. To Avoid Social Aversion and/or Evaluation, FIII, Factor III. To Pursue Attention, FIV, Factor IV. To Pursue Tangible Reinforcement, PA, Positive Affect; NA, Negative Affect; n.s., non-significant differences.

SOP, FI, FII and FIII for school refusal and NA as compared to Moderate and Low Aggression. The comparisons between Moderate Aggression and Low Aggression were significant in the case of the three first factors of the SRAS-R and NA. Finally, High Aggression and Moderate Aggression scored significantly lower on PA than Low Aggression.

The magnitude of these differences (Cohen's d index) ranged from 0.24 to 1.36. Differences between High Aggression and Low Aggression revealed the largest effect sizes (see **Table 3**).

### DISCUSSION

Three profiles of child aggressive behavior (High Aggression, Moderate Aggression and Low Aggression) characterized, respectively, by high, moderate and low levels on all AQ dimensions (i.e., Anger, Hostility, Physical and Verbal Aggression) were identified. These results indicate that although such dimensions reflect different components of aggression, a close and positive relationship exists among them (McKay et al., 2016). Thus, the complexity of aggressive behavior is evident, extending beyond the motor component (i.e., Physical and Verbal Aggression) and it also implies the hostile beliefs and cognitions system of the subject as well as their wrathful emotional tendencies.

TABLE 2 | Mean scores, standard deviations and post hoc contrasts between mean perfectionism, school refusal and affect scores obtained by the three classes of aggressive behavior.


Note: FI, Factor I. To Avoid Negative Affectivity, FII, Factor II. To Avoid Social Aversion and/or Evaluation, FIII, Factor III. To Pursue Attention, FIV, Factor IV. To Pursue Tangible Reinforcement, PA, Positive Affect; NA, Negative Affect. <sup>∗</sup>Also when gender was considered as a covariate.

Regarding the results for each class, inter-class differences were found for all of the dimensions considered in this study, with the exception of the FIV on school refusal. High Aggression reported higher scores on all maladjustment dimensions, as well as the lowest levels in PA, proving to be the most maladaptive group. In contrast, Low Aggression obtained the best results in terms of psychological adjustment. As for the third profile, Moderate Aggression emerged as the second most maladaptive profile. These results question the premise that certain levels of aggression may become adaptive and demonstrate that children with a low aggressive behavior profile tend to be better adjusted psychologically. Thus, although certain authors have attributed some benefits to aggression for the resolution of certain social problems (Pellegrini, 2007), aggressive behavior should be considered both dangerous and time-consuming (Nelson and Trainor, 2007).

Similarly, differences in the mean scores in perfectionism, school refusal and affect between the three profiles identified in this study allow for the examination of the relationship between such variables and aggressive behavior.

First, with respect to both perfectionist dimensions, students characterized by high levels of aggression scored significantly higher on SPP and SOP than their peers with low and moderate levels of aggressive behavior. These results are in accordance with Hypothesis 1 and with those studies that found evidence of a positive association between the perfectionist dimensions and the components of aggressive behavior (García-Fernández et al., 2017; Stoeber et al., 2017; Vicent et al., 2017). However, effect sizes of these differences have shown that aggressive behavior is more closely linked to SPP than to SOP. García-Fernández et al. (2017) explained the relationship between SPP and aggression based on the frustration-aggression model. According to these authors, children with high SPP manifest aggressive behaviors, either physical or verbal, for two reasons. On the one hand, as a consequence of the anger experienced toward others after the humiliation resulting from not being able to achieve the imposed expectations. On the other hand, it could be justified as an attempt to defend themselves from an environment that is considered to be highly harsh and critical, in other words, as a consequence of the hostility. Likewise, the frustration resulting from failing to reach the high self-imposed goals and the anger derived from a strong tendency to self-criticism would explain the link between SOP and aggressive behavior (Vicent et al., 2017).

Regarding the relationship between the aggression profiles and the explanatory factors of school refusal, the results partially support Hypothesis 2, since students characterized by high levels of aggression scored significantly higher on the first three factors of the SRAS-R as compared to their peers with low levels of aggressive behavior. The large differences between High and Low Aggression in FIII are expected, given that this third factor of school refusal has been associated with both internalizing and externalizing problems (Kearney, 2002). However, the effect sizes revealed that the largest differences between the High and Low Aggression profiles were associated with the first two factors of the SRAS-R, which present comorbidity with generalized anxiety, social anxiety and depression (Kearney and Albano, 2004; Kearney et al., 2005). These findings contrast with previous research that has associated externalizing problems with FIII and FIV, mainly, with subjects who base their school refusal on FIV due to its relationship with school absenteeism. FIV is more frequent in adolescents than in children. It is not based on anxiety and is developed without parental consent (Yahaya et al., 2010; Kearney, 2016). However, a number of emotional reactions such as excessive anxiety, crying, stress, or excessive somatic complaints may also arise, especially in younger students who base their school refusal on FI and FII (Kearney and Bensaheb, 2006), which could be accompanied by an aggressive behavior response. In fact, in a review performed by Grant et al. (2004) it was noted that in 53 of the 60 studies reviewed, general stress levels predicted high levels of aggressive behavior.

Despite the fact that scientific literature has revealed a positive and significant correlation between the FIII and FIV on school refusal and the presence of externalizing behaviors (Kearney and Silverman, 1993; Higa et al., 2002), none of the previous studies was carried out in a Spanish community sample of participants in the late childhood period (8–12 years of age) and applying an instrument of aggressive behavior based on components at cognitive, emotional and motor levels. Therefore, the particularities of this research may also be explanatory factors for the differences found.

Lastly, according to Hypothesis 3, participants with high levels of aggressive behavior reported significantly lower levels of PA as compared to those having Low Aggression levels. These results also coincide with previous research that has found a negative and significant association between aggressive behavior and PA (Hewig et al., 2004; Dufey and Fernández, 2012; Shachar et al., 2016). Thus, this negative relationship could be explained by the fact that individuals who tend to show positive emotional states usually present greater prosocial behavior (Aknin et al., 2018). In contrast, literature on aggression has concluded that individuals tend to attack others when experiencing negative emotions (Verona et al., 2002; Harmon-Jones, 2003; Dufey and Fernández, 2012; Shachar et al., 2016). In other words, ''when people feel bad, they are too likely to have angry feelings, hostile thoughts and memories, and aggressive inclinations'' (Berkowitz, 2001, p.335). According to this statement, results from this study have found that High Aggression scored significantly higher on NA than the other profiles, with these differences being moderate in size.

### LIMITATIONS AND FUTURE RESEARCH

Certain limitations of this study should be considered. First, the sampling procedure and the sample size ensure the representativeness of the Spanish community population between the ages of 8 and 12. Nevertheless, results from this study must be generalized with caution. Therefore, it would be interesting for future research to replicate this study using other age and cultural groups. Second, the design employed impedes the establishment of causal relationships between aggressive behavior and perfectionism, school refusal and affect. This limitation could be solved by using longitudinal data or with structural equation modeling. Third, this work also has the limitations of using self-report measures (Fernández-Montalvo and Echeburúa, 2006) which could be solved with a multi-source and multi-method assessment. Furthermore, it should be noted that this study does not take into account the potential relationship between perfectionism, school refusal and affect. This limitation should be addressed by future research performing a causal-explanatory model of aggression behavior considering the possible links between the predictor variables (i.e., perfectionism, school refusal and affect). Finally, it should be noted that this study is based on the Buss and Perry model which defines aggressive behavior as a set of three components: emotional, cognitive and motor. Thus, other functions such as proactive/reactive aggression have not been considered.

### PRACTICAL IMPLICATIONS AND CONCLUSIONS

Despite the mentioned limitations, this work is a novel contribution for research on aggression for several reasons. First, it is the first study to examine profiles of aggressive behavior while jointly considering the dimensions of Anger, Hostility, Physical Aggression and Verbal Aggression. Second, no previous study has analyzed the link between aggressive behavior (according to its three components: emotional, cognitive and motor) and SOP and Positive and NA in child population. Last but not least, this is the first work to offer empirical evidence on the relationship between aggressive behavior and school refusal. Thus, in accordance with our results, almost 50% of the child population manifests high levels of Anger, Hostility and Physical and Verbal Aggression. These children also present clear perfectionist trends. They often feel sad, afraid and miserable and rarely experience positive emotions like happiness or pride. Furthermore, they tend to refuse school since attending school causes them great discomfort, because they suffer in social or school evaluation situations, or given that they have difficulties in being separated from their parents. This high aggression profile could be detected by school

### REFERENCES


counselors using the AQ and identifying those students who report high scores in all dimensions evaluated by the mentioned scale.

Finally, if the correlates of aggressive behavior, perfectionism, school refusal and NA with certain problems and psychopathologies are taken into account (e.g., Jaafar et al., 2013; Thornton et al., 2013; Morris and Lomax, 2014; Schütz et al., 2014), High Aggression could be considered a profile having a high risk of psychological vulnerability to be treated without further delay. Therefore, Farrington et al. (2017) described the effectiveness of developmental prevention programs on aggressive behavior in children and adolescents. These programs are defined as community-based programs that may be focused on individual (providing training in social competencies, interpersonal problem solving and other cognitive or behavioral skills), family (providing counseling on child-rearing, coping with family stress or training in parenting skills) or school (improving the school climate, teaching behavior, etc.). Likewise, prevention programs on aggression should be combined with strategies designed to bolster levels of resilience in order to avoid or overcome psychological problems (i.e., perfectionism, school refusal and NA) associated with children having a High Aggression profile.

### AUTHOR CONTRIBUTIONS

MV and CG have participated in conducting a literature search and writing this manuscript. CJI has reviewed this research. RS has participated in performing statistical analyses. JMG-F has designed this research.

### FUNDING

Part of this investigation is supported by a project of the Vice-Rectorate of Research and Knowledge Transference of the University of Alicante (GRE16-07), awarded to the fourth author and a project of the Ministry of Economy and Competitiveness (EDU2012-35124) awarded to the fifth author.


**Conflict of Interest Statement**: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Vicent, Inglés, Sanmartín, Gonzálvez and García-Fernández. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# The Human Anger Face Likely Carries a Dual-Signaling Function

#### Jinguang Zhang\*

Department of Communicology, University of Hawaii, Honolulu, HI, United States

Keywords: interpersonal aggression, anger, facial expression, physical strength, aggressive intent, costly signaling

### INTRODUCTION

Anger is an integral part of interpersonal aggression (Baumeister et al., 1990; Sell et al., 2009b) and has a cross-culturally recognizable facial expression (Ekman, 1973). This expression typically entails simultaneously lowering one's browridge, raising the cheeckbones and mouth, widening the nose, and pressing the lips (Ekman and Friesen, 1978; Sell et al., 2014). Given these species-typical features, recent studies sought to reveal their signaling function. That is, what does the human anger face communicate?

Sell et al. (2014) argued that the anger face mainly enhances facial cues of physical strength, thereby increasing the angry person's perceived fighting ability. In a paper published in the same year, Reed et al. (2014) argued that the anger face communicates the angry person's commitment to carry out threats. We believe that these two hypotheses complement each other to provide a more complete analysis of the signaling function of the anger face. In our discussion, we focus on men because interpersonal aggression is primarily a male activity (Puts, 2010). At the same time, though this opinion piece concerns the signaling function of face, we use research on vocal signals to build up our arguments. This is because (to our knowledge) a major component of this opinion piece, namely aggressive-intent signaling, has been mostly demonstrated with vocal signals.

### Aggressive Signals

Aggressive signals are naturally-selected structures or acts that communicate signalers' threat potential, including their resource-holding potential (RHP; e.g., physical strength) and aggressive intent (i.e., the willingness to escalate in a fight; Hurd and Enquist, 2005). Both types of aggressive signals are prevalent in animals (see below), and the use of those signals helps reduce the cost of combats. For example, adult red deer stags weigh ∼330 pounds on average and carry large, piercing antlers, and both features are capable of causing serious physical damages. However, the annual rate of permanent injuries is ∼6% among stags that engage in rutting fights (Clutton-Brock et al., 1979). This is partly because roaring contests, where two stags stand apart from and take turn to roar at each other, resolve ∼50% of the fights on average (Clutton-Brock and Albon, 1979). The roaring contests can resolve conflicts of interest because stags' roars convey information predictive of their chance of winning a pending fight against each other, and stags use such information to make fight-or-flight decisions.

### Signals of RHP

Stags' roars are an RHP signal because the roaring rate correlates with stag's physical condition (e.g., deterioration caused by aging; Clutton-Brock and Albon, 1979) and the minimum formant frequency of the roars correlates with stags' body weight (Reby and McComb, 2003). With all else being equal, stags that are in better conditions and/or heavier are more likely to win physical fights. Importantly, only stags in better physical conditions can roar faster because roaring fast is energetically demanding. Stags in worse conditions may be able to roar faster than its condition allows, but this cannot last long and will quickly

#### Edited by:

Nelly Alia-Klein, Icahn School of Medicine at Mount Sinai, United States

#### Reviewed by:

Elias Manjarrez, Benemérita Universidad Autónoma de Puebla, Mexico

> \*Correspondence: Jinguang Zhang jzhang6@hawaii.edu

Received: 26 November 2017 Accepted: 07 February 2018 Published: 20 February 2018

#### Citation:

Zhang J (2018) The Human Anger Face Likely Carries a Dual-Signaling Function. Front. Behav. Neurosci. 12:26. doi: 10.3389/fnbeh.2018.00026 exhausts the stags, impairing their ability to make the next move. As such, an energetic cost proportional to signalers' condition prevents weaker stags from faking greater RHP, and the honesty of roaring rates as an RHP signal is maintained (i.e., the handicap principle; Zahavi, 1975). At the same time, because body weight is almost impossible to fake, stags' roars also constitute an "unfakeable" index signal of RHP (Maynard Smith and Harper, 2003).

### Signals of Aggressive Intent

Aggressive-intent signals broadcast one's willingness to escalate in combats, and, as RHP signaling, animals that signal stronger aggressive intent are more likely to win the contested resources without fighting (Searcy and Nowicki, 2005). Much research showed that aggressive-intent signals exist (e.g., Vehrencamp, 2001; Searcy et al., 2006; Akçay et al., 2011), contrary to earlier arguments (e.g., Maynard Smith, 1982) that aggressive-intent signals could not have evolved. Those arguments are based on the observation that the association between the form of most aggressive-intent signals (e.g., song singing) and their content (e.g., aggressive intent) is often arbitrary. This would render the signals prone to bluff and thus useless in the long run in resolving conflicts of interest.

However, the retaliation-cost model (Enquist, 1985) suggests that aggressive-intent signals can be honest if they elicit aggression from signal receivers. Specifically, Enquist considered how an animal can use one of two signals, S (strong), and W (weak), to signal different levels of aggressive intent. S is more intense and more effective in repelling opponents than W is, and both strong and weak animals can use S and W equally well (i.e., the signals do not entail production or maintenance costs).

In the model, the focal strategy is: (1) if strong, signal S; if the opponent responds with S, attack; if the opponent signals W, repeat S, and attack if the opponent does not withdraw; but (2) if weak, signal W and give up if the opponent responds with S, but attack if the opponent signals W. Bluffing (i.e., using S when being weak) may succeed against weak opponents but will solicit attacks from strong opponents. When the cost to a weak animal of being attacked by a strong animal is larger than the benefit of bluffing, the focal strategy that promotes honest signaling (e.g., using S only when strong) can prevail against bluff and be selected.

Thus, the retaliation-cost model suggests that (1) aggressiveintent signals are honest when signal intensity is calibrated to signalers' RHP and (2) a receiver-dependent cost (i.e., retaliation) keeps deception rates low. Supporting the model, Anderson et al. (2012) showed that song birds higher in trait aggressiveness are more likely to approach opponents that emit soft songs, a putative aggressive-intent signal (see also, Popp, 1987; Molles and Vehrencamp, 2001). Recently, Zhang and Reid (2017) showed that men with greater threat potential (e.g., higher in upper-body strength) become more aggressive upon hearing a low-pitched male voice under a mating prime that simulates intrasexual competition. This finding suggests that the retaliation-cost model can be used to study human aggressive interactions.

### The Case of the Face

Facial muscles are highly homologous in anthropoids (Diogo et al., 2009). Like humans, several species of nonhuman primates (e.g., bonobos and chimpanzees) are capable of making facial expressions considered "angry" or threatening (Bard et al., 2011; Waller and Micheletta, 2013). For example, the bulging-lip face of chimpanzees is akin to the human anger face because they share muscle action units, including the chin raiser and lip pressor (Parr et al., 2007). The staring bared-teeth scream face, silent scream face, and the tense face of chimpanzees are threatening because those facial expressions are mostly observed in aggression initiators (Parr et al., 2005). However, whether those expressions signal chimpanzees' RHP, aggressive intent, or both remains unknown.

The human face is a reliable RHP signal, as men can accurately track other men's physical strength by looking at those men's neutral face (Sell et al., 2009a). Adding to the signaling function of the face, Sell et al. (2014) showed that the individual components of a prototypical anger face (e.g., lowered browridge) make the angry person appear physically stronger. The anger face also correlates with the angry person's approach tendencies (e.g., Yik, 1999; Adams et al., 2006). In particular, Reed et al. (2014) argued that the anger face signals the angry person's commitment to carry out threats by showing that proposers in an ultimatum game made more generous offers to a person making an angry face than to a person making a neutral face. Reed et al. further argued that the anger face honestly signals threat commitment because people only make the anger face when angry and that intense outburst of emotion makes its expression (i.e., the anger face) difficult to fake. Second, the complex neurological mechanisms associated with facial expressions also make an anger face difficult to fake. Collectively, this second line of research suggests that the anger face signals aggressive intent (Fridlund, 1994).

### An Integrative Hypothesis

We believe that the cue-enhancement hypothesis (Sell et al., 2014) and the threat-commitment hypothesis (Reed et al., 2014) complement each other. Specifically, an anger face is scary (Fridlund, 1994), but behaviors only enhancing strength cues do not necessarily strike fear. For example, the movements that bodybuilders make in competitions should increase their perceived strength by emphasizing their upper-body muscles. However, audiences unlikely watch bodybuilding competitions with fear, because they know that the bodybuilders are not going to use that increased strength to attack them. In other words, the threat value (Searcy and Beecher, 2009) of aggressive signals will be reduced if intent cannot be reliably signaled and perceived.

We suspect that the anger face is intimidating because it signals the angry person's aggressive intent in addition to increasing the person's perceived strength by sending the message that "I am going to inflict this much physical force on you if you do not back off." This message integrates the signaling functions specified by the cue-enhancement and threatcommitment hypotheses (i.e., fighting ability and intent). It also addresses a limitation of the threat-commitment hypothesis, that is, it is not clear from Reed et al. (2014) what the anger face

commits in a threat. Given the link between the anger face and physical-strength perceptions (Sell et al., 2014), the commitment is likely about the amount of physical force one is ready to deploy in a pending fight.

If this is the case, a critical question is what maintains the honesty of the anger face as an aggressive signal. Reed et al. (2014) elaborated two mechanisms (see above), and we suggest two more here. The first is an energetic cost. Anger, at least when it is genuine, is tiring due to activities such as muscle contraction, blood vessel dilation, and increases in breathing and heart rates. These activities enhance one's physical strength, prepare one for a fight, but are also energetically demanding. It follows that only men in better physical conditions can more effectively mobilize their strength through angering, making their threat genuine and their anger face scary. Men in poorer physical conditions (e.g., being exhausted at the moment) can also assume an anger face, but the face is likely perceived as less intimidating. Compared to men in better conditions, those in poorer conditions should be less effective in mobilizing their physical strength, rendering their threat less genuine.

A retaliation cost can also help maintain the signal honesty of the anger face. To the extent the retaliation-cost model can be used to study human aggressive interactions (Zhang and Reid, 2017), the model predicts that an angrier face (compared to a less angry one) should be more likely to elicit aggression from physically stronger men than from physically weaker men. We are not aware of tests of this prediction. However, angry facial expressions were shown to induce approaching behaviors in observers in some studies (Adams et al., 2006; Wilkowski and Meier, 2010) but avoiding behaviors in others (Marsh et al., 2005; Seidel

### REFERENCES


et al., 2010). The retaliation-cost model flags an unmeasured moderator: respondents' (i.e., signal receivers') physical strength. Approaching behaviors, which is associated with retaliation, should be more common among physically stronger respondents whereas avoidance should be more common among weaker respondents. Testing this moderation effect may help reconcile the seemingly inconsistent findings described above.

**Figure 1** describes our integrative account of the signaling function of the human anger face in terms of its content and the costs that maintain its honesty.

This integrative hypothesis applies to other features of interpersonal aggression, too, such as violent yelling. To the extent that violent yells increase yellers' perceived physical strength (Sell, 2011), the yells may also carry a dual-signaling function like the anger face, that is, to advertise the amount of physical force one intends to inflict on opponents. At the same time, the honesty of violent yelling as an aggressive signal is also likely maintained by the energetic and retaliation costs. Testing these predictions will bridge human and animal models of anger and aggression and attest to the value of integrating those models in studying interpersonal aggression.

### AUTHOR CONTRIBUTIONS

JZ conceived, designed, and authored this research.

### FUNDING

This research is funded by an award (#1551982) of the U.S. National Science Foundation to JZ.

the song sparrow. Animal Behav. 82, 377–382. doi: 10.1016/j.anbehav.2011. 05.016

Anderson, R. C., Searcy, W. A., Hughes, M., and Nowicki, S. (2012). The receiverdependent cost of soft song: a signal of aggressive intent in songbirds. Animal Behav. 83, 1443–1448. doi: 10.1016/j.anbehav.2012.03.016

Bard, K. A., Gaspar, A. D., and Vick, S. (2011). "Chimpanzee faces under the magnifying glass: Emerging methods reveal cross-species similarities and individuality," in Personality and Temperament in Nonhuman Primates, eds A. Weiss, J. E. King, and L. Murray (New York City, NY: Springer), 193–231.


**Conflict of Interest Statement:** The author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Zhang. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Longitudinal Links between Executive Function, Anger, and Aggression in Middle Childhood

Helena L. Rohlf, Anna K. Holl, Fabian Kirsch, Barbara Krahé\* and Birgit Elsner

Department of Psychology, University of Potsdam, Potsdam, Germany

Previous research has indicated that executive function (EF) is negatively associated with aggressive behavior in childhood. However, there is a lack of longitudinal studies that have examined the effect of deficits in EF on aggression over time and taken into account different forms and functions of aggression at the same time. Furthermore, only few studies have analyzed the role of underlying variables that may explain the association between EF and aggression. The present study examined the prospective paths between EF and different forms (physical and relational) and functions (reactive and proactive) of aggression. The habitual experience of anger was examined as a potential underlying mechanism of the link between EF and aggression, because the tendency to get angry easily has been found to be both a consequence of deficits in EF and a predictor of aggression. The study included 1,652 children (between 6 and 11 years old at the first time point), who were followed over three time points (T1, T2, and T3) covering 3 years. At T1, a latent factor of EF comprised measures of planning, rated via teacher reports, as well as inhibition, set shifting, and working-memory updating, assessed experimentally. Habitual anger experience was assessed via parent reports at T1 and T2. The forms and functions of aggression were measured via teacher reports at all three time points. Structural equation modeling revealed that EF at T1 predicted physical, relational, and reactive aggression at T3, but was unrelated to proactive aggression at T3. Furthermore, EF at T1 was indirectly linked to physical aggression at T3, mediated through habitual anger experience at T2. The results indicate that deficits in EF influence the later occurrence of aggression in middle childhood, and the tendency to get angry easily mediates this relation.

Keywords: executive function, anger, relational aggression, physical aggression, reactive aggression, proactive aggression, childhood, longitudinal study

### INTRODUCTION

A meta-analysis that included a wide range of EF measures concluded that EF is negatively associated to antisocial behavior, with varying effect sizes depending on the specific form of antisocial behavior and the occurrence of comorbid problems (Ogilvie et al., 2011). However, there is a lack of longitudinal studies that examined the effect of deficits in EF on the development of aggression, particularly in middle childhood, taking into account different forms and functions of aggression. The present study extends previous research by examining the longitudinal links between EF and different forms (relational and physical) and functions (reactive and proactive) of

#### Edited by:

Rosa Maria Martins De Almeida, Federal University of Rio Grande do Sul (UFRGS), Brazil

#### Reviewed by:

Lin Sørensen, University of Bergen, Norway Stuart F. White, Boys Town, United States

> \*Correspondence: Barbara Krahé krahe@uni-potsdam.de

Received: 27 October 2017 Accepted: 07 February 2018 Published: 27 February 2018

#### Citation:

Rohlf HL, Holl AK, Kirsch F, Krahé B and Elsner B (2018) Longitudinal Links between Executive Function, Anger, and Aggression in Middle Childhood. Front. Behav. Neurosci. 12:27. doi: 10.3389/fnbeh.2018.00027 aggression over 3 years. In addition, previous research has mostly studied direct links between EF and aggression without considering potential underlying mechanisms. The present study addressed this issue by including individual differences in the experience of anger as a mediating variable.

### Executive Function

There is disagreement in the literature over the exact definition of EF. However, EF can be described as an umbrella term that is usually equated with conscious, higher order processes associated with the prefrontal cortex (Hughes et al., 2005). EF governs goal-directed action and planning of behavior, and allows for adaptive responses to novel, complex, or ambiguous situations. As an important aspect of self-regulation, EF is considered vital for autonomous and adaptive psychological functioning (Séguin et al., 2007). Miyake et al. (2000) differentiated between three components of EF in college students, namely inhibition of prepotent responses, working memory updating, and mental set shifting. In a latent-variable analysis, these factors were moderately correlated, but clearly separable, and also had some common underlying mechanisms that contributed to all EF tasks. This unity-but-diversity framework is the most accepted conceptualization of EF, supported also by studies with children and adolescents (e.g., Lehto et al., 2003; Huizinga et al., 2006). Inhibition involves withholding or restraint of a motor response, and is considered central to EF (Miyake et al., 2000). Working memory updating (working memory) is the ability to maintain and manipulate information over brief periods of time (Huizinga et al., 2006). Shifting is the ability to alternate between mental rule sets or tasks (Miyake et al., 2000), and is considered the most complex EF component. An additional EF component that is also frequently mentioned is planning, which is essential to the EF domains of goal setting and goal-oriented behavior (Anderson, 2002). Unlike other cognitive abilities, EF shows a pronounced development after early childhood, paralleling the protracted maturation of the prefrontal cortex (Blakemore and Choudhury, 2006). The single components of EF, however, seem to follow differential courses throughout childhood and adolescence, involving progressive and regressive phases of development (Best et al., 2009; Best and Miller, 2010).

### Aggression

Among the many advantages of EF is the ability to regulate behavior that is prohibited by social norms, such as aggressive behavior. Aggression is defined as "any form of behavior directed toward the goal of harming or injuring another living being that is motivated to avoid such treatment" (Baron and Richardson, 1994, p. 7). In the present study, we distinguished between different forms and functions of aggression. A widely used classification of forms of aggressive behavior is the distinction between physical and relational aggression. Physical aggression refers to behavior that is intended to harm another person through the threat or use of physical force, whereas relational aggression is defined as behavior aimed at damaging another person's social relationships or feeling of social inclusion (Crick and Grotpeter, 1995). Children's use of physical aggression normally decreases during their preschool years, whereas relational aggression tends to increase during middle childhood, particularly in girls (Côté et al., 2007).

The distinction between different functions refers to the motivation that leads a person to act aggressively. Unprovoked aggressive behavior that aims to reach a certain goal, such as social dominance or the achievement of material goals, is described as proactive aggression. Proactively aggressive behavior can also be described as "offensive," "instrumental," and "cold-blooded" (Vitaro et al., 2006). Proactive aggression is conceptually linked to callous-unemotional (CU) traits. Children high on CU traits are characterized by a lack of guilt, reduced empathy, reduced display of emotions, callousness, and uncaring behavior (Vitaro et al., 2006; Blair et al., 2014), and they use aggressive behavior to reach desired rewards or social dominance (Pardini et al., 2003). Reactive aggression, by contrast, refers to aggressive behavior that is displayed in response to a perceived threat or provocation (Dodge and Coie, 1987; Card and Little, 2006). Reactively aggressive behavior can also be described as "defensive," "impulsive," and "hot-blooded" (Walters, 2005). The majority of children seem to follow a stable-low course in reactive and proactive aggression over the course of middle childhood, but some children show substantial changes by either increasing or decreasing their use of reactive and/or proactive aggression (Cui et al., 2016). Taken together, longitudinal evidence suggests that middle childhood is a period of important developmental change for both the forms and functions of aggression.

### The Link between Executive Function and Aggression

A large body of correlational research has shown that EF is negatively related to aggression in preschool-aged children and adolescents. For instance, low levels of EF coincide with preschoolers' externalizing behavior, which includes aggressive behavior (e.g., Hughes and Ensor, 2008). Similarly, preschoolers who were rated as "hard to manage" by their parents showed significantly lower EF than a less problematic comparison group (Hughes et al., 1998). A meta-analysis that covered a broad age range from early childhood to adulthood concluded the negative relation between EF and antisocial behavior to be robust, with one of the largest effects for externalizing behavior disorder (Ogilvie et al., 2011). Similarly, in a meta-analysis on preschoolers, EF, inhibition in particular, was correlated with externalizing behavior with a medium effect size (Schoemaker et al., 2013). Longitudinal evidence demonstrated that 3-year old children with low levels of effortful control, a cognitive construct closely related to EF (Bridgett et al., 2013), showed an increased risk for a chronic pattern of elevated externalizing behavior throughout middle childhood (Olson et al., 2017). By comparison, research on the relation between EF and aggression in middle childhood is limited, although this age range would be important to investigate. As a consequence of the developmental change in the forms (Côté et al., 2007) and functions (Cui et al., 2016) of aggression in middle childhood and the ongoing development of EF (e.g., Blakemore and Choudhury, 2006), this developmental period is of particular interest.

Different theoretical explanations have been proposed for the link between EF and aggression. The frontal-lobe hypothesis of emotional and behavioral regulation suggests that cognitiveneuropsychological functions in the frontal lobe, which are related to EF, appear to be systematically impaired in individuals showing physical aggression (Séguin, 2009). That is, children whose physically aggressive behavior does not decline after preschool as expected (Côté et al., 2007) are thought to have deficits in their EF. As a consequence, they have problems in regulating their behavior and solving social problems (Séguin and Zelazo, 2005; Zadeh et al., 2007). For example, they may not represent a problem adequately, may show deficits in planning a solution, or in reacting flexibly to different kinds of social situations. This is also the focus of another theoretical framework, the Social Information Processing (SIP) model, which proposes that children who show aggressive behavior may have deficits in their social information processing compared to nonaggressive age mates (Crick and Dodge, 1994). These difficulties may be influenced by children's EF (Huesmann et al., 1987; de Castro and van Dijk, 2018).

A further explanation relates to the integration of emotional processes into social-cognitive information processing (Lemerise and Arsenio, 2000). Particularly, anger seems to play an important part in the mediation between social-cognitive processes and aggressive responses (de Castro et al., 2005). EF is involved in the regulation of negative affect already in toddlers (Putnam et al., 2008) and preschool children (Carlson and Wang, 2007). Accordingly, deficits in EF may increase the experience of anger. Indeed, deficits in EF are linked to higher levels of negative affect, including anger (Gagne and Hill Goldsmith, 2011; Healey et al., 2011; Bridgett et al., 2013). In addition, higher cognitive control—a related construct to EF—in adolescence seems to act as a buffer against later, maladaptive outcomes of chronic anger, for example, adult antisocial personality traits (Hawes et al., 2016). Further evidence comes from research on irritability, which is defined as an increased proneness to anger (Leibenluft, 2017). That is, children with higher levels of irritability showed deficits in the processing of emotional stimuli, impaired context-sensitive regulation (Leibenluft and Stoddard, 2013), and neural dysfunction in processes associated to EF, such as error monitoring, reward processing, and emotion regulation (Perlman et al., 2015). Anger, in turn, is an important impelling factor of aggressive behavior (Leibenluft and Stoddard, 2013). The role of anger as an antecedent of aggression can be explained by the anger-related action tendency that is assumed to activate aggression-related motor impulses (Berkowitz and Harmon-Jones, 2004). Further, irritability can be conceptualized as a maladaptive response to frustration or threat (Leibenluft, 2011). Supporting this assumption, children in preschool age and in middle childhood who are prone to anger were found to be more likely to engage in aggressive behavior (e.g., Eisenberg et al., 1999; Arsenio et al., 2000; Olson et al., 2005; Wakschlag et al., 2015). The theoretical and empirical links between EF and anger on the one hand, and anger and aggression on the other hand suggest that the association of EF with aggression may partly be explained by the habitual experience of anger. So far, this assumption has received little attention, especially for the developmental period of middle childhood, and was therefore addressed in the present study.

With regard to different forms of aggression, most research has focused on relations of EF and physical rather than relational aggression. One reason for this may be that—as suggested by the frontal-lobe hypothesis of emotional and behavioral regulation outlined above—impairments in EF appear to be specific to physically aggressive behavior (Séguin, 2009). Those studies that included both forms of aggression revealed mixed findings. In line with the frontal-lobe hypothesis, EF was found to be negatively associated with physical aggression and not related with relational aggression in 3- to 6-year-old children (O'Toole et al., 2017). By contrast, other cross-sectional research has found negative relations of EF to both physical and relational aggression in early childhood; only working memory was positively associated with proactive relational aggression (Poland et al., 2016). However, these studies did not account for overlapping variance of physical and relational aggression, which may have an impact on the respective relations. With regard to middle childhood, previous research has failed to support the assumption that only physical aggression is related to deficits in EF. In a sample of fourth- and fifthgrade children, impaired central executive working memory, an indicator of EF, was associated with both physical and relational aggression (McQuade et al., 2013). Furthermore, in a population sample of 9-year-olds, working memory updating was negatively related only to relational, not to physical aggression (Granvald and Marciszko, 2016). Another study in middle childhood did not find significant paths of impaired EF to physical or relational aggression after controlling for symptoms of attention deficit/hyperactivity disorder (Diamantopoulou et al., 2007). Altogether, the inconsistency among studies that have taken both relational and physical aggression into account points to the need for further research into the role of EF in the development of different forms of aggression.

Regarding functions of aggression, few studies have differentiated between reactive and proactive aggression when examining the link between EF and aggression, particularly in middle childhood. In 9- to 12-year-old children, deficits in EF, particularly in response inhibition and planning, were found to be positively associated with reactive aggression. The relations between planning and reactive aggression, but not between planning and proactive aggression were moderated by hostile attributional biases (Ellis et al., 2009; Rathert et al., 2011). In addition, a measure of self-regulation that included EFcomponents was negatively linked to reactive, but not proactive aggression in 6- to 16-year-old children and adolescents (White et al., 2013). Thus, deficits in EF seem to be more involved in the development of reactive compared to proactive aggression. One explanation for the relation between deficits in EF and reactive aggression may be the potential mediating role of anger, as outlined above. Because anger is a major component of reactive, but not proactive aggression (for a review, see Hubbard et al., 2010), it can be assumed that only reactive aggression is indirectly predicted by poor EF via the experience of anger. Consequently, anger may also mediate between EF and reactive aggression.

### The Current Study

The aim of this study was to examine the prospective paths between EF and different forms (physical and relational) and functions (reactive and proactive) of aggression in a large population-based sample in middle childhood, with the habitual experience of anger considered as a potential underlying mechanism. The study included three measurement time-points covering 3 years. At T1, a latent factor of EF was calculated from measures of inhibition, set shifting, working-memory updating, and planning, which were assessed by using behavioral tasks and a teacher-report measure. Children's tendency to experience anger was assessed via parent reports at T1 and T2, and the forms and functions of aggression were rated by teachers at all three time points. The prospective paths were analyzed via structural-equation modeling, controlling for age, gender, and information-processing capacity.

Based on the theoretical assumptions and previous evidence outlined above, four hypotheses were postulated: First, we expected to find a negative relation between EF at T1 and physical aggression at T3, such that lower EF would predict higher levels of later physical aggression (Hypothesis 1). Considering relational aggression, the existing evidence is mixed, because some research found a negative (McQuade et al., 2013) and other either no (Diamantopoulou et al., 2007) or even a positive relation to EF (Poland et al., 2016). We therefore examined these competing predictions for the relation between EF at T1 and relational aggression at T3 in our model. In addition to potential direct effects, we expected negative indirect effects between EF at T1 and physical and relational aggression at T3 through habitual anger at T2 (Hypothesis 2). Thus, we proposed that lower EF would predict a higher tendency to experience anger at T2, which in turn would predict higher rates of physical and relational aggression at T3. With regard to the functions of aggression, we postulated that EF at T1 would be a negative predictor of reactive aggression at T3 but would be unrelated to proactive aggression at T3 (Hypothesis 3), based on earlier evidence (e.g., Ellis et al., 2009; Rathert et al., 2011). Furthermore, we expected negative indirect effects between EF at T1 and reactive aggression at T3 through habitual anger at T2. Thus, lower EF at T1 would predict a higher tendency to experience anger at T2, which in turn would predict higher rates of reactive aggression at T3 (Hypothesis 4).

In addition, we tested potential gender differences in the postulated paths between EF, anger, and aggression. In previous research, gender differences received little attention, in particular in conjunction with the distinction between forms and functions of aggression. Given the gender-related differences in the occurrence of the two forms of aggression (boys usually show more physical, girls slightly more relational aggression; e.g., Card et al., 2008), it was deemed important to include gender as a potential moderator in the analyses. However, the few studies that have addressed gender differences yielded little support for the assumption that the longitudinal links between EF and forms or functions of aggression might differ by gender (White et al., 2013). Based on these findings, we expected that the proposed associations would hold for boys and girls.

### METHOD

### Participants and Procedure

The sample was part of a large longitudinal study on intrapersonal developmental risk factors in childhood and adolescence based at the University of Potsdam, Germany. The children were recruited from 33 public primary schools in the Federal State of Brandenburg, Germany. At T1, the sample consisted of N = 1,652 children (52.06% girls) aged between 6 and 11 years (M = 8.36, SD = 0.93)<sup>1</sup> . At T2, 1,611 children (51.8% girls) participated again (M = 9.12 years, SD = 0.93, range 7.11–11.89) and at T3, the remaining sample consisted of 1,501 children (51.5% girls; M = 11.07 years, SD = 0.92, range 9.12– 13.76). This corresponds to a high retention rate of 97.5% from T1 to T2 and 92.3% from T2 to T3.

The mean interval between T1 and T2 was 9.14 months (SD = 1.80), and between T2 and T3, it was 23.83 months (SD = 1.66).

Approval for the procedure and the instruments was granted by the Ethics Committee of the authors' university as well as the Ministry of Education, Youth, and Sport of the Federal State of Brandenburg. The EF tasks were administered in individual test sessions by trained project members at the participants' schools. Parents and teachers completed the questionnaires either online or in paper-pencil form. For each child, informed consent was obtained from the parents.

### Measures

### Executive Function

The EF subcomponent inhibition was assessed by the Fruit Stroop task (Archibald and Kerns, 1999; adapted by Röthlisberger et al., 2010), a child-version of a Stroop paradigm with vegetables and fruits as stimulus items. The task consisted of four trials, and in each a page depicting 25 stimuli was presented to the child. Page 1 depicted colored rectangles (blue, green, red, yellow), page 2 showed fruits or vegetables in their typical colors (plum blue, lettuce—green, strawberry—red, banana—yellow). Page 3 depicted the same fruits and vegetables, but all were colored gray. Page 4 displayed the same fruits and vegetables, but all were colored incorrectly. The child was instructed to name the correct color of the stimuli (pages 1 and 2), or to name the color that the fruits and vegetables should have (pages 3 and 4), as quickly as possible. For each page, the time (in seconds) required for giving correct responses for all 25 stimuli was measured. As dependent variable, an interference score was generated based on Röthlisberger et al. (2010): time p.4—[(time p.1 × time p.3)/(time p.1 + time p.3)]. Higher scores indicated a lower ability to successfully inhibit the prepotent response of naming the color in which the stimuli were depicted on page 4.

The EF subcomponent working memory was assessed using the Digit-Span Backward task (Petermann and Petermann, 2007). This is a complex working memory task (Best and Miller, 2010), and measures of complex working memory and updating have been found to be highly correlated in children (e.g., St Clair-Thompson and Gathercole, 2006). In this task, participants were

<sup>1</sup>The original sample size was N = 1,657, however in the present study, 5 children were excluded due to missing values on the cluster variable that was included in the analyses (see section 2.3. Plan of Analyses).

told a sequence of digits, which they had to repeat in reverse order. Each trial consisted of two sequences of equal length. The first two sequences were 2 digits long, and in each of the next trials, the sequences were lengthened by one digit up to a maximum number of eight digits, yielding a total of 7 trials with 14 sequences. Within each trial, at least one sequence had to be answered correctly in order to proceed to the next trial. The dependent variable was the total number of sequences that had been repeated correctly with a potential range of 0 to 14.

The EF subcomponent shifting was assessed using the Cognitive Attention Shifting task (Röthlisberger et al., 2010; adapted from Zimmermann et al., 2002). Participants were presented with a single-colored fish and a multi-colored fish appearing simultaneously on the left- and right-hand side, respectively, of the computer screen. Children were told to feed each kind of fish and to always alternate between the two kinds by pressing one of two keys on a QWERTZ keyboard. Across several trials, the side on which the two kinds of fish appeared changed randomly. This required the children to remember their previous response—that is, which kind of fish they fed—in order to maintain the requirement of alternating feeding. A total of 46 trials (interstimulus intervals ranged from 300 to 700 ms) was separated by a short break during which positive feedback was given. The dependent variable was the number of correct responses for the 22 switch trials, that is, the trials that required children to change their response pattern (i.e., from alternately pressing left/right to repeating left/left or right/right; Austin et al., 2014).

The EF subcomponent planning was measured using items of the Planning and Organizing-scale from the Behavior Rating Inventory of Executive Function (BRIEF; Gioia et al., 2000). Eight of the original 10 items were selected based on their factor loadings and translated into German by two native speakers. The items covered a range of problems that students can face when they need to plan or organize present and future tasks for school (e.g., "does not plan tasks for school in advance"). Teachers indicated planning disability of their students during the past 6 months using a 5-point response scale ranging from 1 (never) to 5 (always). A total score was computed by averaging the item scores. The internal consistency was high with α = 0.93.

### Aggression

At all three time points, aggression was measured using a teacherreport questionnaire that contained subscales with three items each for physical and relational aggression as well as for proactive and reactive aggression.

The teachers first rated the frequency of physical and relational aggression during the past 6 months on a 5-point scale ranging from 1 (never) to 5 (daily) (physical aggression: e.g., "hit, shoved, or pushed peers"; relational aggression: e.g., "spread rumors or gossips about some peers"). The items were adapted from the Children's Social Behavior Scale—Teacher Form (CSBS-T; Crick, 1996). In a next step, teachers were asked to rate the functions of aggressive behaviors, based on the Instrument of Reactive and Proactive Aggression (IRPA; Polman et al., 2009; proactive aggression: e.g., "to be the boss," reactive aggression: e.g., "because someone teased or upset him/her"). The response scale ranged from 1 (never) to 5 (always). The items on the function of aggression were only completed if the total score of the frequency of physical and relational aggression was larger than 1. Thus, the children for whom the teachers reported no physical or relational aggression at all had logical missing values on the measures of proactive and reactive aggression. The handling of these missing values is explained below.

For all four subscales, total scores were created by averaging the corresponding items, based on acceptable to high internal consistencies (physical aggression: αt1 = 0.93, αt2 = 0.94, αt3 = 0.93; relational aggression: αt1 = 0.91, αt2 = 0.92, αt3 = 0.91; proactive aggression: αt1 = 0.80, αt2 = 0.77, αt3 = 0.81; reactive aggression: αt1 = 0.85, αt2 = 0.84, αt3 = 0.88).

### Habitual Anger

At T1 and T2, parents rated the extent to which their children habitually experienced anger with the subscale Anger/Frustration of the Temperament in Middle Childhood Questionnaire (TMCQ; Simonds, 2006). This subscale assesses the extent of anger shown by the child in response to the interruption of ongoing tasks or goal blocking (e.g., "my child gets angry when she or he has trouble with a task"). The scale contains 7 items, and the response scale ranges from 1 (almost always untrue) to 5 (almost always true). A total score was obtained by averaging the item scores. The internal consistency was good with α = 0.80 at both time points.

### Information-Processing Capacity

Information-processing capacity was assessed at T1with the Digit-Symbol Test (DST) of the German version of the Wechsler Intelligence Scale for Children (Petermann and Petermann, 2007). Children were given a worksheet on which they had to assign either common shapes (Version A; ages 6–7) or the numbers 1 to 7 (Version B; ages 8 and older) to various symbols. A key in which a specific shape/number was paired with each of the symbols was presented in the first row of the worksheet. For both versions, the number of correct symbols allocated within 120 s was measured (standardized T-values were calculated). Information-processing capacity was measured to control for basic intellectual ability, which could be confounded with EF.

### Plan of Analysis

SPSS (Version 23) was used for descriptive computations, and the hypotheses were analyzed through structural equation models using Mplus (Version 7.4; Muthén and Muthén, 2012). In all models, the robust Maximum Likelihood estimator (MLR) was used to account for the non-normal distribution of the data. Missing data were handled by the Full Information Maximum Likelihood (FIML) estimation option to avoid a reduction in sample size. To be able to use the FIML approach for the logical missings on the items of the functions of aggression, we included a participant's overall frequency scores of aggression at all three time points as covariates in the models. The frequency of aggression is a perfect predictor of the presence or absence of a data point on the two functions of aggression. Therefore, missing data could be treated as missing at random, which allowed us to use the FIML approach (Enders, 2010).

Because participants were nested within school classes, class was included as a cluster variable in all analyses. Due to the trait-like nature of aggressive behavior, we included a random intercept for both forms and both functions of aggression, following the recommendation of Hamaker et al. (2015). Because habitual anger and the three EF subcomponents were not assessed at all three time points, we were not able to include random intercepts for these variables (at least three time points are required to specify random intercepts). EF was modeled as a latent factor using the measures of working memory, inhibition, shifting, and planning as indicators.

The model fits of the measurement model of EF and of the structural equation models were evaluated based on the criteria of Hu and Bentler (1998), with a comparative fit index (CFI) ≥0.95, a root-mean-square error of approximation (RMSEA) ≤ 0.06, and a standardized root-mean-square residual (SRMR) ≤ 0.08 indicating a good fit. The χ²-statistic was not interpreted as a measure of absolute fit, because it is biased in large samples (Schermelleh-Engel et al., 2003). Bootstrap analyses were used to test indirect effects. If the bootstrapped 95% confidence interval does not include zero, the indirect effect is considered to be significant (Shrout and Bolger, 2002). The potential moderating effect of gender was examined using multi-group analyses. The measurement invariance between the gender groups was assessed based on comparisons between a fully constrained and a fully unconstrained (freed) model. The indicator for measurement invariance was a nonsignificant difference in χ² with scale corrections for the MLR estimator, as proposed by Satorra and Bentler (2001) or a nonsignificant Wald test for invariance in the indirect effects.

### RESULTS

### Descriptive Statistics, Gender Differences, Factor Analysis, and Correlations

The descriptive statistics of all study variables are presented in **Table 1**. Gender differences were analyzed using t-tests for independent samples. If the assumption of homogeneity of variance was violated (as indicated by the Levene's Test for Equality of Variances), the degrees of freedom were adjusted using the Welch-Satterthwaite method. To account for multiple testing, we used a strict alpha level of p < 0.01. Effect size was calculated as Cohen's d. Boys were rated to be significantly more physically aggressive than were girls at all time points, all ts ≥ 11.00, ps < 0.001, ds ≥ 0.60. For relational aggression, no significant gender differences were found. Boys as compared to girls were also rated to show significantly more proactive aggression at T1 and T3, all ts ≥ 3.48, ps < 0.01, ds ≥ 0.25, and more reactive aggression at T2 and T3, all ts ≥ 3.48, ps < 0.01, ds ≥ 0.30. For the measure of anger, a significant gender difference was found only at T1, with higher scores for boys than for girls, t(1332) = 3.64, p < 0.001, d = 0.19.

With regard to EF, girls scored significantly higher than did boys on shifting, t(1535) = 3.79, p < 0.001, d = 0.19, and boys scored significantly higher than did girls on inhibition, t(1534.04) = 3.24, p < 0.01, d = 0.16, and planning, t(1415) = 8.23, p < 0.001, d = 0.56. Furthermore, girls scored significantly higher than did boys on information-processing capacity, t(1642) = 7.18, p < 0.001, d = 0.35. Due to these differences, gender was included as a covariate in the structural equation models (except for the multi-group model).

A latent factor of EF was specified by using the measures of working memory, inhibition, shifting, and planning as manifest indicators. Factor loadings were of moderate size (working memory: 0.55, inhibition: −0.62, shifting: 0.46, planning: −0.51; all ps <.001). The resulting measurement model showed a good fit (χ 2 [3] = 2.76, p = 0.25; CFI = 1.00; RMSEA = 0.02, 90% CI [0.00,0.05]; SRMR = 0.01).

**Table 2** presents the correlations among all study variables and their links with age and information-processing capacity. The following significant correlations were found: EF at T1 was negatively correlated with anger at T1 and T2, and with all aggression measures at T1, T2, and T3. Anger was positively correlated with all aggression measures within and across time points. Age and information-processing capacity were positively correlated with EF. Furthermore, age was negatively correlated with reactive aggression at T3, and information-processing capacity was negatively correlated with physical aggression at all time points and with reactive aggression at T1 and T3. As a consequence, we decided to include age and informationprocessing capacity as covariates in all models.

### Hypothesis-Testing Analyses

Two separate models were specified to examine the proposed links between EF, anger, and aggression: one for the forms of aggression (physical and relational; see **Figure 1**) and one for the functions of aggression (reactive and proactive; see **Figure 2**). Age, gender, and information-processing capacity were included as covariates in both models.

### Links between EF, Anger, and Forms of Aggression

The model for the forms of aggression (**Figure 1**) showed an acceptable model fit (χ 2 [39] = 363.05, p < 0.001; CFI = 0.93; RMSEA = 0.007, 90% CI [0.06,0.08]; SRMR = 0.05). In line with Hypothesis 1, we found that controlling for stable individual differences in aggression, there was a significant negative path from EF at T1 to physical aggression at T3. Regarding relational aggression at T3, our data revealed a significant negative link to EF at T1 as well. Thus, the lower a child's EF, the higher the teacher-rated frequency of physical and relational aggression after the 3-year period. The paths from EF at T1 to physical and relational aggression at T2 were also negative and significant.

A significant negative link between EF at T1 and habitual anger at T2 was found, indicating that the lower children's EF was at T1, the more anger-prone they were rated by their parents at T2. Moreover, there was a significant positive link between habitual anger at T1 and both physical and relational aggression at T2, and there was a significant positive path from habitual anger at T2 to physical aggression at T3, but no significant link to relational aggression at T3.

Hypothesis 2 postulated indirect negative effects between EF at T1 and both forms of aggression at T3 through habitual anger at T2. This hypothesis was only partially confirmed, because an TABLE 1 | Descriptive statistics of the study variables for the total sample and for boys and girls.


EF, executive function; A, aggression; DST, digit-symbol test; <sup>a</sup>Max values are theoretically infinite, thus, table values are sample-specific. Means in bold differ significantly between boys and girls.


EF, executive function; Phy A, physical aggression; Rel A, relational aggression; Reac A, reactive aggression; Proac A, proactive aggression; DST, digit-symbol test. \*\*\*p < 0.001, \*\*p < 0.01, \*p < 0.05.

indirect effect was found only for physical aggression, β = −0.01, 95% CI [−0.021, −0.001]. For relational aggression, the indirect path was not found due to the nonsignificant path from habitual anger at T2 to relational aggression at T3 Links between EF, Anger, and Functions of Aggression.

The model for the functions of aggression (**Figure 2**) also showed an acceptable fit (χ 2 [61] = 407.04, p < 0.001; CFI = 0.91; RMSEA = 0.06, 90% CI [0.05,0.06]; SRMR = 0.05). As predicted in Hypothesis 3, we found that controlling for stable individual differences in aggression, there was a significant negative path from EF at T1 to reactive aggression at T3, whereas the path from EF at T1 to proactive aggression at T3 was nonsignificant. Similarly, there was a significant negative link between EF at T1 and reactive, but not proactive aggression at T2. Finally, Hypothesis 4 postulated an indirect negative effect between EF at T1 and reactive aggression at T3 through habitual anger at T2. This hypothesis was not supported, because no significant link between T2 habitual anger and T3 reactive aggression was found.

### Multi-Group Analyses of Potential Gender Differences

To examine potential gender differences in our first model, considering physical and relational aggression, we compared a fully unconstrained model, in which all paths were allowed to vary between girls and boys [fit: χ²(72) = 396.54, p < 0.001; CFI = 0.92; RMSEA = 0.07, 90% CI [0.07,0.08]; SRMR = 0.05], with a fully constrained model, in which all paths were constrained to be equal [fit: χ²(132) = 650.12, p <0.001; CFI = 0.88; RMSEA = 0.07, 90% CI [0.06,0.07]; SRMR = 0.12]. The difference in χ² was significant, 1χ²(60) = 262.66, p < 0.001, indicating gender differences in specific parts of the model. Therefore, we computed a revised model [fit: χ²(121) = 442.26, p < 0.001; CFI = 0.92; RMSEA = 0.06, 90% CI [0.05,0.06]; SRMR = 0.07], in which we constrained all paths of the model to be equal between boys and girls, but with free estimation of those covariances and intercepts that were found to be different between boys and girls (e.g., the covariances between physical and relational aggression at all three time-points). The revised model had a significantly better fit than the fully constrained model, 1χ²(11) = 150.69, p < 0.001, and did not fit significantly worse than the fully unconstrained model, 1χ²(49) = 66.0, p = 0.053. In the revised model, there were no significant gender differences in the hypothesized paths [all 1χ²s ≤ 3.64, ps ≥ 0.19; all W(1) ≤ 0.13, ps ≥ 0.72].

For our second model, considering reactive and proactive aggression, we followed the same procedure. We also compared a fully unconstrained model, in which all paths were allowed to vary between girls and boys [fit: χ²(116) = 445.94, p < 0.001; CFI = 0.91; RMSEA = 0.06, 90% CI [0.05,0.06]; SRMR = 0.05], with a fully constrained model [fit: χ²(198) = 624.94, p < 0.001; CFI = 0.088; RMSEA = 0.05, 90% CI [0.05,0.06]; SRMR = 0.08]. The difference in χ² was significant, 1χ²(82) = 188.41, p < 0.001. Then, we computed a revised model [fit: χ²(191) = 524.07, p <

0.001; CFI = 0.91; RMSEA = 0.05, 90% CI [0.04,0.05]; SRMR = 0.06], in which we constrained all paths of the model to be equal between boys and girls, but with free estimation of some covariances and intercepts that were found to be different (for instance the covariances between the overall frequency scores of aggression at the three time-points). The revised model had a significantly better fit than the fully constrained model, 1χ²(7) = 86.93, p < 0.001, and did not fit significantly worse than the fully unconstrained model, 1χ²(75) = 91.87, p = 0.090. In the revised model, there were no significant gender differences in the hypothesized paths [all 1χ²s ≤ 2.79; ps ≥.095; all W(1) ≤ 0.63, ps ≥ 0.43].

### DISCUSSION

The aim of the present study was to examine the longitudinal associations of EF (calculated as a latent factor of EF from behavioral measures of inhibition, set shifting, and workingmemory updating, as well as teacher-reported planning), parent-reported habitual anger, and teacher-reported forms of aggression (i.e., physical and relational) and functions of aggression (i.e., proactive and reactive) in middle childhood. The hypotheses were examined in a large population-based sample in a three-wave design over a period of 3 years.

In line with Hypothesis 1, we found that EF was a significant negative predictor of physical aggression, and we also found a significant negative path between EF and relational aggression. This was true for the paths from T1 to T2 as well as from T1 to T3. Thus, the more deficits in EF children showed at T1 the higher was their teacher-rated frequency of both forms of aggression 1 and 2 years later. Our results held after controlling for information-processing capacity, gender, and age in the whole model. Further, we controlled for stable between-person differences by inclusion of a random intercept for forms and functions of aggression. Following the reasoning by Hamaker et al. (2015), this method allowed us to uncover causal relationships in within-persons processes. Therefore, our findings replicate longitudinal findings from other age groups (e.g., Hughes et al., 1998; Hughes and Ensor, 2008; Ogilvie et al., 2011; Schoemaker et al., 2013; Olson et al., 2017), and they extend previous cross-sectional research in middle childhood (e.g., McQuade et al., 2013), showing that lower EF at a mean age of 8 years predicted higher physical and relational aggression at about 9 and 11 years as within-person change. Furthermore, the negative path from EF to physical aggression is consistent with the frontal-lobe hypothesis of physical aggression (Séguin, 2009), and the social information-processing theory of aggression (Crick and Dodge, 1994). Consequently, the children in our study, who showed physically and relationally aggressive behavior at the age of 9 and 11 years might already have manifested significant cognitive deficits in their EF abilities, located in the frontal lobe, at the age of 8 years.

Our finding of significant negative paths from EF at T1 to relational aggression at both T2 and T3 confirms some of the previous cross-sectional findings (McQuade et al., 2013), but contradicts others (no significant path: Diamantopoulou et al., 2007; positive relation: Poland et al., 2016). Nevertheless, our finding is in line with the social information-processing theory of aggression that proposes that physically as well as relationally aggressive children have deficits in their cognitive processing of social situations (Crick and Dodge, 1994). To further examine the role of EF in the development of relational aggression, it may be important to include a more differentiated assessment of relational aggression that considers the complexity of different relationally aggressive behaviors. According to Crick et al. (1999), relationally aggressive behaviors can take different forms ranging from relatively simple, direct types (e.g., threatening to end the friendship) to more complex, indirect types (e.g., mobilizing peer group members against a certain child to make that child feel excluded). The latter requires a higher level of cognitive skills to be used effectively. Thus, it may be that only the direct types of relational aggression are related to deficits in EF, whereas the indirect types of relational aggression are unrelated or even positively related to EF. However, this remains to be tested in future studies.

In Hypothesis 2, we postulated that anger would mediate the link between EF and both forms of aggression. However, this prediction was only confirmed for physical, and not for relational aggression. For physical aggression, the results indicate that the lower the children's EF at T1 the higher their parentreported habitual anger was at T2, which in turn predicted more teacher-rated physical aggression at T3. Even though the size of the indirect effect was small, it was still significant after controlling for age, gender, information-processing capacity, and stable between-person differences of physical and relational aggression. The tendency to experience anger is only one of many intrapersonal factors involved in the complex emergence of aggressive behavior (see Krahé, 2013, for a review). However, our findings extend previous cross-sectional research (e.g., de Castro et al., 2005) and support the role of anger as one explanatory construct in the link between EF and physical aggression in middle childhood. Further, the negative mediation between EF and physical aggression by anger highlights the importance of considering emotions in the social-cognitive information processing of children who display aggressive behavior (Lemerise and Arsenio, 2000). Our study is—as far as we know—the first to uncover this meditational effect over a time-span of 3 years in middle childhood, which underlines the need to consider large time intervals in the development of physical aggression, as well as both emotional and cognitive processes.

For relational aggression, this mediation effect was not found. This finding is inconsistent with theory and previous research that has found anger to be involved in the development of both physical and relational aggression (Crick et al., 1999, 2002). However, the positive path from anger at T2 to relational aggression at T3 only narrowly missed the level of significance (p = 0.054, β = 0.06). This trend tentatively suggests that the tendency to experience anger may be involved in the negative link between EF and later relational aggression. To date, only few studies have differentiated between forms of aggression when examining the link between anger and aggression. Future research is needed to explore potential differences between relational and physical aggression regarding the association with EF and anger, not only in middle childhood.

With regard to the functions of aggression, we found that EF was a significant negative predictor of reactive aggression, but that EF was unrelated to proactive aggression. This pattern is consistent with Hypothesis 3 and was found for the paths from EF at T1 to functions of aggression both at T2 and at T3. It is in line with the few previous cross-sectional studies regarding the association between EF and the functions of aggression in middle childhood (e.g., Ellis et al., 2009; White et al., 2013). Moreover, the nonsignificant path from EF to proactive aggression is consistent with research on CU traits, which are conceptually linked to proactive aggression (e.g., Pardini et al., 2003). Several studies have shown that CU traits are unrelated to deficits in different domains of EF such as inhibition (Tye et al., 2017) or set shifting (Mitchell et al., 2002). Furthermore, our findings support the theoretical differentiation of the two functions of aggression. Higher EF enables children to behave in a planned and deliberate fashion, which is characteristic of proactive aggression. In contrast, reactive aggression refers to impulsive aggressive acts that do not require sophisticated planning. Thus, the inability to plan and to inhibit behavioral responses, both components of low EF, may explain the direct negative paths between EF and reactive aggression at 1 or 2 years later found in this study.

Contrary to Hypothesis 4, no indirect link between EF and reactive aggression via anger was found. We did find a negative link between EF and anger, confirming previous evidence that EF is involved in the regulation of negative affect in children (e.g., Carlson and Wang, 2007). This finding is also consistent with research on deficits in processes associated with EF, for instance error monitoring, in children with chronic irritability (Perlman et al., 2015). However, there was no link between anger and reactive aggression, which is surprising given that anger is assumed to be a major impelling factor of reactive aggression during childhood (e.g., Eisenberg et al., 1999; Arsenio et al., 2000; Olson et al., 2005; Leibenluft and Stoddard, 2013). One explanation may be that the relation between anger and reactive aggression depends on moderating variables. These may include the ability to regulate the behavioral impulses related to anger, or the tendency to act impulsively, which relates to less voluntary and more reactive aspects of control (Rothbart and Rueda, 2005). Another reason could be that until now the relations of chronic anger, irritability, and disruptive behavior symptoms were mainly investigated in clinical samples with no differentiation between reactive and proactive aggression (e.g., Wakschlag et al., 2015). In addition, further variables besides anger may act as mediators of the link between EF and reactive aggression. These may include social information-processing variables, such as theory of mind (Renouf et al., 2010) or hostile attribution bias (de Castro et al., 2002).

In addition to the examination of the longitudinal links between EF and aggression, our study contributes to previous research by considering potential gender differences. We did find gender differences on some of the study variables, in particular on the frequency of physical aggression, with boys scoring higher than girls at all three time points. This finding is in line with a meta-analysis on gender differences in aggressive behavior (Card et al., 2008). However, the multi-group analyses revealed that the predictive paths from EF to aggression and from EF to habitual anger as well as the indirect paths from EF over anger to aggression did not vary by gender, which confirms and extends previous cross-sectional research (White et al., 2013). Consequently, the processes and mechanisms that lead from EF to aggressive behavior seem to be equivalent in girls and boys in middle childhood.

### Strengths and Limitations

We believe that the present study has several strengths. These include the large sample size, the longitudinal design with three time points covering about 3 years across middle childhood, the differentiation of forms and functions of aggression, and the examination of potential gender differences. Furthermore, we used three different sources to assess the study constructs, a procedure that is known to reduce common method biases (Podsakoff et al., 2003). A final strength is the inclusion of a random intercept, by which we controlled for the stability of individual differences in the forms and functions of aggressive behavior. This new procedure is recommended to overcome the limitations of the traditional cross-lagged model by separating within-person change, which is the focus of our models, from stable between-person differences (Hamaker et al., 2015).

Despite these strengths, some limitations have to be noted. One refers to the distinction between "cool" and "hot" aspects of EF during childhood and adolescence (Zelazo and Carlson, 2012). Cool EF is usually associated with the lateral prefrontal cortex and operates in relation to more abstract and decontextualized problems. In contrast, hot EF is associated with the orbitofrontal cortex and operates in more motivationally and emotionally significant situations (Zelazo and Müller, 2002). The measures we used to assess EF covered only the cool component. Nevertheless, our findings demonstrate that cool EF alone seems to play an important role in the prediction of physical, relational, and reactive aggression in middle childhood. A further limitation of our study is that aggression was only assessed by teacher reports. This may have led to an underestimation of relationally aggressive behavior because it includes more indirect forms of aggression that may be less obvious for teachers. Thus, peer reports or peer nominations could provide important additional information because aggressive behavior usually takes place within the peer group. Finally, it is important to mention that our sample was a community sample. Whether similar associations are also found for clinical populations of youth with serious levels of aggression and/or chronic symptoms of irritability is a question for future research.

### REFERENCES


### CONCLUSION AND IMPLICATIONS

Our study extends the existing literature about the relation between EF and aggression in middle childhood by taking the habitual tendency to experience anger as a potential mediator into account, by considering different forms and functions of aggressive behavior, and by analyzing gender differences. We found that EF predicted physical, relational, and reactive aggression over the course of middle childhood, and that the link between EF and physical aggression was partly mediated by habitual proneness to anger. Although there were gender differences in the frequency of aggressive behavior, these predictive paths were not moderated by gender. Our findings highlight the importance of addressing EF in programs that aim to reduce aggressive behavior. In the last years, an increasing number of programs to promote EF has been developed, and the effectiveness of these programs has been demonstrated (Diamond and Lee, 2011), making EF a promising candidate for the prevention of aggression. For the prevention of physical aggression in particular, teaching strategies for coping with anger should also be considered. Regarding gender, our findings indicate that both gender groups should be addressed in prevention programs. Although the mean level of physically aggressive behavior in particular may be higher among boys, the paths between EF, anger, and aggression seem to be similar for girls and boys. Our study underlines the need to promote the development of EF not only in early, but also in middle childhood to prevent later physical, relational, and reactive aggression and its negative consequences.

### AUTHOR CONTRIBUTIONS

All authors have contributed substantially to the conceptualization and the design of the work. HR, AH, and FK: Have primarily collected, analyzed and interpreted the data; HR and AH: Have written the first draft of the paper, and all authors contributed to revise the paper; BK and BE: Have provided input and supervision to the analyses. All authors agreed to be accountable for all aspects of the work.

### FUNDING

This research was funded by the German Research Foundation as part of the Research Training Unit Intrapersonal developmental risk factors in childhood and adolescence: A longitudinal perspective (GRK 1668). We also acknowledge financial support from the Open Access Publishing Fund of the University of Potsdam.


Child Psychiatry Hum. Dev. 42, 609–621. doi: 10.1007/s10578-011- 0236-3


mediating role of language. J. Abnorm. Child Psychol. 35, 141–152. doi: 10.1007/s10802-006-9052-9


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Rohlf, Holl, Kirsch, Krahé and Elsner. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Exogenous Testosterone Enhances the Reactivity to Social Provocation in Males

Lisa Wagels 1,2,3\*, Mikhail Votinov 1,2,3 , Thilo Kellermann<sup>1</sup> , Albrecht Eisert 4,5 , Cordian Beyer 3,6 and Ute Habel 1,2,3

<sup>1</sup>Department of Psychiatry, Psychotherapy and Psychosomatics, Medical Faculty, Uniklinik RWTH Aachen University, Aachen, Germany, <sup>2</sup> Institute of Neuroscience and Medicine 10, Research Center Juelich, Juelich, Germany, <sup>3</sup>JARA-Institute Brain Structure Function Relationship, Research Center Juelich and RWTH Aachen University, Aachen, Germany, <sup>4</sup> Institute of Pharmacology and Toxicology, Medical Faculty of RWTH Aachen University, Aachen, Germany, <sup>5</sup>Hospital Pharmacy, University Hospital RWTH Aachen, Aachen, Germany, <sup>6</sup> Institute of Neuroanatomy, Medical Faculty, RWTH Aachen University, Aachen, Germany

Testosterone affects human social behavior in various ways. While testosterone effects are generally associated with muscular strength and aggressiveness, human studies also point towards enhanced status–seeking motives after testosterone administration. The current study tested the causal influence of exogenous testosterone on male behavior during a competitive provocation paradigm. In this double blind, randomized, placebo (PL)-controlled study, 103 males were assigned to a PL or testosterone group receiving a colorless PL or testosterone gel. To induce provocation, males played a rigged reaction time game against an ostensible opponent. When participants lost, the opponent subtracted money from the participant who in return could subtract money from the ostensible opponent. Participants subjectively indicated anger and self-estimated treatment affiliation (testosterone or PL administration). A trial-by-trial analysis demonstrated that provocation and success during the repeated games had a stronger influence on participants' choice to reduce money from the opponent if they had received testosterone. Participants who believed to be in the testosterone group were angrier after the experiment and increased monetary reductions during the task course. In line with theories about mechanisms of testosterone in humans, provocation is shown to be necessary for the agency of exogenous testosterone. Thus, testosterone reinforces the conditional adjustment of aggressive behavior but not aggressive behavior per se. In contrast undirected frustration is not increased by testosterone but probably interferes with cognitive appraisals about biological mechanisms of testosterone.

Keywords: aggression, testosterone administration, status hypothesis, challenge hypothesis, males, placebo effect

### INTRODUCTION

The influence of testosterone (T) on aggression has been studied across a variety of species. Animal research overall supports the assumption of increased aggression associated with high T plasma levels (Gleason et al., 2009). The basis of such a relationship has been unveiled via the Challenge Hypothesis (Wingfield et al., 1990) based on the observation that T plasma levels in male birds would rise as a function of social challenges. This in turn increased aggressive

#### Edited by:

Nelly Alia-Klein, Icahn School of Medicine at Mount Sinai, United States

#### Reviewed by:

David Shum, Griffith University, Australia Carlos Tomaz, Universidade Ceuma, Brazil

#### \*Correspondence:

Lisa Wagels lwagels@ukaachen.de; l.wagels@fz-juelich.de

Received: 27 November 2017 Accepted: 19 February 2018 Published: 02 March 2018

#### Citation:

Wagels L, Votinov M, Kellermann T, Eisert A, Beyer C and Habel U (2018) Exogenous Testosterone Enhances the Reactivity to Social Provocation in Males. Front. Behav. Neurosci. 12:37. doi: 10.3389/fnbeh.2018.00037 reactions. A second hypothesis, the Status Hypothesis was synthesized upon the Challenge Hypothesis later on, suggesting that T effects in humans depend on the challenge of the social status and that the direction of the effect will support statusseeking behaviors (Eisenegger et al., 2010). Thus, both prosocial and aggressive behaviors can be promoted via T depending on the context.

Different mechanisms, traditionally divided in organizational and activational effects, may explain the role of T in aggression. First, organizational effects of T in humans might determine the development of aggressive traits (Turanovic et al., 2017). Organizational effects are actions of steroid hormones which occur during early critical periods to organize neural pathways which could be responsible for certain behaviors such as aggression. While organizational effects occur early in life and are permanent, activational effects are transient and occur throughout life. Activational effects may thus influence human behavior via rapid changes in the neural circuit of aggression (Goetz et al., 2014). Despite of numerous studies investigating activational effects of T in the context of aggression and competition (Carré and Olmstead, 2015), causal evidence in humans is still rare. One reason for this is the predominantly correlational nature of earlier human studies supporting a weak but positive correlation of aggression with basal T (Archer et al., 2005). For instance, in inmates overt confrontations have been associated with higher salivary T (Dabbs et al., 1995). Critically, T might increase violent and aggressive behavior, however, it is equally possible that frequent aggressive acts increase T levels over time.

Thus, aggression research has further focused upon effects of exogenous T in humans (for an overview see Bos et al., 2012). While numerous studies investigated subtle effects of T on social-emotional behavior, initially primarily females were investigated due to the existence of an appropriate administration paradigm in females (Tuiten et al., 2000). Later on, studies were conducted in males mainly applying T via dermal administration (Zak et al., 2009; Cueva et al., 2015, 2017; Bird et al., 2016; Kopsida et al., 2016; Welling et al., 2016; Carré et al., 2017; Panagiotidis et al., 2017; Wagels et al., 2017a). Partly, studies investigated aggressive or antisocial behavior in males (Zak et al., 2009; Dreher et al., 2016; Kopsida et al., 2016; Carré et al., 2017; Cueva et al., 2017; Panagiotidis et al., 2017).

Two paradigms were most frequently applied to investigate the T-aggression relationship: The Ultimatum Game (UG) or the Point Subtraction Aggression Paradigm (PSAP). The UG represents a negotiation between two players. Usually, player one has a certain amount of money and is asked to make an offer to a second player sharing some of the money. The second player can accept the offer, which means that the deal is carried out, or reject the offer which means that both players will not receive any money. In the UG mostly rejections of low offers, which to some degree reflect social provocation, have been studied. While there is some evidence for T administration enhancing the likelihood of rejecting low offers (Zak et al., 2009) other studies neither support this in males nor females (Zethraeus et al., 2009; Eisenegger et al., 2010; Dreher et al., 2016; Kopsida et al., 2016; Cueva et al., 2017). Instead, in females sublingual T administration increased fair bargaining behavior (Eisenegger et al., 2010). Considering that fair behavior might avoid conflict—or social threat—the T effect here might be in line with the Status Hypothesis. Interestingly, females in the same study were less fair when believing to be in the T group. Cognitive appraisal about hormonal effects thus might be relevant and potentially contribute to the mixed findings that are reported in the UG. Nevertheless, underlining the assumption that T predominantly supports status-seeking behavior which depends on the context, males were shown to administer higher punishments to low offers but also higher rewards to generous offers in a modified UG (Dreher et al., 2016). Critically, antisocial behavior—rejecting an offer, or punishing the opponent—influenced the actual earnings of an individual. This conflict of reward-seeking and punishment or fairness motivation may be another reason for the divergent findings in the UG and impede the direct investigation of aggression.

Probably more closely investigating aggression, the effect of T administration has been studied during the PSAP several times. The PSAP involves an ostensible opponent who can steal points from the participant. The participant, on the other hand, can repeatedly choose between three buttons: A money button, a protection button or a counterattack button. Stolen points from attacking the opponent usually are not added to the participants' earnings. First studies showed that long-term T administration of a supraphysiologic doses over several weeks (2 and 6) increased aggressive responses (the number of attack button presses) towards social provocation in the PSAP (Kouri et al., 1995; Pope et al., 2000). Newer findings which tested the effect of a single T administration in males did not confirm enhanced aggressive reactions towards social provocation after T administration (Carré et al., 2017). However, the authors found that T increased aggressiveness in men who were highly dominant or low in self-control. Again, the effect of T on the reaction towards a social challenge seems to depend on individual (status-seeking) motives. A disadvantage of the PSAP is that provocation frequency of the ostensible opponent depends on the participant's behavior (frequent protect or attack decisions will result in less provocations as the program usually blocks provocations for a certain time after these decisions). Although researchers try to control for task variability, the initial situation participants are confronted with might differ strongly and thereby disguise effects of T.

In order to gain a better understanding of how exogenous T influences males behavior during a provocation task, we applied a modified version of the Taylor Aggression Paradigm (TAP; Giancola and Parrott, 2008). We primarily aimed to circumvent two limitations of the above reviewed paradigms: First, earnings should be independent from the punishment decision of the participant; second, predefining the opponent's behavior, provocation situations were fixed. During the TAP participants play repeated rounds of reaction-time games against an ostensible opponent. Both players can

ostensibly punish the other player by reducing money, when winning the round which is not added to their actual earnings. Previous research showed that participants in this task act with ''tit-for-tat'' like behavior towards the punishment of the feasible opponent (Krämer et al., 2007). In detail, this means that participants adjust their punishment levels to the preceding provocation. Such behavioral adjustments might reduce conflict potential and protect the social status.

Assumptions of the current study were based on the Challenge Hypothesis and the Status Hypothesis. First, exogenous T should enhance punishment behavior during the modified TAP due to the social competition. Especially losing would constitute a social challenge thus promoting aggressive behavior. Moreover, we expected that exogenous T would increase tit-for-tat like behavior compared to the placebo (PL) group in order to gain a high social status. Since this has not been investigated before, we also investigated in an exploratory way the temporal course of punishment behavior comparing T and PL.

A secondary goal of the study was to investigate if aggression is related to the subjective belief of having received T (or PL). Since a previous study demonstrated that the belief to have received T leads more rejections in the UG (Eisenegger et al., 2010), we expected that individuals who believe to be in the T group would react more aggressively independent of the provocation.

### MATERIALS AND METHODS

### Sample

The study included 103 male participants recruited in Aachen via online advertisements and postings. For their participation, participants received a fixed amount of 70 Euros and additionally the money they won in two further paradigms they performed in the study. All participants had normal or corrected vision, no contraindications for magnetic resonance imaging (MRI), and no history of traumatic brain injury, psychiatric or neurological illness and were right handed (according to Oldfield, 1971). Participants were between 18 years and 35 years old (M = 24.17, SD = 3.76). Ethnicity was not explicitly ascertained. A more detailed description of the sample is reported elsewhere (Wagels et al., 2017b). Written and informed consent was obtained from all participants in accordance with the recommendations of the Declaration of Helsinki. After the scanning session, participants were fully debriefed about the study aims and the cover story around the paradigm. The study was approved by the internal ethics committee of the RWTH Aachen medical faculty and was not evaluated as clinical trial. We therefore did not

during the game period the participant plays a reaction time game against an ostensible opponent; during the feedback period, the participant sees if he won or lost and how much money he lost due to the ostensible opponents decision. The last period was assumed to influence the behavior in the following trial.

register as a randomized controlled trial in any official online register.

### Procedure

The TAP was part of a large study including several tasks related to aggression (non-social and social aggression) and risk-taking of which two were performed in an MRI environment (for an overview see **Figure 1**). The TAP was the first task of the scanning session. One major aim of the complete study was to investigate the interaction of T administration and genetic variability in the serotonergic system (MAOA VNTR polymorphism, and serotonin transporter polymorphism) regarding neural responses. Results concerning this gene-hormone interaction have been published for the risk task (Wagels et al., 2017b). In order to have a clear focus on the influence of T administration in a new social aggression paradigm this manuscript only presents behavioral data and will focus on the gene-hormone interaction on a neural level elsewhere.

In order to reach a stable hormonal level at the baseline measurement, sessions started between 12:00–14:00 and took about 6.5 h. After taking a first blood sample to determine baseline serum levels, participants received either 5 g TestimTM corresponding to 50 mg T, or an equivalent amount of sonography gel (PL). The gel was applied on the upper part of the back and the shoulders of participants by a blinded experimenter. Participants performed several short tasks, filled out personality questionnaires, provided saliva samples for genotyping analysis and had about 1 h break before the scanning session.

To improve the credibility of the aggression paradigm, before the scanning session individuals were introduced to an ostensible male opponent who was supposedly guided to a separate test room. Before and after the task blood samples were taken to test for task effects on T plasma levels. The task was followed by another experimental task on risk-taking, a restingstate measurement and an anatomical scan. After scanning participants were asked if they believed to have received T or PL.

## Task: Modified Taylor Aggression Paradigm (TAP)

The TAP is a well-validated aggression task (Giancola and Parrott, 2008) usually disguised as a reaction-time game against a real opponent (for the modified version applied here, see **Figure 2**). Participants were instructed to react as fast as possible to a target (fast moving soccer ball) appearing in any corner of the screen. If they were faster than their opponent they would win 50 cents, otherwise they would lose 0–100 cents. The amount of money they would lose was ostensibly determined by their opponent on an 11-ary scale and was presented at the end of the trial in an actually predefined pseudo-randomized order. Individuals could decide how much money they would reduce from their opponent in case they would win at the beginning of each trial. It was stated clearly, that neither the opponent nor the participant would earn the money they subtracted but only the 50 cents they earned in win trials. Monetary reductions thus were not related to reward but consistent with the definition of aggression as a goal-directed behavior with the intent to harm another individual who is motivated to avoid such a treatment (Baron and Richardson, 1994).

In total, there were 54 predefined lost trials (23 high provocation trials: 80–100 cents, 25 low provocation trials: 0–20 cents, 6 medium provocation trials: 30–70 cents) and 30 trials in which they won (always 50 cents). Minor variations could emerge when individuals' reaction time was below 600 ms. In these cases, individuals lost the trial followed by a medium provocation trial (50 cents). Overall, the paradigm lasted 25 min.

An important advantage of the current task compared to the PSAP is the control of provocations which can be gradually modified and which are the same for each participant. Moreover, since the course of the task is the same for each participant it is possible to study both adaptive behavioral changes (punishment adjustments depending on the strength of the provocation) and accumulative frustration (a time-dependent increase of punishment levels) and if these are influenced by the administration of T.

### Hormonal Levels

T and cortisol (C) levels were analyzed with immunologic in vitro quantitative determination of T/C in human serum and plasma (Electrochemiluminescence immunoassay, ECLIA; Roche<sup>r</sup> Diagnostics GmbH, Mannheim, Germany)<sup>1</sup> . In order to verify the treatment success and task influence on T levels, a repeated-measures ANOVA was performed with time as withinsubject variable and treatment group as between-subject variable. The same procedure was performed for C levels.

### Behavior and Emotions

To investigate task-related behavior, we fitted a general linear model on a trial-by-trial basis, which aimed at predicting aggressive behavior of the volunteer using the amount of money he subtracted from his ostensible opponent at each single trial as surrogate. Hence, the amount of money participants reduced in a specific trial (trial x, where x denotes the trial number) was the dependent variable reflecting the participants' aggressive behavior within the trial. As predictors, we included the outcome of the game (win = 1 vs. lose = 0) in the preceding trial (x − 1) and the amount of money reduced by the opponent (0–100) in the preceding trial (x − 1) as well as the trial number x (1–84) modeling linear temporal shifts in aggressive behavior. An intercept was also included in order to account for individual aggression levels across the whole task. All parameters were estimated for each participant and included in a full factorial analysis with the between-subject factor treatment (T, PL). We also included the subjective treatment believe (bT, bPL) as covariate to control for a potential influence. Outliers were excluded if the deviation was more than 2 standard deviations above the mean (**Supplementary Figure S1**).

Parallel to the behavioral responses, emotional effects of the task were tested measuring state anger. Therefore, the difference score (post task—pre task) of the State-Trait-Anxiety Inventory (STAXI, Schwenkmezger et al., 1992) was estimated. Treatment group was added as between-subject factor and the subjective belief about the received treatment was included as covariate.

### Additional Exploratory Analyses

In order to test the relationship of task related anger and task related aggression, individual parameter estimates of the model (temporal course, outcome, provocation, intercept) were included in a step-wise regression analysis. As dependent variable, the STAXI state score (post task-pre task) was added.

Assuming that the relationship of anger and aggression could be influenced by treatment and treatment belief, significant parameters were applied to a moderated moderation model. In detail, we tested a model in which aggression parameters were applied as dependent variable, the anger increase as predictor variable and treatment as well as treatment belief as moderator variables. This procedure was performed with the PROCESS tool of SPSS (Hayes, 2012) applying model 3. This model assumes a three-way interaction of treatment, treatment belief and anger.

### RESULTS

### Group Characteristics

Participants belief to have received PL or T was independent of the received treatment, X 2 (1,N = 92) = 0.25, p = 0.397, η 2 <sup>p</sup> = 0.052, see **Table 1**. The other comparisons of trait aggression between treatment groups did not indicate any group differences, F(4,95) = 0.32, p = 0.880, η 2 <sup>p</sup> = 0.012.

### Hormone Plasma Levels

The analysis included 97 participants since blood samples at T3 could not be gathered of six participants. T plasma levels (**Figure 3**) differed between groups, F(1,95) = 14.38, p < 0.001, η 2 <sup>p</sup> = 0.131, between measurement time points, F(1.30,94) = 5.90, p = 0.01, η 2 <sup>p</sup> = 0.058, and as a function of group by measurement time point, F(1.30,94) = 45.71, p < 0.001, η 2 <sup>p</sup> = 0.325. T plasma levels differed between the T and PL group at time 2 (p < 0.001) and time 3 (p < 0.001), but not at baseline (p = 0.63). In the PL group, T plasma levels significantly decreased from T1 to T2 (p = 0.002), and T1 to T3 (p = 0.009), but did not differ between T2 and T3 (p = 0.100). In the T group, T plasma levels increased from T1 to T2 (p < 0.001), from T2 to T3 (p = 0.002), and from T1 to T3 (p < 0.001). C levels decreased over time in both groups, F(1.5,93) = 54.02, p < 0.001, η 2 <sup>p</sup> = 0.365, but did not differ between groups (p = 0.170).


<sup>1</sup>www.roche-diagnostics.com

### Behavior and Emotions

The analysis included 88 participants in total due to outliers and missing information about believing to have received a PL or testosterone gel. Task related aggressive behavior, measured by the amount of subtracted money from the ostensible opponent, could be significantly explained by prior provocation, F(1,85) = 36.78, p < 0.001, η 2 <sup>p</sup> = 0.302, prior outcome, F(1,85) = 19.83, p < 0.001, η 2 <sup>p</sup> = 0.185, but not by the time course of the task when controlling for subjective believe in having received treatment, F(1,85) = 2.66, p = 0.087, η 2 <sup>p</sup> = 0.030. Significant differences of the T and PL group were noticed for provocation, F(1,85) = 5.61, p = 0.020, η 2 <sup>p</sup> = 0.062, and outcome parameters, F(1,85) = 11.90, p = 0.001, η 2 <sup>p</sup> = 0.120 (**Figure 4A**). The mean effect of provocation in the T group was larger than in the PL group (T: M = 0.18, CI [0.13; 0.23] PL: M = 0.09, CI [0.04; 0.15]) and the mean effect of outcome was smaller in the T group than in the PL group (T: M = −7.76, CI [−10.17; −5.36] PL: M = −1.49, CI [−4.19; 1.20]). Please note that a negative effect for outcome represents that losses coded with 0 lead to higher monetary reductions than winnings coded with 1. The intercept which reflected the individual estimate of overall task aggression did not differ significantly between the T and PL group, F(1,85) = 0.05, p = 0.830, η 2 <sup>p</sup> = 0.001. Finally, the time course was significantly influenced by the believe to have received T or PL, F(1,85) = 8.72, p = 0.004, η 2 <sup>p</sup> = 0.093. In order to follow-up this covariate effect, a t-test for independent samples was performed. The post hoc test indicated that bT participants, compared to bP participants, became more aggressive during the task, t(85) = 2.95, p = 0.004 (**Figure 4B**).

The analysis of anger ratings included 91 participants due to missing information about believing to have received a PL or testosterone gel. In order to evaluate if the task induced negative emotions, state anger (STAXI) pre and post task was compared (**Figure 5**). The repeated measures ANCOVA revealed a significant effect of task with increased anger after the task, F(1,88) = 30.00, p < 0.001, η 2 <sup>p</sup> = 0.254. There was no significant difference between the PL and T group (p = 0.961) and no interaction of task and treatment (p = 0.513), however, the interaction of task and the covariate treatment belief was significant, F(1,88) = 4.00, p < 0.030, η 2 <sup>p</sup> = 0.052 and there was a main effect of the covariate treatment belief, F(1,88) = 4.64, p < 0.034, η 2 <sup>p</sup> = 0.050. Post hoc analyses demonstrated that the bT group was overall more angry, F(1,89) = 4.69, p = 0.033, η 2 <sup>p</sup> = 0.05 (bT: M = 13.91 ± 0.71, bP: M = 12.13 ± 0.46) while groups did not significantly differ before the task (p = 0.233). Thus, after the task, the bT group was significantly more angry than the bP group, F(1,89) = 5.40, p = 0.022, η 2 <sup>p</sup> = 0.057 (bT: M = 16.44 ± 1.1, bP: M = 13.51 ± 0.63).

### Relationship Anger and Aggression

First, associations of emotions and behavior within the social context were tested. Including all task-related model parameters in a step-wise regression model, the temporal course of aggressive behavior and the aggression level were selected as significant predictors for task related anger increase, F(2,96) = 11.47, p < 0.001. Both predictors were positively associated with the anger increase meaning that a temporal aggression increase during the task predicted higher post task anger, b = 4.66 ± 1.03, p < 0.001, and similarly a higher general aggression level (intercept) predicted higher post task anger, b = 0.030 ± 0.012, p = 0.020. Neither outcome related behavior nor provocation related behavior were significant predictors of task related anger and were thus excluded for the model estimation.

Two moderated moderation analyses were performed separately for the temporal aggression parameter and the intercept. The intercept-model was not significant, R = 0.16, p = 0.957. The temporal-model was significant, R = 0.56, p < 0.001. Significant model contributions are given in **Table 2**. Notably, the interaction of treatment, treatment belief and anger was not statistically significant—but there was a trend for the interaction of treatment belief and anger (p = 0.060) as well as treatment and treatment belief (p = 0.051). Conditional effects indicated that the relationship of anger and temporal aggression increase was significant for the PL group both for bT, b = 0.028 ± 0.01, p = 0.036 and bPL, b = 0.074 ± 0.03, p = 0.004. In the T group, the relationship was only significant for bT, b = 0.069 ± 0.02, p = 0.002 but not for bPL, b = 0.022 ± 0.01, p = 0.114.

### DISCUSSION

The current study investigated the causal role of T on males' punishment behavior following social provocation. Exogenous T led to an increased tit-for-tat behavior (Krämer et al., 2007) of male participants: Compared to the PL group, the T group responded with larger punishments after high but smaller punishments after low provocations and wins. Replicating similar findings, we demonstrate that T is context sensitive, likely supporting status-seeking behaviors (Dreher et al., 2016). The T effect does not seem to be a result of overall enhanced frustration since post-task anger and overall aggression levels were unaffected by T administration. However, participants who believed to have been treated with T were angrier after the task and dispensed increased punishment levels to their apparent opponent.

### Provocation as Necessary Trigger for Aggressive Reactions

Provocation as described via money reduction by an ostensible opponent or by losing a costly competition against this

opponent, in the current study led to higher punishment of the opponent. Participants did not profit from monetary reductions but knew that these reductions would diminish the earnings of their opponent. Thus reducing money from the ostensible opponent had the intention to harm another individual who is motivated to avoid this harm—defining an aggressive act (Baron and Richardson, 1994). The punishments participants chose may quantify the harm participants want to inflict on their opponent and thus characterize their aggressiveness. We assessed aggressiveness in relation to several factors, the outcome of the reaction-time games, the provocation by the ostensible opponent, and the time course of the task. Aggressive behavior did not differ between T and PL measured as general level, and the time course of aggressive actions was comparable in both groups. Thus, we conclude that T administration did not affect aggressiveness per se.

Aggressiveness of participants that had received T more strongly varied towards the provocation of the opponent and the game outcome compared to participants in the PL group. T thus did not affect overall aggression but made participants more sensitive for contextual changes. Relative to the PL group participants acted with a stronger tit-for-tat behavior, acting less aggressively in low but more aggressively in high provoking situations. Aggressive reactions especially if oriented towards a provoking opponent may constitute retaliation. T administration thus seems to shift the need for retaliation: In low provoking situation this need is rather reduced but in high provoking situations it increases.

Considering that overall aggression did not differ but instead behavior more strongly followed the opponents' provocation, this may provide some support for the Status Hypothesis. Assuming that losses as well as high provocations would constitute a status attack, T administration increases retaliatory behavior especially in situations in which the social status of males is endangered. Interestingly, it concerns about the social status seem to drive retaliatory behavior more so in men than in women (Geniole et al., 2015). We cautiously suggest that T might be an underlying factor for such gender-specific motivations of retaliating provocative acts. Status-seeking motives as underlying factor for a context dependent enhancement or reduction of aggressiveness, would be in line with numerous studies explaining bidirectional T effects in social interactions (Eisenegger et al., 2011).

Research of the past years demonstrated that T effects underlie both individual and contextual characteristics. With regard to aggression, especially high individual trait dominance may be needed to observe enhanced aggression as a non-genomic effect of increased T (Carré et al., 2017). It may be speculated that highly dominant males would have reacted with higher punishments throughout the TAP, but this has to be confirmed in future studies. We here find some evidence for the influence

significantly in the testosterone (T) and placebo (PL) group. (B) Anger increased significantly more in participants that believed to have received testosterone (bT) than those who believed to have received placebo (bPL). <sup>∗</sup>Significant effect (p < 0.05).

of the context, which has mostly been observed in economic gambling paradigms such as the UG (Eisenegger et al., 2010; Dreher et al., 2016). In contrast to the UG, aggressive decisions in this study were neither costly nor rewarding for the individual and thus might not be driven by the motivation to earn (or lose) more money via high punishments.

While the findings replicate the context effect, it remains open if T modified the perception of the context or the decision of how to act in a corresponding context. Several studies suggest that exogenous T modifies the perception of social threat (van Honk et al., 2000; Wirth and Schultheiss, 2007; Wagels et al., 2017a). Neural processing of reward and social threat seems to be altered under T (Hermans et al., 2008, 2010; Radke et al., 2015). In turn this might produce a shifted sensitivity to reward and punishment as suggested before (van Honk et al., 2004). T administration in this study thus might have increased the feeling of being treated fairly in low provoking situations and the feeling of being treated unfairly in high provoking situations. Partly this may be supported when including the data of the previous experiment in the non-social environment (see Supplementary Table S1 for further details). Anger reactions in the non-social provocation task towards provocation, corresponded to the tit-for-tat behavior in the social aggression task. While, due to time constraints, specific emotional reactions to provocation and outcome could not be assessed, the data of the non-social context provide some support for an enhanced emotional reaction specifically to provocation which might reflect a shift in the perception under T administration.

On the other hand, it is still possible that the perception of fairness remained stable comparing T and PL groups, but the decision how to react in such a more or less fair situation shifted. While to our knowledge no study explicitly investigated this question regarding activational effects of T, organizational effects seem to alter punishing decisions without affecting the perception of the situation (Ronay and Galinsky, 2011).

Though the current results are in line with the Challenge Hypothesis (Wingfield et al., 1990), they emphasize the need of a precise characterization of the challenge. The decision of what is perceived as a challenge might be highly individual and context sensitive and is probably driven by status motives.

### Success vs. Failure

The influence of the competition outcome in the current paradigm has to be interpreted with caution, since all lost trials were confounded with provocation. Primarily, the observed effect of enhanced aggression after lost compared to won trials might be related to the provocation aspect since there was no provocation in the win trials. Nevertheless, the effect may also refer to the game outcome itself (e.g., participants after T administration simply act less aggressively after winning a competition). Competitive situations can modulate T levels (Geniole et al., 2017) and T levels can influence the willingness to engage in future competitions (Carré et al., 2013). In women, high T levels after winning a competition can predict prosocial behavior (Casto and Edwards, 2016). Future tasks should separate outcome and provocation phases to assess complete independent effects.


∗ trend level p < 0.10; ∗∗significant p < 0.05.

### Expectation Effect

While T administration may promote a tit-for-tat behavior but not aggressiveness per se, such anger driven reactions may be triggered by subjective beliefs about T. Similar to the current results, the subjective belief about the effects of T has been shown to promote antisocial egoistic bargaining behavior previously (Eisenegger et al., 2010). While not all studies observed assumed stereotypical behavior due to treatment expectations (Dreher et al., 2016; Cueva et al., 2017) some demonstrated that the belief to have received a steroid can increase physical performance and aggressive behavior (Björkqvist et al., 1994; Maganaris et al., 2000).

In the current study, we observed separate effects of T and the belief about the received treatment. We suggest that the belief in having received a T treatment is associated with stronger frustration as reflected by higher anger ratings after the task and also increasing aggressiveness. This may be a result of a conscious cognitive attribution process. On the one hand, believing that T increases aggression may work as a self-fulfilling prophecy and thereby make participants act more aggressively. On the other hand, attributing antisocial behavior to biological mechanisms may be a self-serving coping strategy excusing socially undesirable behavior. Notably, such a following explanation is rather unlikely since participant that believed to have received T indicated that they were angrier after the task and thus were not able to improve their coping. Possibly, the belief about having received a PL may attenuate the frustration elicited by the task. Interestingly, more participants believed to have received PL gel (66%). Possibly, participants expected physiological or behavioral changes if they received the testosterone gel. When they did not realize any changes they might have concluded to have received the PL gel. The results on the association of treatment belief and anger or aggressiveness therefore are not unambiguous: Either, an expectation might have elicited anger and aggression, or the retrospective evaluation of the behavior might have led to a conclusion about the treatment. Nevertheless, the results underline the importance of the cognitive influence which can be created via associations with the treatment. Thus, it is highly necessary to assess beliefs about treatment. Moreover, studies that systematically investigate a PL effect of T are needed to avoid interactions of the actual treatment and expectations.

### Limitations

The current study investigated a homogenous sample of young, healthy, male participants. It remains unclear if the results can be generalized to other populations e.g., females. Currently, administration studies are confronted with the difficulty of different permissions for males and females and the comparability of the T increases due to biological endowment.

The investigation of aggression is challenging due to many problems: Aggression is mainly defined as a social act and hence requires a social partner. For organizational purposes, the actors mimicking the opponent were different (all young male) persons possibly influencing the perception of the social communication. However, the random actor variance should not result in a systematic bias. In order to improve the credibility, the study was planned as between-subjects design not as a within-subject cross-over design which would be the preferred method to enhance power and reduce interindividual influences. The task was performed in a MRI environment limiting spatial and temporal freedom potentially inhibiting aggression.

### CONCLUSION

T administration does not generally increase aggression per se, but increases tit-for-tat behavior: The higher the provocation the higher the punishment, the lower the provocation, the lower the punishment. Supporting the Status Hypothesis, this again underlines the context sensitive effectiveness of exogenous T. In contrast, the belief to have received T might actually enhance frustration and leads to an increase of aggressive behavior. Alternatively, the assumption of having received T might provide a retrospective explanation and excuse of usually socially undesirable behavior.

### AUTHOR CONTRIBUTIONS

LW was involved in designing the study, analyzing the data, performing the literature research and drafting the first version of the manuscript. MV was involved in designing the study, analyzing the data and correcting the manuscript. TK was involved in analyzing the data, and correcting the manuscript. AE and CB were involved in designing the study and correcting the manuscript. UH was involved in designing the study, interpretation of the data and correcting the manuscript.

### ACKNOWLEDGMENTS

We thank the laboratory (Labordiagnostisches Zentrum, LDZ, Aachen) for blood serum analyses and Dr. Hansen for his consultation. In addition, we want to thank the pharmacy of the University Hospital RWTH Aachen, especially Ms. Griesel. For their assistance during measurement, we thank Laura Westerhoff and Despina Panagiotidis. This study was supported by the Interdisciplinary Center for Clinical Research within the Faculty of Medicine at the RWTH Aachen University (IZKF Aachen, Grant number N-N 7) and the German Research Foundation IRTG 2150. The funding sources had no role in study design, in collection, analysis, and interpretation of data, in the writing of the report, and in the decision to submit the article for publication.

### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fnbeh. 2018.00037/full#supplementary-material

FIGURE S1 | Mean values of aggressiveness separated for preceding provocation levels for the four outliers. Abnormal responding is mainly shown after medium provocation (indicated within the red frame).

## REFERENCES


in healthy men: a novel two-step pharmacological challenge paradigm. Biol. Psychiatry 76, 324–331. doi: 10.1016/j.biopsych.2014.01.016


dominance. Psychoneuroendocrinology 64, 136–142. doi: 10.1016/j.psyneuen. 2015.11.016


testosterone on economic behavior. Proc. Natl. Acad. Sci. U S A 106, 6535–6538. doi: 10.1073/pnas.0812757106

**Conflict of Interest Statement**: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Wagels, Votinov, Kellermann, Eisert, Beyer and Habel. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Tantrums, Emotion Reactions and Their EEG Correlates in Childhood Benign Rolandic Epilepsy vs. Complex Partial Seizures: Exploratory Observations

#### Michael Potegal <sup>1</sup> \*, Elena H. Drewel <sup>2</sup> and John T. MacDonald<sup>3</sup>

<sup>1</sup>Program in Occupational Therapy, Center for Allied Health Professions, University of Minnesota, Minneapolis, MN, United States, <sup>2</sup>Department of Neuro and Behavioral Psychology, St. Luke's Children's Center for Autism and Neurodevelopmental Disabilities, Boise, ID, United States, <sup>3</sup>Department of Neurology, Medical School, University of Minnesota, Minneapolis, MN, United States

We explored associations between EEG pathophysiology and emotional/behavioral (E/B) problems of children with two types of epilepsy using standard parent questionnaires and two new indicators: tantrums recorded by parents at home and brief, emotion-eliciting situations in the laboratory. Children with Benign Rolandic epilepsy (BRE, N = 6) reportedly had shorter, more angry tantrums from which they recovered quickly. Children with Complex Partial Seizures (CPS, N = 13) had longer, sadder tantrums often followed by bad moods. More generally, BRE correlated with anger and aggression; CPS with sadness and withdrawal. Scores of a composite group of siblings (N = 11) were generally intermediate between the BRE and CPS groups. Across all children, high voltage theta and/or interictal epileptiform discharges (IEDs) correlated with negative emotional reactions. Such EEG abnormalities in left hemisphere correlated with greater social fear, right hemisphere EEG abnormalities with greater anger. Right hemisphere localization in CPS was also associated with parent-reported problems at home. If epilepsy alters neural circuitry thereby increasing negative emotions, additional assessment of anti-epileptic drug treatment of epilepsy-related E/B problems would be warranted.

#### \*Correspondence:

Reviewed by: Omid Kavehei,

Qing Yun Wang, Beihang University, China

Edited by: Klaus A. Miczek,

Michael Potegal poteg001@umn.edu

Received: 01 December 2017 Accepted: 20 February 2018 Published: 09 March 2018

Tufts University, United States

University of Sydney, Australia

#### Citation:

Potegal M, Drewel EH and MacDonald JT (2018) Tantrums, Emotion Reactions and Their EEG Correlates in Childhood Benign Rolandic Epilepsy vs. Complex Partial Seizures: Exploratory Observations. Front. Behav. Neurosci. 12:40. doi: 10.3389/fnbeh.2018.00040 Keywords: anger, anxiety, fear, hemispheric laterality, interictal epileptiform discharges, sadness

### INTRODUCTION

Pediatric epilepsy is among the most common neurological diseases of childhood. Besides cognitive deficits, pediatric epilepsy is associated with significant emotional/behavioral (E/B) problems, some of them occurring more frequently than they do in other serious childhood diseases, such as asthma, diabetes or cardiac conditions (Hoare, 1984; Rodenburg et al., 2005). These problems often include ''internalizing'' anxiety and depression, but ''externalizing'' inattention, hyperactivity, anger and aggression are also reported. Seizure factors known to contribute to E/B risk include earlier onset, multiple types and higher frequency (Austin and Caplan, 2007). Nevertheless, we still don't know why only some children experience E/B problems. This uncertainty can be distressing to parents and puzzling to physicians (Smith et al., 2007), so E/B problems associated with epilepsy may be under-diagnosed and under-treated (Hanssen-Bauer et al., 2007; Verrotti et al., 2014). Although some anti-epileptic drugs (AEDs) appear useful in treating E/B problems (Pressler et al., 2005), physicians are unlikely to use them to treat such problems until a clear and direct connection is made between children's psychopathology and their epilepsy.

Some systematic associations between epilepsy type and particular E/B symptoms have been reported. Two studies of 5–16 years old with complex partial seizures (CPS) found greater internalizing, but not externalizing problems on the Child Behavior Checklist (CBCL, Schoenfeld et al., 1999; Caplan et al., 2004). Conversely, executive dysfunction and impulsivity are more characteristic of frontal than temporal lobe epilepsy (Culhane-Shelburne et al., 2002). In two of six diagnostically more specific studies, children with Benign Rolandic Epilepsy (BRE) or ''central-temporal spikes'' were rated by parents as higher than controls on ''distractibility, concentration and impulsiveness'' with the most significant difference being in ''temper'' (on laboratory-specific Swedish questionnaires, Croona et al., 1999) or as having significantly higher CBCL subscale scores including Social Problems and Delinquent Behavior (in German translation, Weglage et al., 1997). In contrast to these whole-group comparisons, the other four studies identified subgroups of children with BRE, varying from 7.6% (Tovia et al., 2011) through about 30% (Yung et al., 2000; Massa et al., 2001) to 50%–64% (Saeed et al., 2014) of their respective samples, who had behavioral problems (inattention/distractability, hyperactivity, oppositional and/or aggressive behavior) based on one or another mix of parent and teacher checklists, neuropsychological or school psychology evaluations, and formal psychiatric diagnoses. However, these studies used different methodologies that are difficult to compare; a direct comparison of CPS vs. BRE behavioral profiles using the same methods would be most informative.

Results of the BRE studies underscore the importance of identifying and characterizing subgroups with E/B problems. Indeed, the affected subgroup in Massa et al. (2001) was differentiated by six pathophysiological features including intermittent slow-waves, multiple asynchronous spike-wave foci, and many interictal abnormalities during wakefulness and sleep. These results agree with findings that behavior problems are associated with interictal epileptiform discharges (IEDs), e.g., sharp waves, spikes and spike-wave complexes (Pressler et al., 2005). Important as these results are, they were obtained through extended (and sometimes repeated) 2–24 h recordings and labor-intensive reading/counting of EEGs. It would be more clinically useful to have a simplified system for relating children's behaviors to abnormal waveforms routinely scrutinized in an office EEG.

Another source of E/B variability is the laterality of seizure foci. In adults, a preponderance of studies implicate left hemisphere foci in psychiatric problems, typically anxiety (Smith et al., 2007). Similar results, albeit mixed and weak, are reported in older studies of children's inattention and internalizing problems. More recently, 6–15 years old with right localized partial epilepsies reportedly showed more emotional impairments (anger, disruptive behavior) than those with left hemisphere foci (Mathiak et al., 2009). Our related observations are noted below.

Information sources for E/B function are also at issue. To date, most behavioral measures have been parent (or teacher) report via standard questionnaires or structured interviews. These are retrospective, generalized over a child's history, and potentially biased by parents' own issues. Here, we introduce two other sources of information; directly observed emotion responses elicited in episodes from the Laboratory Temperament Assessment Battery (Lab-TAB) and parent's on-the-spot observations of temper tantrums at home. Lab-TAB episodes are widely used for eliciting children's emotional reactions to assess temperament and psychopathology (Gagne et al., 2011; Clifford et al., 2015). Akin to the physiological challenges used in clinical EEGs, we used three brief Lab-TAB episodes to elicit mild fearfulness, frustration and disappointment during EEG recording (all episodes did end happily for the child).

Tantrums are an overlooked but important and frequently troublesome aspect of E/B dysregulation that can degrade quality of life for both parent and child. Older studies found increased prevalence of tantrums in children with epilepsy (Keating, 1961; Nuffield, 1961; Elithorn, 1974). Our study was inspired by a pilot survey of 5–12 years old in the University of Minnesota's Pediatric Neuropsychology Clinic that revealed a remarkable 22 of 23 (96%) of those with epilepsy had tantrums (unpublished observations). Note that tantrums are not normative in this age range; tantrum prevalence estimates for these ages vary from 11%–30% (Potegal, 2018). We, like Mathiak et al. (2009), noted that right hemisphere localization was associated with longer and more severe tantrums.

Frequency and duration are significant tantrum parameters, but tantrum content, what children actually do, is equally important. Our work in other clinical (Potegal et al., 2009) and non-clinical populations (Potegal and Davidson, 2003; Green et al., 2011) shows that tantrums are a mix of anger (indicated by, e.g., stamping, hitting, screaming) and sad distress (indicated by whining, crying, comfort-seeking). Tantrum anger to distress ratios are clinically useful, e.g., identifying child psychiatry inpatients with anxiety diagnoses by the lower anger and greater distress of their on-ward ''rages'' (Potegal et al., 2009). Similarly, in our Neuropsychology Clinic survey, tantrums of children with CPS had the lowest levels of anger relative to distress, consistent with their ''internalizing'' symptoms.

Tantrum emotions may have lateralized hemispheric substrates. In a comparison of tantrum prone vs. non-tantrum prone 4 years old, parent reported and Lab-TAB measures of anger were associated with greater EEG activation (lower alpha power) in left temporal lobe whereas sadness was associated with right frontal activation (Potegal and Stemmler, 2010). These finding are in keeping with the now standard model of anger being associated with left hemisphere activation (Harmon-Jones et al., 2010) vs. anxiety and depression being related to right hemisphere activation (Thibodeau et al., 2006; see Andrews et al., 2000 for the corresponding laterality of seizures triggered by anger vs. fear).

The presence of one or another type of epilepsy, the lateralization of seizure foci, and the extent of EEG pathophysiology might account for the occurrence and characteristics of children's tantrums and/or the variability of their E/B problems. To begin resolving E/B variability, we explored four specific hypotheses about the relationship between emotion reactions in the clinic, tantrums at home, and EEG characteristics observed in 1 h recording sessions of children with BRE or CPS. Family factors were controlled by comparing these children to their unaffected siblings. Our hypotheses were:


### MATERIALS AND METHODS

### Participants

Thirteen 5–12 years old with diagnoses of CPS, 6 with BRE, and 11 of their siblings (nine from the CPS group, two from the BRE group) in the same age range were recruited from the practices of participating pediatric neurologists (Drs. J. T. MacDonald, S. Rothman and the Minnesota Epilepsy Group). Diagnoses of CPS or BRE were based upon classical clinical criteria and EEG recordings (International League Against Epilepsy, 1989). Exclusion criteria for all groups included: IQ < 70; major speech/language, sensory, or motor impairments, autism, or brain tumor; major medical problems or seizures secondary to other medical treatment.

Parents consenting to a verbal question or letter from their physician were contacted by telephone. Those indicating interest were sent a written consent form; consent was obtained after they had reviewed the form and had all questions answered. Age of seizure onset, seizure types, current seizure frequencies, estimates of lifetime seizures; EEG, MRI and CT results; and medication history were extracted from medical records. Families were compensated \$100 for each child who participated. This study was carried out in accordance with the recommendations of the Belmont Report, University of Minnesota Institutional Review Board (IRB) with written informed consent from all subjects. All subjects gave written informed consent in accordance with the Declaration of Helsinki. The protocol was approved by the Faculty Social-Behavioral Committee of the University of Minnesota IRB.

### Measures

### Tantrum Calendar/Check-In Calls

Parents noted the onset and duration of each tantrum on a 6 week calendar. Weekly check-in calls asking about the most recent tantrum fostered compliance and allowed assessment of calendar reliability. Tantrum frequency and mean duration were calculated from the completed calendars.

### Post-Tantrum Checklist

A Post-Tantrum Checklist was completed by parents immediately following each of 1–3 tantrums. It assessed the event that triggered the tantrum as well as 14 anger behaviors (e.g., hit, kick, scream) and five distress behaviors (e.g., whine, cry, comfort-seeking) each scored as 0, 1 (one or two occurrences), or 2 (three or more occurrences). Parents also rated overall tantrum severity on a 1–5 scale (3 was ''average'' severity; 1 and 5 were ''among the least severe'' in the last month and ''among the most severe'', respectively). Choices for the child's post-tantrum behavior/mood included ''returned to his/her usual activities as if nothing happened'' and ''seemed in a bad mood'' (Checklist available from MP).

### Questionnaires

Parents completed the 6–18 years Child Behavior Checklist (CBCL; Achenbach and Rescorla, 2001) and the Parenting Stress Index (PSI; Abidin, 1990). Socio-economic variables, i.e., parent age, education, family income and number of children in the family were recorded.

### EEG Recording

Children had a 1 h waking EEG using a 128 channel sensor net (Geodesic EEG System 300, Michel et al., 2004). Electrode impedances were set < 50 KΩ; data were bandpass filtered (0.5–50 Hz) and sampled at 1–200 Hz referenced to vertex. The recording sequence was: 10 min baseline (child sitting still), hyperventilation challenge (<3 min), 10 min baseline, three mild emotion challenges (2–4 min each), final 10 min baseline.

### Emotion Challenges

Parents were informed beforehand about the challenges and observed them from the control room. They were told that they could terminate testing at will. None did so. At the end of each challenge, parents noted their child's overall response on a colorcoded 1–10 scale that also had facial expression cartoons and icons representing emotion intensity (scales available from MP). Children's behavior was videotaped.

### **Spiders**

This episode first tapped phobic-like responses by asking the child to take a large (toy) spider out of its cage and match it to one of several spider pictures (spiders are the most frequent item on children's free self-report of scary things, Muris et al., 1997). Social anxiety was then tapped by telling the child that s/he had 2 min to prepare a speech on spiders to ''some people waiting down the hall''. After 2 min, s/he was told that the people had left, so no speech was needed.

### **Unsolvable puzzles**

The child was asked to assemble three Wechsler Intelligence Scale for Children-III Object Assembly puzzles (switching of pieces made assembly impossible). At the 60 and 90 s points in each 2 min trial they were prompted to ''Work a little faster''. After the third puzzle, the experimenter exclaimed ''Oh my! These puzzles all have missing pieces''; the child was then given a final, easily solvable puzzle and praised for solving it.

### **Worst prize**

Children were told that they had worked so hard that when they were done they would get the prize they had pre-selected from among a set of prizes. Unwrapping the box left by an assistant, they found the least desirable prize, a scribbled on and torn rag doll. After 1 min, the assistant returned and said ''That's not the toy you wanted''. The assistant left again but returned 1 min later having ''found'' the desired prize, which was given to child. Following completion of all episodes, children reported the intensity of their feelings during each one on 1–10 scales like those given to parents.

### Data Analyses

#### Groups

The behavior scores of the two siblings of children with BRE were not statistical outliers on the score distributions of the nine CPS siblings, so all sibling data were pooled into a single Sib group (n = 11).

### Anger/Distress Index

Relative proportions of anger and distress in each checklisted tantrum was quantified with the Anger/Distress Index (A/D-I). A/D-I = (6A − 6D)**/**(6A + 6D) where 6A and 6D are the respective sums of anger and distress behaviors. Sample A/D-I values: 1.0 = anger only, 0.0 = equal mix, −1.0 = distress only.

### EEG Pathophysiology

### **Abnormality score**

All subjects were awake during recording. Digitally stored EEG was visually scored offline by a blinded pediatric neurologist (JTM). Abnormal EEG patterns were defined as rapid shifts in normal waking background to: (1) higher voltage bursts of rhythmic theta or delta activity; and/or (2) IEDs including spikes and spike-wave complexes (Westmoreland, 1996; Geyer et al., 1999; Koutroumanidis et al., 2004; Michel et al., 2004; Jing et al., 2010). **Table 1** quantifies abnormalities: Brief bursts of moderate to high voltage theta were scored as 1; higher voltage theta with sharp or saw-tooth morphology was scored as 2. For IEDs, brief spikes were scored as 1, spike-wave patterns as 2 (Blume et al., 1984; Devinsky et al., 1988; Williamson et al., 1993; Bare et al., 1994). To avoid over-counting abnormal waveforms, the total Abnormality score for each EEG was the sum of the highest scores for each of the three waveforms (range 0–5).

### **Laterality score**

Focal abnormal waveforms involving at least three ipsilateral leads were scored for laterality as in **Table 1**: 1 (exclusively left), 3 (equally distributed) or 5 (exclusively right). Left or rightweighted scores of 2 or 4 were given when abnormal waveform amplitude in one hemisphere was at least twice that in the other hemisphere (Sharbrough, 1993; Kellaway, 2000; Lee et al., 2000; Foldvary et al., 2001).

Differences between groups were tested with univariate and multivariate analysis of variance (MANOVA) and covariance (MANCOVA). Siblings of children with epilepsy often have EEG abnormalities (Croona et al., 1999; Ottman, 2001; Bali et al., 2007). These children, as well as those with EEGs within normal limits, were included in correlations with abnormality. Post hoc linear regression was used to distinguish the contributions of the various EEG abnormalities to behavior scores. Correlations with laterality included only children with some EEG abnormality (regardless of level). We report statistical tests only for correlations with r > 0.4 and for which visual inspection of scatter plots showed general trends that were not due to a few outliers.

### RESULTS

### Family Demographics

Family demographics are in **Table 2**. One CPS family was Asian, another African-American, all others were white. For the CPS group, 55% of fathers and 77% of mothers completed college; for the BRE group, 75% of both mothers and fathers completed college. Families reported a range of incomes; most were middle class. Out of 19 possible family stressors, about 40% of families in both groups reported none, 80%–90% reported ≤ 3.

### Tantrum Calendar Reliability and Representativeness of Checklist Tantrums

Overall, 84% of check-in calls were completely consistent with subsequently submitted calendars. Inconsistencies were generally


IED are interictal epileptiform discharges.

#### TABLE 2 | Group composition and family demographics.


#### TABLE 3 | Tantrum characteristics by group.


A/D Index is the Anger Distress Index defined in "Materials and Methods" section.

minor differences between telephone and calendar reported date, duration or intensity. One tantrum reported by phone was unnoted on the calendar.

For 44% of children, every calendar tantrum was also reported on the checklist. Overall, 69% of all calendar tantrums were reported on the checklist. There was no significant difference in mean duration of calendar and checklist tantrums (t(21) = 1.02) reported for each child. Similarly, 86%–91% of overall severity ratings ranged from 2 to 4 for calendar and checklist tantrums, respectively; there was no difference in the severity distributions (χ 2 (3) = 0.54). Thus, tantrums reported by checklist were representative of a child's tantrums overall.

### BRE vs. CPS Differences

Progressively greater percentages of sibs, CPS and BRE children had tantrums (**Table 3**); statistically, there was a trend for a higher percentage of children in the two epilepsy groups to have tantrums (χ 2 (1) = 3.19, p < 0.08). **Table 3** also shows the BRE group had more frequent tantrums, the CPS group had the longest.

Generally, children with BRE experienced and/or expressed more anger than children with CPS who showed more sadness and distress. Thus, a MANCOVA with age as covariate showed an overall difference in tantrum frequency, duration and A/D-I among the three groups (Wilk's Lambda = 0.45, F(6,32) = 2.6, p < 0.05). Among the three variables, the A/D-I differed significantly across groups (F(2,21) = 5.8, p < 0.02). Simple post hoc contrasts showed a significant difference between BRE vs. CPS in A/D-I (p < 0.01) and a trend for a difference in duration (p < 0.09). In fact, the combination of brevity and

greater anger of BRE tantrums vs. longer, sadder CPS tantrums distinguished the two epilepsy groups, as shown by their nearly non-overlapping distributions on the duration-A/D-I plane (**Figure 1**). Additionally, considering the two most common post-tantrum categories, the CPS group were more likely to be in a ''bad mood'' than to ''return to usual activities'' while the BRE group was the reverse; siblings were intermediate (**Table 3**).

Trends in EEG emotion challenges were similar. A univariate ANOVA showed a trend for among-group differences in childrated anger (F(2,27) = 3.2, p < 0.06); a post hoc test showed higher BRE than CPS anger (p < 0.03). Siblings were intermediate. No such trends were found for child-rated fear or sadness (Parent and child ratings of the child's emotion correlated significantly for two of the three challenges in each group: CPS: Fear r = 0.65, Anger = 0.59; BRE: Fear r = 0.52, Sadness r = 0.41). The videos showed that the majority of siblings smiled at least once during both Puzzle and Worst Prize episodes (55% and 78%, respectively), fewer BRE group children smiled (17% and 50%), and only one or two CPS group children smiled (10% and 17%). This difference in smiling was statistically significant for Worst Prize (χ 2 (2) = 7.88, p < 0.02).

Because the CBCL has no anger or sadness scales, per se, we compared the ratio of Aggression to Withdrawn scale scores across groups. They differed significantly. The BRE group had higher Aggression than Withdrawn scores on average (mean ratio = 1.1 ± 0.17) while the CPS and Sib groups had lower Aggression than Withdrawn scores (mean ratios = 0.96 ± 0.11 and 0.94 ± 0.09, respectively, F(2,27) = 4.2, p < 0.05).

### Emotion/Behavior-EEG Correlations

#### Abnormality

Four siblings of children with CPS had abnormal waveforms. Across all subjects, EEG Abnormality score correlated significantly with emotion ratings by one measure for each

of the challenges: child self-rated fear for spider speech (r = 0.49) and parent ratings of observed anger for puzzles (r = 0.49) and sadness for the bad prize (r = 0.56, **Figure 2**). The three abnormal waveforms affected emotion responses differentially. Linear regressions showed that higher levels of IEDs and/or theta predicted significant increases in parent-rated anger (F(3,27) = 4.24, p < 0.02) and sadness (F(3,27) = 5.83, p < 0.005); delta had no significant effect (see **Table 4** for the breakdown by abnormality).

#### Laterality

Across all subjects, the laterality of EEG abnormality was significantly, but oppositely, associated with Fear and Anger ratings. Fear correlated with left hemisphere localization (laterality index vs. child-rated speech fear, r = −0.69; vs. parentrated fear, r = −0.47), but child-rated Anger correlated with right hemisphere localization (r = 0.56). The most striking correlation of right hemisphere localization was with the PSI scales for children with CPS. All but one of the PSI scales had r's ≥ 0.36; the highest correlations were with Reinforce Parent (r = 0.69) and Acceptability (r = 0.58). **Figure 3** shows the overall increase of Total Child Problems with right hemisphere localization. The addition of BRE group data reduced the PSI


B, regression coefficient; Beta; standardized regression coefficient; SE, Standard Error.

correlations, so these latter results may pertain only to children with CPS.

### Other Variables

Family demographics had no significant across-groups correlation with any emotion-response variables. Although several variables correlated moderately with younger age at first seizure, only two were statistically significant. Across all children with epilepsy, age at first seizure correlated with child-reported sadness, r = −0.54, p < 0.05. For children with CPS, age at first seizure correlated with tantrum frequency, r = −0.69, p < 0.01. AED status of the two epilepsy groups is shown in **Table 2**. Two thirds of children with CPS and all of those with BRE were on AEDs at time of testing. Of the AED-treated children, all but two were on monotherapy. There were no differences among AED subgroups in **Table 2** on any variable reported here.

## DISCUSSION

Associations between pediatric epilepsy and children's E/B problems are well established. Some major risk factors are known, but why some children experience internalizing problems, others display externalizing problems, while still others have only average problems remains unclear. Our direct observations of emotion responses in the laboratory and parent-reported tantrums at home clarify which children are likely to be affected and what problems their EEG might predict. Results suggest that children with BRE react with more anger and aggression; those with CPS react with more sadness and withdrawal. Across both groups, left hemisphere localization of abnormal EEG waveforms was associated with greater fear of public speaking, while right hemisphere localization was associated with greater anger. Right hemisphere localization was also associated with parent-reported problems at home in children with CPS. Increased EEG waveform abnormality correlated with heightened negative emotions in all groups.

The trend for increased tantrums in our sample is consistent with older reports (Keating, 1961; Nuffield, 1961; Elithorn, 1974). The 60% prevalence of tantrums among siblings is also higher than 25%–50% estimates reported in the literature. A higher proportion of sibling tantrums were reportedly triggered by squabbles with siblings (29% vs. 4%–7% for the epilepsy groups). This may indicate that having a sibling with epilepsy increases conflict among children in the family.

Children with BRE tend to have more frequent, but briefer, angry tantrums and then get over it; children with CPS have longer, sadder tantrums and are more likely to be in a bad mood afterward. Similarly, children with BRE showed more anger in response to challenge while children with CPS were least likely to smile. That BRE is associated with more intense anger while CPS associates with stronger distress supports previous findings of externalizing and internalizing problems in BRE vs. CPS, respectively (Weglage et al., 1997; Croona et al., 1999; Yung et al., 2000; Massa et al., 2001, vs. Schoenfeld et al., 1999; Caplan et al., 2004).

The localization and degree of abnormality of EEG waveforms may now help explain why only some children with epilepsy have E/B problems. Our correlational analysis confirmed previous comparisons between epilepsy and control groups (Mathiak et al., 2009) in that children's abnormal right hemisphere waveforms are associated with greater anger, behavioral difficulties and conflict with parents (Note that PSI items are phrased in terms of complaints about child misbehavior and therefore maximize parent report of such problems).

Across diagnostic groups, greater interictal waveform abnormalities were associated with more intense emotion responses. These results resemble previous subgroup differences (Massa et al., 2001), and also highlight EEG waveforms that may predict E/B problems. Theta activity and spikes predicted emotion reactions, but delta did not, which is consistent with observations that delta range slowing is a more generalized phenomenon that may be found in individuals without seizures (Reiher et al., 1989; Hughes and John, 1999). Use of different AEDs within a study can complicate interpretation. Here, we saw no E/B differences among children treated with levetiracetam, other AEDs or no AEDs. In any event, our correlations are between behavior and on or off-medication EEG, i.e., brain activity as it was measured, however it came to be that way. Thus, the important but separate issue of how AEDs may affect behavior is irrelevant to the results as presented.

The systematic relationship between neurological effects, represented by abnormal waveforms, and E/B function highlighted here contrasts with findings that life stress (low socio-economic status, poor parental health) contributes to mental health problems associated with pediatric epilepsy (Ott et al., 2003; Baum et al., 2010). However, our participating families were relatively high SES and reported few stressors. These demographics, together with the lack of family SES effects on emotion variables and the use of sibling controls suggests that environmental stress is probably limited in this study and neurological factors more likely to be causal. As has been noted, the effect of adverse biological factors becomes clearer when environments are benign (Raine, 2002).

Epilepsy-related pathophysiology may shift the neural circuitry of emotions in one or another direction. Our EEG findings suggest that emotion changes in epilepsy fall along two dimensions, laterality and abnormality. Lateralization shifts the balance between anxiety and anger, with left foci associated with excessive anxiety and right foci with excessive anger. If seizure foci reduce ipsilateral interictal function (via, e.g., hypometabolism, Benedek et al., 2006; Badawy et al., 2009), thereby disinhibiting the contralateral hemisphere (Reggia et al., 2001), our results would be entirely consistent with the standard model of left hemisphere activating anger and right hemisphere activating anxiety (Thibodeau et al., 2006; Harmon-Jones et al., 2010).

Along the abnormality dimension, intensity of all negative emotions may rise with increasingly abnormal EEG waveforms. The general tendency to experience negative emotions more frequently and intensely has been called ''Negative affectivity'', and is a major childhood precursor to clinical anxiety and depression (Lonigan and Vasey, 2009). Thus, epilepsyrelated brain pathophysiology, indicated by abnormal EEG theta and IEDs, may increase negative affectivity, leading to clinical psychopathology. Indeed, animal models show that the ''kindling'' of limbic seizures increases interictal fearfulness (Depaulis et al., 1997; Adamec, 2003). Finally, the respective shifts toward anger vs. anxiety in BRE vs. CPS emotion reactions suggests that epilepsy type determines which and/or how neural circuits are affected.

Practically, the direct and possibly causal connection between epilepsy and E/B problems suggested here implies that attention to behavioral issues and referral for appropriate psychotherapy and behavior management training for parents should be a routine part of clinical care for children with epilepsy. Because behavior related abnormal waveforms detected in an office visit time-frame appear to predict E/B problems, clinical EEGs might enable physicians to anticipate what problems a child is likely to have and explain this to parents. Lastly, while some AEDs adversely affect behavior (Ronen et al., 2000), lamotrigine and oxcarbazepine are already used as mood stabilizers in children (Reijs et al., 2004). Our findings provide additional rationale for assessing treatment of epilepsy-related E/B problems with select AEDs (Pressler et al., 2005).

Our scoring of abnormal EEG waveforms is novel and requires replication. The stability of the hemispheric localizations is also unknown. The frequency of IED ''migration'' in children, leading to ''false'' localizations, has long been debated (Andermann and Oguni, 1990 vs. Blume, 1990). Reports of interhemispheric migration range from never (Blume and Kaibara, 1991) through 0.1% (Lundervold and Skatvedt, 1956) to 10% (Lee et al., 2010) and 20% of foci (calculated from Table 1 and Figure 3 in Konishi et al., 1994). Whether apparent migration results from seizure propagation, multifocality or other processes, future studies should determine which, if any, foci migrate and whether this is associated with any behavioral change. However, the chief limitation of this study is modest

### REFERENCES


sample sizes, as acknowledged by the phrase ''Exploratory observations. . .'' in our title. Some effects do replicate earlier work (more severe problems correlate with earlier age of seizure onset) and significant differences were found with small numbers of subjects, suggesting consistency with the literature and robustness of results. Nonetheless, our conclusions must remain working hypotheses until experimental effects are replicated.

### AUTHOR CONTRIBUTIONS

MP designed the study, analyzed the data and wrote the manuscript. EHD added the public speaking component to the Spiders episode and supervised testing. JTM reviewed the EEG literature, designed the recording protocol and analyzed the EEG data. EHD and JTM edited the manuscript.

### DISCLOSURE

We confirm that we have read the Journal's position on issues involved in ethical publication and affirm that this report is consistent with those guidelines.

### ACKNOWLEDGMENTS

This study was supported by a Grant-in-Aid from the University of Minnesota Graduate Faculties and NICHHD grant R01-HD055343 to MP. We thank Michael Frost, M.D. and his staff, Sheryl Andersen, Carol Hoskin and Lindsey Reese at the Minnesota Epilepsy Group, for help in recruitment. Jeri Ann Miller organized test sessions, collected and entered data. We thank Susan Callaghan, Adele Dimian, Alana Feijo and Matt Helgeson for help in testing. Profs. Miguel Fijol and Steven Rothman provided advice and guidance at various stages (and we thank Prof. Frances Lawrence for her forbearance). Most of all, we thank the children and parents who made this study possible (and also Anna Sofia Clement-Potegal, the first author's then 6 year old daughter who served bravely as the first pilot subject).


sensitivity, specificity and predictive value. Can. J. Neurol. Sci. 16, 398–401. doi: 10.1017/s0317167100029450


**Conflict of Interest Statement**: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Potegal, Drewel and MacDonald. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Aggression, Social Stress, and the Immune System in Humans and Animal Models

Aki Takahashi 1,2,3 \*, Meghan E. Flanigan<sup>2</sup> , Bruce S. McEwen<sup>3</sup> and Scott J. Russo<sup>2</sup>

<sup>1</sup> Laboratory of Behavioral Neuroendocrinology, University of Tsukuba, Tsukuba, Japan, <sup>2</sup> Fishberg Department of Neuroscience and Friedman Brain Institute, Icahn School of Medicine at Mount Sinai, New York, NY, United States, <sup>3</sup> Laboratory of Neuroendocrinology, The Rockefeller University, New York, NY, United States

Social stress can lead to the development of psychological problems ranging from exaggerated anxiety and depression to antisocial and violence-related behaviors. Increasing evidence suggests that the immune system is involved in responses to social stress in adulthood. For example, human studies show that individuals with high aggression traits display heightened inflammatory cytokine levels and dysregulated immune responses such as slower wound healing. Similar findings have been observed in patients with depression, and comorbidity of depression and aggression was correlated with stronger immune dysregulation. Therefore, dysregulation of the immune system may be one of the mediators of social stress that produces aggression and/or depression. Similar to humans, aggressive animals also show increased levels of several proinflammatory cytokines, however, unlike humans these animals are more protected from infectious organisms and have faster wound healing than animals with low aggression. On the other hand, subordinate animals that receive repeated social defeat stress have been shown to develop escalated and dysregulated immune responses such as glucocorticoid insensitivity in monocytes. In this review we synthesize the current evidence in humans, non-human primates, and rodents to show a role for the immune system in responses to social stress leading to psychiatric problems such as aggression or depression. We argue that while depression and aggression represent two fundamentally different behavioral and physiological responses to social stress, it is possible that some overlapped, as well as distinct, pattern of immune signaling may underlie both of them. We also argue the necessity of studying animal models of maladaptive aggression induced by social stress (i.e., social isolation) for understanding neuro-immune mechanism of aggression, which may be relevant to human aggression.

Keywords: aggression, social stress, immune system, depression, humans, animal models

### INTRODUCTION: INTERPLAY BETWEEN IMMUNE SYSTEM AND CENTRAL NERVOUS SYSTEM

The immune system is the body's primary active defense against physical injury and pathogens. These insults cause activation of leukocytes that produce cytokines to promote multiple kinds of inflammatory responses. Cytokines are known to produce an array of sickness behaviors such as reductions in activity, food intake, and social interaction, along with increased sleep

#### Edited by:

Xiao-Dong Wang, Zhejiang University, China

#### Reviewed by:

Millie Rincón Cortés, University of Pittsburgh, United States Michael Arthur Van Der Kooij, Johannes Gutenberg-Universität Mainz, Germany

#### \*Correspondence:

Aki Takahashi akitakahashi@kansei.tsukuba.ac.jp

> Received: 06 December 2017 Accepted: 06 March 2018 Published: 22 March 2018

#### Citation:

Takahashi A, Flanigan ME, McEwen BS and Russo SJ (2018) Aggression, Social Stress, and the Immune System in Humans and Animal Models. Front. Behav. Neurosci. 12:56. doi: 10.3389/fnbeh.2018.00056 and anhedonia (Larson and Dunn, 2001; Dantzer et al., 2008). Psychological stress can trigger cytokine release, and growing evidence has shown an important role for the immune system in regulating negative emotional states as well as personality (Black, 2003; Zalcman and Siegel, 2006; Dantzer et al., 2008; Koolhaas, 2008; Maes et al., 2009; Réus et al., 2015).

To produce deleterious behavioral effects in response to stress, peripheral cytokines must enter and act upon brain circuitry controlling mood and emotion (Menard et al., 2017b). There are two main pathways for peripheral cytokines to affect the central nervous system (CNS): the neural pathway via vagus nerve and the humoral pathway via crossing of the blood brain barrier (for more detail, see review from Hodes et al., 2015; Pfau and Russo, 2015). Within the CNS, activated microglial cells, astrocytes, neurons, and endothelial cells have all been shown to produce several cytokines and express many cytokine receptors (Hopkins and Rothwell, 1995; Allan et al., 2005). Thus, brain cytokines have important roles beyond inflammatory processes and can act as neuromodulators to regulate neuronal transmission and plasticity. For example, one of the major proinflammatory cytokines, interleukin-1β (IL-1β), has been shown to increase metabolism of norepinephrine and serotonin (5-HT) (Dunn, 1992; Linthorst et al., 1994, 1995; Zalcman et al., 1994; Brebner et al., 2000; Anisman et al., 2008). In addition, IL-1β increases the production of corticotrophinreleasing factor (CRF) from the hypothalamus and therefore activates the hypothalamus-pituitary-adrenal (HPA) stress axis (Berkenbosch et al., 1987; Linthorst et al., 1994, 1995; Angeli et al., 1999). IL-1 activates the nuclear factor kappa B (NF-κB) signaling pathway (Osborn et al., 1989), which is wellestablished to regulate synaptic plasticity (Schneider et al., 1998; Russo et al., 2009; Boersma et al., 2011; Christoffel et al., 2011).

The immune system can also be modulated by the CNS via top-down mechanisms involving nervous and endocrine systems. For example, stress or negative emotions such as anger activates both the HPA axis and the sympatheticadrenal-medullary (SAM) axis to induce the release of pituitary and adrenal hormones such as adrenocorticotropic hormone, glucocorticoids, prolactin, growth hormone, noradrenaline, and adrenaline. These hormones directly modulate the activity of many immune cells, which express a variety of hormone receptors (for review see Glaser and Kiecolt-Glaser, 2005). For example, glucocorticoids strongly suppress immune cells, and they are widely used in the treatment of inflammatory and autoimmune diseases (Boumpas et al., 1993). Glucocorticoids inhibit the production of pro-inflammatory cytokines by acting directly on glucocorticoid receptors on leukocytes (Lew et al., 1988; Angeli et al., 1999; Dhabhar and McEwen, 1999) (**Box 1**). The sympathetic nervous system, which regulates physical responses to fight/flight situations, also innervates the hematopoietic stem cell niche located within lymphoid organs and bone marrow to modulate leukocyte differentiation and release through a β-adrenergic receptor mechanism (Elenkov et al., 2000; Bierhaus et al., 2003; Tan et al., 2007) (**Box 1**).

## AGGRESSION AND THE IMMUNE SYSTEM IN HUMANS

Aggressive behavior in humans is complex in its expression (i.e., physical and verbal) as well as its causes (i.e., provoked emotion) and consequences on victims (i.e., trauma induced psychopathology) (**Box 2**). Several mental disorders, such as schizophrenia, psychosis, antisocial personality disorder, impulse control disorder, depression, attention deficit disorder, and autism spectrum disorders accompany escalated forms of aggression toward others (violence) or one's self (self-mutilation) (Connor et al., 2002; Raine et al., 2002; Volavka et al., 2005; Comai et al., 2012; Matson and Jang, 2014; Das et al., 2016).

Increasing evidence shows that there is an important relationship between aggression traits and the immune system. For example, it has been reported that immunotherapy to treat patients with hepatitis C by chronic administration of interferon alpha (IFN-α) increases irritability and anger/hostility in some patients (McHutchison et al., 1998; Kraus et al., 2003; Lotrich et al., 2013). Also, hostile marital relationships are associated with slower wound healing and dysregulated cytokine production at wound sites (Kiecolt-Glaser et al., 2005). Pathological levels of aggression in intermittent explosive disorder (IED) or psychosis appear to be linked to heightened proinflammatory cytokines (Coccaro et al., 2014; Das et al., 2016). A similar relationship is observed in healthy individuals with high aggression as a personality trait. As summarized in **Table 1**, these individuals consistently showed higher circulating cytokines such as interleukin-6 (IL-6) and tumor necrosis factor-α (TNF-α), as well as Creactive protein (CRP) than non-aggressive individuals (**Box 3**). However, these correlations varied depending on the subtype of aggression scale used. In one recent study, it was shown that the behavioral, but not cognitive nor affective, subscale of hostility positively correlated with IL-6 and CRP (Marsland et al., 2008). Therefore, elevations in basal levels of circulating proinflammatory cytokines, i.e., IL-6 or CRP, appear to be related to the behavioral aspect of aggression traits. Whether this relationship is direct or indirect must be addressed using animal models.

High states of anger (acute episodes of anger) also induce proinflammatory cytokine release. Marital couples show increases in plasma IL-6 and TNF-α after conflict interactions compared to after supportive interactions, and these increases in cytokines were larger in couples who showed higher hostile behaviors during their conflict interactions (Kiecolt-Glaser et al., 2005). Interestingly, the expectation of an aggressive encounter can also increase circulating proinflamatory cytokine levels as well. Rugby athletes displayed increased IL-1β levels in their blood 2 h before a match, when state anger is high, compared to basal levels (Pesce et al., 2013). Furthermore, IL-1β levels 2 h before the match were positively correlated to anger score. Thus, both an aggressive experience and an expectation of an aggressive event are accompanied by state-related increases in inflammatory cytokines. It is possible that the readiness to provoke anger and aggressive behavior depends on these

#### BOX 1 | Biphasic effects of stress on immune response.

Experiences leading to activation of the autonomic nervous system and HPA axis, as well as other mediators, have biphasic effects on immune responses. One example of this is seen with delayed type hypersensitivity (DTH), an antigen-specific cell-mediated immune response, where acute stress enhances the DTH response while chronic stress has an immunosuppressive effect on the DTH response (Dhabhar and McEwen, 1999). Immune cells "move to their battle stations" and traffic via the circulation to places in the body where they can fight an infection or heal a wound (Dhabhar et al., 2012). Besides adrenaline and cortisol as mediators of this trafficking, IFN-γ, and chemokines are involved (Dhabhar et al., 2000, 2010). IFN-γ plays a role in short-term stress induced immune-enhancement of cell-mediated immunity (Dhabhar et al., 2000), the primary acquisition of immune memory (Dhabhar and Viswanathan, 2005), and in anti-tumor immunity (Dhabhar et al., 2010). However, the actions of IFN-γ may largely come into play after leukocytes have been trafficked to potential sites of immune activation by corticosterone, epinephrine, and norepinephrine (Dhabhar et al., 2012). Parasympathetic and sympathetic activation have synergistic and somewhat opposing effects on immune activation; while activation of the sympathetic response increases inflammatory cytokine production, activation of the parasympathetic response has anti-inflammatory effects (Borovikova et al., 2000; Bierhaus et al., 2003; Matteoli et al., 2014).

#### BOX 2 | Aggression study in human and animal models.

#### Human

Human aggression can be largely categorized into reactive aggression (impulsive and hostile) and premediated (instrumental) aggression. Quantification of aggression in human is often conducted by psychometric inventories. Buss-Durkee Hostility Inventory (BDHI, Buss and Durkee, 1957) or its modified Buss and Perry Aggression Questionnaire (BPAQ, Buss and Perry, 1992) contains subscales for hostility (cognitive aspect of aggression including feelings of ill-will and injustice), anger (affective aspect of aggression including physiological arousal and preparation for aggression), and physical and verbal aggression (behavioral aspect of aggression including instrumental or motor components of behavior). The Cook-Medley Hostility (Ho) Scale is another commonly used questionnaire (Cook and Medley, 1954) that measures cognitive (cynicism and hostile attribute), affective (hostile affect), and behavioral (aggressive responding) aspects of aggression. Other studies use methods to provoke anger in experimental settings and quantify participants' aggressive responses toward fictitious competitors (for review, see Miczek et al., 2002).

#### Animal models

Aggressive behavior in animals is species-typical behavior and it differs depending upon the social system of the species, and includes factors such as territorial aggression, dominance-related aggression, and maternal aggression (Miczek and Fish, 2006). Moreover, aggressive behavior in animals has been traditionally separated into offensive vs. defensive forms of aggression. Offensive aggression is motived by resource control and threat to those resources, whereas defensive aggression is motived by danger of harm to the individual itself (Blanchard and Blanchard, 2006). In the rodent model, aggressive behavior is often quantified by using the resident-intruder test. This test is conducted in the home-cage of a resident male (or sometimes resident female or dam) where an intruder male is introduced. Latency to first attack behavior (often a bite) is measured and used as an index of readiness to initiate aggressive behavior. Animals with shorter attack latency are considered to have higher aggression. In addition, the overall frequency and duration of aggressive acts are measured. Importantly, while dominance is a trait often associated aggression, for the purpose of this review it is important to highlight some distinctions. Dominance-related aggression typically occurs more often in species with defined social hierarchies, such as non-human primates, and is measured by ethological observation to record animals' interactive behaviors with other members in their habitat (either captivity or the field). Importantly, social dominance does not simply reflect trait aggressiveness (Buwalda et al., 2017) but can result from other factors such as ability to mobilize support or anxiety in monkeys and rats (de Waal, 1998; van der Kooij et al., 2017).

individual differences in cytokine production. But also, since noradrenaline triggers activation of monocytes to produce inflammatory cytokines (Bierhaus et al., 2003), sympathetic activation by aggressive events or their expectation might be the cause of this increased cytokine production.

Immune responses are largely divided into two categories: a rapid general immune response (innate immunity) and an acquired delayed immune response (adaptive immunity). Lipopolysaccharide (LPS) is an endotoxin that composes the outer surface membrane of gram-negative bacteria, acts as a pathogen-associated molecular pattern (PAMPs), and stimulates the innate immune response. Studies examining the role of the innate immune response on aggression show that there is a positive correlation between LPS-stimulated monocyte TNF-α expression and aggression (in hostility and behavioral, but not anger, subscales) in healthy males (Suarez et al., 2002). Similarly, monocytes isolated from females with high hostility released more IL-1 and IL-8 than those isolated from low hostility females after LPS-stimulation (Suarez et al., 2004). There have also been reports of increased natural killer (NK) cell cytotoxicity in highly hostile individuals (Christensen et al., 1996; Miller et al., 1999). Thus, individuals with high aggression traits tend to have high innate immune responses, though it is still unclear whether these are causally linked to aggression.

Only a few studies have examined the adaptive immune system in anger/hostility. Of note, one study found that there was a significant positive correlation between the frequency of T and B lymphocyte numbers and past aggressive acts; however, this relationship was only clear in individuals with moderate aggression, but not in highly aggressive individuals (Granger et al., 2000). Another study found that hostility is positively correlated with the release of both pro-inflammatory (TNF-α and IL-2) and anti-inflammatory (IL-4 and IL-10) cytokines from isolated T-cells (Mommersteeg et al., 2008). T-cell driven IL-6, however, was negatively correlated with hostility in the aforementioned study, which opposes the results observed in studies where cytokines were measured from whole serum/plasma or directly in monocytes. This discrepancy suggests that the relationship between aggression traits and inflammatory response is different depending on the leukocyte cell types studied.

Studies measuring cytokines from cerebrospinal fluid (CSF) contrast with those measuring them from peripheral blood. There is no observed correlation between IL-6 levels in the CSF



TABLE

1


Continued

#### BOX 3 | IL-1, IL-6, CRP, TNF-α.

IL-1 (interleukin-1) is a potent pro-inflammatory cytokine first identified as an endogenous pyrogen due to its ability to affect the hypothalamic thermoregulatory center. Currently, there are 11 cytokines in the IL-1 super family (for review, see Allan et al., 2005). Two major subtypes of IL-1 ligands, IL-1α and IL-1β, bind to IL-1 receptors (IL-1R) to activate intracellular cascades such as NF-κB and mitogen-activated protein kinases (MAPKs), and trigger the transcription of multiple inflammation-associated genes including IL-6 and TNF-α. There is also a ligand known as IL-1RA that antagonizes IL-1R to inhibit downstream signaling. Many types of cells in both the peripheral and central immune system produce IL-1 and express IL-1 receptors, including leukocytes, endothelial cells, adipocytes, fibroblasts, neurons, and glial cells.

IL-6 (interleukin-6) is a cytokine that can exhibit either anti-inflammatory or pro-inflammatory properties depending on whether the IL-6 receptor and glycoprotein 130 (gp130) signal transducer are soluble or membrane bound. As is the case with IL-1, IL-6 is produced in many cell types. It was originally identified as B-cell differentiation factor, but it also has a variety of additional functions outside of B cells such as production of acute-phase proteins from liver, angiogenesis, T-cell differentiation, bone metabolism, and neuronal growth (for review, see Hodes et al., 2016).

CRP (C-reactive protein) is one of the acute-phase proteins from the liver activated in by pro-inflammatory cytokines as early a response to inflammation. CRP acts as a pattern recognition molecule that binds to the surface of several microbes and dead cells, and it has been used as a sensitive but non-specific marker of inflammation and infection (Pepys and Hirschfield, 2003).

TNF-α (tumor necrosis factor alpha) is a pro-inflammatory cytokine that was originally identified as a cytotoxic factor produced by lymphocytes and macrophages. More recently TNF-α has been shown to trigger the induction of an array of pro-inflammatory cytokines to regulate cell proliferation, differentiation, and cell death (Aggarwal et al., 2012).

and aggression (Coccaro et al., 2015). Instead, there is a positive correlation between levels of soluble IL-1 receptor II (sIL-1R2) in the CSF and aggression history (Coccaro et al., 2015). IL-1R2 and its soluble form sIL-1R2 act as decoy receptors for IL-1 and inhibit IL-1 mediated signal transduction (Allan et al., 2005). The sIL-1R2 binds to IL-1β with high affinity, and thus the level of IL-1R2 was used as an indirect measurement of IL-1β in the CNS in their study.

These studies in humans highlight the correlational relationship between aggression and the immune system. In the later section, we discuss findings from animal models where more causal relationships between the immune system and aggressive behaviors are beginning to be examined.

### LINK BETWEEN AGGRESSION, DEPRESSION, AND THE IMMUNE SYSTEM IN HUMAN STUDIES

Similar to the findings from aggression studies, increased circulating IL-6 has been observed in humans suffering from major depression (Maes et al., 1997; Kiecolt-Glaser et al., 2003; Hodes et al., 2014; Kiraly et al., 2017). Thus, increased circulating IL-6 seems to be one of the important endophenotypes in depressive-like behaviors as well. Although aggression (violence) and depression are phenotypically very different behavioral outputs, both aggression and depression are triggered by social stress. In fact, suicidal behavior, the most problematic consequence of depression, can be considered as a form of escalated aggression toward the self, and a high comorbidity of suicide and aggression has been observed in human patients (McCloskey and Ammerman, 2018). Thus, it is possible that aggression and depression share certain biological mechanisms.

Repeated immunotherapy to treat patients with acquired immune deficiency syndrome, autoimmune disease, or hepatitis C increases depression as well as anger/hostility (McHutchison et al., 1998; Kraus et al., 2003). In addition, epidemiological evidence has shown that individuals with psychological traits of either depression or hostility have a greater risk of developing coronary heart disease (CHD), which is a known inflammatory disease (Rozanski et al., 1999; Smith and Ruiz, 2002; Betensky and Contrada, 2010). Both male and female individuals with high hostility displayed higher plasma IL-6 levels than non-hostile controls only when they concurrently suffered high depression symptoms (Suarez, 2003). This positive correlation between hostility and IL-6 levels was absent among individuals with low depression symptoms. A similar pattern was observed in another study, where the authors found a significant interaction between hostility and depression symptoms for serum IL-6 and CRP (Stewart et al., 2008). A much larger study of both males and females confirmed a positive correlation between hostility and CRP in highly depressed individuals, but also found no relationship between these measures in people with low levels of depression (Brummett et al., 2010). In contrast, other studies observed a positive correlation between hostility and IL-6 or TNF-α only in individuals with low levels of depression (Miller et al., 2003). This study reported that all individuals with high levels of depression showed high levels of IL-6 regardless of hostility, and thus the relationship between hostility and circulating cytokine levels was not observed. In other experimental designs, the relationship between hostility and cytokines remained even after correcting for depression phenotypes (Marsland et al., 2008). Given the inconsistent results from these correlational studies in humans, far more work is needed to elucidate whether disruptions within the immune system are a common endophenotype for both depression and anger/hostility traits in humans.

### ANIMAL MODELS OF AGGRESSION AND THE IMMUNE SYSTEM

In contrast to human aggression research, which mainly focuses on examining anger and hostility as negative emotional states with pathological aspects that could be the matter of clinical concern, aggression research in animals must consider its ethological and evolutionary importance. Aggression has an adaptive significance for most animal species and is critical for acquiring and protecting territory, food, reproductive mates, and offspring. In animals with hierarchical societies, aggressive behavior is thought to help individuals gain and maintain higher social status (**Box 2**). It has been shown that aggressive behavior, especially the experience of winning, has rewarding properties in animals and repeated aggressive experience may lead to compulsive, pathological aggression that is highly reinforcing (Fish et al., 2002; Falkner et al., 2016; Golden et al., 2016, 2017). Since aggressive behavior poses a strong risk of injury, it is reasonable to assume that animals with high levels of aggression would have stronger immune responses in order to actively recover from injury and to protect themselves from infection.

As with humans, differences in peripheral immune function have been observed between high and low aggressive nonhuman primates and rodents. For example, baboons with higher hierarchical status within the group showed faster wound healing than subordinate individuals in the wild (Archie et al., 2012). Male cynomolgus monkeys with the lowest social status had rates of infection by adenovirus five times greater than monkeys with higher social rank (Cohen et al., 1997). Also, more aggressive cynomolgus monkeys had higher lymphocyte numbers than less aggressive monkeys when they were infected by herpes B virus (Line et al., 1996). In mice, female BALB/c mice with high aggressive behavior were less vulnerable to tumor induction by murine sarcoma virus than low aggression BALB/c females (Amkraut and Solomon, 1972). In agreement with this, highly aggressive C57BL/6 and CBA male mice show stronger immune response (increase in plaque- and rosetteforming cell numbers) toward immunization with protein antigen than submissive males (Devoino et al., 1993). There was also a positive correlation between aggression traits and experimental autoimmune encephalomyelitis (EAE) response in a wild rat population whereby aggressive male rats were more susceptible to experimentally-induced autoimmune disease (Kavelaars et al., 1999), suggesting that aggressive individuals have highly activated immune systems (**Table 2**). These data are in some conflict with data obtained from humans, as highly aggressive humans tend to have higher concentrations of proinflammatory cytokines and slower wound healing than less aggressive humans (Kiecolt-Glaser et al., 2005). It is possible that activation of the immune system is adaptive in aggressive or dominant individuals but can become maladaptive in extreme cases of pathological aggression, which make up most of the cases in human studies. This hypothesis needs further testing.

Individual differences in peripheral immune responses are also reported in forward genetic models of aggression in which animals are selectively bred for aggressive behavior. High aggression NC900 mice and low aggression NC100 mice have been selected over generations from the ICR outbred founder population (Cairns et al., 1983; Gariepy et al., 1996). Interestingly, the NC900 line showed reduced vulnerability to tumor development after calcinogen treatment than the low aggression NC100 line (Petitto et al., 1993). Furthermore, splenic NK cytotoxic activity was also higher in NC900 mice, and exposure to a T cell mitogen caused greater splenic T cell proliferation and increased production of proinflammatory cytokines IL-2 and intereferon-gamma (IFN-γ) in NC900 mice (Petitto et al., 1993, 1994). Thus, aggressive NC900 mice have stronger NK cell and T cell immune responses than nonaggressive NC100 mice. These differences were observed without having any aggressive experience, suggesting that they may reflect trait-like immunity differences.

Studies in constitutive gene knockout mice support the involvement of the immune system in aggressive behaviors. Deletion of TNF receptors, TNF-R1 and TNF-R2, reduced the duration of aggressive behaviors in the resident-intruder test in male mice (Patel et al., 2010). This is in line with findings from human studies in which TNF-α is increased in highly aggressive individuals (**Table 1**). On the other hand, IL-6 knockout mice showed shorter attack latency and increased frequency of aggressive behaviors in the resident-intruder test (Alleva et al., 1998). This same study showed that overexpression of IL-6 had no effect on inter-male aggression, but increased nonagonistic social interaction behaviors such as anogenital sniffing. Although interpretation of the results from these knockout mice is complex because of possible compensatory changes in other cytokines throughout the developmental period, these results strongly indicate functional involvement of the immune system in aggressive behaviors.

Although there are relatively few studies that have investigated the role of brain cytokine signaling in aggression, a handful of studies performed in a cat defensive rage model suggest that it functionally promotes defensive aggression (Zalcman and Siegel, 2006). In this model, electrical activation of either the periaqueductal gray (PAG) or medial preoptic area/hypothalamus causes cats to express a range of defensive aggressive behaviors in response to threat such as hissing, pupillary dilatation, retraction of the ear, as well as increases in blood pressure and heart rate (Siegel et al., 1999). It has been shown that IL-1β, IL-2, and their receptors are localized in a variety brain regions including the PAG and the medial hypothalamus (Bhatt et al., 2005; Hassanain et al., 2005). The local administration of IL-1β into the medial hypothalamus caused an enhancement of the defensive rage response (reduction of attack latency) after PAG stimulation (Hassanain et al., 2003). This pro-aggressive effect of IL-1β injection into the medial hypothalamus was blocked by the 5-HT<sup>2</sup> receptor antagonist LY-53857. Importantly, strong co-localization of 5-HT2C receptors and IL-1 type 1 receptors (IL-1RI) in the medial hypothalamus may reflect the fact that IL-1β and 5-HT are activating the same population of neurons in this region to enhance defensive rage responses (Hassanain et al., 2005). In contrast to IL-1β, microinjection of IL-2 into the medial hypothalamus suppressed defensive rage behavior (Bhatt et al., 2005). The suppressive effect of IL-2 was blocked by pretreatment with a GABA<sup>A</sup> receptor antagonist into the medial hypothalamus, suggesting that the effect of IL-2 in the medial hypothalamus is mediated through GABA<sup>A</sup> receptors (Bhatt et al., 2005). However, IL-2 also facilitated defensive rage behavior when it was microinjected into the PAG (Bhatt and Siegel, 2006), and thus the function of cytokines for defensive rage depends heavily on the brain area in which it is expressed. Also, the function of cytokines might be different depending upon the type of aggression. In mouse territorial aggression models using a resident-intruder



TABLE

2


Continued

test, it has been shown that systemic injection of IL-1β causes a strong increase in attack latency concomitant with a reduction in the total duration of aggressive behaviors (Cirulli et al., 1998), indicating that systemic IL-1β has a suppressive effect on intermale offensive aggression. However, this study used only systemic treatment of IL-1β and its effect in the brain has to be studied in offensive aggression.

In summary, both peripheral cytokines and cytokines in the brain have important modulatory roles in both offensive and defensive aggression. Further studies to examine the complex neural circuitry in which cytokines act to affect aggressive behavior will be necessary to understand the extent of neuroimmune interactions in aggression.

### ANIMAL MODEL OF DEPRESSION AND THE IMMUNE SYSTEM

A large body of evidence shows that both chronic and repeated exposure to social defeat stress, which leads to depression-like behaviors in stress susceptible individuals, has a significant effect on the immune system. Currently, we understand far more about the detailed biological mechanisms underlying the link between social defeat stress and immune activation than we do between aggression and immune activation (for reviews, see Hodes et al., 2015; Pfau and Russo, 2015; Ménard et al., 2017a; Weber et al., 2017). Chronic social defeat stress has been used as an animal model of depression with high ethological and face validity (Miczek et al., 2008; Golden et al., 2011). Disruption of established social hierarchy in the home cage by repeated intrusions of a large dominant male has been shown to cause intensive stress and increases circulating corticosterone in male C57BL/6 mice (Avitsur et al., 2001, 2002). Despite the known immuno-suppressive effect of corticosterone, animals who underwent this social disruption procedure displayed an increased number of splenic monocytes and elevations in IL-6, IL-1β, and TNF-α release after endotoxin LPS stimulation compared to unstressed controls (Stark et al., 2001; Avitsur et al., 2003, 2005; Bailey et al., 2009). Interestingly, these studies also found that social stress caused glucocorticoid receptor (GR) desensitization in splenic monocytes, making them insensitive to inhibition by glucocorticoids and further exacerbating the proinflammatory effects of stress (Stark et al., 2001; Avitsur et al., 2002; Jung et al., 2015). Furthermore, monocytes expressing GR lost the ability to efficiently translocate GR into the nucleus, and thus were unable to suppress NF-κB activity (Quan et al., 2003). These changes in the properties of spleen monocytes in socially disrupted animals were mediated by increased norepinephrine and epinephrine release in the blood and spleen (Hanke et al., 2012). These results suggest that repeated social stress results in abnormal activation of the immune system through a loss of negative feedback signaling via corticosterone.

The peripheral immune system has also been implicated in determining individual vulnerability to social stress. For example, Hodes et al. (2014) examined the levels of circulating cytokines in susceptible and resilient C57BL/6 male mice after 10 days of repeated social defeat stress. Susceptible animals that developed social avoidance after repeated defeat experiences, exhibited higher IL-6 in their serum compared to stress-resilient mice as well as non-defeated control mice. Transplantation of bone marrow from susceptible donor males into host control males caused an increased social aversion following acute social stress, indicating that leukocytes are at least partly responsible for stress susceptibility. Interestingly, there was a preexisting difference in leukocytes between susceptible and resilient males such that susceptible animals had more circulating leukocytes and produced more IL-6 after LPS stimulation than resilient animals (Hodes et al., 2014). Increased peripheral IL-6 levels were also observed in susceptible female mice in a newly developed female social defeat stress model, indicating that IL-6 is a common mechanism mediating social stress susceptibility among the sexes (Takahashi et al., 2017). In addition to exhibiting differences in the immune response to stress, we also show that permeability of the blood brain barrier is different between resilient and susceptible male mice (Menard et al., 2017b). Stress susceptible mice display damaged and leaky blood vessels in the nucleus accumbens (NAc) due to loss of Claudin 5 expression, a molecule that helps form tight junctions between endothelial cells that make up the blood brain barrier (BBB). This impairment in BBB permeability allows more blood IL-6 to enter the brain and promotes depression-like behaviors following social defeat. Future work is needed to understand at the molecular level how social stress primes the neuro-immune axis to open the BBB and promotes a permissive environment for immune-CNS interactions.

### BIOLOGICAL MECHANISMS LINKING AGGRESSION, DEPRESSION, AND IMMUNE FUNCTION

Broadly, the studies described above support the notion that both aggressive and socially-defeated animals display increased production of peripheral cytokines. However, mounting evidence suggests that this simple view does not capture the complex dynamics of cytokine levels in socially stressed animals. For example, repeated aggressive encounters over 20 days caused increases of serum proinflammatory cytokines IL-6, IL-7, and IL-15 in loser, but not winner, male C57BL/7 mice (Stewart et al., 2015). Importantly, there was an increase in the antiinflammatory cytokine IL-10 only in winner males. Further studies are required to understand whether these changes in cytokines observed in winners and losers represent common responses to social stress, or if there are specific immune responses that differ between winners and losers (such as specific activation of anti-inflammatory IL-10 in winners) that function to alter subsequent behavioral outputs. One hypothesis is that dominant individuals have well-balanced activation of both proinflammatory and anti-inflammatory cytokines to actively cope with injury or infections caused by aggressive interactions. In contrast, the experience of repeated social defeat stress, which causes a depression-like phenotype, may induce dysregulation of the immune system leading to a pro-inflammatory state of the animal.

The stress hormone corticosterone increases in both dominant and subordinate animals during an aggressive encounter (Covington and Miczek, 2005), but only subordinate animals show over-activation of HPA axis and long-term hypercortisolism after continuous subordination (Ely and Henry, 1978; Sapolsky, 1989). For example, dominant baboons showed normal activation of cortisol secretion by corticosteronereleasing factor (CRF) and suppression by glucocorticoid negative feedback. By contrast, continuously subordinated baboons displayed hypercortisolism, a blunted response to CRF, and resistance to glucocorticoid-induced negative immune feedback (Sapolsky, 1990). Similarly, in human air traffic controllers with high competence and satisfaction showed a positive correlation between plasma cortisol and amount of workload (number of airplanes under their control), but individuals with low competence showed blunted or disregulated cortisol response (Rose et al., 1982). However, recent work suggests that both the stability of the hierarchy as well as the species under investigation influences findings of whether subordinate status is associated with the highest rate of physical and psychological stressors (Abbott et al., 2003). The immune system also responds to agonistic encounters (acute social stress) in both aggressive dominant and submissive defeated individuals. However, as we have discussed in former sections, dominant animals mount strong adaptive immune responses that protect them from infection or enhance their recovery from injury, while defeated animals develop more prolonged dysregulation of the immune system that leads to pathological physiology and behavioral phenotypes. While causal data linking corticosterone to immune-mediated aggression is not yet available, it is one of the key mediators of brain-immune interactions (Spencer et al., 1991; Dhabhar and McEwen, 1999; Marques-Deak et al., 2005; also see **Box 1**) making it an attractive candidate for further study.

5-HT is one of the most well-studied neurotransmitters involved in the pathophysiology of both depression and escalated aggression (Olivier, 2015; Manchia et al., 2017). Serotonin reuptake inhibitors (SSRIs), the most widely used pharmacological treatment for depressed patients, suppress escalated forms of aggression in rodent models (Pinna et al., 2003; Caldwell and Miczek, 2008; Mikics et al., 2017) and owner-directed aggression in dogs (Dodman et al., 1996). It has been shown in a variety of studies that peripheral cytokines are capable of altering central 5-HT neurotransmission. For example, endotoxin LPS administration activated catecholamine metabolism and increased tryptophan in the brain of mice and rats (Mefford and Heyes, 1990; Dunn, 1992; Linthorst et al., 1995). Furthermore, peripheral IL-1β administration increased levels of 5-HT or its metabolite 5-HTIAA in several brain areas and increased 5-HT1B and 5-HT2C receptor expression in the hippocampus (Gemma et al., 1997; Connor et al., 1998; Dunn, 2006; Anisman et al., 2008), whereas IL-1 receptor antagonists reduced extracellular 5-HT in the hypothalamus (El-Haj et al., 2002). Conversely, in the periphery, 5-HT has immunoregulatory functions via its actions on 5-HT receptors expressed on immune cells (Mössner and Lesch, 1998; Herr et al., 2017). It has been shown that antidepressants, such as imipramine and fluoxetine, have anti-inflammatory effects and reduce the production of the pro-inflammatory cytokine IFN-γ while increasing production of the anti-inflammatory cytokine IL-10 (Maes et al., 1999; Kubera et al., 2001; Ramirez et al., 2015; Köhler et al., 2017). However, these findings need to be interpreted with caution, given that other studies have reported that SSRI treatment increases several pro-inflammatory cytokines in the brain, and that this elevation is necessary to produce the anti-depressant effects of SSRIs (Warner-Schmidt et al., 2011).

In most species, the expression of aggression is generally not pathological but adaptive. Therefore, the link between immune function, aggression, and depression that is observed in humans cannot be simply studied in ethologically relevant animal models of aggression. Rather, it will be appropriate for future studies to utilize animal models of escalated or pathological aggression when investigating the specific role of the immune system in aggression that is relevant to human psychiatric disease. For example, social isolation stress has been shown to induce escalated aggression as well as depressive-like behaviors in rodents (Haller et al., 2014). There are other animal models of escalated aggression considered to be relevant to the study of human aggression (Miczek et al., 2013). These animal models should be used in future studies aimed at understanding the immunobiology of aggression and its relation to depression.

## CONCLUSION AND FUTURE PERSPECTIVES

Findings from studies using animal models of social stress have uncovered numerous neurobiological mechanisms of immune brain interactions that have opened up new avenues for antidepressant drug discovery. In contrast, studies to define the neurobiological mechanism of immune—brain interactions in pathological aggression have lagged far behind, which some have argued reflects the relative paucity in novel drug targets for the treatment of psychiatric conditions with high aggression. Thus, we need a much better understanding of the neural circuits affected by the immune system involved in aggressive behavior. The use of animal models of escalated pathological aggression will help us understand relevant immunobiological mechanisms driving aggressive behavior and provide insight into the possible causes of pathological human aggression. Additionally, it is important to better define how the immune system interfaces with brain circuitry to control aggression. Future studies will have to consider whether individual differences in the immune system are causally linked with aggression, or whether the differences in immune function are simply a consequence of differing amounts of aggressive behaviors.

We currently have a very limited understanding of the neuroimmune interactions mediating aggression in females. In human studies, positive correlations between aggression and peripheral cytokine levels are observed in both males and females, yet rodent studies have not tested causal immunological mechanisms of aggression in females. Future studies will be required to clarify the degree of overlap in the immunological mechanisms that mediate male and female aggression. Given that that there are significant basal sex differences in the immune system (Klein and Flanagan, 2016), it is hypothesized that the immunological mechanisms driving aggression in males and females are different and highly dependent upon gonadal hormone status.

### AUTHOR CONTRIBUTIONS

AT, MF, BM, and SR discussed about the content of this review and wrote the manuscript. All authors read and approved the manuscript.

### REFERENCES


### FUNDING

This research was supported by JSPS KAKENHI Grant Number 17H04766 and 15K12773 (AT), US National Institutes of Mental Health (NIMH) grants T32 MH096678, F31 MH111108-01 (MF), R01 MH114882, R01 MH104559, and R01 MH090264 (SR), and National Center for Complementary and Integrative Health grants P50 MH096890 and P50 AT008661 (SR).


stress-induced synaptic and behavioral plasticity. J. Neurosci. 31, 314–321. doi: 10.1523/JNEUROSCI.4763-10.2011


are attenuated by antiinflammatory drugs in mice and humans. Proc. Natl. Acad. Sci. U.S.A. 108, 9262–9267. doi: 10.1073/pnas.1104836108


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Takahashi, Flanigan, McEwen and Russo. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Heightened Salience of Anger and Aggression in Female Adolescents With Borderline Personality Disorder—A Script-Based fMRI Study

Marlene Krauch<sup>1</sup> , Kai Ueltzhöffer 1,2,3, Romuald Brunner <sup>4</sup> , Michael Kaess 4,5 , Saskia Hensel <sup>6</sup> , Sabine C. Herpertz <sup>1</sup> and Katja Bertsch<sup>1</sup> \*

<sup>1</sup> Department of General Psychiatry, Center for Psychosocial Medicine, University of Heidelberg, Heidelberg, Germany, <sup>2</sup> Department of Psychology, Goethe University Frankfurt, Frankfurt am Main, Germany, <sup>3</sup> Bernstein Center for Computational Neuroscience, University of Heidelberg, Mannheim, Germany, <sup>4</sup> Department of Child and Adolescent Psychiatry, Center for Psychosocial Medicine, University of Heidelberg, Heidelberg, Germany, <sup>5</sup> University Hospital of Child and Adolescent Psychiatry and Psychotherapy, University of Bern, Bern, Switzerland, <sup>6</sup> Department of Psychosomatic Medicine, Central Institute of Mental Health, Mannheim, Germany

Background: Anger and aggression belong to the core symptoms of borderline personality disorder. Although an early and specific treatment of BPD is highly relevant to prevent chronification, still little is known about anger and aggression and their neural underpinnings in adolescents with BPD.

#### Edited by:

Nelly Alia-Klein, Icahn School of Medicine at Mount Sinai, United States

#### Reviewed by:

Gianluca Serafini, Ospedale San Martino (IRCCS), Italy Etsuro Ito, Waseda University, Japan

\*Correspondence: Katja Bertsch

Katja.bertsch@med.uni-heidelberg.de

Received: 14 December 2017 Accepted: 09 March 2018 Published: 26 March 2018

#### Citation:

Krauch M, Ueltzhöffer K, Brunner R, Kaess M, Hensel S, Herpertz SC and Bertsch K (2018) Heightened Salience of Anger and Aggression in Female Adolescents With Borderline Personality Disorder—A Script-Based fMRI Study. Front. Behav. Neurosci. 12:57. doi: 10.3389/fnbeh.2018.00057 Method: Twenty female adolescents with BPD (age 15–17 years) and 20 female healthy adolescents (age 15–17 years) took part in this functional magnetic resonance imaging (fMRI) study. A script-driven imagery paradigm was used to induce rejection-based feelings of anger, which was followed by descriptions of self-directed and other-directed aggressive reactions. To investigate the specificity of the neural activation patterns for adolescent patients, results were compared with data from 34 female adults with BPD (age 18–50 years) and 32 female healthy adults (age 18–50 years).

Results: Adolescents with BPD showed increased activations in the left posterior insula and left dorsal striatum as well as in the left inferior frontal cortex and parts of the mentalizing network during the rejection-based anger induction and the imagination of aggressive reactions compared to healthy adolescents. For the other-directed aggression phase, a significant diagnosis by age interaction confirmed that these results were specific for adolescents.

Discussion: The results of this very first fMRI study on anger and aggression in adolescents with BPD suggest an enhanced emotional reactivity to and higher effort in controlling anger and aggression evoked by social rejection at an early developmental stage of the disorder. Since emotion dysregulation is a known mediator for aggression in BPD, the results point to the need of appropriate early interventions for adolescents with BPD.

Keywords: anger, aggression, emotion regulation, borderline personality disorder, adolescence, social threat sensitivity, functional magnetic resonance imaging, group x age interaction

## INTRODUCTION

Borderline Personality Disorder (BPD) is a life-span mental disorder, which causes a high burden for the affected individuals and their social environment. Patients with BPD are characterized by instability in affect and interpersonal relationships, identity disturbance, and heightened impulsivity. Furthermore, deficits in emotion regulation as well as intense feelings of anger and difficulties in anger control belong to the core symptoms of BPD [American Psychiatric Association (APA), 2013]. In an effort to cope with strong negative feelings, such as anger, many BPD patients show self-destructive behaviors such as self-injury that commonly begins already in early adolescence (Zanarini et al., 2008). In addition, aggressive outbursts against significant others are frequent reactions in patients with BPD. These often result from intense feelings of anger provoked by potential signals of interpersonal threat, such as interpersonal provocation, rejection, or exclusion and may therefore be described as predominantly reactive in nature (Newhill et al., 2009, 2012; for review, also see Mancke et al., 2015; Zanarini et al., 2017).

Previous cross-sectional and longitudinal studies have revealed emotion dysregulation and high levels of trait anger as important mediators of increased reactive aggression in BPD (Newhill et al., 2012; Scott et al., 2014; Mancke et al., 2017), thus supporting the theory of aggressive behavior in BPD being a dysfunctional effort to control intense feelings of anger. Typical situational triggers for feelings of anger are interpersonal situations where the patients feel rejected or excluded by others. Consistently it is assumed that patients with BPD are hypersensitive for interpersonal rejection and threats (Barnow et al., 2009; Bertsch et al., 2013, 2017; Veague and Hooley, 2014) and it has been described that this heightened interpersonal threat sensitivity is one of the core mechanisms for reactive aggression in BPD (Mancke et al., 2015). The deficit in coping with feelings of anger is supposed to be due to the patients' more general deficit in regulating intense negative emotions. Several studies have investigated the neural mechanisms that underlie these emotion regulation difficulties in adult patients with BPD and revealed structural as well as functional frontolimbic abnormalities. A meta-analysis of Schulze et al. (2016) confirmed enhanced activation in the left amygdala and the left hippocampus to emotionally negative stimuli in medication-free adult patients with BPD compared to healthy controls. Furthermore, enhanced activation in the left posterior insula could be found in adult patients with BPD (Schulze et al., 2016). While these findings on limbic hyperactivation reflect a heightened emotional responding to negative emotional stimuli, further findings on a hypoactivation in the anterior cingulate cortex (ACC) and prefrontal regions, such as the orbitofrontal cortex (OFC) (Silbersweig et al., 2007; Koenigsberg et al., 2009; Krause-Utz et al., 2014) support the assumption of deficits in regulating limbic activation. Using a script-driven imagery paradigm to study the neural correlates of imagined physical aggressive reactions to rejection-related anger, we recently found elevated right lateral orbitofrontal and right dorsolateral PFC activations in adult male BPD patients compared to both adult female BPD patients and healthy male controls (Herpertz et al., 2017). While female groups did not differ in neural responses to anger induction in this study, enhanced activations in the dorsal anterior cingulate cortex (dACC), medial prefrontal cortex (mPFC), precuneus, and insula were found in female adult BPD patients compared to healthy controls when feelings of social exclusion were induced with a cyberball paradigm (Domsalla et al., 2013).

Although empirical research has confirmed the reliability and validity of BPD among adolescents (Kaess et al., 2014; Winsper et al., 2016), until now only few studies on emotion dysregulation have focused on BPD in childhood and adolescence. However, this group is of particular interest since the investigation of deficits in emotion regulation in adolescents with BPD could give essential insights in the etiology and the course of the disorder and might hence minimize confounding influences in adult samples, such as a long history of illness and comorbidities as well as treatment exposure. It could therefore not only contribute to a better understanding of BPD in general but also to an improvement of early interventions that may attenuate the full manifestation of the disorder.

Interestingly, Lawrence et al. (2011) found that being excluded triggered the same negative emotions in adolescents and young adults with BPD (15–24 years) as in healthy controls, but that the subjective intensity of these negative emotions (amongst others anger) was elevated in the BPD group. Hence, although triggering negative emotions across groups, already in adolescence experiences of social exclusion are associated with stronger arousal and higher emotional intensity in adolescents with BPD. Besides more intense negative emotional reactions to experiences of social exclusion, deficits in the regulation of emotions have been found in adolescents with BPD compared to healthy controls but also adolescents with other psychiatric disorders, using psychometric (Ibraheim et al., 2017) as well as ambulatory assessment methods (Santangelo et al., 2017). On a neural level, the few studies that investigated structural differences in adolescents with BPD point to volume reductions in the OFC (Chanen et al., 2008; Brunner et al., 2010) and the ventral ACC (Whittle et al., 2009; Goodman et al., 2011), but not in the amygdala or hippocampus as reported for adult BPD patients (Schulze et al., 2016). Interestingly, structural brain differences in partly overlapping brain regions could also be revealed at a very early stage of development in other psychiatric disorders related to emotion dysregulation, such as unipolar and bipolar depressive disorders (Serafini et al., 2014). Functional brain alterations in adolescents with BPD have so far only been addressed in one pilot study. Comparing neural responses to emotionally negative pictures in six adolescent patients with BPD and six healthy controls, this study revealed increased activations in the amygdala, hippocampus, superior frontal gyrus, and precentral gyrus in adolescents with BPD (LeBoeuf et al., 2016).

Taken together, previous fMRI studies on emotion dysregulation as well as anger and aggression in BPD have by the majority focused on adult samples and only little is known about the neural correlates of these aspects of the symptomatology in the early stage of the disorder. Considering the formerly reported structural alterations as well as deficient self-reported emotion regulation in adolescents with BPD, we also assume functional abnormalities in brain circuits involved in the regulation of emotions, such as anger, in this early stage of BPD. In improving our understanding of the neural correlates associated with the symptomatology, we hope to contribute to an early implementation of appropriate therapeutic interventions for adolescents with BPD.

The aim of the present study was to investigate neural correlates of rejection-related feelings of anger and of subsequent other-directed or self-directed aggressive reactions in female adolescent BPD patients in order to contribute to a better understanding of the etiology of disturbed emotion regulation in BPD. Given the fact that most of the psychiatric in- and out-patients with BPD is female and most of previous research has focused on female adult and adolescent BPD patients, we decided to focus on an all-female sample in this first fMRI study on anger and aggression in adolescents with BPD knowing the importance of investigating sex-dependent effects in further studies.

We used a script-driven imagery paradigm with scripts describing a social rejection/exclusion situation and the elicited intense feelings of anger followed by descriptions of self-directed aggressive reactions or aggressive reactions against the rejecting person. Based on previous findings in adolescents with BPD and similar to adult BPD patients, we expected a heightened emotional responding in adolescents with BPD compared to an age-matched healthy control group. We hypothesized that the latter would be reflected in a hyperactivation in the amygdala and the insula in adolescents with BPD compared to adolescent healthy controls when listening to the scripts describing anger inducing rejection situations as well as aggressive behavior. Furthermore, we hypothesized to find reduced activations in prefrontal regions, such as the OFC reflecting deficits in the regulation of highly arousing states of negative emotion in adolescents BPD compared to healthy adolescents. Consistently, we expected adolescents with BPD to score higher on the ratings for feelings of anger after listening to the scripts.

### MATERIALS AND METHODS

### Participants

Twenty female adolescents with BPD (Y-BPD; age 15–17 years) and 20 female adolescent healthy controls (Y-HC; age 15–17 years) took part in the study. The study also comprised 34 female adults with BPD (A-BPD; age 18–50 years) and 32 female adult healthy controls (A-HC; age 18–50 years). The adult sample was almost identical with the female sample reported by Herpertz et al. (2017), therefore, the current analyses and results focus on the adolescent sample and age-related differences.

All adolescent and adult BPD patients currently met at least 5 out of 9 BPD criteria according to the Diagnostic and Statistical Manual of Mental Disorders, 4th edition (DSM-IV, American Psychiatric Association, 2013) and all adolescent and adult healthy controls had never received a psychiatric diagnosis or undergone a psychotherapeutic/psychopharmacological treatment (see **Table 1** for details regarding the sample characteristics). For the two adolescent groups, participants were included with at least 15 years and at most 17 years of age, for the two adult groups participants had to be 18–50 years old. General exclusion criteria comprised: Neurological disorders, current alcohol/drug abuse (urine toxicology screening), alcohol/drug abuse in the last 2 months (interview) or severe medical illness. Additional exclusion criteria were lifetime diagnoses of schizophrenia, schizoaffective or bipolar disorder, as well as reported alcohol/drug dependence in the last 12 months. In the adolescent group, N = 4 patients took antidepressant medication (N = 3 SSRIs, N = 1 Agomelatin), while all adult patients had been free from any psychotropic medication-use for at least 2 weeks prior to participation.

Recruitment was done by the central project of the KFO 256 (Schmahl et al., 2014), a Clinical Research Unit funded by the German Research Foundation (DFG) dedicated to investigating mechanisms of disturbed emotion processing in BPD. The study was approved by the Ethics Committee of the Medical Faculty of the University of Heidelberg. Participants and caregivers (in case of minors) provided written informed consent.

### Measures

All patients and healthy controls took part in an extensive on-site diagnostic interview to assess BPD and other current and lifetime psychiatric disorders. Interviews consisted of the Structured Clinical Interview for DSM-IV disorders (SCID-I; Wittchen et al., 1997) and the International Personality Disorder Examination (IPDE; Loranger, 1999) for assessing the diagnosis of BPD and axis I and II comorbidities. Interviews were performed by experienced diagnosticians who held at least a Master's degree in Psychology or M.D. and underwent standardized training resulting in high inter-rater reliabilities (ICC ≥ 0.91 for both, the number of BPD criteria and the dimensional score assessed by the ZAN-BPD scale). BPD symptom severity was assessed with the Zanarini Rating Scale (ZAN-BPD; Zanarini et al., 2003). Additionally the following trait measurements were assessed: impulsivity with the Barratt Impulsiveness Scale (BIS; Patton and Stanford, 1995), dissociation with the "Fragebogen zu Dissoziativen Symptomen" (FDS; Freyberger et al., 1999; German version of the Dissociative Experience Scale by Bernstein and Putnam), and emotion dysregulation with the Difficulties in Emotion Regulation Scale (DERS; Gratz and Roemer, 2004), as well as intelligence based on Raven's progressive matrices (Raven et al., 2003). Moreover, trait anger was assessed with the State-Trait Anger Expression Inventory (STAXI; Spielberger, 1991), a factor derived instrument that comprises 44 items measuring the experience as well as the expression of anger. Discriminant and convergent validity have been supported (Deffenbacher et al., 1996). Trait aggressiveness was measured with the Aggression Questionnaire (AQ; Buss and Warren, 2000), which comprises the four scales Anger, Hostility, Verbal Aggression and Physical Aggression. It is a wellestablished instrument for the assessment of aggression, the four scales have been reported to have moderate to high internal TABLE 1 | Sample description and self-report data.


Lifetime Current Lifetime Current Lifetime Current Lifetime Current


Y-BPD, female adolescents with BPD; Y-HC, female adolescent healthy controls; A-BPD, female adults with BPD; A-HC, female adult healthy controls. ZAN-BPD, Zanarini Rating Scale for Borderline Personality Disorder; AQ, Buss and Perry Aggression Questionnaire; STAXI trait anger, trait scale of the State-Trait Anger Expression Inventory; DERS, Difficulties in Emotion Regulation Scale; BIS, Barratt Impulsiveness Scale; FDS, Fragebogen für dissoziative Symptome (FDS), which is the German version of the Dissociative Experiences Scale (DERS); PTSD, Posttraumatic Stress Disorder; PD, personality disorder.

consistencies and to be stable over 7 months of testing (Harris, 1997).

### Script-Driven Imagery Task

We used a script-driven imagery task with participants listening to eight standardized scripts, each consisting of four separate phases: baseline, anger induction, other-directed/self-directed aggression, relaxation. The anger induction phase was based on narratives of interpersonal rejection, the other-directed aggression phase comprised narratives of directing physical aggression toward another person, the self-directed aggression phase narratives of self-harming behavior. Each participant listened to four scripts containing narratives of aggressive behavior against others and four scripts describing aggressive behavior against oneself, with order of presentation of the two script types being pseudo-randomized. Within each script, the duration of each of the four phases was 25 s and the duration of the inter-phase interval was 8 s. The scripts were lively read by professional actors and participants were instructed to imagine the described scenes as vividly as possible in order to provoke intense emotional responses. Each of the eight scripts was followed by self-ratings on 5-point Likert scales asking for feelings of anger after the anger induction phase, feelings of anger after the aggression phase as well as for levels of dissociation, derealization, and vividness of imagination. The self-ratings were followed by a 20 s inter-script interval.

### Data Acquisition

Data acquisition was performed in a 3T Tim Trio wholebody scanner (Siemens, Erlangen, Germany) equipped with a 32-channel head coil. Forty transverse slices were acquired in each volume using a T2<sup>∗</sup> -sensitive gradient EPI sequence (TR = 2.350 s, TE = 27 ms, voxel size 2.3 × 2.3 × 2.3 mm). Additionally, isotropic high-resolution (1 × 1 × 1 mm) T1 weighted coronal-oriented structural images were recorded. The course of the experiment as well as acquisition of the data during the experiment was controlled by Presentation 14.2 (Neurobehavioral Systems). Audio texts were presented using an in-ear sound system (Sensimetrics).

## Data Analysis

### Self-Report and Self-Rating Data

Self-report and self-rating data were analyzed using SPSS 20.0 using t-tests for independent groups (Y-BPD vs. Y-HC) with a two-tailed p < 0.05. We additionally performed 2 × 2 analyses of variance (ANOVAs) to analyze specific characteristics of the adolescent BPD sample (group by age interactions). Please note that varying degrees of freedom are due to missing data from one A-HC for ZAN-BPD, from four Y-BPD, one Y-HC, and one A-HC for all other questionnaire data, and from one Y-HC and one A-HC for the self-ratings during the experiment.

### FMRI Data

FMRI data were preprocessed and analyzed in SPM8 under Matlab R2012b. Standard data preprocessing comprised temporal adjustment for differences in slice time acquisition, motion correction, co-registration of EPI images with T1 weighted structural images, segmentation of structural images, normalization into MNI space, and spatial smoothing with an 8-mm full-width-half-maximum (FWHM) kernel. On the first level we set up a general linear model (GLM) for each participant with baseline, anger, other-directed aggression, self-directed aggression, relaxation and rating as regressors as well as 6 motion regressors; we defined the contrasts baseline, anger, other-directed aggression and self-directed aggression for each participant. On the second level, we entered these contrasts into a group (BPD, HC) × age (adolescent, adult) × phase (baseline, anger, other-directed aggression, self-directed aggression) fullfactorial model. Since we were primarily interested in neural correlates of anger and aggression in adolescents with BPD and differences between adults with and without BPD are reported elsewhere (Herpertz et al., 2017), the reported analyses focus on differential contrast between Y-BPD and Y-HC (i.e., Y-BPD vs. Y-HC for anger>baseline, other directed aggression > baseline, and self-directed aggression > baseline). The specificity of the reported effects for the adolescent sample was addressed in additional group (BPD vs. HC) by age (adolescent vs. adults) interaction analyses. To protect against false positive activations, we used a double-threshold approach in all of our fMRI analysis, combining a voxel-based threshold (p < 0.001, uncorrected) with a minimum cluster size of k ≥ 88 (Hayasaka and Nicols, 2004). For this purpose we first used the AFNI program 3dFWHMx to estimate the parameters of an extended, mixed model spatial autocorrelation function (ACF) on the basis of our data. We entered the resulting parameters (a, b, c) 0.313954, 5.59137, 10.2879 into 3dClustSim and obtained a minimum cluster size of 87.7 voxels for our current data, given an uncorrected single voxel threshold of p < 0.001 and a corrected cluster based threshold of p < 0.05.

### RESULTS

### Self-Report Data

Adolescents with BPD reported significantly higher levels of BPD symptoms, aggressiveness, trait anger, emotion dysregulation, impulsivity, and dissociation than healthy adolescent controls [all t(33) ≥ 3.21, p ≤ 0.003]. The group by age interaction was significant for symptom severity [F(1, 101) = 15.64; p ≤ 0.001; η 2 <sup>p</sup> <sup>=</sup> 0.13; higher symptom severity in Y-BPD vs. A-BPD; t(52) = −3.81; p ≤ 0.001]. Further details on demographic, diagnostic, and self-report data are provided in **Table 1**.

### Self-Ratings

Adolescents with BPD did not differ significantly from healthy adolescent controls in their subjective anger ratings after the rejection-based anger induction phase or the aggression phase, the vividness of imagination, or subjective derealization [t(39) ≤ 1.92, p ≥ 0.062], but Y-BPD reported significantly stronger dissociation than Y-HC [t(39) ≤ 2.50, p ≥ 0.017]. There were no significant group by age interactions except for subjective anger after aggression [F(1, 114) = 4.26; p = 0.041] with significantly higher anger ratings in healthy adolescents than healthy adults [t(46) = −2.07; p = 0.045], but no significant difference between adolescents and adults with BPD [t(40.87) = 0.11; p = 0.913]. Details about self-rating data are provided in **Table 2**.

### fMRI Data

### Anger-Induction Phase

Y-BPD showed higher activation in a large cluster comprising parts of the left insula, putamen and claustrum (peak voxel [x, y, z]: −32, −10, 10; T = 3.62, k = 394, p < 0.001) compared to Y-HC. The reverse contrast (Y-HC>Y-BPD) did not result in any significant effects (see **Table 3** and **Figure 1A**). The group by age interactions revealed a significant cluster in the left postcentral gyrus and the left precuneus ((Y-BPD>Y-HC)>(A-BPD>A-HC) peak voxel [x, y, z]: −30, −36, 68; T = 4.20, k = 102, p < 0.001).

### Other-Directed Aggression Phase

Y-BPD showed higher activation in a large cluster including parts of the left insula, putamen, opercular part of the inferior frontal gyrus, middle and superior temporal gyri, pallidum, precuneus, thalamus, and hippocampus (peak voxel [x, y, z]: −34, −48, 8; T = 5.03, k = 1963, p < 0.001; see **Table 4** and **Figure 1B**) compared to Y-HC. The group by age interaction revealed a similar cluster ((Y-BPD>Y-HC)>(A-BPD>A-HC) peak voxel [x, y, z]: −30, −12, 8; T = 4.11, k = 1050, p < 0.001). In addition, Y-HC compared to Y-BPD showed higher activation in a cluster that included the right caudate (peak voxel [x, y, z]: 22, 28, 0; T = 4.04, k = 100, p < 0.001) as well as in a cluster which comprised parts of the right middle and superior temporal gyri (peak voxel [x, y, z]: 42, −48, 10; T = 4.48, k = 430, p < 0.001). Similar to the latter finding, the group by age interaction revealed higher activation in the right middle and superior temporal gyri ((Y-HC>Y-BPD)>(A-HC>A-BPD), peak voxel [x, y, z]: 52, −46, 8; T = 3.69, k = 126, p < 0.001).

### Self-Directed Aggression Phase

When comparing Y-BPD to Y-HC, we found higher activation in a cluster including the left putamen and insula (peak voxel [x, y, z]: −32, −10, −2; T = 3.74, k = 178, p < 0.001) and, additionally, in the left middle temporal gyrus (peak voxel [x, y, z]: −50, −28, 0; T = 3.68, k = 93; p < 0.001). The reverse contrast (Y-HC>Y-BPD) as well as the group by age interactions did not reveal any significant effects (**Table 5** and **Figure 1C**).

#### TABLE 2 | Self-rating data.


Y-BPD, female adolescents with BPD; Y-HC, female adolescent healthy controls; A-BPD, female adults with BPD; A-HC, female adult healthy controls.

TABLE 3 | Full-Factorial Analysis, whole brain results during anger induction phase, p < 0.001 and cluster size k ≥ 88.


Y-BPD, female adolescents with BPD; Y-HC, female adolescent healthy controls; A-BPD, female adults with BPD; A-HC, female adult healthy controls; k, cluster size; MNI, location of the peak voxel according to the Montreal Neurological Institute brain atlas.

### DISCUSSION

This is the first fMRI study investigating neural correlates of rejection-related feelings of anger and reactive aggression in adolescent BPD patients. Using a script-driven imagery setting to induce feelings of anger, and descriptions of subsequent aggressive reactions, we found increased activations in the left posterior insula and left dorsal striatum as well as in the inferior frontal gyrus and parts of the mentalizing network in female adolescents with BPD compared to female age-matched healthy controls. At least for the other-directed aggression phase, this pattern of activation could only be found in the adolescent sample suggesting specific alterations for adolescents with BPD. Together with previous studies, these findings suggest an enhanced emotional reactivity to interpersonal threat- or rejection-related situations early in the development of BPD. Since deficient emotion regulation has emerged an important mediator for aggression in BPD, the current findings support the need of early and specific interventions for affected adolescents.

Listening to anger-inducing descriptions of interpersonal rejection resulted in stronger activations in large clusters including the left posterior insula and the left dorsal striatum, mainly the putamen, in adolescents with BPD compared to healthy adolescents. Similar patterns of increased activations in the left insula and putamen, but also the middle temporal gyrus, were also found in adolescents with BPD during the descriptions of self-directed aggressive behaviors. Furthermore, the imagination of acting out aggressively against the rejecting person also caused elevated activations in large clusters including the left posterior insula and putamen, the middle temporal gyrus reaching into the superior temporal gyrus, the pallidum, precuneus, thalamus, hippocaumpus, and the inferior frontal gyrus. Notably the results of the group by age interaction indicate that the latter effect is characteristic for adolescents with BPD.

Increased left posterior insula activation indicates enhanced emotional reactivity to anger-inducing descriptions of interpersonal rejection and subsequent behavioral responses in terms of aggression directed against the own or the rejecting person in adolescents with BPD. The insula has been found to be crucially involved in the detection and processing of emotionally salient stimuli (Mühlberger et al., 2010). Therefore, it is not surprising that elevated insular activations are commonly reported in response to social exclusion or rejection in adults (Peyron et al., 2000; Eisenberger et al., 2003; Lieberman and Eisenberger, 2009) and adolescents (Masten et al., 2009). Since the posterior insula is part of the affective pain network, it's activations in the context of social rejection may be regarded as a neural correlate of social pain. Following this, the current findings suggest that interpersonal rejection leads to higher levels of social pain in adolescents with BPD than in healthy adolescents. In adult patients with BPD similarly elevated posterior insula activations in response to emotional stimuli have been interpreted as a neural correlate of deficient emotion regulation (Niedtfeld et al., 2010; Schulze et al., 2011). This is of interest since adolescents with BPD did not only show rejectionrelated increased, but also prolonged posterior insula responses



Y-BPD, female adolescents with BPD; Y-HC, female adolescent healthy controls; A-BPD, female adults with BPD; A-HC, female adult healthy controls; k, cluster size; MNI, location of the peak voxel according to the Montreal Neurological Institute brain atlas.

that was also present during the description of subsequent other-directed and self-directed aggressive reactions suggesting deficits in emotion regulation capacities.

Support for the interpretation that interpersonal rejection may be more painful and hence salient for adolescents with BPD than healthy adolescents also comes from enhanced activations in the left dorsal striatum (putamen) as this region is involved in the coding of stimulus saliency (Zink et al., 2003). In addition, elevated activations in the thalamus and hippocampus in response to descriptions of other-directed aggressive reactions speak for enhanced elevated bottom up emotion generation (Reiman et al., 1997) and may indicate deficits in the processing of contextual information (Holland and Bouton, 1999; Liberzon and Abelson, 2016). Interestingly, stress effects on the hippocampus and the dorsal striatum are well documented and have been associated with alterations in related contextual and habitual memory processes (for review see de Quervain et al., 2017).



Y-BPD, female adolescents with BPD; Y-HC, female adolescent healthy controls; A-BPD, female adults with BPD; A-HC, female adult healthy controls; k, cluster size; MNI, location of the peak voxel according to the Montreal Neurological Institute brain atlas.

Contrary to our hypotheses and to previous studies in adults with BPD (Schulze et al., 2016), we did not find increased amygdala responses to the descriptions of interpersonal rejection, other-directed aggression, or self-directed aggression in female adolescents with BPD in the current study. As Herpertz et al. (2017) indeed found a heightened amygdala activation in male but not in female adults with BPD it remains unclear whether the lack of increased amygdala activation in the female adolescent sample in the present study is due to sex differences in amygdala responsivity to the current experimental manipulation or due to a lack in statistical power related to the rather small sample of adolescents (see below). It should also be taken into consideration that the lack of heightened amygdala activation in our paradigm might be specific to the adolescent sample. Previous structural MRI studies also could not find amygdala gray volume alterations (Chanen et al., 2008; Brunner et al., 2010) that have commonly been reported in samples of adult patients with BPD (see Schulze et al., 2016). Additionally, the fact that we found higher anger ratings not only in the adolescent patients, but also in the adolescent healthy controls may indicate a likewise increased emotional reaction—and amygdala activation—in the sample of the healthy controls as a reason for the lack of significant differences in this region between the two adolescent groups.

Interestingly, adolescents with BPD showed a heightened activation in the inferior frontal gyrus in response to descriptions of aggressive reactions against others. As the inferior frontal gyrus has been associated with different forms of self-control and self-regulation, including the regulation of emotion in general (Lieberman, 2007; Tabibnia et al., 2011) and of anger in particular (Fabiansson et al., 2012), this suggests strong efforts to control intense feelings of anger as well as aggressive impulses in adolescents with BPD. Furthermore, elevated activations in the middle temporal gyrus, superior temporal gyrus, and precuneus, regions that belong to the mentalizing network (Saxe and Powell, 2006), in adolescents with BPD could reflect a neural correlate of the tendency to hypermentalize or over-attribute emotions and intensions of others in adolescents with BPD (Sharp et al., 2011). Heightened activation in the precuneus might additionally reflect a stronger emotional reaction to interpersonal rejection in the adolescents with BPD in the present study, as a hyperactivation in the precuneus recently has been associated not only with mentalizing but also with the experience of social exclusion in adult healthy volunteers (Beyer et al., 2014) and with a heightened interpersonal rejection sensitivity in healthy adolescents (Masten et al., 2009).

On a behavioral level, the rating data did not reveal significant differences for the anger ratings between adolescents with BPD and healthy adolescents, which is contrary to our hypothesis. Interestingly healthy adolescents compared to adult healthy controls showed higher anger ratings when listening to descriptions of aggressive behavior, indicating a heightened emotional responding not only in adolescents with BPD but also in healthy adolescents. This might explain why the two adolescent groups did not differ in their anger ratings on the behavioral level, although they showed significant differences in neural activation.

Taken together, the fMRI data suggest a stronger salience of interpersonal rejection in adolescents with BPD associated with higher levels of social pain. Elevated activations in the thalamus and hippocampus suggest an even stronger activation of bottom-up emotion generation and memory retrieval despite high effort to mentalize and control feelings of anger. Considering heightened anger ratings also in healthy adolescents, the fMRI results indicate a level of emotion dysregulation in adolescents with BPD that goes beyond regular emotion regulation difficulties in adolescence (Guyer et al., 2016) suggesting deficits in the interplay of brain regions involved in the generation and regulation of negative emotions, which has been previously proposed as an important neural correlate of emotion dysregulation in BPD. Importantly, this is the first time that this has been shown in adolescents with BPD suggesting an early onset of a failure in emotion regulation.

The experimental paradigm and the possibility to compare the findings in adolescents to those of an adult sample are major advantages of the current study. Nevertheless, several shortcomings need to be mentioned. First, we were only able to include N = 20 female adolescents with BPD aged 15–17 years and future studies with larger samples of (medicationfree) adolescents, a broader age spectrum and/or a longitudinal design are needed before strong conclusions can be drawn. Second, mainly for reasons of feasibility we only included female adolescents in the present study and thus were not able to investigate possible sex differences in neural correlates of anger and aggression in adolescents. As we recently reported distinct sex differences in the same task (Herpertz et al., 2017), further studies with male and female adolescents are needed.

Third, a clinical control group would be needed to address the specificity of the current findings for adolescents with BPD. Fourth, although the majority of our participants were medication-free, we cannot rule out that antidepressant medication in N = 4 adolescents with BPD may have affected amygdala activation (see Schulze et al., 2016 for negative effects of medication on amygdala activity in adult BPD patients). Fifthly a longitudinal investigation of the adolescent sample would be necessary to draw conclusions regarding the development of BPD symptomatology. Finally, we do not have information on feelings of loneliness, shame, guilt, or other negative emotions in response to our rejection-based anger induction and description of aggressive behaviors.

Our results suggest a stronger salience of interpersonal rejection and subsequent aggressive reactions in female adolescents with BPD compared to age-matched healthy female controls. A heightened emotional reactivity to interpersonal rejections might thus be already apparent at early developmental stages of BPD. A question that arises from the current findings is if the stronger emotional reaction to interpersonal rejection in adolescents with BPD is related to real experiences of former peer rejection. This aspect should be further addressed in future studies. In a therapeutic context it could be helpful for adolescents with BPD to develop functional strategies to regulate negative emotions, such as intense feelings of anger. So far, interventions from dialectical behavioral therapy for adolescents (DBT-A; Rathus and Miller, 2002), such as reality check or emotion regulation and stress tolerance skills, or mentalization-based interventions (MBT-A; Bateman and Fonagy, 2010; Rossouw and Fonagy, 2012) could be helpful to down-regulate high levels of emotional arousal already at a very early stage of BPD symptomatology. Learning adaptive emotion regulation strategies at an early stage of the disorder may reduce self-destructive and aggressive behaviors and thus increase the likelihood for positive interpersonal relationships and social functioning in general. Since interpersonal dysfunctions belong to the most persistent symptoms of BPD

### REFERENCES


(Gunderson, 2007), specific and early treatments are of particular importance.

### AUTHOR CONTRIBUTIONS

MaK has contributed substantially to acquisition, analysis, and interpretation of the data. She has drafted the article. KU has contributed substantially to analysis and interpretation of the data. RB and MiK have contributed to conception and design as well as the interpretation of data. They have revised the manuscript critically for important intellectual content. SH has contributed substantially to acquisition of the data. She has revised the manuscript critically for important intellectual content. SCH and KB have contributed substantially to conception and design as well as the interpretation of data. They have revised the manuscript critically for important intellectual content. All authors gave final approval of the version to be published.

### ACKNOWLEDGMENTS

The study was part of the Clinical Research Group KFO 256 supported by the German Research Foundation (Schmahl et al., 2014; www.kfo256.de; HE2660/12-1,HE2660/12-2; HE2660/7-2;HE2660/14-1;HE2660/16-2;BE5292/2-1;BE5292/ 3-2). We acknowledge financial support by Deutsche Forschungsgemeinschaft within the funding programme Open Access Publishing, by the Baden-Württemberg Ministry of Science, Research and the Arts and by Ruprecht-Karls-Universität Heidelberg.

with first-presentation borderline personality disorder. Psychiatry Res. 163, 116–125. doi: 10.1016/j.pscychresns.2007.08.007


others in patients with borderline personality disorder and axis II comparison subjects. Psychiatry Res. 252, 134–138. doi: 10.1016/j.psychres.2017.02.054


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Krauch, Ueltzhöffer, Brunner, Kaess, Hensel, Herpertz and Bertsch. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# The Endocannabinoid System, Aggression, and the Violence of Synthetic Cannabinoid Use, Borderline Personality Disorder, Antisocial Personality Disorder, and Other Psychiatric Disorders

#### Nathan J. Kolla1,2,3 \* and Achal Mishra1,2

*<sup>1</sup> Department of Forensic Psychiatry, Centre for Addiction and Mental Health, Toronto, ON, Canada, <sup>2</sup> Department of Psychiatry, Faculty of Medicine, University of Toronto, Toronto, ON, Canada, <sup>3</sup> Waypoint Centre for Mental Health Care, Waypoint Research Institute, Penetanguishene, ON, Canada*

#### Edited by:

*Klaus A. Miczek, Tufts University, United States*

#### Reviewed by:

*Gabriella Gobbi, McGill University, Canada Maria Morena, University of Calgary, Canada Margaret Haney, Columbia University Medical Center, United States*

> \*Correspondence: *Nathan J. Kolla nathan.kolla@camh.ca*

Received: *26 October 2017* Accepted: *21 February 2018* Published: *27 March 2018*

#### Citation:

*Kolla NJ and Mishra A (2018) The Endocannabinoid System, Aggression, and the Violence of Synthetic Cannabinoid Use, Borderline Personality Disorder, Antisocial Personality Disorder, and Other Psychiatric Disorders. Front. Behav. Neurosci. 12:41. doi: 10.3389/fnbeh.2018.00041* Endogenous and exogenous cannabinoids bind to central cannabinoid receptors to control a multitude of behavioral functions, including aggression. The first main objective of this review is to dissect components of the endocannabinoid system, including cannabinoid 1 and cannabinoid 2 receptors; the endogenous cannabinoids anandamide and 2-arachidonoylglycerol; and the indirect cannabinoid modulators fatty acid amide hydrolase and monoacylglycerol lipase; that have shown abnormalities in basic research studies investigating mechanisms of aggression. While most human research has concluded that the active ingredient of marijuana, 19-tetrahydrocannabinol, tends to dampen rather than provoke aggression in acute doses, recent evidence supports a relationship between the ingestion of synthetic cannabinoids and emergence of violent or aggressive behavior. Thus, another objective is to evaluate the emerging clinical data. This paper also discusses the relationship between prenatal and perinatal exposure to cannabis as well as use of cannabis in adolescence on aggressive outcomes. A final objective of the paper is to discuss endocannabinoid abnormalities in psychotic and affective disorders, as well as clinically aggressive populations, such as borderline personality disorder and antisocial personality disorder. With regard to the former condition, decreased anandamide metabolites have been reported in the cerebrospinal fluid, while some preliminary evidence suggests that fatty acid amide hydrolase genetic polymorphisms are linked to antisocial personality disorder and impulsive-antisocial psychopathic traits. To summarize, this paper will draw upon basic and clinical research to explain how the endocannabinoid system may contribute to the genesis of aggressive behavior.

Keywords: endocannabinoid system, aggression, synthetic cannabinoids, borderline personality disorder, antisocial personality disorder

## INTRODUCTION

Aggression is a multifaceted behavior that leads to harm toward the self or others and whose genesis can be traced to a multiplicity of individual and environmental factors. Essential to understanding mechanisms of aggression or violence (we use these terms interchangeably for the purpose of this review) is a thorough dissection of relevant neurobiological systems. Here, we focus on the neurochemistry of exogenous cannabinoids and features of endocannabinoid system (ECS) signaling that relate to aggressive behavior. This article is divided into five sections. First, we discuss the animal literature probing the ECS and aggression. Second, we discuss developmental effects of cannabis use during prenatal and perinatal periods as well as in adolescence on manifestation of aggression. Third, we examine recent data linking use of synthetic cannabinoids (SCs) to manifestation of aggressive behavior. Fourth, we discuss violence arising from cannabis use in schizophrenia (SCZ) and other common psychiatry disorders. Fifth, we highlight the available evidence pointing to alterations of the ECS in borderline personality disorder (BPD) and antisocial personality disorder (ASPD), two psychiatric conditions characterized by high levels of violence.

### ECS AND AGGRESSION IN ANIMAL MODELS

While a full discussion of the mechanisms and actions of the ECS is beyond the scope of this review, we provide a brief overview. Stimulated by 19-tetrahydrocannabinol (THC), the primary psychoactive ingredient in cannabis, the ECS modulates the activity of a large number of brain neurotransmitters and is a potent modulator of myriad neural circuits influencing human behavior (Basavarajappa, 2007; Crowe et al., 2014). Endogenous cannabinoid ligands (endocannabinoids) are fatty acid amides and monoacylglycerols that function as lipid neuromodulators. Unlike most other neurotransmitters that are stored in vesicles, endocannabinoids exhibit rapid, ondemand synthesis in response to neuronal activation, and once synthesized, undergo retrograde synaptic transmission to the extracellular space where they bind to presynaptic endocannabinoid receptors (Basavarajappa, 2007; Crowe et al., 2014). This style of neurotransmission is known to precisely regulate information flow within most major neurotransmitter pathways and contributes to the synaptic plasticity of brain regions involved in syndromes associated with violent behavior (Basavarajappa, 2007).

Several investigations (Supplementary Table 1, Sheet 1) have documented an anti-aggressive effect of THC in animals (Dorr and Steinberg, 1976; Miczek, 1978). THC binds to cannabinoid CB1 receptors in the central nervous system to exert its psychoactive effects (Matsuda et al., 1990). In a landmark investigation (Miczek, 1978), low-dose THC was administered to resident mice, rats, and squirrel monkeys prior to their participation in a resident-intruder paradigm. Results revealed a decreased frequency of attacks by resident animals in a dose-dependent manner. That is, higher doses of THC were associated with lower attack frequencies. THC was injected intraperitoneally in mice and rats and delivered orally to monkeys. THC administration to intruder mice, rats, and squirrel monkeys did not change defense, submission, or flight reactions when these animals were paired with non-treated attacking resident opponents. Doses ranged from 0.125 to 4.0 mg/kg of THC. Another study (Van Ree et al., 1984) examined social contact behavior in isolated rats that had received low or high doses of THC injected intraperitoneally or cannabidiol, a phytocannabinoid derived from cannabis that is not an addictive drug but may have anxiolytic effects (Crippa et al., 2011). While higher doses of THC (10 mg/kg) had a suppressive effect on social interactions, lower doses (1 mg/kg) decreased aggressive behavior, including fighting, kicking, or biting. Cannabidiol had no effect on social contact behaviors. An investigation conducted in pigeons similarly reported a negative correlation between an injected THC dose (0.5 mg/kg or 1.0 mg/kg) and aggressive responding (Cherek et al., 1980).

On the other hand, some animal studies have failed to detect an effect of THC on aggression [(Sieber et al., 1980) (20 mg THC/kg administered orally), (Cutler and Mackintosh, 1975) (5 mg/kg administered by intraperitoneal injection)] or, alternatively, have reported increased aggressive behavior following cannabinoid or THC administration. On the other hand, providing aggressive, electrically shocked rats with propylene glycol and marijuana (1 mg THC/kg ingested orally) increased aggressive responding (Carder and Olson, 1972). A subsequent investigation (Ueki, 1979) reported that grouphoused rats became significantly more aggressive—fighting between cage mates and muricide emerged—by chronic daily doses of THC (6 mg/kg injected intraperitoneally). Aggression manifested approximately two weeks into treatment. Thereafter, even a single dose of THC elicited an attack response or muricide if rats were isolated. Aggressive behavior was maintained as long as rats were held in isolation. Once transferred to group housing, however, muricide decreased by 50% and attacks were reduced. Why does THC/cannabis administration appear to provoke aggression in some settings but not others? Several possibilities can reconcile these apparent discrepancies. One, the dose and delivery of administered THC appear to be important variables. In general, studies that used smaller doses of THC/cannabis were less likely to report the emergence of aggression. In some instances, aggression even decreased. Two, aggressive responding may be related to the chronicity of THC exposure in animals. Three, adverse environmental manipulations could also impact aggressive behavior when combined with THC intake. Four, the rearing environment, namely whether animals are housed in isolation, as a group, or transferred from one setting to another, may impact tendency toward aggressive responding. One could also speculate that THC withdrawal as opposed to THC administration may increase aggression. It is important to note, however, that these animal results do not necessarily translate to humans, given the complex interplay of biopsychosocial factors that can precipitate aggression in homo sapiens.

Since CB1 receptors are the most abundantly expressed G protein-coupled receptors in the central nervous system (Herkenham et al., 1990) and transduce signals upon binding to THC, research examining this component of the ECS and its relation with aggression could yield important information. CB1 receptors are located on serotonergic, noradrenergic, dopaminergic, GABAergic, and glutamatergic nerve terminals (Hermann et al., 2002; Häring et al., 2007; Oropeza et al., 2007; Azad et al., 2008; Kano et al., 2009; Morozov et al., 2009), with signaling effects most prominent at GABAergic and glutamatergic synapses (Katona and Freund, 2012). It is worth noting that dose-dependent effects of THC and cannabinoids in aggression are likely linked to the fact that CB1 agonists at low doses increase serotonin (5-HT), while at lower doses induce an abrupt decrease of 5-HT (Bambico et al., 2007). One study employing the resident-intruder paradigm compared CB1 knockout (CB1KO) mice with wild-type. Results indicated that mice devoid of CB1 receptors were more aggressive toward intruders than wild-type but only during the first testing session (Martin et al., 2002). In a subsequent study that analyzed social encounters with conspecifics, group-housed CB1KO mice, when compared with wild-type, were found to spend more time in threat and attack behaviors, exhibit aggressive behavior sooner, and engage in longer periods of aggression during a social interactions test (Rodriguez-Arias et al., 2013). Interestingly, administering the CB1 agonist arachidonyl-2′ chloroethylamide (2 mg/kg injected intraperitoneally) to singlehoused aggressive mice decreased aggression. These results highlight the importance of CB1 neurotransmission as a potential anti-aggressive signaling pathway.

A subsequent investigation looked at whether cannabinoid CB2 receptor knockout (CB2KO) mice similarly displayed increased aggression compared with wild-type during the social interaction test and resident-intruder paradigm (Rodriguez-Arias et al., 2015). CB2 receptors are mainly localized to immune cells (Pertwee, 2005) but have also been detected in several areas of the rat brain, including cerebral cortex, striatum, amygdala, thalamus, cerebellum, spinal nucleus, olfactory nucleus, and hippocampus (Gong et al., 2006; Onaivi et al., 2008). It has been suggested that CB2 receptors can only be measured in the brain during situations of neuroinflammation (Benito et al., 2008). Study results (Rodriguez-Arias et al., 2015) indicated that group-housed CB2KO mice devoted more time to threat and attack behaviors, engaged in threat and attack activities for longer periods, and launched more attacks than wild-type group-housed mice. Increased aggression manifested during both tasks. The authors also reported that acute administration of a CB2 agonist (1, 2, and 4 mg/kg of JWH133 injected intraperitoneally) to isolated Oncins France 1 (OF1) mice—a strain that had been selectively bred for aggressive behavior—decreased aggression. However, pre-treatment of OF1 mice with a CB2 antagonist (2 or 4 mg/kg of AM630 injected intraperitoneally) and then application of JWH133 (2 or 4 mg/kg) resulted in animals spending more time attacking than mice treated with the CB2 agonist alone. The results of these CB1 and CB2 receptor KO studies suggest that decreased activation of this receptor system may be linked to emergence of aggressive behavior in certain animals. Conversely, stimulation of CB1 and CB2 receptors appears to exert pacifying effects.

One of two major endogenous ligands and primary molecular targets of CB1 receptors is anandamide (AEA). AEA is synthesized on-demand in postsynaptic membranes (Kano et al., 2009) and then feedbacks in retrograde fashion onto presynaptic CB1 receptors, whose activation inhibits afferent neurotransmitter release (Ohno-Shosaku and Kano, 2014). Similar to studies reporting a biphasic effect of THC on aggression, there is some evidence that AEA influences aggressive behavior in a dose dependent manner (Sulcova et al., 1998). In one model of agonistic behavior, singly housed mice were divided into two groups based on whether they attacked opponents (e.g., aggressive mice) or exhibited defensive-escape behavior without any attacks (e.g., timid mice). Principal study findings included the observation that lower doses of AEA (0.01– 0.1 mg/kg) administered systemically did not affect agonistic behavior in aggressive mice. However, the highest dose tested (10 mg/kg) significantly reduced aggression in aggressive mice, while stimulating timidity in aggressive mice. On the contrary, the lowest dose of AEA (0.01 mg/kg) roused aggressive behavior in timid mice, while the highest dose (10 mg/kg) had no effect on aggressive behavior. In addition to dose-related effects, these results point to possible predisposing factors, such as temperament, that may mediate expression of aggression through endocannabinoid signaling.

2-arachidonoylglycerol (2-AG) is another endogenous ligand of cannabinoid receptors with similar properties as AEA (Morena et al., 2016). 2-AG is metabolized by the hydrolytic enzyme monoacylglycerol lipase (MAGL), which is located near CB1 receptors in presynaptic terminals (Gulyas et al., 2004). The impact of 2-AG neurotransmission on aggressive behavior was evaluated in a recent study that employed both a MAGL inhibitor (JZL184) and CB1 receptor antagonist (AM251). MAGL inhibition (JZL184; 8 and 16 mg/kg injected intraperitoneally) was shown to reduce the number of bites delivered and increase the amount of bites received by resident CD1 mice during the resident-intruder paradigm. At higher doses of the MAGL inhibitor (16 mg/kg), mice received more bites than they initiated, while level of defensiveness was unchanged. Adding a CB1 receptor antagonist (AM251; 0.5 mg/kg injected intraperitoneally) did not dampen the effects of MAGL inhibition, suggesting that the results achieved by administering a MAGL inhibitor were not mediated through CB1 receptors (Aliczki et al., 2015). Among intruders treated with the MAGL inhibitor, more bites were received than delivered and mice also engaged in greater defensive behavior versus offensive strategies. The authors noted that previous findings describing a link between cannabinoids and aggression were largely dependent on the testing conditions (e.g., stressfulness of experimental manipulation, timing of testing, and duration of treatment) and were generally low in magnitude. By contrast, MAGL inhibition differed from other cannabinoid treatments, as both biting and offensive behavior were abolished in treated mice. Notably, these observations were present in both residents and intruders, which occupied different hierarchical positions in the paradigm employed. Replicating these results in the same and other species by employing a variety of experimental manipulations would be important next steps to establishing MAGL inhibitors as potent negative modulators of aggression.

The ECS has also been probed in male Syrian hamsters (Moise et al., 2008). The study in question investigated the effect of CB1 receptor blockade (rimonabant 5 mg/kg injected intraperitoneally or AM251 5 mg/kg injected intraperitoneally) and fatty acid amide hydrolase (FAAH) inhibition (URB597; 0.3 or 3 mg/kg injected intraperitoneally) on a variety of paradigms, including models of conditioned and unconditioned social defeat. FAAH is a membrane-derived lipid modulator that is detectable on intracellular membranes in postsynaptic neurons. The enzyme degrades AEA, presumably leading to decreased AEA-induced CB1 neurotransmission. Although conditioned and unconditioned social defeat paradigms are not specifically designed to assay aggression in the same way as the resident-intruder paradigm, the results obtained can inform on behavior that may be compatible or incompatible with aggressive responding. In this experiment, acquisition of unconditioned social defeat was achieved by re-locating an experimental hamster to the home cage of a known resident intruder for 15 min on day 1. Experimental animals exhibited submissive and defensive behavior toward the resident aggressor. Conditioned defeat was achieved 24 h after the initial acquisition of unconditioned defeat in experimental hamsters. A nonaggressive intruder was then transported to the home cage of the experimental hamster for 5 min on day 2. This interval enabled an adequate sampling of behaviors and ensured that behaviors were consistent between tests. For example, defeated hamsters in these models typically display defensive behavior, circumvent social encounters, and exhibit a lack of natural territorial aggression (Jasnow et al., 2005). Study results in the present investigation confirmed that experimental hamsters exhibited unconditioned social defeat on day 1 when exposed to a dominant hamster; that is, they displayed heightened submissive/defensive behavior compared with any other category of behavior. On day 2, they subsequently manifested conditioned defeat, or behavior similar to unconditioned social defeat, upon exposure to smaller, non-aggressive stimulus hamsters. Although diazepam (2 or 6 mg/kg administered intraperitoneally), a benzodiazepine, decreased submissive and defensive behavior during conditioned defeat, aggressive behavior did not differ by dose. Neither FAAH nor CB1 receptor blockade altered expression of conditioned defeat. Furthermore, none of these pharmacologic manipulations transformed behavior acquired on day 1 during the unconditioned social defeat model. Future studies examining the potential role of FAAH in relation to aggressive behavior should involve paradigms especially designed to measure indices of aggression (e.g., resident-intruder aggression assay) in addition to investigating other species.

In summary, although many studies have reported reduced aggression following THC administration, others have reported opposite effects. This discrepancy could be related to dose, chronicity of exposure, or concurrent environmental manipulations. CB1 agonism appears to exert anti-aggressive effects, while CB1KO models are associated with increased aggression. Effects regarding CB2 receptors appear similar: CB2KOs manifest heightened aggression, while agonists at the CB2 receptor lessen aggression. AEA exerts biphasic effects on aggression, which may depend on the temperament of the animal. MAGL inhibition appears to reduce aggression in mice and at higher levels makes them more vulnerable to attack. Results of a conditioned and unconditioned social defeat paradigm in hamsters were unaffected by FAAH administration. However, future studies could discover a role for FAAH in paradigms specifically designed to elicit aggressive responding.

### DEVELOPMENTAL EFFECTS OF CANNABIS EXPOSURE

The effect of cannabis exposure during prenatal and perinatal periods as well as in adolescence has been studied in humans (Supplementary Table 1, Sheet 2). Among prenatally cannabisexposed males and females, girls, but not boys, scored higher on measures of aggression and inattention at age 18 months, although these effects were no longer present at 36 months (El Marroun et al., 2011). Exposure to cannabis during gestation has also been correlated with altered scores on the Self Report Delinquency Scale (Loeber et al., 1989), which includes violence as a subscale, during adolescence (Day et al., 2011). Some research has also found that attention deficits present by age six years can intensify to delinquency and externalizing behaviors in children with prenatal exposure to cannabis (Goldschmidt et al., 2000). Regarding exposure to cannabis in adolescence, preliminary findings indicate that use of cannabis is associated with property and violent crime, especially during the ages of 14–15 years (Fergusson et al., 2002). While these studies provide initial evidence of a link between cannabis exposure at different developmental windows and aggression, more research is required to mechanistically discern how exposure at different time points can lead to aggressive outcomes.

### SCs AND AGGRESSION

SCs are a heterogeneous group of compounds that are sold as herbal matter to be smoked or consumed in other ways. SCs share some properties in common with THC but also show important differences. They are highly lipophilic and cross the blood-brain-barrier easily (Dhawan et al., 2006). In contrast to THC that has weak partial agonist activity at the CB1 receptor, most SCs exhibit full CB1 receptor agonist activity (Elsohly et al., 2014). Moreover, THC displays only modest affinity for the CB1 receptor, whereas SCs show higher affinity and intrinsic activity at this same receptor. Thus, SCs exert stronger agonist action at the CB1 receptor in terms of efficacy (Hillard et al., 2012; Van Amsterdam et al., 2015). As SCs do not contain cannabidiol or cannabivarin, they may lack some of the intrinsic antipsychotic and anxiolytic properties of natural cannabis (Iseger and Bossong, 2015), although it is not universally accepted that cannabidiol or cannabivarin have antipsychotic or antianxiety effects. Furthermore, since there is no standardization of SC products, the concentration of active ingredients can vary significantly within batches (Sedefov et al., 2009; Vandrey et al., 2012). Consequently, the clinical effects of SCs can be highly unpredictable, even among people who have smoked the same batch together (Kronstrand et al., 2013). Acute effects of SCs can include a wide variety of symptoms, some of which may resemble cannabis intoxication. Characteristic symptoms of SC not seen in cannabis intoxication include agitation, seizures, hypertension, emesis, and hypokalemia. SC consumption often presents with a constellation of psychiatric symptoms, including agitation, anxiety, irritability, hallucinations, cognitive impairment, and psychosis (Brents et al., 2011; Hurst et al., 2011; Hermanns-Clausen et al., 2013; Fattore, 2016). Repeated use of SC can induce tolerance, while discontinuation after prolonged use can lead to withdrawal symptoms (Vandrey et al., 2012). Very little is known about the long-term sequelae of chronic SC use, although selfharm, including self-inflicted burns, have been reported (Meijer et al., 2014).

Case reports have chronicled aggressive behavior accompanied by delirium, anxiety, and psychosis in people with no previous medical history following consumption of SCs (Schwartz et al., 2015). The higher incidence of psychotic symptoms among SC users, including agitation and aggression, suggests that the active ingredients contained within SCs may affect the neural pathways implicated in the manifestation of psychotic symptoms. In a chart review of patients who had been admitted to a dual disorders psychiatric unit, individuals who used both THC and SCs were rated as more aggressive than patients using SCs alone, THC alone, or neither, while those using SCs but not THC had the highest levels of agitation (Bassir Nia et al., 2016). A final study compared risk taking and violent behaviors among youth using SCs and cannabis in a nationally representative sample of students from grades 9–12. The investigation found that those who had ever used SCs, compared with students who had only experimented with THC, were more likely to have engaged in sexual violence during dating, physical violence during dating, forced sexual intercourse upon another individual, injuring someone with a weapon on school property, physical fights, and carrying a weapon (Clayton et al., 2017). However, these results remain silent on whether use of SCs leads to aggression and agitation or if youth with pre-established aggressive tendencies are more prone to experiment with SCs. Clearly, the effects of SCs will depend on their constituent components, which differ between the various SCs.

### CANNABIS, OTHER PSYCHIATRIC PRESENTATIONS, AND AGGRESSION

Use of cannabis has also been implicated in the violence of other psychiatric conditions. Here, we provide brief highlights of these relationships. For example, in a sample composed mainly of patients with affective disorders who had been recently discharged from hospital, persistent use of cannabis (cannabis use was coded dichotomously) was associated with violent behavior at several time points (Dugré et al., 2017). The combined use of cannabis (no specific amounts reported) and alcohol was also found to predict violence in SCZ (Koen et al., 2004). Another population study reported an odds ratio of over 18 for cannabis dependence (average amount used not reported) and comorbid SCZ-spectrum disorder with violent offending (Arsenault et al., 2000). However, not all investigations have reported connections between cannabis use and violence in SCZ (Arango et al., 1999). Heterogeneity of results is likely due to variable amounts of cannabis ingested between groups; whether a patient is an inpatient vs. outpatient, thus affecting his or her ability to use; and overall frequency of use.

## BPD, VIOLENCE, AND THE ECS

BPD is a debilitating psychiatric condition that affects 2% of the general population, 10% of psychiatric inpatients, and 20% of psychiatric outpatients (Widiger and Weissman, 1991; Torgersen et al., 2001). Core symptom clusters of BPD include extreme dysphoric moods; self-destructive, impulsive behavior; and disinhibited anger (Leichsenring et al., 2011). These symptoms drive the recurrent, self-directed aggression and suicidal behavior that account for the exceedingly high morbidity and mortality of BPD (Leichsenring et al., 2011). The gravity of self-directed violence in BPD is reflected by several disturbing statistics: over 60% of BPD patients self-mutilate (e.g., cutting wrists or burning skin), 60–70% of patients with BPD attempt suicide, and 8–10% successfully commit suicide (Soloff et al., 1994; Gunderson, 2001; Gunderson and Ridolfi, 2001; Oldham, 2006).

One report found that serum levels of AEA were higher in BPD (Schaefer et al., 2014). Whole blood samples were obtained from 26 patients with BPD, some of whom had comorbid posttraumatic stress disorder (PTSD); 21 individuals with PTSD only; and 30 healthy controls. However, members from each group were using cannabis and a significant proportion of BPD and PTSD subjects were taking psychotropic medications. Elevated serum levels of AEA were found in BPD compared with the two control groups. Overall, the relationship between plasma ECs and brain function is difficult to interpret (Hillard, 2018). A subsequent study found that AEA was reduced in the cerebrospinal fluid of BPD (Koethe et al., 2014). These subjects were likely taking medications and using illicit substances, making it difficult to reconcile differences in findings between plasma and CSF ECs among BPD patients.

## ASPD, VIOLENCE, AND THE ECS

ASPD is a chronic mental condition that affects 1% of American adults (Lenzenweger et al., 2007). Individuals with ASPD consistently engage in reckless, irresponsible, and impulsive behavior from youth onward, and the interpersonal style of ASPD is characterized by manipulation, deceitfulness, and a callous disregard for the rights of others (Ogloff, 2006). Half of all people with ASPD possess a record of criminal offending, and 85% have a history of violent behavior toward others (Robins and Regier, 1991; Samuels et al., 2004). ASPD is associated with the highest rate of violence toward children, intimate partners, and strangers among all psychiatric disorders (Coid et al., 2006). Unsurprisingly, approximately 50% of incarcerated individuals meet diagnostic criteria for ASPD (Fazel and Danesh, 2002). In addition to violence toward others, ASPD is also associated with increased risk of death by violent suicide (Repo-Tiihonen et al., 2001).

A single genetic association study of 137 alcoholic males found that a polymorphism of the gene coding for FAAH (C385A) was associated with a diagnosis of ASPD. The A/A genotype of the FAAH gene is associated with decreased FAAH expression and activity in humans (Chiang et al., 2004). Multivariate regression analysis additionally revealed that higher impulsiveantisocial psychopathic traits predicted the C/C FAAH genotype (Hoenicka et al., 2007), or greater in vitro levels of FAAH. The A/C and A/A genotypes similarly result in greater AEA (Sipe et al., 2002). Finally, a positron emission tomography study of CB1 receptor availability in 47 healthy individuals reported that novelty seeking, a core feature of ASPD that is highly correlated with aggression (Raine et al., 1998), was inversely correlated with CB1 receptor availability, especially in the amygdala (Van Laere et al., 2009). While these results were obtained in a non-clinical sample, they could shed light on the molecular underpinnings of pathological personality traits in ASPD. Furthermore, there is consistency between lower in vitro levels of AEA in ASPD reported in the study by Hoenicka et al. (2007) and increased CB1 receptor availability among individuals with high antisocial traits. Quantifying brain levels of AEA and CB1 receptor availability in ASPD would be a crucial next step to test this mechanism. Importantly, the mechanism between self-harm and aggression toward others shares similarities (Chester et al., 2015).

### CONCLUSION

This review has explored the bench to bedside work highlighting the role of endogenous and exogenous cannabinoids in relation

### REFERENCES


to aggressive behavior. By far, the largest corpus of evidence originates from preclinical models that have examined the major components of the ECS and aggression. With respect to SCs, their unique pharmacodynamic properties differ from those of THC, which perhaps accounts for discrepancies in side effect profiles, including enhanced aggression and agitation in SCs. Finally, preliminary evidence links ECS abnormalities to aggressive psychiatric conditions, such as BPD and ASPD. The next challenge will be to draw connections between indices of aggressive and violence in these disorders with alterations of the ECS. In conclusion, continued exploration of the highly nuanced ECS will only bolster our understanding of human aggression and violence.

### AUTHOR CONTRIBUTIONS

NK and AM were both responsible for researching and writing the article.

### FUNDING

NK is supported by funds from the Canadian Institutes of Health Research, Ontario Mental Health Foundation, Brain and Behavior Research Foundation, and the American Academy of Psychiatry and the Law Institute for Education and Research.

### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fnbeh. 2018.00041/full#supplementary-material


population of Britain: public health implications. Br. J. Psychiatry 189, 12–19. doi: 10.1192/bjp.189.1.12


behavior in male Syrian hamsters. Psychopharmacology (Berl.), 200, 333–346. doi: 10.1007/s00213-008-1209-5


exposure to the novel synthetic cannabinoid ADB-PINACA. J. Emerg. Med. 48, 573–580. doi: 10.1016/j.jemermed.2014.12.038


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Kolla and Mishra. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Pleiotropic Contribution of MECOM and AVPR1A to Aggression and Subcortical Brain Volumes

Marjolein M. J. van Donkelaar 1,2 , Martine Hoogman1,2 , Irene Pappa<sup>3</sup> , Henning Tiemeier 3,4,5 , Jan K. Buitelaar 2,6,7 , Barbara Franke1,2,8† and Janita Bralten1,2 \* †

<sup>1</sup>Department of Human Genetics, Radboud University Medical Center, Nijmegen, Netherlands, <sup>2</sup>Donders Institute for Brain, Cognition and Behaviour, Radboud University, Nijmegen, Netherlands, <sup>3</sup>Department of Child and Adolescent Psychiatry/Psychology, Erasmus Medical Center, Rotterdam, Netherlands, <sup>4</sup>Department of Psychiatry, Erasmus Medical Center, Rotterdam, Netherlands, <sup>5</sup>Department of Epidemiology, Erasmus Medical Center, Rotterdam, Netherlands, <sup>6</sup>Department of Cognitive Neuroscience, Radboud University Medical Center, Nijmegen, Netherlands, <sup>7</sup>Karakter Child and Adolescent Psychiatry, Radboud University Medical Center, Nijmegen, Netherlands, <sup>8</sup>Department of Psychiatry, Radboud University Medical Center, Nijmegen, Netherlands

Reactive and proactive subtypes of aggression have been recognized to help parse etiological heterogeneity of this complex phenotype. With a heritability of about 50%, genetic factors play a role in the development of aggressive behavior. Imaging studies implicate brain structures related to social behavior in aggression etiology, most notably the amygdala and striatum. This study aimed to gain more insight into the pathways from genetic risk factors for aggression to aggression phenotypes. To this end, we conducted genome-wide gene-based cross-trait meta-analyses of aggression with the volumes of amygdala, nucleus accumbens and caudate nucleus to identify genes influencing both aggression and aggression-related brain volumes. We used data of large-scale genome-wide association studies (GWAS) of: (a) aggressive behavior in children and adolescents (EAGLE, N = 18,988); and (b) Magnetic Resonance Imaging (MRI)-based volume measures of aggression-relevant subcortical brain regions (ENIGMA2, N = 13,171). Second, the identified genes were further investigated in a sample of healthy adults (mean age (SD) = 25.28 (4.62) years; 43% male) who had genome-wide genotyping data and questionnaire data on aggression subtypes available (Brain Imaging Genetics, BIG, N = 501) to study their effect on reactive and proactive subtypes of aggression. Our meta-analysis identified two genes, MECOM and AVPR1A, significantly associated with both aggression risk and nucleus accumbens (MECOM) and amygdala (AVPR1A) brain volume. Subsequent in-depth analysis of these genes in healthy adults (BIG), including sex as an interaction term in the model, revealed no significant subtype-specific gene-wide associations. Using cross-trait meta-analysis of brain measures and psychiatric phenotypes, this study generated new hypotheses about specific links between genes, the brain and behavior. Results indicate that MECOM and AVPR1A may exert an effect on aggression through mechanisms involving nucleus accumbens and amygdala volumes, respectively.

#### Keywords: aggression, genetics, MRI, brain imaging, neurobiology

#### Edited by:

Nelly Alia-Klein, Icahn School of Medicine at Mount Sinai, United States

#### Reviewed by:

Zhen Yuan, University of Macau, China Sara Palumbo, Università degli Studi di Pisa, Italy

#### \*Correspondence: Janita Bralten janita.bralten@radboudumc.nl

†These authors share last authorship.

Received: 01 December 2017 Accepted: 15 March 2018 Published: 03 April 2018

#### Citation:

van Donkelaar MMJ, Hoogman M, Pappa I, Tiemeier H, Buitelaar JK, Franke B and Bralten J (2018) Pleiotropic Contribution of MECOM and AVPR1A to Aggression and Subcortical Brain Volumes. Front. Behav. Neurosci. 12:61. doi: 10.3389/fnbeh.2018.00061

## INTRODUCTION

Aggression is a common but heterogeneous phenotype often associated with psychiatric disorders that may be harmful to others (Baron and Deborah, 1994; Miczek et al., 2002). The term covers a wide range of human behaviors, varying from verbal aggression and bullying to physical violence. Together, these behaviors have been associated with a large emotional and financial burden on society, while interventions typically still have small effects (McGuire, 2008; Bakker et al., 2017). To address aggression-related negative outcomes more successfully, a better understanding of the genes and neural mechanisms that control this behavior is essential.

Although heritability estimates differ as a function of the population and the type of aggression that is investigated, twin studies show that about 50% of the variance in aggression can be explained by genetic influences, implicating a role for genetics in the development of aggressive behavior (Tuvblad and Baker, 2011; Veroude et al., 2016). Despite this considerable heritability of aggression, the identification of specific genetic risk factors has been difficult. One factor complicating gene-finding is the largely polygenic nature of aggression. While some monogenic disorders leading to aggression phenotypes do exist (the most well-known example perhaps being Brunner syndrome; Brunner et al., 1993), multiple genetic variants, each with a small effect size, contribute to the aggression phenotype in most individuals. Because of the hypothesized polygenic model of multiple common variants with small effects underlying aggression, studies have investigated the role of these common variants by conducting association studies. Next to early candidate genetic approaches relying on a priori biological hypotheses, several genome-wide hypothesis-generating approaches to gene finding have now also been conducted. The main focus of candidate gene studies of aggression has been on genes related to brain neurotransmitter function, in particular to serotonergic and dopaminergic genes, and on genes related to neuroendocrine signaling, like sex-steroid receptors and stress-related circuitry (Iofrida et al., 2014; Fernàndez-Castillo and Cormand, 2016; Waltes et al., 2016). Genome-wide association studies (GWAS) of aggression have investigated a wide range of aggression related phenotypes. Interestingly, two genes showed evidence for association based on more than one GWAS, NFKB1 and A2BP1. NFKB1 encodes the nuclear factor of kappa light polypeptide gene enhancer in B-cells 1, a transcription regulator involved in axonal regeneration and degeneration. A2BP1 (also called RBFOX1) encodes the RNA binding protein, fox-1 homolog (C. elegans) 1, a neuron-specific RNA splicing factor that regulates the expression of large genetic networks during early neuronal development (Fernàndez-Castillo and Cormand, 2016). Recently, a large-scale GWAS meta-analysis was conducted within the framework of the early genetics and lifecourse epidemiology (EAGLE) consortium, including nearly 19,000 subjects. The researchers combined GWAS data on childhood and adolescent aggression from nine population-based cohorts, and found suggestive evidence of association for a region on chromosome 2, near a gene involved in the regulation of excitatory synapse development (Pappa et al., 2015). While most other GWASs of aggression were relatively small-scaled, top-finding of these studies together with bioinformatics approaches have highlighted the importance of neurodevelopmental and synaptic plasticity genes for aggression risk (Fernàndez-Castillo and Cormand, 2016).

Investigation of the neural correlates of aggression has highlighted the involvement of several brain phenotypes in aggression. The main forms of cognitive dysfunction that have been recognized in the context of aggression are decreased empathy, an increased acute threat response and impaired decision-making. Different neural systems are thought to underlie these cognitive impairments. In short, the main neural substrates of empathic processing are thought to be the ventromedial prefrontal cortex (vmPFC) and the amygdala, while the acute threat response is mediated by the amygdalahypothalamus-periaqueductal gray neural system, and poor decision-making in individuals with aggressive behavior has been related the striatum and vmPFC (Blair et al., 2016). Imaging studies point towards an important role for subcortical brain regions in the neurobiology of aggressive phenotypes (Siever, 2008). Of specific interest in the context of aggression are the amygdala and the striatal subregions nucleus accumbens and caudate nucleus (Blair et al., 2016). The amygdala has been strongly linked to aggression through its role in emotion processing and threat reactivity (Fusar-Poli et al., 2009; Mobbs et al., 2010). A large number of studies have reported differences in the size of the amygdala between aggressive and comparison subjects, predominantly volume reductions (Sterzer et al., 2007; Fairchild et al., 2013; Zhang et al., 2013; Pardini et al., 2014; Wallace et al., 2014; Caldwell et al., 2015; Thijssen et al., 2015; Noordermeer et al., 2016). The striatum has been associated with aggression through its central role in reward sensitivity, processing of punishment and regulation of avoidance behaviors (Finger et al., 2008, 2011; Crowley et al., 2010; White et al., 2013). Impairments in these functions are thought to be the basis of poor decision-making in individuals with aggressive behavior (Fairchild et al., 2009; Blair et al., 2016). Both volume reductions and volume increases of the striatum have been related to aggressive phenotypes, especially for the caudate nucleus and nucleus accumbens (Nosarti et al., 2005; McAlonan et al., 2007; Ducharme et al., 2011; Schiffer et al., 2011; Fairchild et al., 2013; Cha et al., 2015). Brain volume has been shown to be heritable, and the Enhancing Neuro Imaging Genetics through Meta-Analysis (ENIGMA) consortium recently conducted a GWAS meta-analysis on volumes of seven subcortical brain structures and intracranial volume, to identify genetic variants that influence brain structure (Hibar et al., 2015). Identification of such genetic variants may help to uncover mechanisms underlying neuropsychiatric disorders.

Since both aggression risk and brain volumes are heritable, one may hypothesize that part of the genes contributing to aggression neurobiology do so by influencing aggression-related brain volumes. Identification of these genes may highlight specific pathways from gene to aggressive behavior via the brain. However, research into the underlying genetic and neurobiological mechanisms of aggression is complicated by the fact that aggression is a behaviorally and etiologically complex phenomenon. Efforts have been made to recognize different subtypes of aggressive behavior, presumed to differ in their underlying neurobiology. A frequently used system divides aggression into three subtypes; proactive aggression, reactive aggression due to external provocation or threat, and reactive aggression due to internal frustration (Dodge and Coie, 1987; Raine et al., 2006; Brugman et al., 2017; Smeets et al., 2017). Proactive aggression has been related to psychopathic traits and delinquent behavior (Cima et al., 2013). In this subtype, dysfunction in neural circuitry involving (venteromedial) prefrontal and striatal areas is thought to underlie observed difficulties with decision making and reinforcement learning, while a decreased responsiveness of the amygdala to distress cues is thought to reflect deficits in emotional empathy (Blair, 2013). Reactive subtypes of aggression, on the other hand, have been associated with impulsivity, anxiety and hostile interpretation bias (Bubier and Drabick, 2009; Brugman et al., 2015). It has been suggested that reactive forms of aggression are mediated by an overly responsive amygdalarelated threat response circuitry, which is dependent on regulation by cortical brain regions (Blair, 2013). Hence, different pathways to the maladaptive behavior are thought to exist. An added complication for the identification of genetic and neurobiological mechanisms underlying aggression are the marked sex-differences in aggressive behaviors. Sex-differences in aggression are pronounced with respect to prevalence and with respect to type of aggression displayed (Hill, 2002; Collett et al., 2003; Stephenson et al., 2014). For example, males are overrepresented among patients with aggressionrelated disorders such as conduct disorder (Hill, 2002), and are more prone to display physical aggression compared to females (Baillargeon et al., 2007). Identification of subtypeand sex-dependent genetic association may help in elucidating specific links from gene to brain to behavior.

Based on the above, the aim of the current study was two-fold. First, we sought to identify genes influencing both aggression and aggression-related brain volumes. To this end, we conducted genome-wide gene-based cross-trait meta-analyses of aggression and amygdala, nucleus accumbens and caudate nucleus volume, using GWAS meta-analysis data of two large-scale consortia (EAGLE, N = 18,988; ENIGMA2, N = 13,171). Second, we aimed to assess subtype- and sex-specificity of association for identified genes. For this, we conducted gene-wide association analyses with aggression subtypes for these genes in a population sample of healthy adults with available genome-wide genotyping and questionnaire data on aggression subtypes (Brain Imaging Genetics, BIG, N = 501).

### MATERIALS AND METHODS

### Samples

#### EAGLE

GWAS Meta-Analysis (GWAS-MA) data on aggression were obtained from the EAGLE consortium which investigated childhood aggressive behavior using nine population-based studies with a total of 18,988 subjects (mean age = 8.44 years, SD = 4.16; Pappa et al., 2015). Different well-validated parent-report questionnaires were used to assess aggressive behavior. Depending on study sample, aggressive behavior was assessed with the aggression scale of the Childhood Behavioral Checklist (CBCL), the conduct problem scale of the Strengths and Difficulties Questionnaire (SDQ), or comparable items in general questionnaires. Scores derived from SDQ and CBCL questionnaires were shown to be highly correlated and interchangeable for the assessment of children's behavior problems (Goodman and Scott, 1999). Genomic data were imputed to the HapMap reference panel (release 22) and comprised only samples of European ancestry. GWAS was performed for each cohort, followed by removal of single nucleotide polymorphisms (SNPs) with low minor allele frequency (<0.05) and imputation quality (RSQ < 0.3 or INFO < 0.4). Results were combined using the sample-size weighted z-score method as implemented in METAL (Willer et al., 2010), controlling for genomic inflation. Access to the summary statistics was requested through http://www.tweelingenregister.org/EAGLE. All sites involved in this study obtained approval from local research ethics committees, and written parental consent was obtained for all participants.

### ENIGMA2

GWAS-MA data on the aggression-related subcortical volumes of nucleus accumbens, amygdala and caudate nucleus were obtained from the ENIGMA consortium. The ENIGMA consortium conducted GWAS-MA on intracranial volume (ICV) and seven subcortical brain volumes, to identify common genetic variants contributing to volume differences. They used MRI brain scans and genome-wide genotype data of 13,171 subjects of European ancestry from 28 cohorts (discovery sample). Brain scans were examined and processed at each site following a standardized protocol. Subcortical volumes had been adjusted for ICV to identify specific genetic contributions to individual volumes. Genomic data comprised only European samples and were imputed to the 1000 Genomes, v3 phase1 reference panel using MaCH for phasing and minimac for imputation (Fuchsberger et al., 2015). GWAS was performed at each site, and SNPs with an imputation score of RSQ < 0.5 and minor allele count <10 were removed. Results were combined using an inverse-variance-weighted model as implemented in the software package METAL (Willer et al., 2010), controlling for genomic inflation. Further details of the original analysis can be found in Hibar et al. (2015). Access to the summary statistics of ENIGMA was requested through the ENIGMA website<sup>1</sup> . All sites involved in this study obtained approval from local research ethics committees or Institutional Review Boards, and all participants gave written informed consent.

#### BIG

To assess subtype-specific association of identified genes and for mediation analysis, data from the BIG study was used. This study was conducted at the Donders Institute for Brain, Cognition and Behavior (Franke et al., 2010), and consists of self-reported

<sup>1</sup>http://enigma.ini.usc.edu/download-enigma-gwas-results/

healthy adults who participated in smaller-scale imaging studies at the institute. Participants gave consent to use their acquired brain data, donated saliva and performed online testing. In the current study, a sub-sample of 501 subjects with available Reactive Proactive Questionnaire (RPQ) data (Raine et al., 2006), genome-wide genotype data, and structural MRI data was used (age range 18–45 years, 215 male/286 female).

All participants were of Caucasian descent and were screened using a self-report questionnaire for the following exclusion criteria before study participation: a history of somatic disease potentially affecting the brain, current or past psychiatric or neurological disorder, medication (except hormonal contraceptives) or illicit drug use during the past 6 months, history of substance abuse, current or past alcohol dependance, pregnancy, lactation, menopause and magnetic resonance imaging contraindications (Gerritsen et al., 2012). All participants gave written informed consent, and the study was approved by the regional ethics committee (Commissie Mensgebonden Onderzoek/CMO).

### Behavioral and Genetic Measures in BIG

### Aggression Questionnaire

The RPQ was used to assess subtypes of aggression in the BIG study (Raine et al., 2006). The RPQ is a self-report questionnaire consisting of 23 items. For each item, subjects are asked to indicate, how often they have engaged in a given type of behavior, like ''had temper tantrums''. Items are rated on a three-point Likert scale (''never'' = 0, ''sometimes'' = 1, ''often'' = 2). Responses were summed to yield the three factors that best described the RPQ in earlier exploratory factor analysis (Brugman et al., 2017; Smeets et al., 2017) as well as in the current sample (van Donkelaar et al., 2017): ''proactive aggression'' (range 0–12), ''reactive aggression due to internal frustration'' (range 0–9), and ''reactive aggression due to external provocation or threat'' (range 0–10). RPQ proactive aggression scores were dichotomized into high- and low-scoring (score ≥ 2 and score ≤ 1, respectively), because of a highly positively skewed distribution in both males and females (Supplementary Figure S1). An overview of RPQ items can be found in Supplementary Table S1.

### Genotyping and Imputation

Genetic analyses for the BIG study were carried out at the Department of Human Genetics of the Radboud university medical center. Saliva samples were collected using Oragene kits (DNA Genotek, Kanata, Canada), and genomic DNA was extracted as specified by the manufacturer. Genome-wide genotyping was performed on three different genotyping platforms; Affymetrix Genome-Wide Human SNP Array 6.0 (Affymetrix Inc., Santa Clara, CA, USA), Infinium PsychArray-24 v1.1 BeadChip<sup>2</sup> , and Infinium OmniExpress-24 array<sup>2</sup> . Quality control steps and imputation were performed using the Ricopili Rapid Imputation Consortium Pipeline<sup>3</sup> . Pre-imputation quality control included pre-filtering of SNPs with call rate <0.95, filtering of individuals with a genotyping rate <0.98 or inbreeding coefficient >0.02, filtering of SNPs with a call rate <0.98 or Hardy-Weinberg p-value < 1e-06, removal of invariant SNPs, and removal of population ancestry outliers. SHAPEIT<sup>4</sup> and IMPUTE2 (Howie et al., 2009) software were used for haplotype phasing and imputation with 1000 Genomes Phase 3.v5a reference data. For the current study, best-guess genotypes were inferred with a minimum probability threshold of 0.8. Post-imputation quality control included a strict SNP imputation quality threshold ≥0.8, removal of duplicated and related individuals (pi hat > 0.25), removal of individuals with a call rate below 95%, and removal of SNPs with a call rate below 95%, a minor allele frequency of less than 1%, or failing the Hardy-Weinberg equilibrium test at a threshold of p ≤ 1e-6.

### Analyses

### Genome-Wide Gene-Based Cross-Trait Meta-Analyses

We first conducted genome-wide cross-trait meta-analyses of aggression and three different aggression-related brain volume measures (amygdala, nucleus accumbens and caudate nucleus). We used summary statistic data of two large-scale GWAS of: (1) aggressive behavior in children and adolescents (EAGLE, N = 18,988); and (2) MRI-based volume measures of the aggression-relevant brain regions (ENIGMA2, N = 13,171). First, four separate genome-wide gene-based analyses with a 50 kb flanking region around genes were performed for the summary statistic data of aggression and the volumes of amygdala, nucleus accumbens and caudate nucleus using MAGMA v1.06 (de Leeuw et al., 2015). SNPs were mapped onto genes using 1000 Genomes Phase 3 reference data followed by computation of gene p-values by aggregating the effect of common variants within the genes. Next, fixed-effects meta-analyses were performed of aggression with amygdala volume, aggression with caudate nucleus volume and aggression with nucleus accumbens volume, using the weighted Stouffer's Z method as implemented in MAGMA software. Results were considered significant if they reached the Bonferroni-corrected P-value-threshold for testing 18,310 genes (p < 2.731e-6). Significant genes with stronger association p-values in meta-analysis, compared to the separate analyses of aggression and brain volume, were reported and selected for further investigation.

### Gene-Wide Association Analyses for Aggression Subtypes

Gene-wide association of selected genes with three subtypes of aggression, reactive aggression due to internal frustration, reactive aggression due to external provocation or threat and proactive aggression, was assessed. One phenotypic outlier (>4 SD) was removed for all analyses. Gene-wide analyses again included a 50 kb flanking region. Three base gene analysis models are available in MAGMA, each of them sensitive to different genetic architectures: Principal Component Analysis, mean-SNP and top-SNP models. For the current analysis, a multi-model approach was used, combining the results from the base analysis

<sup>2</sup>www.illumina.com

<sup>3</sup>https://sites.google.com/a/broadinstitute.org/ricopili/home

<sup>4</sup>https://mathgen.stats.ox.ac.uk/genetics\_software/shapeit/shapeit.html

models into an aggregate p-value. Separate analyses were run for subjects genotyped on the three different genotyping arrays. Age and four population components derived from multidimensional scaling analysis were included as covariates. Sex was included as an interaction term in the model, yielding a gene p-value for main and interaction effects combined (P\_full). This was followed by meta-analysis of the full model output of the three genotyping arrays, using the weighted Stouffer's Z method as implemented in MAGMA software (de Leeuw et al., 2015). To protect against type I error, the conventional significance threshold (0.05) was lowered to correct for multiple comparisons, testing two genes and using the effective number of independent tests (Meff, see Li and Ji, 2005) for the aggression subtype outcomes, calculated to be 2.5 (taking into account the correlation matrix of the three aggression measures). Hence, results were considered significant if they reached a significance threshold of 0.01. Given the strong sex effects in aggression research (Collett et al., 2003; Stephenson et al., 2014) explorative analysis of sex are reported for reasons of completeness.

### RESULTS

### Genome-Wide Gene-Based Cross-Trait (Aggression-Brain) Meta-Analyses

The MDS1 and EVI1 complex locus gene (MECOM) was significantly associated with the cross-trait construct of aggression and nucleus accumbens volume (p = 4.94e-07), and the Vasopressin Receptor 1A gene (AVPR1A) showed significant association with the cross-trait construct of aggression and amygdala volume (p = 1.64e-06). Both associations were more significant compared to the separate analyses of aggression and the respective brain volume (**Table 1**).

### Gene-Wide Association Analyses for Aggression Subtypes

The general characteristics of the 501 participants from the BIG sample included in the aggression subtype analysis are shown in Supplementary Table S2. Gene-wide association analyses with three aggression subtypes were conducted for AVPR1A and MECOM to identify gene-behavior relationships, including sex as an interaction term in the model. Association of the AVPR1A gene with the score for reactive aggression due to external provocation or threat was nominally significant (P\_full = 0.016), but did not reach the corrected significance threshold (**Table 2**). Explorative analyses in males and females separately showed a stronger contribution to this effect for males (P\_males = 0.037; P\_females = 0.517). The MECOM gene was not associated with any of the aggression subtypes in the population sample.

### DISCUSSION

Using cross-trait meta-analyses of gene-wide association statistics, this study identified two genes as potentially pleiotropic loci for aggression and aggression-related subcortical brain volumes. We identified MECOM as a gene potentially contributing to both aggression risk and nucleus accumbens volume, and we identified AVPR1A as a gene potentially contributing to both aggression risk and amygdala volume. Subsequently, we investigated subtypespecific and sex-dependent association of these genes in an independent sample. No associations with aggression subtypes could be confirmed in this sample, although sex-dependent association of AVPR1A with reactive aggression due to external provocation/threat reached the nominal significance level.

The MDS1 and EVI1 complex locus gene (MECOM) codes for a protein known as transcriptional regulator and oncoprotein (Yoshimi and Kurokawa, 2011). MECOM plays an important role in early development, with Evi1 homozygous mutant mouse embryos dying approximately 10.5 days post coitum showing disrupted cell proliferation and disrupted development of cardiovascular and neural systems (Hoyt et al., 1997). The association p-value for MECOM in the study of aggression improved in the cross-trait meta-analysis of this behavioral trait with nucleus accumbens volume. According to our hypothesis, this might indicate that it exerts its effect on aggression through mechanisms involving the nucleus accumbens. However, we did not observe the association of MECOM with aggression when investigating specific subtypes of aggression in our own, smaller sample of adults. To our knowledge, little is known about MECOM in relation to psychiatric behavioral phenotypes so far, and future work needs to investigate this association in more detail.

Our study provides further evidence for a role of candidate gene AVPR1A in aggression. The Arginine Vasopressin Receptor 1A gene (AVPR1A) codes for the primary receptor of AVP in the brain. AVP is a neuropeptide strongly implicated in complex social and emotional behaviors, including aggression, through a host of animal studies (Ebstein et al., 2010). Also in humans, AVP was shown to play a role in enhancing aggressive behavior. For example, evidence exists for a positive correlation between aggression and cerebro-spinal fluid AVP in humans (Coccaro et al., 1998). Additional evidence comes from


Displayed are genes showing genome-wide significant association in the cross-trait meta-analysis of aggression with an aggression-related brain volume. These genes show more significant association in meta-analysis compared to gene-wide association with aggression and brain volume phenotypes separately. <sup>∗</sup>Bonferroni-corrected P-value-threshold for testing 18,310 genes: p < 2.73e-6.


<sup>∗</sup>P-value for main and sex interaction effect combined. Chr, Chromosome. N SNPs, number of single nucleotide polymorphisms.

genetic association studies. The original aggression GWAS-MA that we used for cross-trait meta-analysis reported gene-wide association of the AVPR1A gene with childhood aggression (P = 1.61e-03), using VEGAS gene-based analysis and correcting for 21 candidate genes tested (Pappa et al., 2015). Using the MAGMA multi-model approach, which has the advantage of yielding a more even distribution of statistical power and sensitivity for a wider range of different supposed underlying genetic architectures compared to other methods (de Leeuw et al., 2015), an even lower p-value was reported in the current study. The cross-trait meta-analysis of aggression and amygdala volume resulted in gene-wide genome-wide significance. Other human genetic association studies of variants in the AVPR1A gene reported association with anger (Moons et al., 2014), gender-specific nominally significant association with pervasive aggression (Malik et al., 2014), but no association in an early study of antisocial traits (Prichard et al., 2007). While we report nominally significant subtype-specific gene-wide association of the AVPR1A gene with reactive aggression due to external provocation or threat in a sample of healthy adults, this finding is not significant after correcting for multiple testing. We can speculate that association of AVPR1A to this subtype of aggression, which specifically measures social responses to threat and provocation by others or actions of self-defense in response to others, would be in line with existing data highlighting the importance of AVP in social context and social communication. Vasopressin signaling is thought to be an important determinant of the intensity and range of social responses displayed in different social situations (Albers, 2012). For example, AVP can alter the extent to which social stimuli are threatening, by modulating sensory information (Thompson et al., 2006). Sex-dependent effects of AVP have also been found in animal research, finding opposite effects of both vasopressin and V1a receptor blockade on aggressive behavior. For example, AVP injection increases aggression in male hamsters but decreases it in females, while injection of V1a receptor antagonists has the opposite results (Gutzler et al., 2010). This data suggests that there may be a difference between males and females in the effects of vasopressin signaling on aggression. Less is known about sex differences in V1a receptor expression. Nevertheless, research in a number of species indicates that receptor distribution might vary in a sex-dependent manner as well (reviewed in Albers, 2015). Moreover, gonadal hormones can modulate the expression of vasopressin and vasopressin receptors (e.g., Dubois-Dauphin et al., 1994; Young et al., 2000), thus partly explaining sex differences in the vasopressin system. Hence, reducing phenotypic heterogeneity and taking into account sex-related heterogeneity may facilitate the search for genes involved in the etiology of aggression.

Our cross-trait meta-analysis results indicate that amygdala volume might serve as a (proxy for related) mechanisms through which the vasopressin receptor could influence aggressive behavior, and that MECOM may exert its effect on aggression through mechanisms involving the nucleus accumbens. Thus, we provide specific hypotheses about shared genetic risk and generated specific hypotheses about links from gene to brain to behavior for future studies to focus on. It is often assumed in imaging genetics research that genetic risk for a neurodevelopmental disorder passes through the brain phenotype to behavior. However, another possibility is that genetic factors influencing behavior also influence the brain in a way that is independent of the behavioral phenotype of interest (Kendler and Neale, 2010). Different mediation analysis methods have been developed to explore these relationships between genetic variants, brain phenotypes and behavior. Historically, the causal steps approach has been popular in mediation analysis. It uses a set of tests of significance for each path in a causal system, although a more powerful way to determine significance of a mediated effect is bootstrapping (Hayes and Scharkow, 2013). Other methods use causal modeling to describe the direction of association between different variables, using exploratory structure learning algorithms to find conditional independencies, which makes it possible to infer parts of the structure of a structural equation model (SEM) and make predictions about causation (Sokolova et al., 2017). Only a few imaging genetics studies have investigated this issue of causality earlier. Those studies showed that only part of the brain regions showing genotype effects actually do mediate between genetics and the behavior under study (Sokolova et al., 2015; van der Meer et al., 2015), proving the importance of such multilevel investigations to elucidate the biological mechanisms, by which brain alterations may be involved in aggression etiology. Currently, available methods for making causal inferences focus on SNP-level investigations, and future studies would benefit from the development of approaches for aggregating common genetic variant data to gene- or gene-set-level in mediation frameworks.

This study has several strengths and limitations. The current study used the largest data-sets available to investigate pleiotropic genetic factors for aggression and brain volumes at gene-level. Cross-trait meta-analysis of brain measures and psychiatric phenotypes is a useful way of detecting shared genetic risk and generating new hypotheses about specific links between genes, the brain and behavior (Franke et al., 2016). We were also able to use a well-phenotyped population cohort to investigate specific subtypes of aggression. While the limitations of using saliva in this cohort as a source of DNA for genotyping may be particularly relevant for the analysis of copy number variants (Fabre et al., 2011), replication of our work using DNA isolated from other sources would further strengthen the confidence in our results.

The added value of using data from such smaller cohorts over large consortium based data lies in the possibility of in-depth phenotyping and reducing sources of heterogeneity that come with the use of pooled data-sets. Nevertheless, we did not detect significant sex- and subtype-specific associations of MECOM and AVPR1A with aggression in this sample. Possibly, we still lacked power to detect small genetic effects due to sample size, or effects might be larger in non-healthy samples. Hence, future studies are needed to further investigate the specific hypotheses about genebrain-behavior relationships generated by the current study. As this study only investigated selected subcortical MRI measures, future work should also be extended to include cortical regions as well as connectivity measures that have been shown to play a role in aggression (Meyer-Lindenberg et al., 2006).

In summary, we identified MECOM and AVPR1A as genes contributing to aggression risk in conjunction with nucleus accumbens and amygdala brain volume, respectively. Future studies may elucidate causality of gene-brain-behavior relationships. Comprehension of sex-specific physiological pathways associated with aggression subtypes is needed to enhance our understanding of the determinants of aggression, and only by understanding the mechanisms underlying different forms of aggression will we be able to develop effective treatment approaches and minimize the social costs of aggression.

### AUTHOR CONTRIBUTIONS

MMJD, MH, JKB, BF and JB contributed to the conception and design of the study. MMJD, IP and HT contributed to data analysis and/or interpretation. MMJD wrote the first draft of the manuscript. All authors contributed to manuscript revision, read and approved the submitted version.

### REFERENCES


### ACKNOWLEDGMENTS

This works was supported by the Netherlands Organization for Scientific Research (Nederlandse Organisatie voor Wetenschappelijk Onderzoek, NWO), i.e., the NWO Brain & Cognition Excellence Program (grant 433-09-229) and the Vici Innovation Program (grant 016-130-669 to BF). Additional support was received from the European Community's Seventh Framework Programme (FP7/2007–2013) under grant agreements n◦ 602805 (Aggressotype), n◦ 602450 (IMAGEMEND), and n◦ 278948 (TACTICS) as well as from the European Community's Horizon 2020 Programme (H2020/2014–2020) under grant agreements n◦ 643051 (MiND) and n◦ 667302 (CoCA). The work was also supported by grants for the ENIGMA Consortium (Foundation for the National Institutes of Health (NIH); grant number U54 EB020403) from the BD2K Initiative of a cross-NIH partnership.

This work was carried out on the Dutch national e-infrastructure with the support of SURF Cooperative.

The Enhancing NeuroImaging Genetics through Meta-Analysis (ENIGMA) Consortium provided summary statistics of the consortium findings to this project. The original publication of those findings as well as the list of contributing samples and people can be found on the ENIGMA website: http://enigma.ini.usc.edu.

The EArly Genetics and Lifecourse Epidemiology (EAGLE) consortium provided summary statistics of the consortium findings to this project. Contributing samples and people can be found in the original publication of those findings (Pappa et al., 2015).

### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fnbeh.2018.00 061/full#supplementary-material

problems. Neurosci. Biobehav. Rev. doi: 10.1016/j.neubiorev.2016.10.022 [Epub ahead of print].


with ADHD: multivariate analysis of morphometry and tractography. Neuropsychopharmacology 40, 1717–1725. doi: 10.1038/npp.2015.18


to predict human emotional reactions to stress. Emotion 14, 562–572. doi: 10.1037/a0035503


**Conflict of Interest Statement**: BF discloses having received educational speaking fees from Shire and Medice. JKB has been in the past 3 years a consultant to/member of advisory board of/and/or speaker for Janssen Cilag BV, Eli Lilly, Lundbeck, Shire, Roche, Novartis and Servier. He is not an employee of any of these companies, and not a stock shareholder of any of these companies. He has no other financial or material support, including expert testimony, patents, royalties.

The other authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 van Donkelaar, Hoogman, Pappa, Tiemeier, Buitelaar, Franke and Bralten. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Anger Modulates Influence Hierarchies Within and Between Emotional Reactivity and Regulation Networks

Yael Jacob1,2,3\*, Gadi Gilam2,3,4 , Tamar Lin2,3 , Gal Raz 1,2,5 and Talma Hendler 1,2,3,6 \*

<sup>1</sup>Sagol School of Neuroscience, Tel Aviv University, Tel Aviv, Israel, <sup>2</sup>Tel Aviv Center for Brain Functions, Wohl Institute for Advanced Imaging, Tel Aviv Sourasky Medical Center, Tel Aviv, Israel, <sup>3</sup>The School of Psychological Sciences, Tel Aviv University, Tel Aviv, Israel, <sup>4</sup>Systems Neuroscience and Pain Laboratory, Department of Anesthesia, Perioperative and Pain Medicine, School of Medicine, Stanford University, Palo Alto, CA, United States, <sup>5</sup>Film and Television Department, Tel Aviv University, Tel Aviv, Israel, <sup>6</sup>Sackler Faculty of Medicine, Tel Aviv University, Tel Aviv, Israel

Emotion regulation is hypothesized to be mediated by the interactions between emotional reactivity and regulation networks during the dynamic unfolding of the emotional episode. Yet, it remains unclear how to delineate the effective relationships between these networks. In this study, we examined the aforementioned networks' information flow hierarchy during viewing of an anger provoking movie excerpt. Anger regulation is particularly essential for averting individuals from aggression and violence, thus improving prosocial behavior. Using subjective ratings of anger intensity we differentiated between low and high anger periods of the film. We then applied the Dependency Network Analysis (DEPNA), a newly developed graph theory method to quantify networks' node importance during the two anger periods. The DEPNA analysis revealed that the impact of the ventromedial prefrontal cortex (vmPFC) was higher in the high anger condition, particularly within the regulation network and on the connections between the reactivity and regulation networks. We further showed that higher levels of vmPFC impact on the regulation network were associated with lower subjective anger intensity during the high-anger cinematic period, and lower trait anger levels. Supporting and replicating previous findings, these results emphasize the previously acknowledged central role of vmPFC in modulating negative affect. We further show that the impact of the vmPFC relies on its correlational influence on the connectivity between reactivity and regulation networks. More importantly, the hierarchy network analysis revealed a link between connectivity patterns of the vmPFC and individual differences in anger reactivity and trait, suggesting its potential therapeutic role.

Keywords: emotion regulation, fMRI, graph theory network analysis, ventromedial prefrontal cortex

### INTRODUCTION

Anger is an omnipresent human phenomenon, with aggression being its prototypical behavioral expression (Averill, 1983; Berkowitz and Harmon-Jones, 2004; Rosell and Siever, 2015). While having an adaptive role in survival, anger may lead to unnecessary violence as well as to personal harmful consequences for health and wellbeing (Johnson, 1990; Chang et al., 2002). Anger regulation is therefore crucial in order to foster prosocial behaviors and avoid negative ramifications

#### Edited by:

Klaus A. Miczek, Tufts University, United States

### Reviewed by:

Klaus Mathiak, RWTH Aachen Universität, Germany Xin Di, New Jersey Institute of Technology, United States

#### \*Correspondence:

Yael Jacob yaelja@gmail.com Talma Hendler hendlert@gmail.com

Received: 27 November 2017 Accepted: 15 March 2018 Published: 04 April 2018

#### Citation:

Jacob Y, Gilam G, Lin T, Raz G and Hendler T (2018) Anger Modulates Influence Hierarchies Within and Between Emotional Reactivity and Regulation Networks. Front. Behav. Neurosci. 12:60. doi: 10.3389/fnbeh.2018.00060 for the individual (Davidson et al., 2000; Gilam and Hendler, 2017). Yet despite extensive psychological work on possible regulation strategies and their related processes the neural mechanism that underlies anger regulation is still debated. Unveiling the neural mechanism of anger regulation could serve future diagnosis and therapy in psychiatry, but also improve adaptive prosocial behavior and well-being of individuals prone to anger inducing incidents. In general, emotion regulation refers to a range of strategies humans can voluntarily or involuntarily utilize to modulate emotional experiences, their intensity and expression (Phillips et al., 2008; Gyurak et al., 2011; Etkin et al., 2015). Indeed, humans' neural circuitry is embedded with networks subserving the regulation of emotional reactivity (e.g., Etkin et al., 2015). To date, it is however, unclear if and how information flow within and between these networks impacts emotion regulation in general and anger regulation in particular.

Recent accounts point to widespread neural activations involved in emotional reactivity and regulation (Ochsner et al., 2012; Etkin et al., 2015). Accordingly, emotional reactivity (e.g., the perception and generation of a threat response) mainly involves subcortical areas, such as the amygdala and periaqueductal gray (PAG), but also cortical regions such as the insula and dorsal anterior cingulate cortex (dACC). Some of these regions are involved in the detection of salient stimuli (Seeley et al., 2007) and others in their rapid evaluation (LeDoux, 1996). Explicit and effortful regulatory processes such as reappraisal (i.e., altering the semantic representation of an emotional stimulus in order to change its affective impact) implicate regions that are typically involved in executive control, such as the dorsolateral PFC (dlPFC)/middle frontal gyrus (MiFG) and superior parietal lobule (SPL), but also inhibition related regions such as the ventrolateral PFC (vlPFC)/inferior frontal gyrus (IFG) and pre-supplementary motor area (pre-SMA). In contrast, implicit and automatic regulation processes (e.g., fear extinction learning in context of a fear conditioning paradigm) has been associated with the ventromedial prefrontal cortex (vmPFC), a major player in self-referential, subjectivevaluation and visceromotor control processes (Roy et al., 2012; Hiser and Koenigs, 2018).

Current postulations suggest that successful emotion regulation relies on interactions between the emotion reactivity and regulation systems (e.g., Kober et al., 2008; Lindquist et al., 2012; Etkin et al., 2015). Nevertheless the complex relationships within and between different parts of the proposed systems with regard to the dynamic unfolding of an emotional episode such as anger, remains largely unclear. Indeed, to our knowledge only two studies focused on functional connectivity patterns in regards to anger and its potential regulation. Resting-state fMRI data revealed a positive association between amygdala-orbitofrontal connectivity and the tendency to try to control expressions of anger (Fulwiler et al., 2012). However, this study was limited since it examined only the amygdala's connectivity with the rest of the brain and did not do so during an actual experience of anger. Moreover, it implemented standard functional connectivity methods that do not address the influencing relationships between brain regions (Hutchison et al., 2013). Effective connectivity methods (Friston, 1994) such as dynamic causal modeling (DCM) and Granger causality analysis (Goebel et al., 2003) were developed to assess such influences. Applying DCM to analyze the brains of individuals watching angry actors demonstrated an increase in the ipsilateral forward connection from the right insula to the right superior temporal gyrus and suppression of the same contralateral connection (Mazzola et al., 2016). Here again participants were not reporting about their anger experience, rather passively observing angry actions of others. Moreover, due to DCM's inherent limitation, the analysis was conducted on a small subset of regions in the emotion reactivity and regulation networks, precluding a fully comprehensive account of the dynamics within and between these networks.

To examine network organization and hierarchy related to implicit emotion regulation, in this fMRI study participants passively viewed an anger-provoking movie. Previous studies conducted on this dataset have shown that during the viewing of this movie excerpt stronger functional connectivity between the salience network and the limbic medial amygdala network was associated with more intense ratings of emotional experience (Raz et al., 2016), and that activity in participants' limbic regions was highly synchronized across subjects using inter-subject correlation analysis (Lin et al., 2017).

Here we applied our recent graph-theory based Dependency Network Analysis (DEPNA; Jacob et al., 2016) on data acquired while individuals viewed anger-provoking movie excerpts. The DEPNA analysis applied computes each brain region's importance in a given network, according to its effect on the correlations between all other pairs of regions. In this way, the DEPNA is able to capture the network's hierarchy of influence between an a priori defined set of regions during different task conditions, and thus to depict the within and between network hierarchies (Jacob et al., 2016; see also Supplementary Material 1 for details on the method). Here we applied DEPNA to assess the hierarchy within and between the emotional reactivity and regulation networks, with respect to changes in self-reported anger intensities induced by a political documentary movie excerpt.

Based on emotional ratings provided by participants following the passive viewing of the movie excerpt in the scanner, DEPNA indices were calculated separately during two epochs in the movie defined as high and low anger experience. The anger rating indicated that the selected movie was highly effective in inducing a dynamic anger experience in which anger intensity culminates just before the end of the scene (**Figure 1**). To note, participants were passively watching the movie excerpt and were not explicitly instructed to apply a specific emotion regulation strategy (as commonly done during explicit emotion regulation paradigms, e.g., Buhle et al., 2014). We therefore assume that implicit emotion regulation was spontaneously applied, designated by the level of emotional experience which unfolded throughout the cinematic scene. This assumption stems from work in psychology showing that there is variability among people in how they rate high intensity stimuli and it has been suggested to be partly related to their regulation tendency (Gross, 1998; Gross and John, 2003).

Based on the previously described idea of dual network involvement and the relation between emotion intensity and tendency for regulation, we therefore hypothesized that the reported increase in anger intensity would involve modulations in connectivity within and between neural networks that are implicated in emotion reactivity (including the insula and the amygdala), and executive control (including IFG and MiFG; Hypothesis 1). Furthermore, the vmPFC which is considered a major player in the implicit emotional regulation network (Etkin et al., 2015) is expected to be particularly influential on network organization during high anger moments (Hypothesis 2). Lastly, difference in network hierarchy between the anger conditions was expected to correlate with both self-reported angry feelings during the movie (i.e., anger state) and anger and emotional regulation tendencies (i.e., anger trait; Hypothesis 3).

### MATERIALS AND METHODS

### Participants

Valid data were collected from 74 subjects (19.51 ± 1.45 years, males only, plus 27 dropouts) with no known history of neurological or psychiatric disorders and 12 years of education, who volunteered to participate in this study. The study was carried out in accordance with the recommendations of the Declaration of Helsinki and approved by the ethics committee of the Tel Aviv Sourasky Medical Center with written informed consent from all subjects. Part of the dataset was used in previous publications (see Raz et al., 2016; Lin et al., 2017).

### fMRI Data Acquisition

All scans were obtained by a GE 3T Signa Excite echo speed scanner with an 8-channel head coil located at the Wohl Institute for Advanced Imaging at the Tel-Aviv Sourasky Medical Center. Structural scans included a T1-weighted 3D axial spoiled gradient echo (SPGR) pulse sequence (TR/TE = 7.92/2.98 ms, slice thickness = 1 mm, flip angle = 15◦ , voxel size = 1 mm<sup>3</sup> , FOV = 256 × 256 mm). Functional whole-brain scans were performed in interleaved order with a T2<sup>∗</sup> -weighted gradient echo planar imaging pulse sequence (time repetition (TR)/TE = 3000/35 ms, flip angle = 90◦ , voxel size = 1.56 × 1.56 × 3 mm<sup>3</sup> , FOV = 200 × 200 mm, slice thickness = 3 mm, 39 slices per volume). Active noise canceling headphones (Optoacoustics) were used.

### fMRI Data Preprocessing

Preprocessing was performed using Brain Voyager QX version 2.4. Head motions were detected and corrected using trilinear and sinc interpolations respectively, applying rigid body transformations with three translation and three rotation parameters. The data were high pass filtered at 0.008 Hz. Spatial smoothing with a 6 mm FWHM kernel was applied. To avoid the confounding effect of fluctuations in the whole-brain BOLD signal, for each TR, each voxel was scaled by the global mean at that time point. Anatomical SPGR data were standardized to 1 × 1 × 1 mm and transformed into Talairach space. SPGR images were then manually co-registered with the corresponding functional maps.

### Anger Inducing Film Excerpt

All subjects underwent fMRI while passively viewing an excerpt from an Israeli documentary film ''Avenge But One of My Two Eyes'' (Mograbi, 2005), that was previously used in our lab to induce anger (Raz et al., 2016; Lin et al., 2017). The excerpt introduces a fierce political confrontation between the director and soldiers at a checkpoint in the West Bank. The duration of the clip was 5:21 min, and the display was preceded and followed by a 30 s epoch during which the participants passively gazed at an all-black slide.

### Continuous Self-Reporting of State Anger Intensity

A continuous rating of anger intensity was obtained retrospectively (**Figure 1A**): 70 of the 74 participants watched the movie excerpt for a second time once again outside the scanner and continuously reported on shifts in intensity of anger experienced during the first viewing in the scanner. Retrospective rating was performed over online recording during the fMRI session to avoid the interference of deliberate introspection with the anger experience. The resemblance between the retrospective and online rating was validated in a previous study (Raz et al., 2012). In this study, participants watched the same movie three times. During the second and third viewings, the participants were asked to provide a continuous retrospective report on their emotional experience during the first viewing. The average correlation between these two reports (test-retest reliability) was 0.93. In a complementary test, emotional ratings obtained during first and second (retrospective report) viewings were compared. The correlation between these ratings (construct validity) was 0.89.

The data were acquired via designated in-house software. By using the computer-mouse, subjects indicated changes in their felt intensity of anger in relation to a vertical scale continuously presented on the screen. The scale included seven levels of anger, from neutral to very strong, each containing 3◦ of change (21◦ in total). The feedback was sampled at a rate of 10 Hz. The individual overall average intensity was computed by the area under the curve (AUC) of the continuous anger intensity. The higher the AUC the more anger the subject reported over the entire movie excerpt.

Two emotional states of high and low anger were differentiated based on the continuous rating of anger intensity (**Figure 1B**). Median ratings across all subjects were calculated for all data points. Based on these values, a Matlab script identified one high anger epoch and one low anger epoch. The pair of epochs was defined so that the epochs: (1) were equal in length; (2) significantly differed in median ratings across subjects at p < 0.01 in the Wilcoxon signed rank test; and (3) were as long as possible (starting with half of the time points and iterated downwards). This procedure yielded a unique solution of two intervals of 132 s in which the median anger rating across all participants reached its maximum (195–327 s; hereby termed high-anger period), and minimum (36–168 s; hereby termed low-anger period).

### Trait Measure of Anger

The gold standard state-trait anger expression inventory-2 (STAXI-2; Spielberger et al., 1999) was used to assess trait anger, calculated as the sum score in 10 items rated on a 4-point frequency scale from 1 (not at all) to 4 (very much). Subjects are asked to report on the frequency of angry feelings experienced over time. Trait anger measures were assessed from 54 out of the 74 participants.

### Trait Measures of Emotional Regulation

To assess the habitual tendency to use emotion regulation strategies we used the emotion regulation questionnaire (ERQ; Gross and John, 2003). This 10-item questionnaire assesses individual differences in the use of two emotion regulation strategies: reappraisal and suppression. Items are measured on a seven-point Likert scale, from 1 (strongly disagree) to 7 (strongly agree). Emotional regulation traits were assessed in 54 of the 74 participants.

### Emotional Reactivity and Regulation Networks of Interest

The networks of interest were adopted from Etkin et al. (2015). The emotion regulation network consisted of eight regions of interest and the coordinates of these regions were extracted from a meta-analysis on cognitive reappraisal (Buhle et al., 2014). The coordinates for the vmPFC, which was added to the regulation network, and the reactivity network, consisted of six regions extracted from a meta-analysis of neuroimaging studies of emotion (Kober et al., 2008). For details regarding these regions of interest (ROIs) see **Table 1**. For each ROI we created a spherical mask (radius = 3 mm) centered on the peak x, y, z Talairach coordinates. The averaged BOLD signal (time series) was then extracted for each ROI mask image and each subject.

### Dependency Network Analysis (DEPNA)

During each of the low and high movie periods we applied the DEPNA method to probe relationships of influence between the network nodes. The DEPNA and its implementation to fMRI was previously described extensively (Jacob et al., 2016). The DEPNA steps needed to calculate the networks' ROIs influence are described in **Figure 2**. Further details on the DEPNA features, characteristics and interpretations are described in Supplementary Table S1 in the Supplementary Material 1.

Briefly, the ROI-ROI correlations were calculated by Pearson's formula (Rodgers and Nicewander, 1988). First we normalized the correlation coefficients by using a Fisher Z transformation. Next, we used the resulting normalized ROI correlations to compute partial correlations (Baba, 2004; **Figure 2A**). The partial correlation coefficient is a statistical measure indicating how a third variable affects the correlation between two other variables (Shapira et al., 2009). The partial correlation between nodes i and k with respect to a third node j—PC (i,k|j) is defined as:

$$PC(i,k|j) = \frac{C(i,k) - C(i,j)C(k,j)}{\sqrt{[1 - C^2(i,j)][1 - C^2(k,j)]}} \tag{1}$$

Where C(i,j), C(i,k) and C(k,j) are the ROI-ROI correlations. The relative effect of the correlations C(i,j) and C(k,j) of node j on the correlation C(i,k) (Kenett et al., 2010; **Figure 2A**), is given by:

$$d(i,k|j) \equiv C(i,k) - PC(i,k|j) \tag{2}$$

This quantity is large only when a significant fraction of the correlation between nodes i and k can be explained in terms of node j. To avoid cases where we sum over positive and negative influences, we reset all negative values to zero.

We then define the total influence of node j on node i, or the dependency D(i,j) of node i on node j to be (**Figure 2B**):

$$D(i,j) = \frac{1}{N-1} \sum\_{k \neq j}^{N-1} d(i,k|j) \tag{3}$$

As defined, D(i,j) is a measure of the average influence of node j on the correlations C(i,k), over all nodes k. N is the number of TABLE 1 | High vs. low anger conditions t-test results.


<sup>∗</sup>p < 0.05, ∗∗q < 0.05, FDR corrected.

nodes in the network. The node activity dependencies define a dependency matrix D whose (i,j) element is the influence of node j on node i.

Next we sorted the nodes according to the system level influence of each node on the correlations between all other node pairs (**Figure 2C**). The system level ''Influencing Degree'' of node j is simply defined as the sum of the influence of node j on all other nodes i, that is:

$$\text{InfluenceDegree}\left(j\right) \tag{4}$$

$$\sum\_{1 \neq j}^{N-1} D(i, j) \tag{4}$$

The DEPNA ''Influencing Degree'' measure indicates the hierarchy of efferent (out-degree) impact of the node on the entire network (or sub-network). The higher this measure is, the greater its impact on all other connections in the network and the more likely it is to be generating information flow in the network.

To create network graph visualization we used the pairwise dependency connectivity matrix. A two-tailed t statistic was computed to compare the two conditions (e.g., high vs. low anger epochs). We then connect only pairwise ROIs with dependencies that were significantly different between the two conditions (p < 0.05 level) creating a simple graph visualization of the differences between the conditions across all subjects. The brain visualization of the graph was conducted with the BrainNet Viewer (Xia et al., 2013) 1 .

The DEPNA was computed for each subject for each period (i.e., high- and low-anger) resulting in an ''Influencing Degree'' for each region (**Figure 2C**). We then conducted a betweenperiods paired t-test for each region's ''Influencing Degree'' (total of 14 ROIs). The results were corrected for multiple comparisons

<sup>1</sup>http://www.nitrc.org/projects/bnv/

node dependencies define a dependency matrix D, whose (i, j) element is the influence of node j on node i. Step 3: (C) "Influencing Degree"—We then define the influences of node j as the sum of the influence D(i, j) of j on all other nodes i. The higher this measure the more this node influenced all other connections in the network. Step 4: (D) Graph Visualization –Each ROI is color-coded according to its influencing or influenced degrees. All pairwise ROIs with dependency elements D that are significantly different between two conditions (or groups) at the p < 0.05 level are plotted as edges. Each edge is color-coded according to the t-test sign as light or dark gray. The arrows represent the direction of influence. (E) Intra-Network influence—The influence within the sub-network is computed for each node as the sum of its influences on the nodes within its network. (F) Inter-Network influence—The influences between the sub-networks is calculated for two different options: (1) as the sum of the influences of a node from one network only on the connections within the second network; and (2) as the sum of the influences of a node from one network only on the connections between the first (kj) and the second (ki) networks. (G) Total inter-network—total influences between the sub-networks were computed as the sum of all inter-network influences from one network on the nodes within the second network.

using FDR (Benjamini and Hochberg, 1995) correcting for 14 tests.

In addition, we further investigated the intra and internetwork influence hierarchies for the reactivity and regulation networks as two separate sub-networks. The intra-network influence was computed for each network node's influence on the connections within its network (**Figure 2E**). We then conducted a between-periods paired t-test for each of the regulation network regions' (eight ROIs) intra-network influence degree. The results were corrected for eight multiple comparisons using FDR correction.

The inter-network influences were divided into two cases: (1) the sum of the influences of a node from one sub-network only on connections within the second sub-network (i.e., interinfluence 1, **Figure 2F**); and (2) the sum of the influences of a node from one sub-network only on connections between the two sub-networks (i.e., inter-influence 2, **Figure 2F**). Finally, the total network influence on the second network was calculated as the sum of the internetwork influences. This feature was calculated for each inter-network influence option separately. Next, for each network configuration (i.e., inter influence 1 and 2), we conducted a between-periods paired t-test for each region's inter-influence degrees (total of 14 ROIs). The results were corrected for 14 multiple comparisons tests using FDR correction.

Hypothesis 1 was tested by the total inter-network influence of the regulation network on the reactivity network. Hypothesis 2 was tested by the nodal influence on several network levels: (1) influence on the entire two networks' brain regions; (2) intranetwork influence within the regulation network (eight ROIs); (3) inter-network influence of the regulation network regions on the connections within the reactivity network; and (4) internetwork influence of the regulation network regions on the connections between the reactivity network and the regulation network. Hypothesis 3 was tested by correlating tests 1–4 from hypothesis 2 with anger measures (anger intensity and trait anger) and trait emotional regulation (reappraisal and suppression) correcting for the two hypotheses within each test using FDR. Pearson correlations were performed to assess the association between the results that were found to be significantly different between the high and low anger epochs (i.e., low-anger minus high-anger ''Influencing Degree'') and subjects' anger or emotional regulation measures. Subjects whose values exceeded the mean by three standard deviations were excluded from the analysis.

## Estimating the Spatial Specificity of the Results

In order to examine whether our DEPNA findings are specific to the networks and to control for whole brain effects of physiological parameters such as respiration, heartbeat, or head motion, we performed a bootstrapping procedure. The observed t and r values resulting from the previous analyses were compared with corresponding values generated by identical analyses of random sets of gray matter regions. The DEPNA findings were obtained using different network configurations with 8 or 14 regions. Accordingly, depending on the relevant configuration, the randomized networks contained 8 or 14 regions. The coordinates of these regions were randomly selected from a sampling space, defined based on ICBM 452 probability map<sup>2</sup> . The mask was created by thresholding ICBM 452 map to exclude voxels with probability lower than 50% of being classified as gray matter.

We performed a post hoc assessment of the specificity of the DEPNA paired t-test findings by comparing the observed t-value with the results of the randomized networks. The t-value background distribution was generated by repeating this procedure 1000 times. The p value of the bootstrapping test was defined as the fraction of the number of random cases which obtained t statistic values smaller than the observed finding, given as

$$p = \frac{\sum \text{RandomNetvals} \le \text{Originalt vals}}{k+1} \qquad \text{(5)}$$

where k is the number of random tests (k = 1000).

Since the observed correlation to behavior effect was tested only on a single network node, one region was randomly selected from each of the 1000 random networks and its DEPNA measure was correlated to the individual behavioral indices. This procedure was repeated for each behavioral index separately (i.e., anger intensity, trait anger and trait emotional regulation), resulting in a background distribution of 1000 correlation coefficients.

The p value of the random correlations test was defined as the fraction of the number of random regions which obtained higher/lower correlation coefficient r values than the observed findings.

### RESULTS

### Within and Between Networks' Hierarchy With Respect to Anger

To test our first hypothesis and evaluate if the total influence of the regulation network on the reactivity network was higher during the high anger epoch compared to the low anger epoch, we calculated the DEPNA total inter-network influence measures of all the regulation network regions. At odds with our expectations, we found that these measures were not significantly different between the low vs. high anger epochs both in the inter-network influence analysis on the connections within the reactivity network (p > 0.1), and for the inter-network influence analysis on the connections between the reactivity and regulation networks (p > 0.2).

To test our second hypothesis about whether the implicit regulation related region (i.e., the vmPFC) exhibited higher influence during the high anger epoch, we applied the DEPNA influencing index on three different network configurations: (1) the entire set of regions (i.e., emotional reactivity and regulation networks), by calculating the ''Influencing Degree'' measure; (2) the within emotion regulation network influence, by calculating the ''Intra-network Influence'' measure; and (3) the emotional regulation influence on the emotional reactivity network, by calculating the ''Inter-Networks Influence'' measure.

The difference in each node's ''Influencing Degree'' (low vs. high anger; **Figure 3** and **Table 1**) indicated the vmPFC (t = 3.11, p < 0.003, q < 0.05 FDR corrected) and right insula (t = 2.4, p < 0.02) had higher influence on both reactivity and control networks in the high anger condition compared to low anger (**Figure 3C**). We note that the right insula result did not withstand correction for multiple comparisons, therefore, we consider it only as a trend. The vmPFC t value result was also found to be significant (p < 0.05) using a control permutation test conducted on 1000 random networks each consisting of 14 random regions (see Supplementary Figure S1A in Supplementary Material 1). The graph visualization showed that the vmPFC had higher influence specifically on the right Insula (t = 2.30, p < 0.03), right IFG (t = 2.62, p < 0.02), left MiFG (t = 2.31, p < 0.03), left SPL (t = 2.06, p < 0.05) and pre-SMA (t = 3.32, p < 0.002; **Figure 3B**).

In order to further investigate the vmPFC's specific impact on the connections within the regulation network, we conducted an intra-network influence analysis (see **Figure 2E**). The vmPFC exhibited higher influence on the regulation network during the high anger epoch (t = 2.9, p < 0.005, q < 0.05 FDR corrected; **Figure 4**). The spatial specificity of this result was p < 0.05 (see Supplementary Figure S1B in Supplementary Material 1).

In order to specifically investigate communication between the regulation and reactivity networks we conducted the DEPNA inter-network analysis. Analysis of the influence of a region on the connections within the second network (i.e., inter-network option 1, see **Figure 2F**) found that the right insula had higher influence on the connections within the regulation network regions (t = 2.01, p < 0.05) in the high anger compared to the low anger condition (**Figure 5A**). We note that this result did not withstand correction for multiple comparisons, therefore, we consider it only as a trend. Analysis of the influence of a region on the connections between the two networks (i.e., inter-network option 2, see **Figure 2F**) found that the vmPFC had significantly higher influence on the connections between the reactivity and regulation network regions (t = 3.03, p < 0.003) in the high anger compared to the low anger condition (**Figure 5B**). The spatial specificity of this result was p < 0.05 (see Supplementary Figure S1C in Supplementary Material 1).

<sup>2</sup>http://www.loni.usc.edu/atlases

### The Relation Between Individual Anger Related Networks' Hierarchy and Behavior

To test our third hypothesis regarding the relation between networks' hierarchy and behavioral indications of anger state and tendencies we investigated the correlation between the level of vmPFC influence indices and the level of anger intensity, trait anger and trait emotional regulation. The vmPFC intra-network influence negatively correlated with anger intensity during the movie excerpt (r = −0.24, p < 0.04, n = 70, **Figure 6A**) and with trait anger (r = −0.31, p < 0.03, n = 51, **Figure 6B**). The spatial specificity of these results was p < 0.003 (see Supplementary Figure S2 in Supplementary Material 1). In other words, higher subjective reported anger intensity and higher tendency to become angry were associated with lower vmPFC influence on regions of the regulation network.

### DISCUSSION

The goal of this study was to utilize the DEPNA to investigate the properties of the hierarchy between the emotion reactivity and regulation networks during an anger experience induced by a movie excerpt. Applying the DEPNA during high- and low anger states depicted by individual self- report, revealed related modifications in network hierarchy. Specifically, DEPNA identified the vmPFC regulation-related region as a central node during the high anger episode based on its impact both on whole system connectivity, and on intra- and inter network connectivity. We demonstrated that lower vmPFC influence within the regulation network was associated with higher anger intensity reported during anger induction, and higher trait anger, emphasizing its relation to the direct experience of induced anger and to the habitual tendency to be angry. Finally, we demonstrated that higher vmPFC impact on internetwork connections was associated with a higher tendency to apply a suppression regulation strategy, linking vmPFC's inter-network influence to an independent measure assessing the engagement of implicit emotion regulation strategies. Together these results support and replicate our previous findings in which vmPFC played a key role in spontaneous anger regulation during an interpersonal induction of anger, as well as correlating with trait suppression (Gilam et al., 2015),

and extend them by demonstrating a unique directional connectivity pattern with regulation and emotional reactivity related networks. Of note however, this study was limited by being conducted on relatively young male participants (19.51 ± 1.45 years), thus it may not reflect the general population. More so, anger was provoked using only one film excerpt and was not compared to other emotional experiences, thus precluding specific claims on anger processing per se.

As might be expected from previous imaging work on implicit emotion regulation (Etkin et al., 2006), the regulation network as a whole, did not demonstrate higher total influence on the reactivity network during high- compared to lowanger epochs as tested by the total inter-network influence analysis. One possible explanation for this result is the fact that in this study subjects were not specifically instructed to regulate their emotions. The only brain region within the regulation network related to implicit emotional regulation processes is the vmPFC. Examining the total influence may have masked more subtle vmPFC influences related to implicit regulation effects. In line with this notion and confirming our second hypothesis, comparison of each network region's ''Influencing Degree'' between the high- and low-anger conditions revealed that the influence of the vmPFC was increased in the high anger condition (**Figure 3**, **Table 1**). The vmPFC also exhibited higher intra-network influence within the regulation sub-network during the high anger epoch (**Figure 4C**). This may indicate a central role of vmPFC during the experience of anger and potentially the inherent involvement of regulatory processes in such an experience (Gilam and Hendler, 2017). The nature of our methodology is important to emphasize. Rather than standard BOLD (co)activation analyses which provide information regarding the general involvement of various brain regions during the emotional experience, DEPNA provides us with a clue regarding the brain regions that drive the emotional regulation process. The significant change in vmPFC impact on regions of the reactivity and regulation network may reflect its involvement in a process that aims to exert control over the anger that stems from the provocation. This may be supported by the negative relationship between the level of vmPFC influence on regulation regions and the level of reported anger intensity (**Figure 6A**). However, since participants were not directly instructed to regulate their emotions, we

cannot know for sure whether this relation is indeed due to greater emotion regulation while viewing the anger-emotive film.

Indeed the vmPFC has been indicated in several other affect related processes such as reward valuation processes (Levy and Glimcher, 2012) and social information processing (Bechara et al., 2000; Rolls, 2004; Adolphs, 2009; Mitchell, 2009).

The broad network perspective as captured by the DEPNA method here, extends our knowledge regarding emotional regulation far beyond anger. It demonstrates that influence hierarchies within and between reactivity and regulation networks rather than activity or co-activation may reflect emotional state modification, even at the individual level. This pattern of results thus challenges the common concept of emotion regulation as a one way down regulation effect of a control network over a reactivity network region (Buhle et al., 2014; Morawetz et al., 2017). Our findings attribute a more active role of the vmPFC in shaping the regulation network organization as well as in its relation with the reactivity network

in support of effective emotion regulation. Network hierarchy is one of probably several other metrics that might capture dynamics in network organization with respect to changes in emotional experience. Further studies should test different network features which have been shown to be sensitive to emotional brain states such as network integration (Kinnison et al., 2012) or modularity (Ben Simon et al., 2017) during the anger experience. It should also be noted that while DEPNA can be used to make inferences regarding the influence hierarchy within a network, it does not infer a causal influence in the true sense, as correlation does not imply causation (Friston, 2009). We therefore suggest that the DEPNA results may target the critical regions to delineate the specific model of connectivity required for causality testing methods such as the DCM. Our results thus indicate that further studies using DCM should test the causality between the vmPFC, right IFG and right Insula (see **Figure 3B**).

One caveat that should be mentioned here is the inherent difficulty of making decisive scientific interpretations when using complex and naturalistic experimental stimuli such as movies. Given the fact that the brain has evolved to cope with/to process a continuous flux of multisensory input, naturalistic experiments may provide important insights into processes that are not optimally captured by classically controlled studies (Adolphs et al., 2016; Gilam and Hendler, 2016), however they are also prone to misinterpretation due to confounding variables. In our specific case, the low- and high anger phases of the movie possibly differ in audiovisual parameters that potentially contribute to the emotional intensity, but may not be considered as emotional per se (e.g., loudness and optical flow). In this case, our findings on differences in influence between the phases may be related to perceptual features rather than emotional factors. However, this caveat is addressed in our analysis of the relations between the influence levels on the one hand, and the reported acute anger intensity and trait anger on the other. In these cases, the variance of the dependent variable (influence level) is explained by independent parameters that are clearly related to emotion. By correlating the influence levels with these emotion measures, we examined their effect on the neural data while keeping the cinematic stimulus constant across subjects (see Raz et al., 2013, for a description of this rationale). An additional claim can be that since in our study the low anger period precedes the high anger period, differences could be attributed to order effects. However, post hoc DEPNA analyses of two low- and two high subsequent time frames of anger periods do not support such a concern (see Supplementary Figure S3 in Supplementary Material 1). Nevertheless, future controlled experimental designs could potentially contribute to understanding our reported findings on the overall difference of vmPFC influence levels between the more and less emotional epochs in light of low-level changes not related to the induced emotion.

To conclude, using the DEPNA approach we demonstrated that high- and low anger states can be characterized by different graph hierarchy of the reactivity and regulation networks. In particular the DepNA metrics depicted the central influence of the vmPFC; a major implicit regulation related node on these networks during the high anger experience. Our findings add to existing research on emotional regulation in general and anger regulation in particular by inspecting the operation and organization of relatively large-scale brain networks, while considering individual differences in state and trait. Specifically, higher anger intensity and trait-anger scores were associated with less vmPFC influence on the regulation network. Assuming the anger experience as provoked by the movie clip indeed initiated an implicit anger regulation process, we suggest that adaptive anger regulation could involve higher vmPFC influence on the regulation network. While anatomical and lesion studies have long indicated the causal role of the vmPFC and its interactions with limbic structures in emotion regulation (e.g., McDonald et al., 1996; Barrash et al., 2000), the system level influence degree measure may comprise a more dynamic fMRI index of this process. We hope that such a measure could be utilized as a target for future neuromodulation therapies aimed to alleviate the negative implications of anger on people's lives.

### AUTHOR CONTRIBUTIONS

YJ analyzed the data and wrote the manuscript. TL and GG acquired all data. GR, GG and TH helped interpret the results and wrote the manuscript. TH coordinated and directed the project. All authors reviewed the manuscript.

### FUNDING

This work was supported by the University of Chicago's Arete Initiative—A New Science of Virtues Program (Grant No. 39174-07 to TH); the Israeli Centers of Research Excellence (Grant No. 51/11 to TH); the Israeli Ministry of Science, Technology and Space (Grant No. 211580 to TH). GG was partially supported by NIH grant R01DA035484.

### REFERENCES


### ACKNOWLEDGMENTS

We are very grateful to Dr. Tal Gonen and Dr. Neomi Singer for helpful discussions, Vicki Myers and Aliya Solski for assistance in manuscript preparation, and Avi Mograbi, the director of the film that was used in this study. TH would like to thank the Sagol Foundation.

### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fnbeh.2018.000 60/full#supplementary-material


**Conflict of Interest Statement**: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Jacob, Gilam, Lin, Raz and Hendler. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Aggression in Women: Behavior, Brain and Hormones

#### Thomas F. Denson<sup>1</sup> \*, Siobhan M. O'Dean<sup>1</sup> , Khandis R. Blake<sup>2</sup> and Joanne R. Beames <sup>1</sup>

<sup>1</sup>School of Psychology, University of New South Wales, Sydney, NSW, Australia, <sup>2</sup>Evolution & Ecology Research Centre, School of Biological, Earth & Environmental Science, University of New South Wales, Sydney, NSW, Australia

We review the literature on aggression in women with an emphasis on laboratory experimentation and hormonal and brain mechanisms. Women tend to engage in more indirect forms of aggression (e.g., spreading rumors) than other types of aggression. In laboratory studies, women are less aggressive than men, but provocation attenuates this difference. In the real world, women are just as likely to aggress against their romantic partner as men are, but men cause more serious physical and psychological harm. A very small minority of women are also sexually violent. Women are susceptible to alcohol-related aggression, but this type of aggression may be limited to women high in trait aggression. Fear of being harmed is a robust inhibitor of direct aggression in women. There are too few studies and most are underpowered to detect unique neural mechanisms associated with aggression in women. Testosterone shows the same small, positive relationship with aggression in women as in men. The role of cortisol is unclear, although some evidence suggests that women who are high in testosterone and low in cortisol show heightened aggression. Under some circumstances, oxytocin may increase aggression by enhancing reactivity to provocation and simultaneously lowering perceptions of danger that normally inhibit many women from retaliating. There is some evidence that high levels of estradiol and progesterone are associated with low levels of aggression. We highlight that more gender-specific theory-driven hypothesis testing is needed with larger samples of women and aggression paradigms relevant to women.

#### Edited by:

Nelly Alia-Klein, Icahn School of Medicine at Mount Sinai, United States

#### Reviewed by:

Lesley J. Rogers, University of New England, Australia Gennady Knyazev, Institute of Physiology and Basic Medicine, Russia

#### \*Correspondence:

Thomas F. Denson t.denson@unsw.edu.au

Received: 17 November 2017 Accepted: 16 April 2018 Published: 02 May 2018

#### Citation:

Denson TF, O'Dean SM, Blake KR and Beames JR (2018) Aggression in Women: Behavior, Brain and Hormones. Front. Behav. Neurosci. 12:81. doi: 10.3389/fnbeh.2018.00081

#### Keywords: women, aggression, brain, hormones, intimate partner violence

''. . .females. . .are not passive victims of violence. Rather, they respond to provocation and are active participants in aggressive interactions.'' (Richardson, 2005, p. 245)

Aggression is a complex social behavior with many causes and manifestations. Over the past several decades, scholars have identified the many forms that aggression can take. Aggression can be physical (e.g., slapping), or verbal (e.g., shouting abuse). It can be direct in nature (e.g., directly retaliating against a co-worker) or indirect with aim of inflicting reputational harm (e.g., spreading rumors about a co-worker behind their back). Aggression can be impulsive, elicited by anger in response to provocation (known as reactive or hostile aggression) or it can be premeditated, less emotional, and used as a means to obtain some other end (known as proactive or instrumental aggression). Aggression that is physically extreme is referred to as violence (e.g., aggravated assault, homicide). Despite their apparently different surface characteristics, these instantiations of aggression all conform to the scholarly definition of aggression as behavior intended to cause harm to someone who is motivated to avoid that harm (Berkowitz, 1993; Baron and Richardson, 1994; Geen, 2001; Anderson and Bushman, 2002).

The aim of this review is to synthesize what is known about women's aggression from behavioral and neurobiological perspectives. We first focus on the behavioral research on women-perpetrated aggression including women's behavior in laboratory aggression paradigms, intimate partner violence (IPV), alcohol-related aggression and sexual violence. We then review data on prenatal and postnatal influences, the central nervous system, and neuroendocrine mechanisms. **Figure 1** summarizes these factors. We conclude by identifying gaps in the knowledge base, and provide suggestions for future research.

### PART 1: AGGRESSIVE BEHAVIOR

Compared to our knowledge of men's aggression, relatively little is known about women's aggression. Indeed, aggression and violence are usually considered male problems. There is some truth to this assumption. Globally, men are more violent than women (UN Office on Drugs and Crime, 2013). However, women frequently engage in other forms of aggressive behavior (Richardson, 2005). Research consistently reports that women use indirect aggression to an equivalent or greater extent than men (Archer and Coyne, 2005). Indirect aggression occurs when someone harms another while masking the aggressive intent (Björkqvist et al., 1992; Arnocky et al., 2012). Specific examples of indirect aggression include spreading false rumors, gossiping, excluding others from a social group, making insinuations without direct accusation, and criticizing others' appearance or personality. Girls' use of indirect aggression exceeds boys' from age 11 onward (Archer, 2004). This difference persists into adulthood; compared to men, adult women use more indirect forms of aggression across various areas of life (Björkqvist et al., 1994; Österman et al., 1998). Indeed, in a large cross-cultural survey of female aggression across 317 societies, Burbank (1987) found that female aggression was mostly indirect and rarely inflicted physical injury. Thus, in the real world aggression is common in women and girls, but the form it takes is largely indirect compared to men's aggression.

Numerous theorists have attempted to explain sex differences in aggression. Because human aggression is a complex social phenomenon elicited by multiple factors operating throughout the lifespan, one must consider how social influences interact with neurobiological mechanisms to influence aggression. Wood and Eagly's (2002) biosocial approach suggests that sex differences in behavior (including aggression) are caused by sex differences in physical attributes that interact with cultural values and customs. They note that sex differences in physical attributes and reproduction often make it more efficient for women to perform certain tasks and for men to perform others. For instance, in their discussion of women's historically limited involvement in warfare, they note that in most hunter-gatherer societies, men engaged in warfare more than women because men are physically larger and stronger and unable to nurse infants. Moreover, essential practices such as

FIGURE 1 | Graphical summary of the present review of factors associated with aggression in women. The left portion displays prenatal and early developmental influences known to affect aggression. The center portion shows neural and hormonal process associated with aggression in women. The right box indicates the different forms of aggression that women engage in and their relative frequencies. Green text indicates uncertainty regarding the robustness of the relationship with aggression in women. We note that this figure summarizes the current review only and that many additional factors do not appear here (e.g., genetic influences, neurotransmitter systems, societal factors). DLPFC, dorsolateral prefrontal cortex; DMPFC, dorsomedial prefrontal cortex; DACC, dorsal anterior cingulate cortex.

nursing, childcare and vegetal food production made it unlikely that women would travel far to engage in warfare (Wood and Eagly, 2002). This division of labor becomes reflected in social norms and values which are transmitted via socialization practices.

According to biosocial interactionist perspectives, social norms become relevant because most cultures endorse warfare as a means to gain status and because most cultures are patriarchal (i.e., men hold more power and status than women). Thus, most cultures reward men for being warriors and punish women for becoming aggressive. Indeed, social norms proscribe physical aggression in women (Eagly and Steffen, 1986) and girls can vocalize these norms from an early age (Crick et al., 1996). However, when women do behave aggressively and are dominant, they often face backlash against them (Barber et al., 1999). In this way, the interaction between biologically specified sex differences and sociocultural construction interact to produce lower direct aggression in women relative to men nearly everywhere in the world. In the next section, we review research on women's laboratory aggression, alcohol-related aggression, intimate partner aggression and sexual aggression.

### Aggressive Behavior in Laboratory Studies

Psychologists have been studying aggressive behavior with laboratory aggression paradigms since the 1960s. The primary strength of laboratory aggression paradigms is that researchers can manipulate variables that might influence aggression while eliminating much of the complexity of the outside world. Researchers can then quantify the observed aggressive behavior. The most commonly used paradigms are variations of the Taylor (1967) aggression paradigm (TAP), and the point subtraction aggression paradigm (PSAP; Cherek, 1981). We first review these paradigms in order to facilitate understanding of gender differences in laboratory aggression.

### Laboratory Aggression Paradigms

In the TAP (sometimes called the competitive reaction time task; Giancola and Zeichner, 1995b), participants are typically provoked in some manner, often through receiving electric shocks or bursts of white noise from another participant (who may be real or bogus; e.g., Giancola and Parrott, 2008; Jones and Paulhus, 2010). Participants may also be provoked by receiving negative feedback on a laboratory task such as an essay or short speech, or by being ignored, rejected, or ostracized by another person (Bushman and Baumeister, 1998; Warburton et al., 2006; Blake et al., 2018). Following provocation, participants are given the opportunity to retaliate against the provocateur to varying degrees, or respond non-aggressively. In the TAP, aggressive behavior is operationalized as the intensity and/or duration of noise blasts directed at the provocateur.

For the PSAP, participants ostensibly play a game against a real or bogus participant to earn points that may be exchanged for money. In modern versions of the paradigm, during each trial participants are given the option to either steal points, defend their points, or earn points (Geniole et al., 2017). Provocation is induced when the focal participant has points stolen from them by the other participant, and aggression is observed when the focal participant steals money from the other participant. As in the TAP, participants may also be provoked via insulting feedback or ostracism. In addition to the TAP and PSAP, aggression in the laboratory can also be operationalized by giving the experimenter a poor recommendation for a coveted job (e.g., Denson et al., 2011) and giving hot sauce to a participant who is known to dislike spicy foods (Lieberman et al., 1999). However, the TAP and PSAP are the most widely studied.

Some researchers have criticized laboratory aggression paradigms on the grounds of poor external validity (e.g., Tedeschi and Quigley, 1996; Ritter and Eslea, 2005). It is true that laboratory paradigms lack a superficial similarity to the real world (i.e., mundane realism). However, several researchers have quantitatively shown that laboratory paradigms possess both strong psychological realism and external validity (Anderson and Bushman, 1997; Giancola and Chermack, 1998; Giancola and Parrott, 2008). For instance, female parolees with a violent criminal history steal more points in the PSAP than non-violent parolees (Cherek et al., 2000). Importantly, all laboratory aggression paradigms are consistent with the widely accepted definition of aggression as behavior intended to harm another person (Anderson and Bushman, 2002). However, few studies were specifically designed to externally validate laboratory aggression paradigms with women.

### Meta-Analytic Evidence

To date, there have been three large scale meta-analyses of gender differences in laboratory aggression paradigms (Eagly and Steffen, 1986; Bettencourt and Miller, 1996; Bettencourt and Kernahan, 1997) 1 . Consistent with the social psychological Zeitgeist at the time, Eagly and Steffen (1986) favored a social learning explanation of gender differences over biological explanations. They concluded that women are less aggressive than men because social roles encourage aggression in men but not women. They found a small-to-medium effect such that men were more physically aggressive than women (d = 0.40), but this effect was greatly reduced for non-physical forms of aggression such as verbal aggression (d = 0.18). A separate group of 200 men and women coded how they would feel if they were to aggress in each study included in the meta-analysis. Relative to men coders, women coders anticipated that experiencing greater guilt, anxiety, and danger would be the consequences of aggressing. Thus, women may be less likely to aggress in the laboratory due to fear of retaliation and an unwillingness to harm others.

In what still remains the most comprehensive meta-analysis to date of gender differences in laboratory aggression, Bettencourt and Miller (1996) examined 107 effect sizes from 64 experiments. Overall, they found a small gender effect (d = 0.24) such that men were somewhat more aggressive than women. When unprovoked, women were less physically and verbally aggressive than men. However, provocation attenuated the

<sup>1</sup>The Eagly and Steffen (1986) and Bettencourt and Miller (1996) meta-analyses also included field studies that contained a behavioral measure of aggression.

gender difference in physical aggression and ameliorated the gender difference in verbal aggression.

Bettencourt and Miller (1996) also examined whether the type of provocation would influence gender differences in aggression. They found that men were more aggressive than women when the provocation induced frustration or insulted participants' intelligence. By contrast, the gender difference in aggression was reduced to zero in studies that manipulated provocation with physical attack (e.g., electric shocks) or an insulting evaluation (e.g., on an essay task). Thus, women and men may be equally aggressive when faced with physical attack or an unjustified insult, at least in the laboratory. Consistent with Eagly and Steffen (1986), Bettencourt and Miller (1996) found that women coders anticipated greater danger than men coders were they to aggress and that men perceived the provocation as more intense than women. These perceptions subsequently predicted a greater male-biased gender difference in aggression. Thus, both meta-analyses converged on perceived danger as one putative psychological gender difference that explains lower aggression observed in women in the laboratory.

During the ''cognitive revolution'' in social psychology in the 1970s and 80s, many researchers were influenced by Berkowitz's (1993) and Berkowitz and LePage's (1967) cognitive neoassociationistic theory of aggression. According to the theory, any aggression-related cues (e.g., weapons, violent images, hostile words, alcohol) prime a cognitive network of aggression-related associations. These primed associations increase aggression when provoked (Carlson et al., 1990). A meta-analysis of the violent cue literature found that men were moderately (d = 0.41) more aggressive than women when participants were not provoked but exposed to violent cues (Bettencourt and Kernahan, 1997). The putative cause is that men may have a more extensively developed violencerelated cognitive network than women, possibly due to gender norms that encourage male aggressiveness (i.e., social role theory). Women may behave less aggressively because female gender roles stipulate that women should not act aggressively when unprovoked. However, as in the previous meta-analyses, this gender difference was reduced to zero when participants were provoked. The authors concluded that once provoked, the influence of gender roles on aggression may become less salient.

#### Alcohol-Related Aggression

Alcohol-related aggression is of interest to neuroscientists because acute and chronic alcohol use is thought to increase risk for aggression via dysfunction in the prefrontal cortex (PFC; Giancola, 2000; Heinz et al., 2011; Gan et al., 2015). Alcohol is also involved in a large proportion of violent crimes. For instance, a study of 11,836 men and women arrested for violent crimes found that 57% and 44% had been drinking prior to committing the crime, respectively (Martin and Bryant, 2001). Similarly, relatively greater alcohol consumption equally predicted fighting in a young group of British men and women (Wells et al., 2005). Although the effects of alcohol on facilitating aggression in men has been an active research area, the same cannot be said for research with women.

In terms of laboratory research, two meta-analyses examined the literature on alcohol-induced aggression that manipulated alcohol administration (Ito et al., 1996; Bushman, 1997). The mean effects of alcohol on aggression were significant for men but not for women; however, both meta-analyses were underpowered to detect effects in women. Nonetheless, several individual experiments have examined alcohol-induced aggression in women.

In two separate experiments, men and women were assigned to a low dose or high dose of alcohol or tonic (Rohsenow and Bachorowski, 1984). To control for expectancies about the effects of alcohol on aggression, half of the participants were told they received alcohol and the other half tonic. They were then insulted by another fictitious participant and given the opportunity to retaliate via a harsh written evaluation (i.e., a measure of verbal aggression). For women, alcohol increased aggression, but only at the low dose. In another experiment, participants consumed alcohol on one day and placebo on another day (Dougherty et al., 1999). On both days, they played the PSAP against a fictitious participant six times from morning to afternoon. Relative to the placebo day, alcohol increased aggression in both men and women for several hours. Furthermore, men and women who were more aggressive on the placebo day showed the greatest alcohol-related aggression. This latter finding suggests that people with dispositions toward aggression when sober are more likely to become aggressive when intoxicated.

Several experiments support this notion that dispositional aggression is related to alcohol-induced aggression. For instance, Giancola (2002a) found that men and women high in dispositional aggressiveness exhibited alcohol-induced aggression in the TAP, but the relationship between trait aggression and aggressive behavior was stronger for men than women. Another analysis of the same sample found that dispositional anger was also positively correlated with alcoholinduced aggression in women, but only at low provocation levels (Giancola, 2002b). For men, trait anger was positively correlated with aggressive behavior and with low and high levels of provocation. Despite a general tendency for alcohol intoxication to increase aggression in men and women, several other studies found that alcohol did not increase aggression in women, but did in men (e.g., Giancola and Zeichner, 1995a; Hoaken and Pihl, 2000; Hoaken et al., 2003; Gussler-Burkhardt and Giancola, 2005).

A recent meta-analysis was able to provide somewhat stronger evidence for the aggression-augmenting effect of alcohol in women (Crane et al., 2017). They examined the 12 available alcohol administration experiments that included women and found a small, but significant effect of acute alcohol intoxication on increased aggression, d = 0.17, CI<sup>95</sup> = 0.03, 0.30. This effect is smaller than that observed in men (i.e., ds = 0.49 and 0.50, in Ito et al. (1996) and Bushman (1997)). However, with only 12 experimental studies on alcohol-related aggression in women, more research is needed both in the laboratory and natural settings.

A follow-up meta-analysis examined gender of the target and provocation as moderators of alcohol-induced aggression in women (Crane et al., 2018). Alcohol increased aggression in women in studies that used relatively more intense provocation (e.g., shocks, insults) but not in studies that used more innocuous provocations (e.g., reading upsetting vignettes). Alcohol increased aggression in studies that included female targets, but not male targets. Although the number of studies was small in number (k = 14), these findings suggest that alcoholrelated aggression in women may be strongest when provoked and retaliating against women targets.

### Summary of the Laboratory Research

The extensive experimental literature on aggression in women and men provides a solid evidence base for the primary conclusion that women are less physically aggressive than men. This finding is consistent with crime statistics showing that men are by far the most violent gender. Nonetheless, women are capable of behaving aggressively, especially when provoked. The gender difference in aggression becomes much smaller when participants are provoked in the laboratory and non-existent when participants are allowed to verbally aggress (Bettencourt and Miller, 1996). Women's relatively lower aggression when unprovoked seems at least partially attributable to greater fearfulness than men when considering behaving aggressively (Eagly and Steffen, 1986; Bettencourt and Miller, 1996). For instance, in one experiment exposure to a laboratory stressor increased aggression in men but decreased aggression in women (Verona and Kilmer, 2007). The authors suggested that women may experience a withdrawal reaction in stressful circumstances whereas men are more likely to experience an approach response.

The research on alcohol-related aggression suggests that intoxication increases aggression in men and women, but the effect tends to be larger in men and people with pre-existing dispositions toward aggressive behavior. Alcoholrelated aggression in women tends to be most severe when provoked and the target of aggression is a woman. One limitation of the laboratory research is that conclusions are based largely on just two direct aggression paradigms: the TAP and PSAP. Although these are well-validated paradigms, the field could benefit from a more diverse set of paradigms. For instance, experimental alcohol research with women and indirect aggression would be informative.

### Intimate Partner Violence

Conflict, especially around romantic jealousy, can elicit aggression between partners, which is known as IPV. Prevalence and victimization rates vary substantially depending on the methodology used and population sampled. Definitions also vary, but in the IPV literature, IPV is frequently considered to be any act of aggression directed toward one's partner, rather than violence specifically (i.e., extreme acts of physical aggression). Lifetime prevalence of IPV victimization was estimated at 37.3% for women and 30.9% for men living in the United States (Smith et al., 2017). Between 8% and 21% of a representative sample of American couples reported experiencing at least one act of IPV in the past year (Schafer et al., 1998). In this section, we focus on heterosexual relationships as relatively little is known about IPV in same-sex attracted relationships in women (for an exception, see Badenes-Ribera et al., 2016).

Some women do use violence against their romantic partners, although the severity and form of the IPV may differ compared to male-perpetrated IPV. Women tend to engage in fewer acts of severe IPV than men, just as women engage in less aggression than men generally. For instance, one study of IPV arrestees reported that women used an average of 1.44 severely violent tactics (as defined by the severe violence scale of the Conflict Tactics Scale; Straus, 1979) during the arrest incident, whereas men used an average of 2.27 severely violent tactics (Busch and Rosenberg, 2004). Women are more likely than men to throw objects at their victim, to use weapons, and to bite their victims (Magdol et al., 1997; Archer, 2002; Melton and Belknap, 2003), whereas men are more likely to beat up, choke or strangle their victims (Archer, 2002).

These gender differences in IPV-related violence are likely due to sexual dimorphism in physical attributes. Because of men's greater size and strength relative to women, on average women can inflict more harm with weapons and thrown objects than their bodies, whereas men can inflict equivalent or greater harm with their bodies. Indeed, IPV causes visibly greater physical and psychological harm in women than men (e.g., Morse, 1995; Archer, 2000; Caldwell et al., 2012). A meta-analysis found that male IPV perpetrators were more likely to cause physical injury than female IPV perpetrators. Over 60% of those injured by their partners in an IPV incident were women (Archer, 2000). Female victims of IPV are not just more likely to suffer physical injury, but also posttraumatic stress disorder, depression and anxiety than their male counterparts (Caldwell et al., 2012). Additionally, an important aspect of IPV is sexual violence, and in this category, women are far more likely to be victims than men (Foshee, 1996; Coker et al., 2002; Black et al., 2011). Similarly, women are far more likely to be victims of IPV-related homicide than men. In Australia, women comprise 76% of IPV homicide victims (Ramsey, 2015). Other countries also show this gender disparity in IPV homicide rates. The World Health Organization (2013) examined 1121 crime data estimates of IPV-related homicides across 65 countries from 1982 to 2011. Of these homicides, the median prevalence of women killed by their partners was 38%, whereas the corresponding rate of murdered men was 6%.

Not all research found lower use of severe violence in women. Some studies using data from the criminal justice system (e.g., police reports, pretrial information and victim statements) of IPV offenders highlight commonalities regarding the use of IPV in women and men. These studies reported that defendants of both genders are equally likely to engage in harassing behavior (e.g., trespassing and stalking), and to have been physically abusive by punching, hitting, slapping, or stabbing (Melton and Belknap, 2003). Findings from these forensic studies suggest women are equally likely to use severe forms of violence as men and to severely injure their partners (e.g., Melton and Belknap, 2003; Busch and Rosenberg, 2004; Henning and Feder, 2004). Several other studies reported that both men and women use coercive and controlling behavior against their partners in equivalent rates (Stets and Pirog-Good, 1990; Stets, 1991; Felson and Outlaw, 2007; Hines et al., 2007; Straus and Gozjolko, 2014), but other studies found that women are less likely than men to engage in controlling behavior (e.g., Johnson, 2006; Hester, 2009).

The criminology literature also highlights important gender differences. For instance, women are more likely than men to be involved in ''dual-arrests'' (i.e., both partners are arrested at the same time; Melton and Belknap, 2003; Henning and Feder, 2004). The authors concluded that dual-arrests might provide evidence for the proposition that many women who commit IPV do so in self-defense (Melton and Belknap, 2003). Women are also much less likely than men to have repeat offenses documented (Hester, 2009), and less likely to have violated an existing protection order (Henning and Feder, 2004). Additionally, based on 16 empirically validated risk factors for criminal recidivism, male IPV perpetrators presented a greater concern for future violence than female perpetrators. Specifically, female perpetrators ranked higher on only three risk factors; younger age, unemployment and severity of offense (i.e., more likely to have used a weapon). By contrast, male offenders ranked higher on the remaining 13 risk factors, including escalation of conflict frequency and/or severity, threats to kill and substance abuse (Henning and Feder, 2004).

Data from the criminal justice system may not generalize to the wider population. IPV offenders who have become involved with law enforcement may differ in numerous ways from those who have not become involved in the criminal justice system. Indeed, large scale studies find that women and men perpetrate IPV (i.e., any form of aggression directed at their partner) at similar rates, although the severity and types of aggressive acts may differ (e.g., Straus, 1980; Archer, 2000; Gass et al., 2011; Desmarais et al., 2012; Renner and Whitney, 2012; Hamel et al., 2015).

These similar rates of IPV perpetration are likely due to the bidirectional aggression that occurs during episodes of IPV. Bidirectional IPV occurs when each partner is both a perpetrator and a victim of IPV (Mennicke and Wilke, 2015). A review of 50 studies examining self-report, police report and archival data studies found that between 49.2% and 69.7% of IPV was bidirectional (Langhinrichsen-Rohling et al., 2012). This bidirectional nature of IPV is also consistent across diverse populations. A review of 111 articles found that female IPV perpetration tends to be highest in clinical populations, with a 41.7% pooled prevalence rate, and lowest in large population studies with a 24.1% pooled prevalence rate (Desmarais et al., 2012), Although the average prevalence rates of women perpetrating IPV differ significantly across different populations (for review see Desmarais et al., 2012; Langhinrichsen-Rohling et al., 2012), the proportion of bidirectional IPV remains consistent across diverse samples, averaging 57.5% (Langhinrichsen-Rohling et al., 2012).

### Motivations for IPV Perpetration

Studies on female perpetrators of IPV show that risk factors and motivations for violence are heterogeneous. One systematic review article focused on motivations for women's IPV perpetration (Bair-Merritt et al., 2010). The review of 23 studies found that self-defense, expressing anger, control, desire for the partner's attention, and retaliation motivated women's IPV perpetration. Indeed, being victimized by an intimate partner is consistently one of the strongest predictors of IPV perpetration for both men and women (O'Leary and Slep, 2012).

One study examined whether motivations for IPV perpetration in women could predict more or less engagement in IPV (Caldwell et al., 2009). Motivations included negative emotion expression, control, jealousy and wanting to portray ''toughness'' to ward off potential victimization. Each motivation significantly predicted physical aggression towards male partners, even when controlling for prior victimization. Likewise, control, toughness portrayals, and negative emotion expression were predictive of psychological aggression perpetration. Jealousy and control motives were also positively predictive of coercive, controlling IPV perpetration (Caldwell et al., 2009).

A more recent study examined the motives for IPV in both men and women arrested for domestic violence offenses (Elmquist et al., 2014). This study found that men and women perpetrators were equally motivated by self-defense, communication difficulties, power/control, and jealousy. Women were, however, significantly more likely to cite negative emotion expression and retaliation as reasons for engaging in IPV than men (Elmquist et al., 2014).

Furthermore, a recent meta-analysis of 580 studies identified 60 risk-factors for IPV perpetration across four different categories; demographic markers, family-of-origin markers, relationship markers and mental health/individual markers (Spencer et al., 2016). Of these 60 risk factors, only three differed by gender. Specifically, alcohol use/abuse was a significantly stronger risk-factor for male than female IPV perpetration. Secondly, a demand/withdrawal relationship communication style was a significantly stronger risk factor for IPV perpetration for men than for women. Finally, experiencing or witnessing domestic abuse as a child was a stronger risk factor for men than for women (Spencer et al., 2016). In sum, the existing literature illustrates more similarities than differences in the motivations and risk factors for IPV perpetration of men and women. These motivations and risk factors could be considered in the development of IPV prevention programs for both men and women.

### Treatment for Women IPV Perpetrators

One out of 10 clients in batterer intervention programs are women (Price and Rosenbaum, 2009), and women often find themselves in batterer intervention programs that were designed for men (Goldenson et al., 2009). These existing programs (e.g., Duluth group therapy and cognitive behavioral therapy) have little to no effect on IPV recidivism in male offenders (Babcock et al., 2004; Stover et al., 2009). Few studies have examined the effectiveness of batterer intervention programs on female perpetrators (Carney et al., 2007). One such study found women were less likely to be physically abusive and passive-aggressive to their partners at the end of treatment; however, their likelihood of using controlling behavior remained unchanged (Carney and Buttell, 2004). Growing evidence for equivalent rates of IPV perpetration among women and men and the lack of studies on batterer intervention programs with women highlights the need for research on interventions for all IPV offenders.

### Summary

There is ample evidence to suggest that women are as likely, if not more likely than men, to commit IPV (e.g., Archer, 2000). However, research also suggests that male perpetrated IPV is more likely to cause physical and psychological injury to women (e.g., Archer, 2000; Caldwell et al., 2012). Studies have also found that male perpetrators of IPV commit a higher number of severely violent acts (Busch and Rosenberg, 2004), and historically have more IPV offences documented than women perpetrators (Hester, 2009). The existing literature emphasizes that IPV is a complex phenomenon that arises from multiple risk and motivational factors (e.g., Elmquist et al., 2014; Spencer et al., 2016). There is little evidence that these factors differ between genders. Thus, it is critical that future research tests what are perhaps simplistic assumptions about male and female IPV perpetration (Richardson, 2005). Namely, the assumptions that men perpetrate IPV to control women, and that women perpetrate IPV only in self-defense (Spencer et al., 2016). Despite potential differences in IPV perpetration by men and women, it is important to also consider women's role in aggressive relationships. Without doing so, there is less room for the development of effective prevention strategies for couples experiencing IPV.

### Sexual Aggression

Like most other forms of aggression, men are more likely to perpetrate sexual aggression than women. One in six American women is raped during their lifetime, the vast majority by men (Centers for Disease Control, 2010). In Australia, 19% of women have experienced sexual violence since age 15 (Parliament of Australia, 2006). Nonetheless, a small minority of women commit acts of sexual aggression against men, women and children. Sexual aggression encompasses numerous sexual activities forced upon a victim without the victim's consent (Krahé and Berger, 2013). As is the case with men, acts of sexual aggression committed by women may include coerced sex, anal or vaginal penetration, oral sex, kissing, exposing genitals and using objects to cause harm (Krahé and Berger, 2013; Cortoni et al., 2017).

A recent meta-analysis examined the prevalence rate of female sexual offending from 2000 to 2013 in 12 countries (Australia, Belgium, Canada, England and Wales, France, Ireland, New Zealand, Norway, Scotland, Spain, Switzerland and the United States; Cortoni et al., 2017). Rather than relying on selected samples, the authors included official government crime statistics and large scale surveys that examined victimization. Results showed that 2.2% of sexual offenders were women. Girls were more likely to perpetrate than adult women. Approximately 40% of victims were men and 4% women. In two-thirds of cases, women were the sole perpetrators. The remaining offenders coperpetrated, mostly with a man (Budd et al., 2017).

As with male-perpetrated sexual aggression, femaleperpetrated sexual aggression is likely to go unreported to police (Stemple et al., 2017). For instance, the Cortoni et al. (2017) meta-analysis reported the prevalence of victimization at approximately 11%, which was 5–6 times higher than the offender prevalence rate derived from the crime statistics. This discrepancy suggests that victims of femaleperpetrated sexual aggression are unlikely to report the crime to police. Victims may fear blame, social sanctions, humiliation, or that their accusations may not be taken seriously by professionals (Fisher and Pina, 2013; Stemple et al., 2017).

Because alcohol is involved in most instances of sexual aggression, victims may blame themselves for drinking. In one large scale survey of German university students, nearly 70% of women perpetrators reported that one or both partners drank alcohol prior to offending (Krahé and Berger, 2013). As do men, women perpetrators report encouraging their victims to use alcohol and take advantage of their victim's intoxicated state (Struckman-Johnson et al., 2003). Thus, alcohol plays a substantial role in women's sex offending and probably underreporting as well.

### Summary

Sexual aggression is primarily perpetrated by males. Nonetheless, a small group of women are sexually aggressive. There is relatively little understanding of why some women perpetrate sexual aggression. Theoretical development should be a priority for this area. Feminist theories of male-perpetrated sexual aggression suggest that men commit rape out of patriarchal concerns and to control women (Brownmiller, 1975). Application of these theories to women may not be appropriate. As the data to date are descriptive and correlational, experimental research with laboratory sexual aggression paradigms is needed to identify causal influences and moderators. Such paradigms exist but to our knowledge have not yet been used with women (for a review, see Davis et al., 2014).

### PART 2: NEUROBIOLOGICAL PATHWAYS TO WOMEN'S AGGRESSION

In this section, we review data on prenatal and postnatal influences, the central nervous system, and neuroendocrine mechanisms that may affect women's aggression.

### Prenatal and Postnatal Influences

Gender differences in aggression emerge during toddlerhood (Archer, 2004). Thus, one approach to understanding these differences is to examine the earliest possible developmental time periods: the prenatal and postnatal periods. The idea is that exposure to certain social or biological risk factors during these sensitive developmental periods can disrupt the normal development of the nervous system, which may predispose offspring to aggression later in life. Here we selectively review a subset of some of the more widely studied factors that have been examined within the context of female aggression in humans and rodents.

Despite showing that exposure to several factors increases risk for aggression during development, there has been limited success in identifying distinct neurobiological pathways to aggression for girls and boys. Liu (2011) reviewed a number of prenatal, perinatal, and postnatal risk factors including smoking during pregnancy, birth complications, maternal depression, malnutrition, lead exposure, head injury, child abuse and maternal stress. Of these, there was only evidence for gender differences in two risk factors: maternal malnourishment and maternal depression. Sons, but not daughters of malnourished mothers were 2.5 times more likely to be classified with antisocial personality disorder as adults. There were also gender differences in the effects of maternal depression on externalizing behavior (i.e., disruptive behavior which includes aggression). For instance, a longitudinal study of over 1,300 children and their mothers found that greater maternal depression was associated with greater externalizing behavior in boys at 2 years of age, but the relationship was stronger for girls at 6 years of age than boys (Blatt-Eisengart et al., 2009). Although the reason is unclear, the disruption to caregiving caused by maternal depression may be particularly difficult for older daughters.

Research with rodents and humans has shown effects of prenatal exposure to psychotropic substances such as cannabis, nicotine, cocaine, and alcohol on aggression in female offspring. For instance, one study found that prenatal cannabis exposure was associated with aggression in 18-month old girls (El Marroun et al., 2011). Another study found that prenatal smoking positively predicted aggression in girls aged 17–42 months, although girls remained less aggressive than boys (Huijbregts et al., 2008). Prenatal cocaine exposure in 5-year olds also increased aggression, but less so in girls than boys (Bendersky et al., 2005). Another study found that prenatal exposure to cocaine predicted heightened aggression in 6–7 year old girls but not boys, and only among girls who had not been exposed to alcohol prenatally (Sood et al., 2005). The rodent literature does not suggest robust gender differences. The increased aggression induced by prenatal cocaine exposure persists into adulthood in both male and female rodents (Williams et al., 2011).

Consistent with these findings, in one study participants played a modified version of the PSAP in which they could not only aggress or earn points, but also temporarily escape. This study included a group of teens with little or no prenatal cocaine exposure and another group of teens with heavy prenatal exposure (Greenwald et al., 2011). The groups did not differ in aggression but the heavy exposure group was more likely to use the escape option. Girls were even more likely than boys to choose the escape option. Thus, prenatal cocaine exposure may alter both flight and fight responses in girls later in life.

As with alcohol use in adulthood, prenatal alcohol exposure has a large body of evidence supporting its role in increasing aggression later in life. For instance in one large-scale study of 625 families, 6–7 year old children who had been exposed to prenatal alcohol were more aggressive (Sood et al., 2001). Girls were less aggressive than boys. The rodent literature suggests that prenatal alcohol exposure increases aggression in male rats but can increase or decrease aggression in female rats (Marquardt and Brigman, 2016).

Prenatal testosterone exposure may also be a developmental mechanism underlying aggression in women. For instance, congenital adrenal hyperplasia is characterized by overproduction of androgens including testosterone in the prenatal environment. Girls and women with this condition are more physically aggressive than girls and women without this condition (Hines, 2010). Studies with rodents also typically show that prenatal testosterone exposure increases aggression in both males and females (e.g., vom Saal, 1979; Mann and Svare, 1983). Other human work has tested the twin testosterone transfer hypothesis, which is the notion that same-sex girl twins should have lower levels of testosterone exposure prenatally than opposite-sex twin pairs (Tapp et al., 2011). This increased testosterone is thought to heighten aggressiveness in the girls who shared the prenatal environment with their brother. One study of 13 year-old twins found support for this notion (Cohen-Bendahan et al., 2005), but robust evidence for this hypothesis is lacking (for a review, see Tapp et al., 2011). Similarly, the ratio of the second finger length to fourth finger length (i.e., 2D:4D ratio) is considered an indirect indicator of prenatal testosterone exposure. Smaller values are thought to indicate higher prenatal testosterone exposure. A meta-analysis showed no relationship between the 2D:4D ratio and aggression in women and only a small but significant effect in men for verbal aggression only (r = 0.035; Turanovic et al., 2017). Thus, the evidence for prenatal testosterone as a risk factor for women's aggression is mixed.

### Summary

Several prenatal and postnatal influences heighten risk for aggression later in life, but most do not differentiate between males and females. Of the risk factors reviewed here, the most evidence for sex-dependent effects is for postnatal maternal depression, prenatal maternal malnourishment, and prenatal exposure to drugs and alcohol. There is some evidence for prenatal testosterone exposure increasing aggression in girls later in life, but the evidence is mixed.

### Brain

In recent decades, researchers have made use of electroencephalography (EEG), brain stimulation, physical body manipulations and functional magnetic resonance imaging (fMRI) to examine the neural mechanisms underlying aggression. We review some of the evidence that examined both women and men or women only. We mention gender differences only when they were reported in the source articles.

### EEG

### **State and Trait Anger/Aggression Correlate With Resting Frontal Asymmetry**

Relatively greater left resting frontal alpha asymmetry is an indicator of approach motivation and greater right asymmetry is an indicator of avoidance motivation (Harmon-Jones et al., 2010). Anger and aggression are considered approach-related phenomenon (Carver and Harmon-Jones, 2009). Several studies indicate that greater individual differences in resting left frontal alpha asymmetry are positively correlated with dispositional anger (e.g., Harmon-Jones and Allen, 1998; Harmon-Jones, 2004; Hewig et al., 2004). Left-sided frontal asymmetry is also positively correlated with trait aggression. For example, in a sample of 15 boys and 11 girls, Harmon-Jones and Allen (1998) found a small, although non-significant, positive correlation between relative left frontal activation and trait aggression. Thus, individual differences in anger and aggression are linked to this neurophysiological indicator of approach motivation.

In a study of 30 men and 35 women, the authors examined the extent to which trait anger and two types of dispositional anger expression styles correlated with resting frontal asymmetry (Stewart et al., 2008). The anger expression styles referred to the extent to which people tend to express anger and aggression (i.e., anger-out) or suppress anger and aggression (i.e., anger-in). Higher trait anger was associated with greater relative left mid-frontal asymmetry. For participants high in trait anger, anger-in (rather than anger-out) positively correlated with relative left activation in regions other than the frontal cortex. Results remained significant even when gender was included as a covariate, suggesting that differences between men and women did not overly influence the correlations in this study. Experimental studies that manipulated state anger conceptually replicated and extended the initial correlational work on trait anger and relative left frontal asymmetry (for a review see Harmon-Jones and Gable, 2017).

### **Trait Anger/Aggression Correlate With Event-Related Potentials (ERPs)**

Additional research investigated the relationship between trait anger or aggression and electrical activity in response to stimuli (for an overview, see Flannery et al., 2007). This electrical activity is known as an event-related potential (ERP). ERP studies of aggression have primarily used oddball or continuous performance tasks and focused on the parietally distributed P300 component in clinical or inmate populations (e.g., Harmon-Jones et al., 1997; Stanford et al., 2003). However, some studies have examined healthy adult populations (e.g., Gerstle et al., 1998; Mathias and Stanford, 1999).

The results from these studies largely suggest that higher self-reported impulsive aggression and hostility were associated with reduced parietal and/or central P300 amplitude (Harmon-Jones et al., 1997; Gerstle et al., 1998; Mathias and Stanford, 1999). The amplitude of the P300 is thought to reflect information processing capacity including stimulus evaluation, attention allocation, and context updating (e.g., Donchin and Coles, 1988; Coles et al., 2000). These results suggest that aggressive individuals may have impairments in these cognitive abilities.

Stewart et al. (2010) extended these results using a sample of 48 men and 54 women. The authors showed that higher anger-out scores were associated with increased P300, N200 (indicating increased response inhibition and/or conflict monitoring), and N400 (indicating increased elaborative stimulus processing) amplitude to negative words. The N200 and N400 are fronto-centrally distributed components of the ERP. These findings suggest that aggressive individuals may exert more effort to override attention to negative information (Stewart et al., 2010). Further, higher anger-in predicted decreased N400 amplitude to negative words, suggesting that these individuals need fewer attentional resources to suppress negative stimuli (Stewart et al., 2010).

### **State Anger/Aggression and ERPs**

Other studies have investigated how inducing anger or aggression affects ERPs (e.g., Krämer et al., 2008; Gable and Poole, 2014). Only one study investigated gender differences (Krämer et al., 2008). In this study, 25 men and 24 women were provoked within the TAP. When participants were deciding on the volume of a shock they would deliver to an opponent, participants high in trait aggression showed enhanced frontal negativity (i.e., N200) when the opponent delivered a high noise blast compared to a low noise blast (Krämer et al., 2008). This effect was greater in participants high in trait aggression who behaved less aggressively in the task. These results suggest that participants higher in trait aggression were more prone to detect conflict and attempted to exert inhibitory control. Men and women did not differ on neurophysiological responses. Another study suggests that women high in trait hostility showed a pattern of EEG data that is compatible with heightened emotional responding to emotional faces but also heightened inhibitory control (Knyazev et al., 2009). Men high in hostility did not show the inhibitory control effect, which is consistent with gender differences in aggressive behavior.

### Brain Stimulation

Frontal cortical asymmetry can be induced with transcranial magnetic stimulation (TMS) and transcranial direct current stimulation (tDCS; for reviews see Angus et al., 2016; Kelley et al., 2017). Slow repetitive TMS (rTMS) can inhibit cortical excitation. Using a small group of 10 healthy women, one study found that inhibiting the right PFC using rTMS caused selective attention toward angry faces, whereas inhibiting the left PFC caused selective attention away from angry faces (d'Alfonso et al., 2000). Similar patterns of frontal activation results have been observed in predominantly female samples (van Honk and Schutter, 2006; Hofman and Schutter, 2009). These results should be interpreted cautiously, however, as another study that used continuous theta-burst magnetic stimulation (a form of TMS) found contrary results in a predominantly male sample. Results showed that inhibition of the left dorsolateral prefrontal cortex (DLPFC) increased aggression compared to inhibition of the right DLPFC (Perach-Barzilay et al., 2013). These results might reflect methodological factors rather than gender effects. For example, the latter study specifically targeted the DLPFC rather than the broader PFC.

Applications of tDCS to the PFC have also found mixed results using healthy samples. One study found that increasing relative left frontal activation increased behavioral aggression after provocation when participants were angry (40 men, 40 women; Hortensius et al., 2012). Riva et al. (2015) found consistent results with Hortensius et al. (2012) in a predominantly female sample (n = 63/80). Specifically, increasing relative right activation in the ventrolateral prefrontal cortex (VLPFC) decreased aggression after social exclusion compared to sham stimulation (Riva et al., 2015). Another study examined 13 men and 19 women and induced right hemispheric dominance via tDCS to the DLPFC (Dambacher et al., 2015a). The stimulation decreased unprovoked aggression, but only in men. The tDCS did not reduce provoked aggression in men or women. A similar study did not find any gender differences when examining the effect of bilateral tDCS on response inhibition or aggression (39 men, 25 women; Dambacher et al., 2015b). Additional research would help to clarify how increasing relative left or right frontal activation impacts anger and aggression and the role of gender in these relationships.

### Bodily Manipulations

Hand contractions and body positioning can induce asymmetric frontal activity. Contracting the left hand increases relative right frontal activity while contracting the right hand increases relative left frontal activity (Harmon-Jones, 2006). In a study of all women (N = 43), following an insult, women who contracted their right hand assigned louder and longer noise blasts to the provocateur than women who squeezed their left hand (Peterson et al., 2008). Relative left frontal activity positively correlated with behavioral aggression for women who squeezed their right hand.

Another study using both men and women found that right hand contractions caused not only greater relative left frontal activity, but also greater self-reported anger in response to ostracism (Peterson et al., 2011). The authors reported that these effects did not differ between men (n = 9) and women (n = 17). Further, in an equal sample of men and women (ns = 23), relative to sitting in an upright position and/or leaning forward, being in a supine position reduced relative left frontal activation in response to an anger-evoking event (Harmon-Jones and Peterson, 2009).

### Summary

Overall, these EEG/ERP and frontal asymmetry manipulation studies provide insight into the neural activation associated with anger and aggression. Although studies often included both men and women, only a select few investigated potential gender differences in these effects. Of those that did, most revealed no differences between men and women and were underpowered. More research is warranted to directly test the nature of gender effects in frontal asymmetry, ERPs, brain stimulation, and bodily manipulations. There is no evidence of robust gender differences in EEG and most studies did not report testing for gender differences.

### Neuroimaging Studies

Several fMRI studies examined neural responses during aggression paradigms in men and women and less commonly, in women only. These studies primarily used the TAP. The methods, analyses, and results differ somewhat from study to study. However, the general consensus is that behaving aggressively activates brain regions associated with negative affect, arousal, cognitive-behavioral control, mentalizing and reward. Specifically, these studies observed activation in the DLPFC, VLPFC, medial prefrontal cortex (MPFC), anterior cingulate (ACC), amygdala, putamen, caudate, thalamus, insula, ventral striatum and hippocampus (Krämer et al., 2007; Lotze et al., 2007; Chester and DeWall, 2016; Emmerling et al., 2016).

None of these studies tested hypotheses about gender differences, but several studies did include both men and women. For instance, one of the first fMRI studies to examine neural activity during the TAP included 11 men and 11 women (Krämer et al., 2007). Another study of 11 women and 9 men found that provocation during the PSAP elicited activation in the ACC, dorsal striatum, insula and PFC (Skibsted et al., 2017). This provocation-related activation correlated with aggressive behavior in the paradigm (i.e., stealing points).

Another study of 30 healthy undergraduate women measured startle eyeblink responses to neutral (e.g., household items) vs. threatening images (e.g., a gun pointed at the participant; Beyer et al., 2014). No men were included in the study. Women with relatively greater startle responses to threatening over neutral images were considered fearful and reactive to threat. Results showed that women with relatively greater startle responses showed lower activation in the brain's mentalizing network, which includes the dorsomedial prefrontal cortex (DMPFC). Corroborating evidence suggests that the DMPFC is positively correlated with aggressive behavior and angry rumination, likely stemming from hostile mentalizing (Lotze et al., 2007; Denson et al., 2009). The authors concluded that women with greater threat reactivity engaged in less mentalizing than women low in threat reactivity. These findings are consistent with meta-analytic reviews showing women's greater feelings of danger and fear when provoked (Eagly and Steffen, 1986; Ito et al., 1996).

Additional fMRI studies examined men and women with borderline personality disorder, which is characterized by reactive aggression (Lieb et al., 2004). A structural MRI study examined the relationships between right and left amygdala volumes with trait aggression in men and women with borderline personality disorder and healthy controls (Mancke et al., 2016). Borderline women reported greater trait aggression than healthy women, but there was no relationship between amygdala volumes and trait aggression in either group of women. By contrast, men showed a positive correlation between right amygdala volume and trait aggression, but only among those diagnosed with borderline personality disorder. Thus, amygdala volume may not be an important factor in aggression among women with borderline personality disorder.

In an fMRI study on borderline personality, women, men and healthy controls engaged in a script-driven imagery task that consisted of two phases (Herpertz et al., 2017). In the anger phase, participants listened to recorded scripts describing harsh interpersonal rejection. Next, in the aggression phase, participants listened to a script describing aggressive behavior. Participants were asked to fully immerse themselves in the scripts. As in the previous study, women with borderline personality disorder reported greater trait aggression and trait anger than healthy women. During both the anger and aggression portions of the task, there were no differences in any of the regions of interest between borderline and healthy women. However, during the aggression phase, women with borderline personality disorder showed positive connectivity with the amygdala and middle cingulate cortex. Men showed the opposite effect; negative connectivity between the amygdala and middle cingulate cortex. Trait anger, but not trait aggressiveness, further strengthened this connectivity in women and weakened it in men. Thus, when imagining an aggressive act, dispositionally aggressive women showed greater amygdala-cingulate connectivity than their male counterparts.

Using another type of social provocation, 15 women and 15 men played a ball tossing game (i.e., Cyberball), ostensibly with two other fictitious same sex participants (Chester and DeWall, 2016). Participants are eventually ignored and left out of the game. This form of ostracism increases anger, aggression and activation in the dorsal anterior cingulate cortex (dACC). In this study, participants completed a measure of trait narcissism followed by playing Cyberball in the scanner. Afterwards outside of the scanner, they were allowed to retaliate via the TAP against one of the two fictitious players. Results showed that the most aggressive participants reported high narcissism and also showed a large increase in the dACC. No gender effects were reported, but they did note that controlling for gender strengthened the effect size of the interaction.

Using the same Cyberball social exclusion method, 20 women and 14 men were either included in the game or excluded (Beyer et al., 2014). Afterwards, participants completed the TAP followed by viewing neutral and emotional scenes. Excluded participants showed heightened activation to emotional social scenes in the brain's mentalizing network, including the DMPFC. In excluded participants, activation in the precentral gyrus in response to viewing emotional scenes mediated the effect of exclusion on aggressive behavior.

#### **Neuroimaging Studies of Substance Use and Aggression**

Researchers are beginning to use fMRI to investigate brain mechanisms responsible for aggression related to alcohol and illicit drugs. Because methamphetamine dependence is associated with increased aggression, Payer et al. (2011) investigated aggression-related neural activity in this population (16 women, 23 men) and healthy controls (18 women, 19 men). Participants completed an affect matching and an affect labeling task. During the affect matching task, participants selected an emotional facial expression that matched a target image. During the labeling task, participants verbally labeled the emotional facial expression. During affect matching, methamphetamine dependent participants showed less activation than controls in the ventral inferior frontal gyrus. During labeling, both dependent and control participants showed increases in the dorsal inferior frontal gyrus and decreases in amygdala activity. Larger amygdala decreases were correlated with lower aggression in the TAP outside of the scanner. Although the authors noted significant gender differences in gray matter volume in the inferior frontal gyrus and amygdala, they did not describe the nature of those differences.

Two fMRI studies investigated the neural correlates of alcohol-related aggression in men and women. In one study, 13 formerly alcohol-dependent participants and 13 controls completed the PSAP in the scanner (Kose et al., 2015). When provoked, control participants showed greater activation in the PFC, thalamus and hippocampus than the formerly dependent group. Independent of group, participants showed negative correlations between the orbitofrontal cortex (OFC), PFC, caudate and thalamus and aggressive behavior. However, these results should be interpreted cautiously as there were only three women in the formerly alcohol dependent group and six in control group.

Another study examined the effects of acute alcohol intoxication on aggression and neural responses (Gan et al., 2015). In that study, 24 healthy young men and 11 women completed the TAP in the scanner once while intoxicated and once after consuming a placebo. Alcohol decreased BOLD responses in the right PFC (i.e., middle frontal and inferior frontal gyri), hippocampus, thalamus, caudate and putamen. Moreover, activity in the amygdala and ventral striatum was not affected by alcohol but was positively correlated with aggression against the provoking opponent. Gender did not influence any of the results, but the authors noted that additional research is needed due to the small sample size. Another study of 12 women and 10 men did not examine aggression but did find that alcohol reduced frontal connectivity in women but not men (Hoppenbrouwers et al., 2010). This frontal dysregulation may be one possible pathway to aggression in women.

### Summary

Neural mechanisms underlying aggression remain poorly understood in women. As most studies did not investigate gender differences and were underpowered, there is not enough evidence of different neural pathways for men and women. The small sample sizes, few women, reliance on the TAP or PSAP, and diverse results preclude firm conclusions at this point. Additional fMRI studies with large samples of men and women and diverse aggression tasks are needed.

### Hormones

In the realm of aggressive behavior, testosterone, cortisol, estradiol, progesterone and oxytocin have been studied extensively in non-human animals, but less so in humans. In this section, we review the evidence on the relationships between these hormones and aggression in women.

### **Testosterone and Cortisol**

In mammalian species, males generally have higher testosterone levels and are more aggressive than females. Similarly, because men are more violent than women globally and men possess much higher testosterone concentrations than women, researchers suspected that testosterone is a strong cause of aggression in men. However, much less research has investigated this possibility in women. One study of 87 women inmates in a maximum-security prison found that testosterone levels correlated with aggressive dominance in prison (Dabbs and Hargrove, 1997). This relationship was reduced among older women, presumably due to lower levels of testosterone. Similarly, a study of a women's rugby team found that the pre-game rise in testosterone was positively correlated with aggressiveness during the game (Bateup et al., 2002). Another correlational study measured testosterone in 155 men and 151 undergraduate women (Harris et al., 1996). Men reported greater aggression than women and had five times more testosterone than the women. Despite these mean differences, the authors found positive correlations between testosterone and self-reported aggression in both women and men. Thus, although aggression and testosterone may be lower in women than men, many studies observed the same positive relationships between testosterone and aggression in women as they do in men (e.g., Prasad et al., 2017; Probst et al., 2018). A study of 12 women in a double-blind placebo-controlled testosterone administration study suggests that testosterone may increase aggression because it reduces sensitivity to punishment and increases reward sensitivity (van Honk et al., 2004).

A meta-analysis revealed that the correlations between testosterone and aggression were small, but significant in both men (r = 0.08) and women (r = 0.13; Archer et al., 2005). Thus, the relationship between testosterone and aggression is not particularly strong in humans. Indeed, a review of the literature suggested that testosterone should be considered as promoting dominance seeking behavior, rather than solely aggression (Eisenegger et al., 2011).

In order to explain these weak correlations between testosterone and aggression, researchers examined cortisol as a moderator of this relationship. The dual hormone hypothesis suggests that low cortisol facilitates the potentiating effect of testosterone on aggressive and dominant behavior, whereas high cortisol blocks this effect (Mehta and Prasad, 2015; for a similar notion using the ratio of testosterone to cortisol, see Terburg et al., 2009). This pattern of data has been observed in forensic samples of men and boys (Dabbs et al., 1991; Popma et al., 2007), but evidence is mixed in women. For instance, one study of 53 healthy undergraduate women found the opposite pattern; women with high concentrations of both salivary testosterone and cortisol showed the most aggression in the TAP (Denson et al., 2013). Other studies failed to find support for the dual hormone hypothesis in women (Cote et al., 2013; Geniole et al., 2013; Welker et al., 2014; Buades-Rotger et al., 2016). However, a recent study of 326 adolescent girls and 134 boys found that testosterone derived from hair samples correlated with self-reported aggression at low levels of cortisol in both boys and girls (Grotzinger et al., 2018). Estimates derived from hair samples may reflect stable trait-like individual differences in cortisol and testosterone more so than values derived from saliva. Thus, these data suggest that interactions between testosterone and cortisol may influence aggression in women. However, more research is needed with large samples and behavioral measures of aggression.

The dual hormone serotonergic hypothesis goes one step further by positing that the interactive relationship between testosterone and cortisol on aggression is further moderated by serotonin availability (Montoya et al., 2012). Specifically, high testosterone, low cortisol, and low serotonin are thought to increase risk for aggression. One study did examine the interactive effects of testosterone and serotonin on trait aggression in 24 women and 24 men (Kuepper et al., 2010). Participants provided testosterone samples over 3 days and subsequently received S-citalopram. The dependent variable was trait aggression. S-citalopram influences serotonin and cortisol. A large vs. small cortisol response to the drug is thought to indicate high vs. low 5-HT availability, respectively. Only men showed the expected high testosterone-low serotonin interaction on trait aggression. Unexpectedly, they also found a low testosterone-high serotonin interaction. Thus, more research is needed to verify the robustness of these results and their applicability to women.

Although most research on hormones and aggression is correlational, some researchers have conducted placebocontrolled experiments. In one such study 24 women and 24 men were administered cortisol or a placebo and subsequently exposed to strong or weak provocation within the TAP (Böhnke et al., 2010). Cortisol increased aggression in women but not men, but only during the most provocative trials of the TAP. Results should be interpreted cautiously due to small cell sizes.

Other research investigated relationships between hormones and neural activity. For instance, Mehta and Beer (2010) found that in a sample of 17 men and 15 women, endogenous testosterone positively correlated with aggression during the Ultimatum Game and negatively with bilateral medial OFC activation. Medial OFC activity statistically mediated the relationship between testosterone and aggression. There were no differences between men and women. However, another fMRI study found a negative relationship between testosterone and aggression in an all-female sample of 39 undergraduates (Buades-Rotger et al., 2016). In that study, participants were exposed to an opponent's angry face or neutral face followed by provoking noise blasts. Testosterone was negatively correlated with amygdala reactivity to the trials with an angry face. Thus, much more research is needed on hormones and neural responses before firm conclusions can be made about these mechanisms in women.

### **Estradiol and Progesterone**

In women, the two ovarian hormones estradiol and progesterone reliably fluctuate during the menstrual cycle. Peak fertility is characterized by high levels of estradiol and low levels of progesterone. Gladue (1991) examined the relationships between estradiol, testosterone, and trait aggression in a matched sample of heterosexual and same-sex attracted men and women. Regardless of sexual orientation, both testosterone and estradiol positively correlated with trait aggression in men; for women, these correlations were negative. Another study of 49 undergraduate women found no relationship between testosterone and trait aggression but replicated the negative relationship between estradiol and trait aggression (Stanton and Schultheiss, 2007).

In another study, 34 undergraduate women kept diaries of competition-related conflict and how they dealt with it (Cashdan, 2003). Women relatively high in testosterone were more likely to resolve the conflict with verbal aggression. Estradiol was unrelated to aggression. Similarly, a study of 33 bulimic women and 23 healthy controls in the early follicular phase of the menstrual cycle reported a positive association between testosterone and trait aggression, but only in the bulimic group (Cotrufo et al., 2000). No correlations were found between estradiol, prolactin and cortisol in either group.

Collectively, these data suggest that endogenous estradiol may be either unrelated or negatively related to aggression in women. However, estradiol may be involved in dominance, assertiveness, and risk-taking in women rather than aggression. Estradiol is positively correlated with implicit power motivation (for a replication see, Stanton and Edelstein, 2009). Similarly, we found that high estradiol and low progesterone was associated with heightened assertiveness in women (Blake et al., 2017). High levels of free estradiol were positively correlated with both aggressive and non-aggressive risk-taking (Vermeersch et al., 2008).

Relatively few studies tested the hypothesis that progesterone would be related to aggression. Ritter (2003) measured trait aggression in 29 healthy undergraduate women during menses and again during the midluteal phase. Progesterone and estrogen are higher during the midluteal phase than during menses. Women reported less trait physical and verbal aggression during the midluteal phase than during menses. However, this study did not directly measure hormones so it is unclear whether the menstrual cycle effect on trait aggression was due to estradiol, progesterone, or both.

Another study measured estradiol, progesterone and testosterone across the menstrual cycle in 15 healthy women (Brambilla et al., 2010). They found positive correlations between estradiol and verbal aggression during the follicular phase, when progesterone and estradiol are low. Testosterone was uncorrelated with hostility and aggression. They also found a negative correlation between progesterone and two components of trait hostility (i.e., suspiciousness and resentment) in the luteal (premenstrual) phase. This finding was conceptually replicated in a larger sample of 122 women (Ziomkiewicz et al., 2012). They found that higher levels of progesterone during the luteal phase were associated with lower self-reports of aggression and irritability. Thus, greater progesterone may reduce hostility and aggression during the luteal phase, whereas low levels of progesterone may increase risk for aggression.

Simultaneously low levels of progesterone and estradiol may increase self-directed aggression. Indeed, one study examined estradiol and progesterone in 281 fertile women within 24 h after attempting suicide (Baca-Garcia et al., 2010). Suicide attempts were more likely during periods of low estradiol and progesterone. Thus, progesterone may be protective against both other-directed and self-directed aggression. One possibility is that progesterone may be associated with improved emotion regulation capacity. In an attempt to determine how high levels of progesterone may aid emotion regulation, 18 healthy women completed an emotion matching task during fMRI with angry and fearful faces (van Wingen et al., 2008). Relative to placebo, a single progesterone administration increased amygdala activity and connectivity between the amygdala and dACC. This latter finding raises the possibility of progesterone assisting emotion regulation via connectivity between the dACC and amgydala (for a review of neuroimaging findings, see Toffoletto et al., 2014).

### **Oxytocin**

Although sometimes referred to as the ''love hormone'' or ''bonding hormone'', the nonapeptide oxytocin may also increase aggressive behavior. Most studies examining oxytocin have either intranasally administered the hormone or a placebo. Less frequently, researchers obtain endogenous levels via lumbar puncture. One study found that oxytocin levels measured in the cerebrospinal fluid were negatively correlated with trait aggression in women (n = 13; Lee et al., 2009). Similarly, Campbell and Hausmann (2013) found that oxytocin relative to placebo lowered aggression on the PSAP, but only among women who were feeling anxious.

Breastfeeding women typically have high levels of oxytocin. One laboratory study using the TAP found that breastfeeding women were more aggressive than formula feeding women and nulliparous women (Hahn-Holbrook et al., 2011). The greater aggression in breastfeeding women relative to the other women was due to lowered stress responses to provocation among the breastfeeding women. Thus, oxytocin may facilitate aggression by lowering perceptions of danger that normally inhibit many women from retaliating (Bettencourt and Miller, 1996). Thus, oxytocin may both increase and decrease aggression via reduced anxiety.

Consistent with this possibility, an fMRI study of 38 women with borderline personality disorder and 41 healthy women were given oxytocin or a placebo (Bertsch et al., 2013). They then classified emotional facial expressions while in the scanner. Relative to the borderline women in the placebo group, borderline women given oxytocin showed reduced threat sensitivity to angry faces and lower amygdala activation. These findings are consistent with the studies showing anxiolytic effects of oxytocin in women and the possibility that oxytocin influences aggression via reduced fear (Campbell, 2008).

In order to make sense of conflicting results of oxytocin on social behavior, Shamay-Tsoory and Abu-Akel (2016) proposed the social salience hypothesis. The idea is that oxytocin enhances the perception of social stimuli; thus, enhancing responses to both positive and negative (e.g., provocation) social stimuli. In this way, provoking individuals should be perceived as more hostile following oxytocin administration. A recent study found support for the social salience hypothesis in a laboratory experiment of 28 men and 20 women (Ne'eman et al., 2016). Using a modified version of the PSAP, participants could behave selfishly, cooperatively, or aggressively. Relative to placebo, oxytocin selectively increased aggressive responses. The authors found no gender differences.

Consistent with the social salience hypothesis, other work suggests that oxytocin may increase IPV. In a placebo-controlled experiment, 46 women and 47 men received oxytocin or placebo, after which they completed a physical pain task and received negative social feedback on a speech (DeWall et al., 2014). Next, they reported on how likely they would be to commit physical IPV against their current partner (or former partner for the single participants). Results showed that oxytocin increased IPV inclinations, but only for those high in trait aggression. Women reported greater IPV inclinations than men, but gender did not interact with the oxytocin manipulation. The authors suggested that people high in trait aggression may engage in more IPV as a controlling tactic when experiencing negative affect. However, there is another plausible alternative explanation that is consistent with the social salience hypothesis. Oxytocin may have enhanced the subjective impact of the pain and negative feedback. Among people high in trait aggression, who tend to have a hostile world view, this greater oxytocin-induced impact may have facilitated greater inclinations towards IPV (Buss and Perry, 1992).

### Summary

This brief review of five hormonal mechanisms underlying aggression in women suggests few clear findings. As with men, the positive relationship between testosterone and aggression in women is small. The dual hormone hypothesis has had some success in predicting aggression in men, but less so in women. The data on estradiol and progesterone are suggestive of the possibility that high levels of these hormones reduce aggression and self-directed harm in women. However, much more work is needed. The literature on oxytocin suggests that the hormone can decrease and increase aggression in women. Increases in aggression are likely due to a combination of the hormone's anxiolytic effects as well as enhanced reactivity to provocation. The social salience hypothesis provides a promising framework from which to test specific predictions about conditions under which oxytocin enhances or inhibits aggression in women.

### DISCUSSION

In this review, we examined the numerous behavioral expressions of aggression that women engage in along with the early developmental, neural, and hormonal correlates. The factors are summarized in **Figure 1**. Our review highlights that relative to men's aggression, we know little of the underpinnings of women's aggression. Most studies on brain and hormonal mechanisms of aggression included only men, did not examine gender differences, or did so in a post hoc manner, and/or relied on small samples. Thus, there is little opportunity to make robust conclusions about how the processes reviewed here influence aggression in women. By contrast, the behavioral data are clear in that women tend to engage in predominantly indirect aggression, IPV with equal frequency but lesser severity than men, and rarely sexual aggression. Thus, our review is in accord with Richardson (2005), who noted that women are quite capable of aggression. Nonetheless, the limitations of the extant data provide opportunities for future research testing novel hypotheses. We urge more theoretical development to derive a priori gender-specific predictions about the mechanisms underlying aggressive behavior in women.

### Future Directions

There are a number of unknown aspects about the causes and nature of women's aggression. For instance, little is known about aggression in same-sex attracted women. Relative to men, the perpetration of sexual aggression in women remains poorly understood as well. Sexual aggression committed by women is a relatively low frequency behavior and victims are unlikely to report its occurrence. These issues make it a difficult phenomenon to study. Nonetheless, both men and women victims of sexual violence show the same negative psychological outcomes, making all forms of sexual violence worthy of further study. Laboratory sexual aggression paradigms developed for women would be informative (see Davis et al., 2014).

Our review of neural correlates of aggression also showed no convincing evidence of divergent pathways for men and women. Most of the EEG/ERP, brain stimulation, and fMRI studies that included men and women did not report testing for gender differences or did not find any. The role of hormones in determining women's aggression was also largely unclear, but worthy of future study as theoretical development in this area is becoming increasingly sophisticated (Mehta and Prasad, 2015; Shamay-Tsoory and Abu-Akel, 2016). Since fear plays a significant role in women's reaction to provocation and subsequent aggression (Eagly and Steffen, 1986; Bettencourt and Miller, 1996), brain regions involved in fear processing and arousal (e.g., amygdala, hypothalamus) seem like promising regions of interest.

One limitation of the laboratory and brain research on women's aggression is the reliance on the TAP and PSAP as the primary measures of aggression. Although well-validated, both involve direct retaliation toward the provocateur. Women tend to engage in indirect aggression to a greater extent than direct aggression. Thus, it is unclear to what extent the laboratory work represents realistic behavior in women. Development of indirect aggression paradigms for the laboratory would facilitate greater understanding as would field experimentation.

We have also left out a discussion of genetic influences. Aggression is highly heritable, and in the past several years, a number of candidate genes such as MAOA and 5-HTTLPR have been identified as conferring risk for aggression, impulsivity, and emotion regulation deficits (Ficks and Waldman, 2014). Similarly, the field of epigenetics has much to offer, especially if we are to understand women's aggression across the lifespan (Waltes et al., 2016). Optogenetic technology in animal models also holds promise. For instance, optogenetic stimulation of neurons in the hypothalamus caused male mice to attack females, males, and inanimate objects (Lin et al., 2011). Using optogenetics holds promise for understanding some of the brain processes that may heighten female aggression.

Although it was outside of the scope of this review, all the mechanisms we discussed here are mediated via neurobiological processes that we did not discuss. For instance, serotonin has been robustly implicated in aggression and is affected by prenatal smoking and maternal malnutrition (Liu, 2011). There are no doubt many mediating processes at various levels of specificity that remain to be explored.

### CONCLUSION

Aggression is a complex social behavior that has been extensively studied in men. Comparatively, women's aggression has been neglected. We suggest that there is a need for more theory-driven research in the investigation of aggression in women. Such work could contribute to the development of more effective evidencebased treatments that target gender-specific risks for aggression.

### AUTHOR CONTRIBUTIONS

TFD drafted the sections on laboratory aggression, sexual aggression, prenatal influences, neuroimaging and hormones. He also drafted the general discussion. SMO drafted the intimate partner violence section. JRB drafted the EEG, brain stimulation and bodily manipulations sections. KRB wrote portions that appear throughout the manuscript. All authors provided critical revisions and contributed to theoretical development.

### REFERENCES


### FUNDING

Preparation of this manuscript was supported by a grant from the Australian Research Council (FT140100291) to TFD. SMO and JRB were supported by Australian Postgraduate Awards.


Drug Alcohol Depend. 118, 470–474. doi: 10.1016/j.drugalcdep.2011. 03.004


disorder. Psychoneuroendocrinology 34, 1567–1573. doi: 10.1016/j.psyneuen. 2009.06.002


Aggress. Behav. 24, 1–8. doi: 10.1002/(sici)1098-2337(1998)24:1<1::aid-ab1>3. 3.co;2-f


prenatal cocaine. Neurotoxicol. Teratol. 27, 191–201. doi: 10.1016/j.ntt.2004. 10.005


**Conflict of Interest Statement**: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Denson, O'Dean, Blake and Beames. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Genetic and Psychosocial Predictors of Aggression: Variable Selection and Model Building With Component-Wise Gradient Boosting

Robert Suchting<sup>1</sup> \*, Joshua L. Gowin<sup>2</sup> , Charles E. Green<sup>3</sup> , Consuelo Walss-Bass <sup>1</sup> and Scott D. Lane1,2

<sup>1</sup>Department of Psychiatry and Behavioral Sciences, McGovern Medical School, University of Texas, Houston, TX, United States, <sup>2</sup>Section on Human Psychopharmacology, National Institute on Alcohol Abuse and Alcoholism, Rockville, MD, United States, <sup>3</sup>Center for Clinical Research & Evidence-Based Medicine, Department of Pediatrics, McGovern Medical School, University of Texas, Houston, TX, United States

Rationale: Given datasets with a large or diverse set of predictors of aggression, machine learning (ML) provides efficient tools for identifying the most salient variables and building a parsimonious statistical model. ML techniques permit efficient exploration of data, have not been widely used in aggression research, and may have utility for those seeking prediction of aggressive behavior.

Objectives: The present study examined predictors of aggression and constructed an optimized model using ML techniques. Predictors were derived from a dataset that included demographic, psychometric and genetic predictors, specifically FK506 binding protein 5 (FKBP5) polymorphisms, which have been shown to alter response to threatening stimuli, but have not been tested as predictors of aggressive behavior in adults.

### Methods: The data analysis approach utilized component-wise gradient boosting and model reduction via backward elimination to: (a) select variables from an initial set of 20 to build a model of trait aggression; and then (b) reduce that model to maximize parsimony and generalizability.

Results: From a dataset of N = 47 participants, component-wise gradient boosting selected 8 of 20 possible predictors to model Buss-Perry Aggression Questionnaire (BPAQ) total score, with R <sup>2</sup> = 0.66. This model was simplified using backward elimination, retaining six predictors: smoking status, psychopathy (interpersonal manipulation and callous affect), childhood trauma (physical abuse and neglect), and the FKBP5\_13 gene (rs1360780). The six-factor model approximated the initial eight-factor model at 99.4% of R 2 .

Conclusions: Using an inductive data science approach, the gradient boosting model identified predictors consistent with previous experimental work in aggression; specifically psychopathy and trauma exposure. Additionally, allelic variants in FKBP5 were identified for the first time, but the relatively small sample size limits

#### Edited by:

Nelly Alia-Klein, Icahn School of Medicine at Mount Sinai, United States

#### Reviewed by:

Joshua W. Buckholtz, Vanderbilt University, United States Dan Zhang, Tsinghua University, China

> \*Correspondence: Robert Suchting robert.suchting@uth.tmc.edu

Received: 27 November 2017 Accepted: 20 April 2018 Published: 07 May 2018

#### Citation:

Suchting R, Gowin JL, Green CE, Walss-Bass C and Lane SD (2018) Genetic and Psychosocial Predictors of Aggression: Variable Selection and Model Building With Component-Wise Gradient Boosting. Front. Behav. Neurosci. 12:89. doi: 10.3389/fnbeh.2018.00089 generality of results and calls for replication. This approach provides utility for the prediction of aggression behavior, particularly in the context of large multivariate datasets.

Keywords: aggression, FKBP5, trauma, psychopathy, boosting, machine learning, data science

### INTRODUCTION

Aggression is a complex multifaceted phenomenon (Anderson and Bushman, 2002; Raine, 2002; Mendes et al., 2009) that is influenced by many factors. Understanding and prediction of aggression must account for this complexity in order to extract a meaningful signal from amidst considerable noise. Key factors include: developmental history—notably childhood trauma (Caspi et al., 2002; Gowin et al., 2013; Milaniak and Widom, 2015) presence of psychopathology (Glenn and Raine, 2009; Alcorn et al., 2013; Anderson and Kiehl, 2014); externalizing personality traits (Gardner et al., 2015; Pasion et al., 2017); emotional and inhibitory dysregulation (Gao et al., 2015; Coccaro et al., 2016; Hsieh and Chen, 2017); biological factors, including genetic variation (Tuvblad and Baker, 2011; Bevilacqua et al., 2012; Takahashi et al., 2012; Dorfman et al., 2014); and contextual/situational factors such as substance use and provocation (Miczek et al., 2002; Cherek et al., 2006; Giancola et al., 2009; Skibsted et al., 2017).

Science has traditionally progressed via isolation of and emphasis on individual variables in the tradition of hypothesis testing and frequentist statistical inference, while fewer studies have utilized discovery-based, data science approaches in the study of aggressive behavior (but see Ang and Goh, 2013; Carré and Olmstead, 2015; Rosellini et al., 2016). As data science has become more established and widely utilized in scientific discovery and prediction (Hastie et al., 2009; Hofman et al., 2017; Wiens and Shenoy, 2018), novel inductive analytic techniques have enabled and advanced the analysis of complex, multivariate data. These approaches include mining of very large datasets, as well as application to smaller datasets where large amounts of information are obtained from each individual, but the dataset contains a relatively small number of subjects. In the present study, we utilized a data science approach to examine predictors of trait aggression, including interpersonal and demographic variables, history of trauma, psychopathology and genetic variations in the FK506 binding protein 5 (FKBP5) protein.

The FK506 binding protein 51 (FKBP5) is a glucocorticoidrelated chaperone and immunophilin protein that plays a role in immune system function. Relevant to the present report, FKBP5 is implicated in emotional dysregulation. Specifically, certain FKBP5 variants appear to modulate clinically relevant aspects of mood and behavior in the context of childhood trauma and post-traumatic stress disorder (Klengel et al., 2013; Klengel and Binder, 2015; Zannas et al., 2016), as well as other stress-related pathologies via interaction with the glucocorticoid receptor (Bevilacqua and Goldman, 2011; Zannas et al., 2016). For example, FKBP5 gene × environment interactions play a role in depression (Gillespie et al., 2009; Appel et al., 2011; Tozzi et al., 2016), and—relevant to the present report—aggressive behavior in children (Bevilacqua et al., 2012; White et al., 2012; Bryushkova et al., 2016). Importantly, genetic variation for FKBP5 has not been tested as a predictor of aggressive behavior in adults. Thus, we examined three FKBP5 single nucleotide polymorphisms (SNPs) commonly implicated in stress-related emotional dysregulation.

As described above, variable selection for the present study was governed by factors with known associations to anger, inhibitory control, and aggressive behavior. However, our data science-informed analytic approach (described below) should be understood as quasi-exploratory rather than driven by traditional hypothesis testing. The primary goals were to: (1) determine which of the known predictors of aggression were most important; and (2) to examine the contribution of a hypothesized genetic variant toward trait aggression. Machine learning (ML) was used to explore these goals without overfitting the trait aggression outcome, measured here by the Buss-Perry Aggression Questionnaire (BPAQ; Buss and Perry, 1992).

### MATERIALS AND METHODS

### Participants

Forty-eight participants were recruited from the greater Houston metropolitan area using local newspaper and radio advertisements, as part of a larger experimental study described in Gowin et al. (2013) and summarized below. This study was carried out in accordance with the recommendations of the Belmont Report and the University of Texas Health Science Center Committee for the Protection of Human Subjects (IRB), with written informed consent from all subjects. All subjects gave written informed consent obtained in person in accordance with the Declaration of Helsinki. The protocol was approved by the University of Texas Health Science Center Committee for the Protection of Human Subjects. For the present analyses, participants provided demographic information, psychometric data and saliva samples. K-nearest neighbors imputation was used to fill in a small amount (<2.5%) of missingness in the data on the child trauma questionnaire (CTQ) and Shipley II predictors.

### Design

The present study was derived from a larger, laboratory-based experimental study in which adult participants were given acute dose of 20 mg cortisol or placebo, and measures of salivary cortisol and state aggression (Point Subtraction Aggression Paradigm) were taken over a 5-h testing period (Gowin et al., 2013). To increase the likelihood of including participants with histories of trauma and heightened aggression, we advertised for individuals on parole or probation. We have used this strategy in several previous studies of childhood trauma and/or aggression (Gowin et al., 2010, 2013; Alcorn et al., 2013). However, we did not specify participant selection based on any DSM diagnostic and psychometrically-established clinical cut-offs for trauma exposure. In addition to the experimental procedures, measures of trait aggression, childhood trauma, and psychopathy were obtained at baseline from all participants. Additionally, at baseline a subset of 48 participants provided demographic information and saliva samples for genetic testing focused on FKBP5 Summarized below, the baseline measures collectively formed the dataset for the present analyses.

### Measures

### Demographics

Following from established associations described in the introduction and based on baseline demographic variables collected in the Gowin et al. (2013) study age, education, ethnicity, sex and smoking status were included as demographic predictors in the present study.

### Buss-Perry Aggression Questionnaire (BPAQ; Buss and Perry, 1992)

This measure of aggression features four subscales derived from factor analysis: physical aggression, verbal aggression, hostility and anger. It is a widely used psychometric measure of aggression, employed across a range of contexts and populations of interest. The dependent variable used in the present analyses was BPAQ total score, calculated by summing the standardized scores on the constituent subscales of the BPAQ. The BPAQ has strong psychometric properties (Buss and Perry, 1992; Harris, 1997), and use of the total score is established in previous studies of aggression (Moeller and Dougherty, 2001; Palmer and Thankordas, 2005; Gowin et al., 2013). The sum of the four factor scores results in a total aggression score. The BPAQ total score was used as the primary outcome.

### Child Trauma Questionnaire (CTQ; Bernstein and Fink, 1998)

The CTQ is a 28 item self-report Likert-type scale of maltreatment during childhood. The instrument consists of five subscales: physical abuse, physical neglect, emotional abuse, sexual abuse and emotional neglect). The CTQ is a 28 item self-report Likert-type scale of maltreatment during childhood. The instrument consists of five subscales: physical abuse, physical neglect, emotional abuse, sexual abuse and emotional neglect). It is perhaps the most common psychometric instrument used in the measurement of childhood trauma exposure (Viola et al., 2016).

### Impulsive/Premeditated Aggression Scale (IPAS;

#### Stanford et al., 2003)

The impulsive/premeditated aggression scale (IPAS) is a 30 item self-report measure that classifies aggression into two sub-scales, premeditated and impulsive. It has measurement sensitivity related to history of violence, trauma and aggression-related personality characteristics (Stanford et al., 2008; Teten et al., 2008). Scores from the two subscales were used as independent predictors in the present analysis.

### Self-Report Psychopathy Scale III (SRP-III; Neumann et al., 2012)

The self-report psychopathy scale III (SRP-III) is a Likert-type scale of psychopathy, measured on a scale from 1 to 5. The measure consists of four subscales: callous affect, erratic lifestyle (ELS), criminal tendencies and interpersonal manipulation. The instrument is sensitive in both normative samples and populations with externalizing psychopathology related to aggression (Alcorn et al., 2013). Scores from each subscale were used as independent predictors in the present analysis.

### Shipley II Test of Cognitive Aptitude (Shipley et al., 2009)

The Shipley II is a measure of cognitive aptitude that correlates highly with general intelligence scales. The test construction used in the present study consisted of one 40-item verbal subscale (vocabulary) and one 20-item non-verbal subscale (block patterns). A composite score is derived from the two subscales and provides an index of overall cognitive ability. The composite score was used in the present data analyses.

### FK506 Binding Protein 5 (FKBP5 Gene)

Genomic DNA was extracted from saliva Oragene DNA collection kits using the prepIT DNA extraction kit (DNA Genotek Inc, Ottawa, ON, Canada). Allelic discrimination for the FKBP5 SNP was performed using the Taqman 5'nuclease assay (Life Technologies, Grand Island, NY, USA). All samples were run in duplicate. Genotypes were determined using the ABI 7900HT SDS 2.2.2 software adapted in the ABI 7900HT Sequence Detection System. Based on previous work outlined in the introduction, the following SNPs were examined: FKBP5\_13 (rs1360780); FKBP5\_92 (rs9296158); and FKBP5\_94 (rs9470080).

### Data Analytic Strategy

The present analysis utilized component-wise gradient boosting to develop an optimal model to predict aggression from the baseline set of 20 predictors (see **Table 1**). The optimal model was then simplified to maximize parsimony using a process called model reduction. Details of these techniques follow. All predictors were standardized by z-score before analysis to place them on a comparable metric and provide estimates of the relative influence of the predictor variables. The trait aggression outcome was left in its raw unstandardized metric to ease interpretability in raw units of the BPAQ score. This two-stage model building process has shown success in determining the best predictors of smoking lapse during a quit attempt (Suchting et al., 2017) as well as choosing the strongest inflammatory markers predicting depression in adolescents over time (Walss-Bass et al., 2018).

### Component-Wise Gradient Boosting

Component-wise gradient boosting is a ML technique for statistical model estimation that iteratively builds a strong prediction model from an ensemble of weak prediction models via gradient descent (Bühlmann and Hothorn, 2007). The technique seeks to model the relationship between some outcome



Frequencies (%) and mean (SD) are provided for each predictor.

(here, aggression) and a set of predictors using an algorithm that optimizes a loss function (e.g., for generalized linear models, the negative log-likelihood function). This algorithm is implemented in the mboost package in R (Hofner et al., 2014; Hothorn et al., 2016). In brief, the algorithm works as follows: (1) initialize an estimate of a function to fit the outcome with offset values; (2) specify a set of ''base learners'' (simple regression estimators); (3) compute the negative gradient of the loss function, fit each of the base learners separately to the negative gradient vector, select the best-fitting base-learner, and update the current function estimate with a shrinkage penalty; and (4) repeat step 3 until a stopping iteration (chosen via bootstrap or cross-validation) is met. While the algorithm could conceivably run until convergence, a stopping iteration mstop is established in order to prevent overfitting and lower prediction accuracy. Tuning mstop to some finite value results in an implicit variable selection property, as only one base learner is selected during each iteration. Further, the use of a shrinkage penalty in model fitting provides L1-penalized model coefficients.

Penalization supplies decreased variability of model estimates at the cost of slightly increased bias and helps alleviate problems of collinearity (Kuhn and Johnson, 2013). More complex models with a large number of predictors P relative to the number of participants in the sample N may have unstable and inflated parameter estimates due to increasing inter-correlations among predictors (collinearity). The mboost algorithm optimizes prediction by removing predictors via variable selection and by using penalization to counter inflated parameter estimates that result from collinearity. The generalized linear/additive model building process also results in readily interpretable models, as opposed to many other ML algorithms that may generate interpretation-resistant or ''black box'' predictions.

#### Model Reduction

The final optimized model chosen via component-wise gradient boosting features regularized parameter estimates and inherent variable selection. This model may then be simplified to maximize parsimony at the expense of pure predictive power and increased bias in estimation in a process called model reduction. To find the most parsimonious model, we engage in backward elimination from the optimized model fit in mboost. Backward elimination is an exploratory stepwise procedure that begins with all of the variables in the optimized model fit by mboost and tests the fit of the model (measured by Akaike information criteria, or AIC) by the deletion of each variable. The variable (if any) that most improves the model by being deleted is then removed. This process is repeated until further deletion does not improve the model. A simplified model that retains around 95% of the fit (e.g., via R 2 ) of the full model may be considered a successful approximation (Ambler et al., 2002; Harrell, 2015). Reduction may also result in a model with a more attractive parameter-to-sample size ratio. For the present analysis, backward elimination is performed using the StepAIC() function in the MASS package in R (Venables and Ripley, 2002; R Core Team, 2017).

### RESULTS

**Table 1** provides summary statistics for all demographic, psychometric and FKBP5 predictors included in the model. The sample was largely male (77%) and African American (77%). FKBP5 allele distributions did not deviate from Hardy-Weinberg equilibrium. The mean BPAQ score was 64.04 (SD = 19.78, range = 32–111). This is comparable to previous studies in our lab examining individuals with a history of SUD and ASPD (Gowin et al., 2010, 2013; Alcorn et al., 2013). Across those studies, the mean BPAQ value = 67.44 (SD = 15.95, range = 40–124).

### Component-Wise Gradient Boosting

The mboost() function was used to derive an optimal model fitting BPAQ total score to a set of 20 candidate base-learners. Tuning the optimal number of boosting iterations by 10-fold cross-validation resulted in mstop = 38. The resultant model retained 8 of the 20 predictors and yielded an R <sup>2</sup> = 0.66. Standardized penalized coefficients for these predictors are included in **Table 2**. These coefficients included smoking status, FKBP5\_13 allelic variants C/T and T/T, and several subscales from the CTQ (trauma) and SRP3 (psychopathy) measures. For this eight-factor model the three strongest predictors were the three retained subscales from the SRP3 psychopathy measure. These measures were related to increases in BPAQ total score of 7.24, 3.27 and 2.25 points for one standard deviation increases in callous affect, ELS and criminal tendencies, respectively.

TABLE 2 | Parameter estimates of the optimized model derived by the mboost algorithm, based on the original 20 predictor variables with Buss-Perry Aggression Questionnaire (BPAQ) total score as the outcome variable, ranked by absolute value.


Predictors were z-scored before estimation; BPAQ total score was measured in raw units. R<sup>2</sup> = 0.651. Note: FKBP5\_13 = rs1360780; CTQ, Childhood Trauma Questionnaire (PA, physical aggression; PN, physical neglect); SRP3, Self-Report of Psychopathy (IM, interpersonal manipulation; CA, callous affect; ELS, erratic lifestyle; CT, criminal tendencies).

### Model Reduction With Elimination

Results of the model reduction using the backwards elimination technique from the full penalized eight-factor model are shown in **Table 3**. For model comparison purposes, the variables selected by the mboost algorithm were refit in an unpenalized model before backward elimination. Backwards elimination shifted R 2 from 71.8 to 71.4, thus approximating 99.4% of the R 2 (the coefficients from the backward elimination process are unpenalized and yield a different basis for R 2 from the boosted model). The model was highlighted by the following relationships: active smoking was associated with higher trait aggression; having the FKBP5\_13 T/T allele was associated with lower trait aggression relative to having the FKBP5\_13 C/C allele (reference contrast); CTQ history of childhood physical abuse was associated with higher trait aggression while history of physical neglect was associated with lower aggression; and SRP3 callous affect was associated with higher trait aggression. While model parameters from stepwise selection are inherently biased (coefficients may be inflated), bootstrap standard errors and 95% confidence intervals are provided to ensure maximum possible robustness of statistical inferences. **Table 3** describes parameter estimates for the reduced model. The strongest effects found in the reduced model demonstrated that a one standard deviation increase in callous affect was related to a 10.7 point increase in BPAQ total score and that presence of the T/T allele (as compared to the C/C allele) was related to a 10.7 point decrease in BPAQ total score.

### DISCUSSION

The present report used the mboost technique with subsequent backward elimination to determine a parsimonious set of predictors of trait aggression, highlighted by associations with callous affect, childhood trauma and FKBP5\_13 alleles. While our analytic approach was not hypothesis-driven, these predictors correspond with the broader extant literature on human aggression. Both childhood trauma and callous unemotional traits are robustly associated with aggression and related conduct problems during adolescence and adulthood (Hare and Neumann, 2009; Frick and Ray, 2015; Milaniak and Widom, 2015; Gillikin et al., 2016). Moreover, there is growing empirical support that the FKBP5 gene plays a key role in the modulation of the stress response and the regulation of emotion, including risk for aggressive behavior (Klengel et al., 2013; Bryushkova et al., 2016), and the present study is the first to demonstrate this relationship in adults, and the first to demonstrate an association between aggression and the T allele of rs1360780. While beyond the scope of the present data, it is possible that the predictive utility of FKBP5 and CTQ abuse variables result from the presence of a gene × environment phenotype (Tuvblad and Baker, 2011).

The mboost technique is a modern hybrid approach that sits in between traditional generalized linear models and ML approaches that model interactions of higher-order complexity (Hothorn et al., 2016). Supervised ML techniques, including ensemble boosting and bagging approaches like mboost (Bühlmann and Hothorn, 2007), offer utility in identifying relationships among complex, multifactorial phenomena that define many human behaviors, such as violence and aggression. Such analytic approaches provide advantages to modern translational research that seeks to integrate across diverse sources of high-dimensional data, for example genetics, neuroimaging and psychometrics. In the present context, these techniques provide automated optimization of a predictive regression model for an outcome of interest, such as aggression. As opposed to traditional statistical analyses, these algorithms can maximize the utility of available data without ''data dredging, '' whereby many relationships between variables



Note: FKBP5\_13 = rs1360780 (PA, physical aggression; PN, physical neglect); CTQ, Childhood Trauma Questionnaire; SRP3\_CA, Self-Report of Psychopathy Callous Affect.

are examined in an exhaustive yet unsystematic fashion, and only the significant relationships are reported. Such research products represent part of the current controversy surrounding poor replication of findings in the behavioral sciences. Here we fully acknowledge the limitations of the modest amount of available data, using ML to optimize the statistical modeling of that data, and providing incremental knowledge gained. Accordingly, the present findings should reinforce previous evidence that childhood abuse and psychopathic traits increase trait aggression, and should also provide preliminary evidence of relationships between the FKBP5 polymorphism and trait aggression in adults. In particular, the strongest predictors (callous affect, FKBP5\_13 T/T allele) were related to approximately 10 point differences in BPAQ total score per standard deviation, as compared to the reference category.

It should be noted that neither the boosting model nor the backwards elimination model should be considered correct. The two complementary models provide different levels of detail regarding the relationships between the predictors and the outcome. To the extent that future samples are similar in nature to the present sample, the penalized boosting model may be a better reference model. Increasingly dissimilar samples may be better represented by the more parsimonious reduced model. Given the high degree of approximated fit obtained here, the reduced model may be sufficient in most contexts; however, this should not be taken to mean that it is superior—only different in applicability.

The limitations of the present project constrain the generality of the results, but they are encouraging in supporting a growing literature linking FKBP5 expression and exposure to stressors (e.g., childhood trauma) to emotional dysregulation. Dysregulation may be expressed in a variety of behavioral manifestations, including psychopathy (callous affect), deficient inhibitory control, and aggressive behavior. In the present case, we show that T carriers of the FKBP5 rs1360780 are tied to trait aggression and hostility (BPAQ); the predictive model accounted for approximately 66% (boosting) and 71% (backwards elimination) of the variance. Previous results using similar data science analytic methods obtained prediction outcomes of AUC = 0.76, 0.74 and 0.77 for cardiac events (Wu et al., 2010), methamphetamine relapse (Gowin et al., 2015), and suicide attempts (Passos et al., 2016), respectively. Putting the accuracy of any such model into proper context requires an understanding of not only the accuracy of prior models that addressed phenomenon of similar complexity (i.e., human aggression), but also of the limits of best performance that can reasonably be expected. Such limits may be constrained by insufficient data (e.g., the small sample size available in the present analysis), model sophistication, and in the phenomenon of interest (Hastie et al., 2009; Hofman et al., 2017). This study did not stratify genetic effects by ancestry, which could lead to occult stratification. However, as the sample was predominantly of African ancestry, stratification seems unlikely, although it remains unclear if the effects of FKBP5 on aggression extend to European or Asian ancestry samples. How well these results generalize to broader populations or clinically diagnosed groups is important, and will need to be ascertained in replication studies involving other populations selected based either on specific clinical criteria or obtained from larger, more heterogeneous samples. Accordingly, the value of the present data will be determined by the ability of future projects to systematically replicate the results with extended and enriched samples.

In the present report, we provide a modest example of the application of modern analytic data science techniques (gradient boosting) to data obtained within the context of an experiment that featured a range of variables selected based on the extant literature. Typically, studies of the present kind do not provide for statistical techniques that validly allow simultaneous examination of all factors. However, via this hybrid approach, we show that approximately two-thirds of the variation in trait aggression (BPAQ) was predicted by an initial combination of eight, and subsequently six key variables. Notably, the final model included psychometric personality variables (callous affect), developmental history (childhood trauma) and genetic variants (FKBP5). While cogent accounts of complex, multifactorial interactions require larger, more detailed, and longitudinal datasets, the results underscore the emerging importance of understanding gene × environment interactions in emotional dysregulation and aggression (Tuvblad and Baker, 2011; Weeland et al., 2015; Holz et al., 2016). The current approach and dataset were underpowered to examine such interactions, but such endeavors are currently planned for larger datasets culled from electronic medical records data. Notably, several of the variables under consideration in this project were previously examined in isolation. These individual variables were identified as predictors in independent studies. One novel feature of this project was the examination these factors in the same individuals. Accordingly, the FKBP5, SRP and CTQ data collectively add value by providing systematic (vs. direct) replication of prior findings. Recent work has highlighted the importance of replication in science (i.e., ''reproducibility''; Aarts et al., 2015; Elliott and Resnik, 2015). Here, we provide preliminary data suggesting these variables are collective predictors of trait aggression.

Access to electronic healthcare system, collaborative multisite and national longitudinal databases has become more common. Accordingly, big data science approaches continue to refine the methods needed to model the complexity in these datasets, and—critically—to interpret the outcomes (Dipnall et al., 2016; Krystal et al., 2017; Wiens and Shenoy, 2018). These rapidly developing tools stand to provide deeper understanding of the relationships among neural, genetic, psychological, and contextual variables in human aggression, moving toward improved prediction and prevention efforts.

### AUTHOR CONTRIBUTIONS

RS performed the primary statistical analyses (component-wise gradient boosting, model approximation) and co-wrote the data analytic strategy and results. JLG helped conceive and develop the original experiments and helped author with the introduction and methods. CEG guided the statistical approach and co-authored the data analytic strategy. CW-B processed and analyzed all the saliva samples to derive the genetic data for the FKBP5 SNPs. SDL helped conceive and develop the original experimental design, and served as senior author on the project, providing oversight over the project and each section of the manuscript.

### FUNDING

This work was supported in part by past National Institutes of Health (NIH) grants NIH DA P50 09262 and NIH DA R01 03166.

### REFERENCES


The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

### ACKNOWLEDGMENTS

We would like to thank Dr. Ahmad Hariri for the valuable suggestion to focus on FKBP5 in our genetic analysis. We would like to thank Joe Alcorn 3rd, Jessica Vincent, Nuvan Rathnayaka, Zahra Kamdar and Gabriel Fries for valuable technical research support in the collection and processing of these data.


**Conflict of Interest Statement**: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

The handling Editor declared a past co-authorship with one of the authors SDL.

Copyright © 2018 Suchting, Gowin, Green, Walss-Bass and Lane. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Genes and Aggressive Behavior: Epigenetic Mechanisms Underlying Individual Susceptibility to Aversive Environments

#### Sara Palumbo<sup>1</sup> , Veronica Mariotti <sup>2</sup> , Caterina Iofrida<sup>3</sup> and Silvia Pellegrini <sup>2</sup> \*

<sup>1</sup>Department of Surgical, Medical, Molecular Pathology and Critical Care, University of Pisa, Pisa, Italy, <sup>2</sup>Department of Experimental and Clinical Medicine, University of Pisa, Pisa, Italy, <sup>3</sup>Department of Pharmacy, University of Pisa, Pisa, Italy

Over the last two decades, the study of the relationship between nature and nurture in shaping human behavior has encountered a renewed interest. Behavioral genetics showed that distinct polymorphisms of genes that code for proteins that control neurotransmitter metabolic and synaptic function are associated with individual vulnerability to aversive experiences, such as stressful and traumatic life events, and may result in an increased risk of developing psychopathologies associated with violence. On the other hand, recent studies indicate that experiencing aversive events modulates gene expression by introducing stable changes to DNA without modifying its sequence, a mechanism known as "epigenetics". For example, experiencing adversities during periods of maximal sensitivity to the environment, such as prenatal life, infancy and early adolescence, may introduce lasting epigenetic marks in genes that affect maturational processes in brain, thus favoring the emergence of dysfunctional behaviors, including exaggerate aggression in adulthood. The present review discusses data from recent research, both in humans and animals, concerning the epigenetic regulation of four genes belonging to the neuroendocrine, serotonergic and oxytocinergic pathways—Nuclear receptor subfamily 3-group C-member 1 (NR3C1), oxytocin receptor (OXTR), solute carrier-family 6 member 4 (SLC6A4) and monoamine oxidase A (MAOA)—and their role in modulating vulnerability to proactive and reactive aggressive behavior. Behavioral genetics and epigenetics are shedding a new light on the fine interaction between genes and environment, by providing a novel tool to understand the molecular events that underlie aggression. Overall, the findings from these studies carry important implications not only for neuroscience, but also for social sciences, including ethics, philosophy and law.

Keywords: epigenetics, aversive environment, aggressive behavior, NR3C1, OXTR, SLC6A4, MAOA

### NEW FRONTIERS IN EPIGENETIC RESEARCH

Studies both in animals (Mosaferi et al., 2015) and humans (McEwen et al., 2012) indicate that the environment, mostly during prenatal stage and infancy, impact significantly on neural development, as several critical periods with lasting consequences on behavior have been documented (Stiles and Jernigan, 2010). Moreover, rodent studies have found that adolescent

Edited by: Aki Takahashi, University of Tsukuba, Japan

### Reviewed by:

Bauke Buwalda, University of Groningen, Netherlands Herb E. Covington, Tufts University, United States

> \*Correspondence: Silvia Pellegrini silvia.pellegrini@med.unipi.it

Received: 09 January 2018 Accepted: 28 May 2018 Published: 13 June 2018

#### Citation:

Palumbo S, Mariotti V, Iofrida C and Pellegrini S (2018) Genes and Aggressive Behavior: Epigenetic Mechanisms Underlying Individual Susceptibility to Aversive Environments. Front. Behav. Neurosci. 12:117. doi: 10.3389/fnbeh.2018.00117 expression of 5-HT1B receptors has a direct impact on later patterns of aggressive behavior (Nautiyal et al., 2015). Therefore, the flexibility of neural programing during critical periods seems to be a significant mediator of long-lasting effects on behavior (Morrone, 2010). In particular, adversities experienced during prenatal life and infancy interfere with the normal processes of cell proliferation and differentiation leading to altered neural circuits that may result in cognitive and emotional deficits. These alterations have been associated with both proactive (e.g., children callous-unemotional traits) and reactive aggressive behavior (e.g., children externalizing disorder spectrum) that may anticipate Antisocial Personality Disorder (Frick and White, 2008; Buchmann et al., 2014; Mann et al., 2015; Rosell and Siever, 2015).

Aggression, throughout evolution, serves an important role in the survival of a species (Darwin, 1859, 1871). Being aggressive gives the best chances for survival and reproduction (Veroude et al., 2016). This is true for all mammalian species, including human. However, when excessive, the consequences of aggressive acts can be maladaptive (Takahashi and Miczek, 2014; Waltes et al., 2016).

Experiencing repeated aversive life events or protracted stress during pregnancy, especially during the first trimester of gestation, results in increased risk of physically-aggressive tendencies, delinquency and conduct disorder, both in early childhood and adolescence (Kvalevaag et al., 2014; Van den Bergh et al., 2017). During the first trimester, the neuroectoderm develops and becomes the source of neural progenitor cells, as well as the foundation of the neural tube (Stiles and Jernigan, 2010). Similar outcomes are predictable by postnatal traumas. The risk of aggressive behavior in childhood is particularly high in infants neglected during their first 2 years of life, when the brain doubles its volume and a massive synaptogenesis occurs (Knickmeyer et al., 2008; Tau and Peterson, 2010). Neglecting to provide early-life basic physical needs and emotional support as a parent can later lead to higher scores of aggression in childhood, measured by the Child Behavior Checklist (Kotch et al., 2008). Moreover, recurrent experiences of emotional abuse or witnessing violence throughout childhood predict physical aggressive behavior in adulthood (Sansone et al., 2012).

Studies in animals support the implication of prenatal and childhood adversities in the origin of aggressive behavior. In juvenile and adult male rats, for example, an increased number of physical attacks toward inoffensive peers and females have been predicted by repeated maternal separation in the first 2 weeks of life or by post-weaning social isolation (Haller et al., 2014). In rats, both prenatal and early postnatal stressors, like physical restraint during pregnancy or repeated maternal separation after birth, interfere with normal cell proliferation and differentiation and with dendritic formation, leading to altered neural circuits that may result in exaggerate aggressive behaviors (Lukas et al., 2010; de Souza et al., 2013). These early aversive experiences affect the functioning of many biochemical pathways (e.g., vasopressin, oxytocin, serotonin and cortisol pathways) that play a crucial role for the development of social skills and for the response to stress; the persistence of these alterations predisposes juvenile rats to excessive offensive play-fighting and then, as adults, to high levels of offensive attacks toward peers (Veenema et al., 2006; Veenema and Neumann, 2009; Veenema, 2009; Lukas et al., 2010; de Souza et al., 2013; Haller et al., 2014).

In addition to prenatal and early postnatal life, adolescence also represents a time-window particularly sensitive to external/environmental events, as in this period the brain concludes its maturation process (Morrone, 2010). As demonstrated by rats, during this period of life, a massive reorganization occurs in specific brain areas—hippocampus, cortex and amygdala—whose morphological and functional alterations have been linked to violence in humans (Isgor et al., 2004; Morrison et al., 2014); an increased amygdala volume, for example, has been observed in incarcerated criminals (Schiffer et al., 2011). Furthermore, peripubertal exposure of rats to fear-inducing stressors, such as the presence or the smell of a predator, predicts the expression of aggressive behavior later in adulthood (Cordero et al., 2013; Márquez et al., 2013).

It has been known for some time that genetic variants, which regulate aminergic signaling in brain, modulate vulnerability to aversive environmental factors resulting in different behavioral phenotypes (for a review see Iofrida et al., 2014; Veroude et al., 2016). More recently, it has emerged that environmental factors stably affect gene expression by producing specific signals to DNA, chromatin and mRNA that do not modify the nucleotide sequence (Bale, 2015). This phenomenon, known as epigenetics, probably mediates the long-lasting effects of aversive experiences on brain and behavior, through the generation of new trajectories of neuronal development (Morrison et al., 2014; Bale, 2015).

### THE EPIGENETIC MECHANISMS: AN OVERVIEW

The main epigenetic changes playing an active role in gene expression regulation are represented by DNA methylation, post-translational histone modifications and post-transcriptional regulation by microRNAs (miRNAs; Dolinoy et al., 2007; Chhabra, 2015).

DNA methylation is carried out by three active isoforms of the DNA methyltransferase family (DNMT-1, -3a and -3b), which are ubiquitous nuclear enzymes, able to transfer residues of methyl groups from the S-adenosylmethionine (SAM) to unmethylated cytosines, preferably cytosine-guanine dinucleotides (CpGs; Chiang et al., 1996). Most CpGs are grouped in specific loci of the genome, the CpG islands, which are located into promoters, exons and, to a lower extent, introns (Schwartz et al., 2009; Gelfman et al., 2013). DNMTs inhibit DNA transcription by blocking the interactions among DNA, RNA polymerase II and transcription factors, by promoting the heterochromatin formation and by interfering with the splicing process (Maunakea et al., 2013). DNMT-1, also called maintenance methyltransferase, methylates the newly replicated strand of DNA by copying the methylation patterns from the parent strand. Its role is to preserve the correct DNA methylation pattern during mitosis in daughter cells (Bird, 2002). DNMT-3a and -3b perform de novo methylation of unmethylated CpGs and produce new DNA methylation marks. The de novo methylation mainly occurs in the early embryonic cells and, not surprisingly, both enzymes are highly expressed in these cells (Okano et al., 1999).

Post-translational histone modifications are covalent modifications of the amino-terminal tails of the histones including acetylation, phosphorylation, methylation and ubiquitylation. Such modifications influence the interaction between DNA and histones, thus modifying the chromatin compacting state (Bannister and Kouzarides, 2011). Histone acetylation is mediated by the histone acetyltransferase (HAT) enzymes that cause chromatin decondensation by transferring acetyl groups from the acetyl-Coenzyme A to lysine residues within the amino-terminal tails of nucleosomal histones. The addition of acetyl groups neutralizes the positive charge of lysines, weakening the interaction between histones and DNA, thus making DNA accessible to the transcriptional machinery (Bannister and Kouzarides, 2011). At the opposite, the histone deacetylases (HDACs) remove acetyl groups from lysine residues to restore their positive charge. Histone deacetylation allows histones to tightly bind DNA, thus favoring a more compact configuration of chromatin and a consequent inhibition of transcription (Lombardi et al., 2011). Histone phosphorylation predominantly occurs on threonine, tyrosine and serine residues and is mediated by kinases that transfer a phosphate group from ATP to the hydroxyl group of the target amino-acid side chain. The addition of phosphate groups negatively charges histones, thus weakening their interaction with DNA. The histone dephosphorylation is catalyzed by phosphatases (Bannister and Kouzarides, 2011). Histone methylation takes place on the side chains of lysines and arginines, within the histone tails (Kouzarides, 2007). Histone Lysine Methyltransferase (HKMT) and Protein Arginine Methyltransferase (PRMT) are the enzymes that catalyze the transfer of a methyl group from SAM to lysine and arginine residues, respectively (Bannister and Kouzarides, 2011). Both lysine and arginine methylations act either as activators or as repressors for transcription (Kouzarides, 2007). Lysine residues are de-methylated by both the lysine-specific demethylase 1 (LSD1) and the jumonji domain 2 protein (JMJD2), whereas the jumonji domain 6 protein (JMJD6) de-methylates the arginine residues (Bannister and Kouzarides, 2011). Histone ubiquitylation consists of a binding between histone lysine residues and ubiquitin, through the sequential action of three enzymes: E1-activating, E2-conjugating and E3-ligating enzymes. Also this histone modification can be either activatory or repressive for transcription. Ubiquitin is removed by specific isopeptidases named de-ubiquitin enzymes (Bannister and Kouzarides, 2011).

miRNAs are untranslated transcripts that originate from MIR genes, located in clusters within the introns of other genes (Bhat et al., 2016). MIR genes are transcribed by RNApol II or III in long primary transcripts called pri-miRNAs that undergo extensive processing to generate mature doublestranded miRNAs. One strand is complementary to the 3'untraslated region of the target mRNA, where it binds, thus blocking gene expression either temporarily, through the mRNA translational repression, or permanently, through the mRNA cleavage (Issler and Chen, 2015).

The above-described chromatin modifications are extremely dynamic and subjected to continuous changes in response to external stimuli. As such, these molecular processes are vulnerable to modifications before and after childbirth, rendering gene expression plastic throughout mammalian life. Thus, they represent promising targets for behavioral treatment strategies. Talking about aggressive behavior predisposition, however, a role has been described to date only for DNA methylation and histone acetylation, whose scientific evidence in literature is reviewed below.

### GENES WHOSE EPIGENETIC MARKS ARE INVOLVED IN HUMAN AGGRESSIVE BEHAVIOR

### Nuclear Receptor Subfamily 3-Group C-Member 1 (Glucocorticoid Receptor; NR3C1)

Nuclear receptor subfamily 3-group C-member 1 (NR3C1) encodes for a nuclear glucocorticoid receptor that interacts with cortisol to control the functioning of the hypothalamic-pituitaryadrenocortical (HPA) axis via a negative feedback that ultimately inhibits cortisol release (Kino and Chrousos, 2002).

According to a meta-analysis published in 2009 (Hawes et al., 2009), a great amount of data shows that cortisol is reduced in antisocial behavior. Low basal levels of blood cortisol, for example, have been associated with externalizing behavior in childhood (Alink et al., 2008) and adolescence (Shoal et al., 2003; Shirtcliff et al., 2005). In adolescence, low plasma concentration of cortisol has been negatively correlated also to low self-control (Shoal et al., 2003), delinquent behavior and proactive and reactive aggression (Poustka et al., 2010). Interestingly, a history of child abuse and neglect predicted lower HPA activity and higher trait and state aggression in adults, suggesting that the HPA hypo-activity may be a mediator between environment and long-lasting aggressive behavior (Gowin et al., 2013). A recent study confirmed this hypothesis; specifically, the hypo-methylation of NR3C1, which translates into augmented inhibitory control of HPA axis, has been shown to be induced by early adverse family environment, and to represent a risk factor for aggressive externalizing behavior in adolescence (Heinrich et al., 2015). As these epigenetic changes are produced during infancy when brain development is maximal, they persist well beyond in life (Radtke et al., 2011). They significantly impact neurodevelopment and predispose to behavioral alterations, including impaired stress response and poor self-regulation (Conradt et al., 2013), which concur in predisposing to aggressive behavior.

### Oxytocin Receptor (OXTR)

Oxytocin is a hypothalamic hormone, also known as the ''social neuropeptide'', that regulates complex social behaviors by promoting attachment and facilitating social interactions (Meyer-Lindenberg et al., 2011). Impaired functioning of the oxytocinergic system has been observed in rodents with aggressive behavior (Lubin et al., 2003; McMurray et al., 2008), and a lower oxytocin concentration in the central nervous system represents a predisposing factor to human aggressive behavior (Lee et al., 2009; Jokinen et al., 2012).

Social environment induces changes in the oxytocinergic system, especially during the early postnatal period and the infancy (Veenema, 2012). Oxytocin secretion (measured in saliva and whole blood), for instance, is stimulated in infants and children by maternal care (Wismer Fries et al., 2005; Tsuji et al., 2015), while childhood maltreatments, especially emotional abuses, result in lower levels of oxytocin in the cerebral spinal fluid of adults (Heim et al., 2009). Similarly, a lower expression of the oxytocin receptor (OXTR) has been detected in rodents and macaques poorly nurtured (Francis et al., 2000; Baker et al., 2017).

DNA methylation of OXTR is an important mechanism linking aversive experiences to susceptibility to abnormal behavior in adulthood (Veenema, 2012; Unternaehrer et al., 2015; Ziegler et al., 2016). A history of repeated early abuses and traumatic experiences, in fact, has been correlated to increased OXTR methylation in depressed and anxious adults (Smearman et al., 2016; Gowin et al., 2017).

OXTR methylation is affected by negative events also before birth. In particular, newborns from women who suffered from drug addiction, psychopathy or showed criminal behaviors during pregnancy, carried hyper-methylated OXTR and had an increased probability of developing callous-unemotional traits (Cecil et al., 2014), indicative of stable and severe aggressive behavior (Frick and White, 2008).

### Serotonin Pathway

Serotonin plays a key role in most of psychiatric conditions and in antisocial/aggressive personality (Nutt, 2008; Seo et al., 2008). Brain serotonin concentration is regulated by serotonin transporter solute carrier-family 6 member 4 (SLC6A4) that controls its reuptake from the synaptic cleft, and by monoamine oxidase A (MAOA) that catabolizes serotonin (Shih et al., 1999).

Hypo-functioning of serotonin neurotransmission has been linked to higher risk of aggressive behaviors (Davidson et al., 2000). For instance, the brain expression of SLC6A4 is reduced in aberrant impulsive-aggressive individuals (Frankle et al., 2005). Consistently, early aversive experiences exert epigenetic regulation of SLC6A4 with implications in the development of such conditions (Provencal and Binder, 2015). Childhood stress, e.g., bullying victimization by peers, increased the saliva methylation of SLC6A4 promoter from age 5 to age 10 (Ouellet-Morin et al., 2013). Moreover, as observed in females, being physically (including sexually) abused by parents from childhood to adolescence predicts, in adulthood, both an increased SLC6A4 methylation in peripheral white cells (Beach et al., 2010) and a higher risk of developing long-lasting antisocial personality disorders (Beach et al., 2011, 2013). An in vivo study in males found a similar link between physical abuses experienced in childhood and SLC6A4 hypermethylation in peripheral lymphocytes correlating with low brain (orbitofrontal cortex) synthesis of serotonin (Wang et al., 2012). These data suggest that SLC6A4 is silenced by early stressors as a protective mechanism aimed at the potentiation of the serotonergic neurotransmission; however, a long-lasting hyper-methylation results in lower cortical thickness (Park et al., 2015; Won et al., 2016) and alters amygdala reactivity (Nikolova et al., 2014), thus probably predisposing to aggressive behavior. For example, adolescents that have been raised in low socioeconomic status show higher methylation of SLC6A4 in peripheral lymphocytes and higher amygdala activation in response to fearful faces (Swartz et al., 2017). As far as the orbitofrontal cortex concerns, an increased activity of this brain area predicted aggressive responses to angry faces (Beyer et al., 2015); moreover, morphological asymmetry of this area has been associated with higher scores at the Lifetime History of Aggression, and Buss-Perry Aggression scales (Antonucci et al., 2006).

Finally, in a rat model of pathological aggression, the exposure to peripubertal stress affected the connectivity between amygdala and orbitofrontal cortex accompanied by a parallel increase of MAOA expression in the frontal cortex in adulthood. Interestingly, an increased H3 acetylation of MAOA was observed in the prefrontal cortex suggesting that the aversive experience has induced a stable epigenetic regulation of the transcription of this gene (Márquez et al., 2013).

### CONCLUSION

In recent years, neuroscientific research has focused more and more on the biological mechanisms that predispose to behavioral disorders as a consequence of the exposure to aversive environments. Specific genetic variants, in interaction with negative environmental experiences during prenatal life, childhood and adolescence, have been shown to affect the development of long-lasting aggressive behavior and psychiatric disorders in adulthood, with significant social, legal and moral implications (Rigoni et al., 2010; Sartori et al., 2011; Jones et al., 2013; Roth, 2013; Iofrida et al., 2014; Rota et al., 2016; Pellegrini et al., 2017). As a matter of fact, recent studies suggest that the same genetic variants that increase the risk of aggressive behavior in combination with a negative environment, may actually act as plasticity variants, making the brain more sensitive also to positive environmental inputs, resulting in increased prosocial behavior (Belsky et al., 2009; Simons et al., 2011; Iofrida et al., 2014).

Aggression actually represents an evolutionary important behavior fostered by stressful life events, fundamental to deal with life threating situations and to preserve one's own life (Stiles and Jernigan, 2010). However, if exaggerate and uncontrolled, it represents a pathological condition characterizing externalizing behavior, conduct disorders, callous-unemotional traits and psychopathy (Beach et al., 2011; Kumsta et al., 2013; Cecil et al., 2014; Heinrich et al., 2015; Kundakovic et al., 2015).

Over the last few years, the epigenetic mechanisms underlying human aggressive behavior have been attracting a growing interest, as they provide a fascinating and reliable explanation of the gene-environment interplay that modulates human violent behavior. Epigenetics, indeed, plays a central role in the adaptation of the human organism to the changing environment. This concept emerged first from studies conducted in monozygotic twins, which showed that different phenotypes may originate from identical genotypes due to epigenetic changes (Poulsen et al., 2007). These differences progressively increase as twins become older, along with the diversification of their lifestyles and living environments (Fraga et al., 2005).

Although the existence of a genetic blueprint underlying brain development is undeniable, the epigenetic control of biological pathways, including the neuroendocrine, serotonergic and oxytocinergic pathways, significantly mediates the behavioral responses to the environment (**Figure 1**; Veenema, 2012; Waltes et al., 2016). Epigenetic changes in these pathways may alter brain morphology and functioning in areas that hold a crucial role in cognitive and emotional processes underlying aggression (Conradt et al., 2013; Ouellet-Morin et al., 2013; Suri et al., 2013; Booij et al., 2015; Puglia et al., 2015; Gowin et al., 2017). Recent data indicate that these epigenetic marks may be, in some extent, reversed by the exposure to an enriched environment therapy; for example, massage therapy significantly reduced aggressive behavior in children and adolescence (Diego et al., 2002; Garner et al., 2008), probably by epigenetic mechanisms (McCreary and Metz, 2016). Alternatively, it is possible to intervene by a pharmacological therapy, as shown in rats: treating aggressive adult rats that had experienced peripubertal stress with a MAOA inhibitor reversed their aberrant behavior (Márquez et al., 2013).

In conclusion, epigenetics is shedding a new light on the fine interaction between nature and nurture, by providing a novel tool to understand the molecular events that underlie the relationship among genes, brain, environment and behavior. Altogether, the results of the studies that we briefly discussed in the present article, clearly indicate that, when it comes to (human) behavior, nature and nurture are not to be regarded as two distinct and separate factors, contrary to the alternating predominance of either one that has been proposed in different historic phases (Levitt, 2013; Moore, 2016). Indeed, distinct genetic backgrounds differentially modulate the individual susceptibility to the environment and at the same time various environmental conditions differentially affect gene expression, in an intimate and fascinating manner that scientists have now begun to disentangle. The findings from this research pave the way to a novel approach to the understanding of human behavior, with important implications also for social sciences, including philosophy, ethics and law. Unveiling the molecular mechanisms that regulate the expression of human behavior will provide a solid scientific basis to what philosophy already sensed since its dawn, suffice it to mention what the great Plato wrote over 25 centuries ago: ''No one is willingly evil, but one can become evil for a bad disposition in his body and for a training without a true education; this is hideous for everyone and happens against his will'' (Timeus, 86e).

### REFERENCES


## AUTHOR CONTRIBUTIONS

SPalumbo, VM and CI searched and reviewed the scientific literature; all the authors discussed the findings from the literature; SPalumbo and VM drafted the manuscript; SPellegrini conceived the work and revised the manuscript.

### FUNDING

This work was supported by a Grant from Fondazione Gio.I.A, Pisa (Italy) and by Fondazione Cassa di Risparmio di Lucca (Grant 2016-2017).


anxiety and aggression in a young adult psychiatric inpatient unit. Aust. N Z J. Psychiatry 42, 414–422. doi: 10.1080/00048670801961131


disorder. Psychoneuroendocrinology 34, 1567–1573. doi: 10.1016/j.psyneuen. 2009.06.002


methylation is associated with bullying victimization and blunted cortisol response to stress in childhood: a longitudinal study of discordant monozygotic twins. Psychol. Med. 43, 1813–1823. doi: 10.1017/s003329171 2002784


for regulating social behavior. Proc. Natl. Acad. Sci. U S A 102, 17237–17240. doi: 10.1073/pnas.0504767102


**Conflict of Interest Statement**: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Palumbo, Mariotti, Iofrida and Pellegrini. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Alcohol, Self-Regulation and Partner Physical Aggression: Actor-Partner Effects Over a Three-Year Time Frame

Brian M. Quigley <sup>1</sup> \*, Ash Levitt <sup>2</sup> , Jaye L. Derrick <sup>3</sup> , Maria Testa<sup>2</sup> , Rebecca J. Houston<sup>4</sup> and Kenneth E. Leonard<sup>2</sup>

<sup>1</sup>Department of Medicine, Jacobs School of Medicine and Biomedical Sciences, University at Buffalo, Buffalo, NY, United States, <sup>2</sup>Research Institute on Addictions, University at Buffalo, The State University of New York, Buffalo, NY, United States, <sup>3</sup>Department of Psychology, University of Houston, Houston, TX, United States, <sup>4</sup>Department of Psychology, Rochester Institute of Technology, Rochester, NY, United States

The question of how individual differences related to self-regulation interact with alcohol use patterns to predict intimate partner aggression (IPA) is examined. We hypothesized that excessive drinking will be related to partner aggression among those who have low self-regulation. In addition, we explored the extent to which differences in self-regulation in one partner may moderate the relationship between alcohol use and partner aggression. A sample of married or cohabitating community couples (N = 280) ages 18–45 was recruited according to their classification into four drinking groups: heavy drinking in both partners (n = 79), husband only (n = 80), wife only (n = 41), by neither (n = 80), and interviewed annually for 3 years. IPA, drinking, and scores on measures of negative affect, self-control, and Executive Cognitive Functioning (ECF) were assessed for both members of the couple. The Actor Partner Interdependence Model (APIM) was used to analyze longitudinal models predicting the occurrence of IPA from baseline alcohol use, negative affect, self-control and ECF. Actor self-control interacted with partner self-control such that IPA was most probable when both were low in self-control. Contrary to prediction, actors high in alcohol use and also high on self-control were more likely to engage in IPA. Partner alcohol use was predictive of actor IPA when the partner was also high in negative affect. Low partner ECF was associated with more actor IPA. These findings suggest that self-regulatory factors within both members of a couple can interact with alcohol use patterns to increase the risk for relationship aggression.

Keywords: alcohol, self-regulation, partner aggression, executive functioning, self-control

### INTRODUCTION

One of the most consistent predictors of intimate partner aggression (IPA) is excessive use of alcohol. Numerous cross-sectional and prospective studies have shown that excessive drinking, particularly male drinking, is associated with the occurrence and frequency of partner violence (see Foran and O'Leary, 2008 for a review). Despite the consistent association, few believe that excessive drinking exerts either a necessary or sufficient condition for partner aggression, but rather

#### Edited by:

Nelly Alia-Klein, Icahn School of Medicine at Mount Sinai, United States

> Reviewed by: Kasia Radwanska, Nencki Institute of Experimental Biology (PAS), Poland

Milen L. Radell, Niagara University, United States \*Correspondence:

Brian M. Quigley bquigley@buffalo.edu

Received: 28 October 2017 Accepted: 08 June 2018 Published: 05 July 2018

#### Citation:

Quigley BM, Levitt A, Derrick JL, Testa M, Houston RJ and Leonard KE (2018) Alcohol, Self-Regulation and Partner Physical Aggression: Actor-Partner Effects Over a Three-Year Time Frame. Front. Behav. Neurosci. 12:130. doi: 10.3389/fnbeh.2018.00130 that it contributes or facilitates the occurrence of aggression. Moreover, there is recognition that any influence of alcohol on partner aggression is conditional, moderated by both situational and individual difference factors. From the beginning, theoretical approaches to understanding the role of excessive drinking on aggression have sought to address the fact that excessive drinking does not lead to aggression in all people or under all circumstances. Taylor and Leonard (1983) argued that alcohol's cognitive disruption ''might facilitate aggressive behavior in the presence of dominant, instigative cues by increasing one's attention to those cues and . . . reducing one's attention to incompatible, inhibitory cues'' (p. 96). This position was expanded upon and formalized as Alcohol Myopia Theory by Steele and Josephs (1990), currently the predominant model of intoxicated behavior. The multiple thresholds model of alcohol-related aggression (Fals-Stewart et al., 2005; Leonard and Quigley, 2017) argued that the impact of alcohol depends on the balance and salience of instigatory and inhibitory cues. Similar to the I<sup>3</sup> model of aggression (Finkel, 2014), these theories all suggest that alcohol is most likely to facilitate aggression among individuals and in situations characterized by high instigation or low inhibition.

Experimental data generally supports the perspective that people with chronically low self-regulatory abilities become aggressive more easily when drinking. Bailey and Taylor (1991) found that alcohol facilitated aggressive behavior among men who were moderate to high in hostility (an instigatory factor), but not among men low in hostility. Alcohol was more likely to increase aggression for people with high levels of trait anger and irritability as compared to low levels (Giancola, 2002a,b, 2004; Parrott and Zeichner, 2002). Similarly, several studies have suggested that executive functioning is an important moderator of the alcohol aggression relationship. Pihl et al. (2003) reported that individuals with low scores on Executive Cognitive Functioning (ECF) were more aggressive than other groups when administered alcohol, but only under low provocation. Giancola (2004) observed that alcohol increased aggression for men who were low on ECF under both low and high provocation.

Studies of IPA have similarly suggested that the alcoholaggression relationship is moderated by the influence of both impelling and inhibiting factors in aggression. Heavy drinking is associated with marital violence only among hostile (Leonard and Blane, 1992) or maritally distressed couples (Leonard and Senchak, 1993; Margolin et al., 1998). Quigley and Leonard (1999) demonstrated that heavy alcohol involvement predicted subsequent aggression only among couples high in verbally aggressive conflict styles. Schumacher et al. (2008) found that excessive drinking by the husband longitudinally predicted IPA in men high on hostility and avoidance coping. Based on these findings, Schumacher et al. (2008) asserted ''alcohol should have a more pronounced effect on individuals with aggressive perceptual and behavioral propensities, and among individuals, who for dispositional or situational reasons, already have some degree of impaired behavioral regulation and control'' (p. 895).

Because aggression is an interactional phenomenon (Tedeschi and Felson, 1994), it is important to examine impelling and inhibitory factors associated with both husbands and wives to fully understand the relationship between alcohol and aggression in couples. Testa et al. (2012) found that husband heavy drinking was associated with violence regardless of wife's drinking; however at lower levels of husband drinking, heavier drinking by the wife predicted husband aggression. Husband and wife instigating and inhibiting behaviors may exacerbate or ameliorate conflicts in the relationship. In a study using daily diary methods, the presence of an instigator (e.g., provocation) on a given day was most likely to lead to IPA when the partner was high on dispositional aggression and low on a measure of ECF (Finkel et al., 2012). Similarly, both partner's lack of inhibitory factors could increase the likelihood of aggression. Individuals involved in a relationship have been found to benefit when both are high in self-control but to experience more problems when neither partner has high self-control (Vohs et al., 2011; Crane et al., 2014; Derrick et al., 2016). Thus, while conflict is present in all relationships, impelling and inhibitory factors may influence the course of that conflict and moderate the impact of excessive drinking on aggression.

In a previous article (Testa et al., 2012), we examined how the interaction between husband and wife alcohol dependence symptoms predicted intimate partner violence in a crosssectional analysis of the current data set at the first measurement point in this study. In the present article, we extend that analysis to examine how husband and wife individual differences in self-regulation may interact with alcohol use to predict partner aggression. Because we wished to examine how individual differences in each partner interacted with their own alcohol use and with the alcohol use of the partner to predict IPA by either member of the couple we chose to analyze the data using the Actor Partner Interdependence Model (APIM; Kenny et al., 2006). In a traditional regression analysis framework, it would have been necessary to conduct separate analyses for each member of the couple, however, in an APIM analysis both husband and wife data are nested within the couple. As is shown in **Figure 1**, which presents the basic APIM model, a dyadic analysis allows the estimation of actor effects and partner effects for each member of the couple. Because observations of each member of the dyad are considered interdependent and nested within couple, we are able to examine the effects of the actor's drinking and self-control, the effects of the partner's drinking and selfcontrol, and the effects of any interactions between actor and partner variables on IPA by each member of the couple simultaneously.

The examination of multiple measures of self-control, both self-report and behavioral measures, provide a unique opportunity to examine how differing aspects of self-regulation may interact with alcohol to predict IPV. Based on our earlier work (Schumacher et al., 2008; Testa et al., 2012; Leonard et al., 2014), we hypothesize that excessive drinking will be related to partner aggression among those who have self-regulatory challenges. In addition, we explore the

extent to which self-regulation factors in one partner may moderate the relationship between heavy alcohol use and aggressive behavior in the other partner. We examine these hypothesizes over a three-year time period. Based on past research (see Quigley and Leonard, 1999), we expected IPA to decrease over time, however, we also hypothesized that the proposed interactions between drinking and self-regulation will predict IPA both cross-sectionally and longitudinally.

### MATERIALS AND METHODS

### Participants

A sample of married or cohabitating couples (N = 280) was recruited from the community via a mail survey of health behaviors in Erie County, NY, USA. A list of households in Erie County, NY, USA likely to contain households with a head of household between the ages of 18 and 45 was purchased from Survey Sampling International. From this purchased list, 21,000 screening questionnaires were mailed to households accompanied by a letter explaining the purpose of the study. The letter stated that the purpose of the study was to estimate the number of different types of families and to determine the eligibility of respondents and their interest in participating in one of the ongoing studies on families and health. We enclosed a non-contingent dollar bill incentive in the questionnaire to improve response rates (see Homish and Leonard, 2009) and provided a stamped envelope to return the questionnaire. We received 5463 responses for a 26% response rate (226 or about 1% were returned due to an incorrect address). Of the 5463 responses, 10.7% were minorities, with 7.6% being African-American, similar to census data for married couples in Erie County.

Responses from the mailed questionnaire were used to assess study eligibility (between the ages of 18 and 45, and married or living together for at least 1 year) and to determine husband and wife heavy episodic drinking (HED) status. Because one aim of the study involved ECF, we excluded couples if either member had a current medical condition that would impair ECF or if either reported having had a seizure, epilepsy, or a 10-min loss of consciousness due to an accident or head injury. In order to ensure adequate numbers of heavy drinking husbands and wives, we utilized disproportionate sampling to recruit couples in which either member of the couple engaged in regular HED. HED was defined as becoming intoxicated or having five or more drinks at one time (four drinks for women) at least weekly. Our goal was to recruit 75 couples in each of four groups; (1) husband and wife both engaged in HED (Both); (2) only the husband engaged in HED (Husband Only); (3) only the wife engaged in HED (Wife Only); and (4) neither engaged in HED (Control).

Of the 5463 responses, 3477 met eligibility criteria. Of those meeting eligibility criteria, three quarters (75%) of the couples were classified as Control. The rates for Husband Only, Wife Only, and Both were 12.3%, 4.1% and 8.5%, respectively. We also asked whether the couple was interested in participating in one of our ongoing studies. Across the four groups, 68% (N = 2347) were interested in participating or hearing more about the studies. The proportion of those who were interested was significantly higher for Husband Only (72%), Wife Only (74%), and Both (76%) than for Control (67%; χ 2 (3) = 16.32, p < 0.01). We sampled from the four groups at different rates in order to achieve the goal of 75 couples in each of the four groups. This disproportionate sampling was by design and has implications for our data analyses. We were able to recruit 80 Control, 80 Husband Only, 79 Both, and 41 Wife Only couples. This was a 43% success rate from those who we attempted to recruit, a rate that did not differ across the four groups (χ 2 (3) = 2.78, p > 0.40). This indicates that the difficulty that we experienced filling the Wife Only cell reflected the rarity of this group in the population, and not any difference in willingness to participate among these couples.

The average age of the final sample at Time 1 was similar between husbands and wives (36.9, SD = 5.8; 35.4, SD = 5.9) respectively. The majority of men and women in the sample were White (91% each), highly-educated (58% of husbands and 67% of wives had completed college education compared to 39% for the county), and most were employed at least part-time (91% of husbands and 80% of wives). The majority of couples were married (87%) as opposed to cohabiting and had been together for an average of 9.84 years (SD = 5.41). Approximately 79% had children. Among those with children, 15% had one child, 38% had two, 19% had three, and 7.5% had four or more. Median income for wives was in the \$20,000–29,999 range, and median income for husbands was in the \$40,000–54,999 range.

Out of the 280 couples who participated at Wave 1, 259 (92.5%) completed the assessment at the 1 year anniversary (Wave 2). At the second anniversary (Wave 3), 243 couples completed the assessment (87% of original sample). The present analysis utilizes statistical procedures (PROC Glimmix in SAS 9.4) that allow for the use of all available data at each time point rather than deleting data listwise if the second or third time point is missing.

### Procedure

Participants completed a series of questionnaires sent and returned through the mail and subsequently attended a laboratory assessment. Both members of the couple provided written informed consent to participate in the research at the time of the mail assessment and again at the time of the laboratory assessment. Mailed questionnaires, sent separately to husband and wife, consisted of background information, attitudes and beliefs about alcohol, and personality measures. Participants were instructed to complete questionnaires independently and not to discuss the questionnaires until both had been returned. At the laboratory assessment, partners independently completed computerized questionnaires that addressed relationship issues and alcohol and drug use. In addition, we administered measures of ECF and conducted a semi-structured face-to-face interview regarding one or more episodes of marital conflict. Participants were assured of confidentiality and that their responses would not be shared with their partners. At the first and second anniversary couples again completed both assessments.

### Measures

### Demographics

Information collected from each partner included age, race, years married and/or years living together, number of children living in the home, and education. Partner reports of age (r = 0.83), total years living together (r = 0.96) and number of children living in the home (r = 0.97) were highly correlated.

### Intimate Partner Aggression (IPA)

Husband and wife perpetrated IPA over the past 12 months was assessed using the physical aggression subscales of Revised Conflict Tactics Scales (CTS-2, Straus et al., 1996). Each partner reported on the frequency of 12 aggressive acts perpetrated by the self (e.g., ''I slapped my partner'') and the same 12 acts as perpetrated by the partner (''my partner slapped me''). Frequency of each act was recorded using the following scale: never (0), once (1), twice (2), 3–5 times (3), 6–10 times (4), 11–20 times (5) and more than 20 times (6). The present analysis examined only the presence or absence of IPA. If either partner reported any act of aggression by the husband toward the wife it was considered IPA by the husband and if either partner reported an act of aggression by the wife toward the husband it was considered an occurrence of IPA by the wife. Thus, in the APIM analysis, each individual nested within couple had their own unique IPA score (1 or 0).

### Alcohol Dependence

The 25-item Alcohol Dependence Scale (ADS, Skinner and Allen, 1982) was used to assess self-reported occurrence of symptoms of dependence such as blackouts and seeing things that weren't really there. As expected, the distribution of ADS scores was highly skewed, with 41.9% of men and 55.2% of women reporting scores of 0, but just 4.7% of men and 1.9% of women scoring greater than 9. Scores were Winsorized to reduce the impact of extremely high scores (Reifman and Keyton, 2010).

### Self-Report Measures of Negative Affect and Self-Control

A general measure of negative affect was assessed with the Multidimensional Personality Questionnaire (Tellegen, 1982). The short version of the Self-Control Scale (Tangney et al., 2004) was also administered as a self-report inhibitory factor. This measure of a person's general tendency toward self-control has good internal consistency (α > 0.89) and has been shown to be predictive of aggression in laboratory studies (DeWall et al., 2007).

### Executive Cognitive Functioning

ECF has often been viewed as an element of cognitive control. Because there is no overall measure of ECF, three types of cognitive tests were used to assess the construct and standardized scores from these tests were combined into a composite ECF score. Past research (Giancola, 2004; Godlaski and Giancola, 2009) has shown alcohol to have stronger effects of aggressive behavior among those low on ECF. During the face to face interview we assessed multiple aspects of ECF assessing cognitive flexibility, attentional selectivity and control, and working memory. These were assessed by the following measures: Stroop Color-Word Task (Stroop, 1935; Golden and Freshwater, 2002). The Stroop Task assesses the ability to inhibit an over-learned response and attentional control. Participants were asked to either read words (Word) or name the ink color (Color, Color-Word) for as many stimuli as they could in 45 s. A Color-Word interference score was calculated as outlined in Golden and Freshwater (2002). Lower scores indicate poorer attentional control. WAIS-III Digit Span (The Psychological Corporation, 1999). A sequence of digits (one digit per second) was read to the participant. For the Forward condition, the participant was asked to repeat the sequence in the same order. For the Backward condition (working memory), the participant was asked to repeat the sequence in the reverse order. The number of digits increased with successful trials. The number of successful trials were summed for total scores. The Backward score was subjected to a square root transformation. Trail Making Test (TMT; Reitan, 1958). The TMT is a measure of cognitive flexibility, visual attention and motor speed (Lezak et al., 2004). In TMT-A, the participant must draw a line connecting a series of numbers in sequential order. TMT-B requires the participant to draw a line connecting a series of letters and numbers alternating between sequential and alphabetical order. Two scores are derived: the completion time (seconds) and number of errors. Completion time scores were subjected to logarithmic transformations while the errors were subjected to inverse transformations.

The measures of ECF, while related to different cognitive functions such as cognitive flexibility, working memory and attention were meant to assess aspects of prefrontal cortex (PFC) functions (Luria, 1980). In order to calculate a composite score of ECF, based on the procedures used by Giancola (2004), who factor analyzed behavioral ECF measures before combining them



Note. A, actor; P, partner. Bold items indicate effect significant at p < 0.05. Italic items indicate effect nearing significance at p < 0.10.

into single index, we standardized the TMT-B completion time score, the Stroop Interference score, and the Digit Span total score and summed the three scores together.

### RESULTS

### Data Analysis Plan

Multilevel analyses were conducted using the Actor-Partner Interdependence Model (Kenny et al., 2006). Models were estimated using SAS PROC Glimmix (Version 9.4; SAS Institute Inc., 2017), and predicted the probability of physical IPA occurrence. Repeated measures at Level 1 were crossed between partners and nested within couple at Level 2 (Laurenceau and Bolger, 2005; Kashy and Donnellan, 2012), allowing for missing data at Level 1. We took into account the disproportionate sampling of the couple drinking groups by weighting the participants in the different groups to reflect their prevalence among the eligible respondents to our mailed survey. Probability weights were used in all models due to the nature of our sample, which oversampled various combinations of heavy drinking partnerships (Pfeffermann, 1993; Korn and Graubard, 1995). The outcome variable, IPA occurrence, was treated as time varying across the three waves of data. Predictor variables were taken from the baseline assessment only and were therefore timeinvariant. Time was centered at baseline and increased yearly across the three waves of data. Linear growth trajectories of IPA occurrence over the three study waves were estimated. Main effects of predictors indicate intercept differences in IPA occurrence, and interactions between predictors and time indicate differences in the linear growth of IPA occurrence. All models allowed for random intercept and error components, and all predictors were grand mean centered and entered as fixed effects.

Preliminary models were run to test the effects of potential baseline covariates including age, children (yes/no), years living together, and education of each partner on the occurrence of IPA. No covariates were significant predictors of IPA in the presence of any of the predictors of interest and were not included in the final models. Preliminary models were also run to test whether men and women were empirically distinguishable in the means, variances, and covariances of the current set of variables (Kenny et al., 2006; Kashy and Donnellan, 2012). Models including gender effects fit significantly worse than indistinguishable models for all three predictors of interest: negative affect (χ <sup>2</sup>diff [10] = −34.04, p < 0.001), self-control (χ <sup>2</sup>diff [10] = −20.80, p = 0.023), and ECF (χ <sup>2</sup>diff [10] = −77.06, p < 0.001). Therefore, models were treated as indistinguishable and gender differences were not tested.

### Rates of Partner Physical Aggression Perpetration and Alcohol Dependence

Because this was a general population sample, the rates of violence were not at the levels of those from a clinical sample but were still frequent enough to allow analysis. At Wave 1, 31.8% of the women and 22.1% of the men perpetrated at least one act of physical aggression based on the maximum reports of both partners. Rates fell over the next two time periods. At Wave 2, 23.3% of the women and 17.0% of the men engaged in at least one act of physical aggression against their partner and at Wave 3 these rates reduced to 21.0% and 12.8% respectively. Although we did not assess treatment seeking for alcohol use or domestic violence, the reduction in IPA over time is consistent with past research on non-treatment seeking samples (Quigley and Leonard, 1996). Alcohol dependence scores stayed relatively stable for men over the 3-year time frame changing from average scores of 3.99 (sd = 4.55) at Wave 1 to 3.88 (sd = 4.43) at Wave 2 and 3.83 (sd = 4.83) at Wave 3 for men and average scores of 2.88 (sd = 4.10) at Wave 1 to 2.58 (sd = 3.76) at Wave 2 and 2.36 (sd = 3.88) at Wave 3 for women.

### APIM Models Predicting Partner Aggression

#### Negative Affect

We first examined growth trajectories of actor IPA occurrence as a function of actor and partner alcohol dependence symptoms and negative affect (χ 2 (273) = 651.66, p < 0.001). There were main effects of both actor and partner alcohol use indicating greater alcohol use was associated with more IPA (see **Table 1**). There was also a partner alcohol by actor alcohol interaction effect, however, as the associated cross-sectional effect was discussed in Testa et al. (2012) we won't discuss it further here. Both actor and partner negative affect interacted with time indicating that those high and low in negative affect diverged in IPV over time with greater negative affect being associated with greater odds of IPA. There was a significant interaction between partner alcohol use and partner negative affect. Although the three-way interaction effect of partner alcohol use by partner negative affect by time did not reach the traditional level of statistical significance (p = 0.051) we find it useful to interpret the significant two-way interaction in light of the moderating effect of time. As shown in **Figure 2**, high partner alcohol dependence predicted actor IPA occurrence relative to low partner alcohol dependence, with the highest probability occurring when the partner was also high in negative affect. Both of these trajectories also decreased over the study period. In contrast, the trajectories of IPA occurrence for low partner alcohol use did not uniformly decrease over time as a function of partner negative affect. As expected, physical IPA occurrence as a function of low partner alcohol dependence and low partner negative affect decreased to the lowest levels over time. However, low partner alcohol dependence and high negative affect was the only combination to not decrease in the probability of IPA occurrence over time. Overall, partner negative affect has a differential effect on actor IPA occurrence at low levels of partner alcohol use but not high levels of partner alcohol use.

### Self-Control

We next examined growth trajectories of actor IPA occurrence as a function of actor and partner alcohol dependence and self-control (χ 2 (325) = 706.83, p < 0.001). As shown in **Table 2**,

TABLE 2 | Tests of baseline associations between drinking and self control on growth trajectories of occurrence of intimate partner violence.


Note. A, actor; P, partner. Bold items indicate effect significant at p < 0.05. Italic items indicate effect nearing significance at p < 0.10.

there was a significant actor alcohol use by actor self-control by time interaction. As shown in **Figure 3**, at baseline, all combinations of actor alcohol dependence and actor self-control, with one exception, predicted relatively similar probabilities of actor IPA occurrence, and similarly declined over time. Low actor alcohol dependence and high actor self-control was notably lower than the other combinations at baseline and did not change over time. Overall, differences in alcohol dependence were not apparent at low levels of self-control, whereas there was a marked difference in IPA occurrence at baseline and over time for low vs. high alcohol dependence for those high in self-control.

There was also a significant actor by partner self-control interaction (see **Figure 4**). Among those with a partner low in self-control, the highest probability of actor IPA perpetration was when the actor was also low in self-control, whereas the lowest probability of actor IPA occurrence was when the actor was low in self-control and the partner was high in self-control. Probabilities of actor IPA occurrence did not differ as a function of partner self-control when the actor was high in self-control and fell between the two trajectories previously described. This interaction was not moderated by time.

#### Executive Cognitive Functioning

We examined growth trajectories of actor IPA occurrence as a function of alcohol use and ECF (χ 2 (295) = 656.31, p < 0.001). There were no moderating effects of time on alcohol use and ECF interactions (i.e., three-way interactions) so time was dropped from these analyses. There was a main effect of partner ECF indicating greater actor IPA when the partner was low on ECF (**Table 3**). Theory would have predicted a significant interaction between actor alcohol use and actor ECF predicting actor IPA occurrence, however, in this sample that interaction did not reach statistical significance (p = 0.062). Because of the importance of that interaction to the threshold model and the fact that

the choice of an alpha level, while a common convention, is somewhat arbitrary we elected to tentatively examine the shape of that interaction. As would be predicted, the highest probability of actor IPA perpetration was found for actors high in alcohol use but low in ECF. No effect of ECF was found among those low in alcohol use. While this finding is consistent with theory, because the effect did not reach the traditional level of statistical significance we suggest caution in interpretation of this effect.

### DISCUSSION

The findings of the current study suggest that a pattern of excessive drinking is more strongly associated with IPA when moderated by factors related to self-regulation. Use of the APIM framework to analyze the data allowed us to examine how both partners' characteristics play a role in relationship aggression. It is clear that IPA by one member of a couple is not only a function of that person's alcohol use and ability to self-regulate but also a function of these same factors within their partner. The findings of the present study were generally consistent with past research examining the moderators of alcohol-related aggression but also help to extend our knowledge regarding the relationship of alcohol use to anger and aggression.

While self-regulation is a broad term that encompasses many constructs, we chose to focus on three aspects that have been found to be related to aggression in past research: ECF, negative affect, and self-control. We examined a number of constructs related to self-control because self-control is a complex behavior involving numerous neurocognitive mechanisms. Past research has suggested that at minimum self-control requires at least two stages that involve different brain centers. When encountering a conflict situation in which self-control may be necessary, recognition of the conflict is usually associated with activity in the anterior cingulate cortex (Kerns et al., 2004). Actual impulse control however, seems mostly controlled within the domain of



Note. A, actor; P, partner. Bold items indicate effect significant at p < 0.05. Italic items indicate effect nearing significance at p < 0.10.

the prefrontal cortex, in particular areas such as the dorsolateral prefrontal cortex (MacDonald et al., 2000) and the ventrolateral prefrontal cortex (Tabibnia et al., 2011).

Actor self-report of self-control interacted with actor alcohol use to predict the occurrence of aggression, however, actors high in levels of self-control had a greatest probability of IPA when they also were heavy drinkers. Other combinations of alcohol use and self-control were moderately associated with the probability of IPA. The reason for this unexpected finding is not clear, however, Imhoff et al. (2014) found a similar interaction between self-control depletion and an individual difference measure of self-control predicting eating restraint and task persistence. In two studies, individuals scoring highly on a self-report measure of self-control similar to the one used here and who had their self-control experimentally depleted showed lower levels of subsequent self-control than those low in self-control who had also been depleted. It was suggested by Imhoff et al. (2014) that individuals scoring high on such a self-report measure of self-control are actually good at avoiding situations in which they must exert selfcontrol. Thus, they often have little experience in dealing with situations in which their self-control is put to a test. When their self-control resources are reduced (either by depletion as in the Imhoff studies or by alcohol use as in our study) and they are put in a situation where they must self-regulate, they are actually very poor at self-regulation due to lack of experience. While this is one possible explanation for our finding there are others as well. An examination of the items in this scale suggests that they may be indicative of a controlling or authoritarian personality for some individuals. Recent research with a population of incarcerated batterers demonstrated that batterers were not more impulsive than other criminals but were more cognitively inflexible (Bueso-Izquierdo et al., 2016). Although the levels of violence in the current sample are low, this finding may be indicative of some of the mechanisms that are functioning at more extreme levels in highly violent couples and in cases of intimate terrorism (Johnson, 2008). The findings suggest that this measure may contain a number of artifacts affecting its validity among certain populations. However, that is an empirical question that should be addressed in future research.

The predicted interaction of actor alcohol use with actor ECF failed to reach the traditional level of statistical significance, however, the form of the interaction suggested that actors who were heavy drinkers and low on ECF may be more likely to commit IPA. Past research has demonstrated that individuals seeking alcohol abuse treatment who have a history of partner violence show more cognitive deficits in attention, concertation and cognitive flexibility than similar treatment seeking men without a history of partner violence (Easton et al., 2008) and laboratory studies have shown that individuals low on ECF are more aggressive when intoxicated (Giancola, 2004; Godleski and Giancola, 2009). While other research has found effects of certain aspects of ECF such as impulsivity (Hoaken et al., 2003; Schumacher et al., 2013) we used a composite score of ECF similar to that used by Giancola (2004) which combined measures of cognitive flexibility, attentional control, and working memory in order to provide a wide-ranging assessment of prefrontal cortex functioning. Previous research using similar ECF measures to ours which did find this interaction were highly controlled laboratory studies, suggesting that our inability to find a significant effect may be due to the significant amount of uncontrolled variability that is endemic of non-experimental research. Although we should interpret our finding with caution due to the failure to reach traditional levels of statistical significance, the pattern is consistent with the theory that heavy alcohol use when combined with a poor self-regulation may be a dangerous combination in the context of a relationship.

The data was analyzed in an APIM framework in order to examine partner and actor effects simultaneously. Two recent studies show showed that deficits in actor self-regulation predicted IPA and that partner alcohol use diminished the effect of actor self-regulation on aggression. In a study by Leone et al. (2016), an aspect of actor impulsiveness (negative urgency) predicted IPA among those whose partners were not heavy drinkers, however, this relationship between actor impulsiveness and IPA was attenuated when the partner was a heavy drinker. Having a heavy drinking partner was associated with actor IPA regardless of the actor's self-regulation. Similarly, Parrott et al. (2017) fund that the relationship between actor emotional regulation and IPA was stronger when the partner was not a problem drinker than when the partner was a problem drinker. Again, partner problem drinking had the effect of predicting IPA but it also reduced the relationship between self-regulation and IPA. We didn't find actor self-regulation by partner alcohol use interactions, but we did find interactions involving partner effects.

First, partner alcohol use interacted with partner negative affect to predict the probability of actor IPA suggesting that actors are more likely to be aggressive when a partner is high on negative affect and also heavy drinker. There are several possible explanations for this finding. One may be that the partner is also more likely to be aggressive and the members of the couple engage in reciprocal aggression. However, the lack of an actor negative affect by actor alcohol use interactions suggests that this is too simple an explanation. It may rather be indicative of a high conflict relationship. Past research has found couple conflict to interact with alcohol use to predict partner aggression (e.g., Quigley and Leonard, 1999). A partner who is frequently angry and intoxicated may more often be responded to with physicality regardless of the characteristics of the actor. The second interaction involving partner effects also suggests that greater conflict leads to more relationship aggression. Actor self-control interacted with partner self-control such that IPA was most probable when both actor and partner were low in self-control and least probable when the actor was low in self-control and the partner was high in self-control. When the actor was high in self-control the partner's self-control did not matter. While it has been established that individuals low in self-control are more likely to enact IPA (Finkel et al., 2009) and that low self-control is a risk factor for experiencing victimization (Pratt et al., 2014) this is the first research to show that both partner's self-control interacts to predict IPA.

The interaction found with time did not suggest that the relationship of self-regulation and alcohol use to IPA changes over time, rather it indicated overall reductions in aggression over time. In the three-way interaction involving time, the patterns suggested that the interaction was due to one condition in which the probability of IPA was low at wave one which stayed low over the next two waves while the probability of IPA in other conditions reduced over time. Partner aggression is known to reduce and even desist over time (Quigley and Leonard, 1996; Walker et al., 2013) and this fact is likely partly responsible for the reduction in IPA over time in this sample.

There are a number of strengths to the present analysis. The sample is unique in a number of ways that help address

### REFERENCES


limitations in past research on alcohol and IPA. First, the sample is older than the college age or newlywed samples that are usually examined in IPA research. Secondly, we oversampled heavy drinking women so that we could properly address the understudied issue of heavy alcohol use by women as a risk factor for their own IPA as well as for being victims of IPA. The oversampling of heavy drinking women may be why the APIM models were indistinguishable by gender as past samples rarely had equivalent numbers of heavy drinking men and women. Our findings were consistent with the multiple thresholds model of alcoholrelated aggression and indicated that the instigating effects of poor self-regulation in combination with the disinhibiting effects of alcohol creates a situation ripe for relationship aggression. This model can help us to understand how alcohol use and factors related to cognitive, emotional and behavioral control in both individuals in the relationship can influence the occurrence of partner violence. Future research should expand the types of self-regulatory factors examined in order to provide more complete understanding of how basic neurocognitive substrates related to self-control direct and moderate intoxicated behavior. Finally, our findings demonstrate the importance of considering the interactional qualities of marital conflict. While we often simply look at characteristics of the perpetrator, situations of partner aggression involve the interaction of two individuals each with their own patterns of alcohol use, personality differences and levels of cognitive functioning.

### ETHICS STATEMENT

The research reported in this manuscript was reviewed and approved by the University at Buffalo's Social and Behavioral Sciences Institutional Review Board.

### AUTHOR CONTRIBUTIONS

KL, BQ, MT and RH contributed to the design of the study. BQ, AL and JD contributed to data analysis. All authors contributed to the draft of the manuscript.

### FUNDING

This research was supported by National Institute on Alcohol Abuse and Alcoholism (NIAAA) grant AA016829 funded to KL.

an examination of actor and partner effects. Aggress. Behav. 40, 440–450. doi: 10.1002/ab.21533


with and without physical violence toward their partners: a preliminary report. Am. J. Drug Alcohol Abuse 34, 29–37. doi: 10.1080/009529907017 64326


neurocognitive substrate. J. Neurosci. 31, 4805–4810. doi: 10.1523/ JNEUROSCI.2859-10.2011


quality. Soc. Psychol. Pers. Sci. 2, 138–145. doi: 10.1177/19485506103 85710

Walker, K., Bowen, E., and Brown, S. (2013). Desistance from intimate partner violence: a critical review. Aggress. Violent Behav. 18, 271–280. doi: 10.1016/j. avb.2012.11.019

**Conflict of Interest Statement**: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Quigley, Levitt, Derrick, Testa, Houston and Leonard. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Differential Roles of the Two Raphe Nuclei in Amiable Social Behavior and Aggression – An Optogenetic Study

Diána Balázsfi1,2, Dóra Zelena<sup>1</sup> , Kornél Demeter<sup>1</sup> , Christina Miskolczi1,2 , Zoltán K. Varga1,2, Ádám Nagyváradi<sup>1</sup> , Gábor Nyíri<sup>3</sup> , Csaba Cserép2,3, Mária Baranyi<sup>4</sup> , Beáta Sperlágh<sup>4</sup> and József Haller1,5 \*

<sup>1</sup> Laboratory of Behavioural and Stress Studies, Institute of Experimental Medicine, Hungarian Academy of Sciences, Budapest, Hungary, <sup>2</sup> János Szentágothai School of Neurosciences, Semmelweis University, Budapest, Hungary, <sup>3</sup> Laboratory of Cerebral Cortex Research, Institute of Experimental Medicine, Hungarian Academy of Sciences, Budapest, Hungary, <sup>4</sup> Laboratory of Molecular Pharmacology, Institute of Experimental Medicine, Hungarian Academy of Sciences, Budapest, Hungary, <sup>5</sup> Institute of Behavioural Sciences and Law Enforcement, National University of Public Service, Budapest, Hungary

#### Edited by:

Rosa Maria Martins De Almeida, Universidade Federal do Rio Grande do Sul (UFRGS), Brazil

#### Reviewed by:

Alexa H. Veenema, Michigan State University, United States Aki Takahashi, University of Tsukuba, Japan Angela Roberts, University of Cambridge, United Kingdom

#### \*Correspondence:

József Haller haller.jozsef@koki.mta.hu; haller@koki.hu

Received: 11 December 2017 Accepted: 11 July 2018 Published: 02 August 2018

#### Citation:

Balázsfi D, Zelena D, Demeter K, Miskolczi C, Varga ZK, Nagyváradi Á, Nyíri G, Cserép C, Baranyi M, Sperlágh B and Haller J (2018) Differential Roles of the Two Raphe Nuclei in Amiable Social Behavior and Aggression – An Optogenetic Study. Front. Behav. Neurosci. 12:163. doi: 10.3389/fnbeh.2018.00163 Serotonergic mechanisms hosted by raphe nuclei have important roles in affiliative and agonistic behaviors but the separate roles of the two nuclei are poorly understood. Here we studied the roles of the dorsal (DR) and median raphe region (MRR) in aggression by optogenetically stimulating the two nuclei. Mice received three 3 min-long stimulations, which were separated by non-stimulation periods of 3 min. The stimulation of the MRR decreased aggression in a phasic-like manner. Effects were rapidly expressed during stimulations, and vanished similarly fast when stimulations were halted. No carryover effects were observed in the subsequent three trials performed at 2-day intervals. No effects on social behaviors were observed. By contrast, DR stimulation rapidly and tonically promoted social behaviors: effects were present during both the stimulation and non-stimulation periods of intermittent stimulations. Aggressive behaviors were marginally diminished by acute DR stimulations, but repeated stimulations administered over 8 days considerably decreased aggression even in the absence of concurrent stimulations, indicating the emergence of carryover effects. No such effects were observed in the case of social behaviors. We also investigated stimulation-induced neurotransmitter release in the prefrontal cortex, a major site of aggression control. MRR stimulation rapidly but transiently increased serotonin release, and induced a lasting increase in glutamate levels. DR stimulation had no effect on glutamate, but elicited a lasting increase of serotonin release. Prefrontal serotonin levels remained elevated for at least 2 h subsequent to DR stimulations. The stimulation of both nuclei increased GABA release rapidly and transiently. Thus, differential behavioral effects of the two raphe nuclei were associated with differences in their neurotransmission profiles. These findings reveal a surprisingly strong behavioral task division between the two raphe nuclei, which was associated with a nucleus-specific neurotransmitter release in the prefrontal cortex.

Keywords: aggression, serotonin, glutamate, GABA, dorsal raphe, median raphe

## INTRODUCTION

fnbeh-12-00163 July 30, 2018 Time: 20:9 # 2

Early studies from the 70's indicated that the serotonergic system plays an important role in aggression control. These studies revealed that the destruction of the main raphe nuclei (dorsal raphe, DR; median raphe region, MRR) increase aggression in mice, serotonin depletion by systemic para-chlorophenylalanine facilitates non-specific killing behavior in rats, and that aggressive behavior in humans is associated with low serotonin levels, the behavior being reversed by serotonin-enhancing drugs (Kostowski et al., 1975; Miczek et al., 1975; Greenberg and Coleman, 1976). The role of serotonin in aggression control was confirmed by subsequent animal and human studies. It was even stated that serotonin is the primary determinant of inter-male aggression, other neurotransmitters affecting it indirectly via serotonin signaling (Nelson and Chiavegatto, 2001). Besides controlling natural manifestations of aggressive behavior (Rosell and Siever, 2015; Sandi and Haller, 2015) deficits in serotonergic neurotransmission are implicated in the development of abnormal animal aggression, i.e., those aggressions that overpass species-specific levels and behavioral patterns (Haller et al., 2005; Haller et al., 2014; Miczek et al., 2015; Sandi and Haller, 2015). Not surprisingly, it was suggested that laboratory research aiming at the development of new psychotropic drugs for the treatment of aggression problems should target the serotonergic system (Olivier, 2015). Research performed in primates and humans support these findings obtained mainly in rodents, including the use of serotonergic compounds for the treatment of aggression-related psychopathologies (Coccaro et al., 2015; de Almeida et al., 2015; Glick, 2015; Zhang-James and Faraone, 2016). However, findings on the role of serotonin in aggression control are in many respects conflicting. Laboratory studies showed for instance that the chronic pharmacological reduction of serotonin availability by a series of serotonergic compounds promoted aggression, but aggression decreased when serotonin release was inhibited acutely (de Boer and Koolhaas, 2005). Some clinical studies show that selective serotonin reuptake inhibitors (SSRIs), decrease aggression in certain aggression-related psychopathologies while being ineffective in others (Coccaro et al., 2015; Glick, 2015); moreover, SSRIs promoted rather than decreased aggression in a series of well documented cases (Bielefeldt et al., 2016; Sharma et al., 2016). The reasons of such discrepant findings are largely unknown.

One possible explanation may reside in the differential involvement of the two main serotonergic nuclei in aggression control, e.g., the MRR and DR. These raphe nuclei send parallel and overlapping projections to limbic structures including the cortex in both animals and humans, but their projection patterns differ, and differences were found with regard to their functional and structural characteristics, including their sensitivity to psychoactive agents (Mulligan and Tork, 1988; Wilson and Molliver, 1991; Hornung, 2003; Hensler, 2006). Perhaps the largest difference between the two nuclei is that the majority of axons originating from the MRR form synapses in the forebrain, whereas DR projections rarely form synapses and exert their effects via volume transmission (Hornung, 2003; Hensler, 2006). Volume transmission (or non-synaptic communication) is typical to monoamine (particularly serotonergic and noradrenergic) and peptidergic neurotransmission. It affects extended brain areas, and targets high-affinity receptors located on extra-synaptic sites, e.g., the soma or dendrites of neurons, and modulate neuron activity rather than transmit information in the way synaptic communication does (Vizi, 2000; Leng and Ludwig, 2008). The findings briefly reviewed above show that the projections of the two raphe nuclei have different anatomical and functional properties; consequently, they may have distinct roles in behavioral control.

To investigate this issue, here we studied the behavioral consequences of MRR and DR stimulation on the social and aggressive behaviors of mice (i.e., non-aggressive and aggressive social interactions, respectively). Stimulations were performed by optogenetic techniques that allow a more precise control over the stimulated brain areas than electric stimulations. Several studies have shown that raphe nuclei are not homogenous neurochemically (Moore, 1980). Therefore, we also studied the impact of stimulations on neurotransmitter release in the prefrontal cortex, a major site of aggression control. In addition to serotonin release, we studied the release of glutamate and GABA, which are expressed by a large share of raphe neurons (Commons, 2009; Jackson et al., 2009; Varga et al., 2009; Sos et al., 2017); moreover, glutamate is often co-expressed with serotonin in the very same raphe neurons (Shutoh et al., 2008; Gagnon and Parent, 2014). We hypothesized that the DR and MRR are different in terms of both behavioral and neurochemical effects.

### MATERIALS AND METHODS

### Animals

Adult C57BL/6N male mice (Charles River, Budapest, Hungary), were used as residents in social encounters. They were 12–14 weeks old at the beginning of the study, e.g., at the time of their surgery. We used 20–25 days old CD1 mice (Charles River, Budapest, Hungary) as opponents in social interaction tests (**Figure 1C**). Animals were housed individually under a standard 12 h light–dark cycle (lights on at 6 am), with food and water available ad libitum. Experiments were approved by the local committee for animal health and care (Animal Welfare Committee of the Institute of Experimental Medicine) and performed according to the European Communities Council Directive recommendations for the care and use of laboratory animals (2010/63/EU).

### Virus Injection and Optogenetics

For the optical control of raphe regions, 40 nL adeno-associated virus vector (AAV; Penn Vector Core, PA, United States) encoding ChR2 (AAV2.5.hSyn.hChR2(H134R)eYFP.WPRE.hGH; 1.3e12 GC/ml; Addgene26973) were injected into the median raphe region (MRR) or dorsal raphe (DR) from glass pipettes (tip diameter 20–30 µm) connected to a MicroSyringe Pump Controller (World Precision Instruments, Sarasota, FL, United States)

under deep anesthesia (intraperitoneal injection of 25 mg/kg xylazine and 125 mg/kg ketamine in 0.9% NaCl) (Balazsfi et al., 2017). The coordinates of the virus injection were the followings: MRR: AP: −4.10 mm, L: 0.0 mm, DV: 4.60 mm; DR: AP: −4.40 mm, L: 0.0 mm, DV: −3.40 mm. Two weeks after the injection mice were implanted with optic fibers (core diameter: 105 µm; flat tip; MRR: 10◦ from dorsal, AP: −4.80 mm, L: 0.0 mm, DV: −4.10 mm; DR: 10◦ from dorsal, AP: −5.20 mm, L: 0.0 mm, DV: −3.35 mm). Optic fibers for implantation and light stimulation were custom made from multimode optical fiber (AFS 105/125Y, NA: 0,22, low-OH, Thorlabs Corp., Munich, Germany) and flanged zirconia ferrule (LMFL-172-FL-C35- OSK, Senko, Hampsire, United Kingdom). Implants were secured by screws and acrylic resin (Duracryl Plus; SpofaDental,

Czech Republic). Behavioral experiments started after 4–7 days recovery. Laser beams (473 nm) were generated by low noise diode-pumped solid-state lasers (Ikecool Corp., Anaheim, CA, United States), then collimated and guided to the implanted optic fiber by fiber-optic patch cords (FT900SM and FT030-BLUE, Thorlabs Corp.). Net energy output was measured by laser power meter (Coherent, LaserCheck, Santa Clara, CA, United States) before and after the experiments. Data were used only when optic fibers transferred 10-20 mW net energy at continuous light emission. The frequency of optogenetic stimulation was 20 Hz (25 ms pulses) in both the behavior and the microdialysis study.

### Experimental Design

Mice were exposed at 2-day intervals to four social interaction tests (see below); i.e., the total duration of the study (including inter-trial days) was 8 days. We used a roman square design. On day one, half of the animals were stimulated and half served as control. Controls were sham stimulated, i.e., they were connected to optic fibers but light was not delivered. The effects of optic stimulation were studied on this experimental day, when all mice were experimentally naïve. The findings of this trial were shown in **Figure 2** (MRR stimulation) and **Figure 3** (DR stimulation).

Three additional social encounters were run to investigate the carryover effects of stimulation. On each of these days, treatments were reversed compared to the previous trial such that each animal was exposed to a total of two control, and two stimulated social interaction tests. By carryover effects, we mean here those effects of stimulation that are detectable on the subsequent, nonstimulation trial. The findings of these trials were shown in **Figure 4**, and were expressed as changes compared to day 1. It was hypothesized that carryover effects, if present, would be independent of the ongoing stimulation. Therefore, the actual stimulation status of mice was not considered when carryover effects were studied. Note that there was no significant interaction between time and stimulation in trials 2–4.

A separate group of mice was used in the in vivo microdialysis study (see below).

### Social Interaction Test

The implanted animal (resident) was equipped with an optic fiber and was placed into the test cage (29 cm × 35 cm × 40 cm) with water and food available ad libitum for a 30 min habituation period. The test started when the intruder (CD1 mouse) was placed into the same cage (**Figure 1C**). The test lasted 21 min and divided to 3 min periods (**Figure 1D**). The 20 Hz optogenetic stimulation was administered in the second (3–6 min), fourth (9–12 min) and sixth (15–18 min) periods or the mice were left for 21 min with the intruder without stimulation.

We videotaped and later scored the behavior of resident (experimental) mice by means of a computer-based eventrecorder software<sup>1</sup> . The experimenter was blind to treatments. We recorded the following behaviors: inactivity/resting (no obvious activity), exploration/walking (walking through the cage or sniffing directed toward the environment), social investigation (sniffing at partner or anogenital sniffing), aggressive grooming

FIGURE 4 | Carryover effects of stimulations. Findings presented here show behavior observed in trials 2, 3, and 4 when all mice had a history of stimulation. The aim of this study was to investigate carryover effects (see Experimental design). (A,B) Differences in the duration of behaviors as compared to trial 1. Values show differences in the time devoted to a particular behavior expressed as the percentage of total test time. (C) The duration of offense in trials 1–4. Here data were shown separately for mice stimulated or non-stimulated within the particular trials. Each cohort of mice (indicated by roman numbers) was submitted to alternating trials of stimulation and non-stimulation. Horizontal line in (A,B), the average duration of behaviors in trial 1; Horizontal bars in (A,B) standard errors of the average duration of behaviors in trial 1; DR, dorsal raphe; MRR, median raphe region; <sup>∗</sup>Significant differences between trials in post hoc tests (p < 0.05 at least).

(pushing down the opponent, while it is standing or trying to escape, nibbling the fur and the skin with quick movements of the head), tail rattling (rapid rattling of the tail while the subject faces its opponent), wrestling (wrestling movements

FIGURE 5 | In vivo release of serotonin (A), glutamate (B), and GABA (C) in the prefrontal cortex of mice stimulated optogenetically in their raphes (median raphe region, MRR; dorsal raphe, DR). The stimulation protocol was identical with that employed for behavioral studies. Vertical blue lines, the timing of stimulations. Note that the first stimulation was started 90 min after the last basal sampling and 15 min before the fourth sampling, whereas the third stimulations started right at the beginning of the fifth fraction. Sample sizes: control n = 6; MRR stimulation n = 9; DR stimulation n = 5. Vertical columns at the right-hand side of graphs, neurotransmitter responses to the infusion of KCl into the raphes. DR, dorsal raphe; MRR, median raphe region; <sup>∗</sup> significant effect of stimulations compared to control levels, same time-point; # significant effect of KCl infusion as compared to baseline levels (the first three time points of each curve).

often associated with biting), chasing (quickly following the opponent which is fleeing; this behavior was subsequent to the delivery of bites to the opponent), defensive upright (trials of keeping the opponent at distance with forepaws while rising on hind legs), avoidance (evading the approaching opponent), and flight (quickly moving away from the chasing opponent). Defensive behaviors (defensive upright, avoidance, and flight) were extremely rare, whereas resting and exploration did not differentiate the groups. Therefore, these behaviors were not shown. We summed up aggressive grooming, tail rattling, wrestling and chasing as offensive behaviors. We recorded both the duration and frequency of all behaviors. For offense, we showed durations only, because frequencies and durations were highly correlated. In the case of bites, we showed frequencies, because these were very brief, and frequencies characterized them better than durations.

### In Vivo Microdialysis

Eight weeks after AAV-ChR2 injection mice were implanted with the optic fiber as described above. After 4–7 days recovery the animals were anesthetized by intraperitoneal 20% urethane (Reanal; Budapest, Hungary) and microdialysis probe [EICOM CX-I Brain Probe (membrane: artificial cellulose, molecular weight cut off: 50,000 Da, OD: 0.22 mm, length: 2 mm)] was inserted into the prefrontal cortex (AP: −4.80 mm; L: 0.0 mm; DV: 5.50 mm), while optic fiber was connected to MRR or DR region. After 2 h equilibration period we collected 9 samples, one in every 30 min. Perfusion rate was 2 µl/min (**Figure 5**) (Goloncser et al., 2017). The first three samples served as baseline. Stimulation started 15 min before the end of the fourth sampling to detect rapid responses. The last stimulation started at the beginning of the fifth sampling period to investigate the habituation of neurotransmitter release to repeated stimulations. The stimulation protocol was identical with that shown in **Figure 1D**. We continued sampling for an additional 1.5 h (samples 6–8). The last sample was collected during the administration of 100 mM KCl for 5 min. This was performed to test the responsiveness of neurons.

### HPLC Analysis of Neurotransmitters Neurotransmitters Serotonin, Glutamate and GABA in Dialysates Were Determined by Using HPLC

Method (Goloncser et al., 2017). The extraction solution (PCA) was 0.1 M perchloric acid that contained theophylline (as an internal standard) at 10 mM concentration. Initial volume of dialysis samples was measured and then diluted with an equal volume of ice cold PCA then supplemented with mobile phase "A" to 300 µL. The sample was centrifuged at 3510 g for 10 min at 0–4◦C and 240 µL was injected onto the enrichment column. The remainder (60 µL) of the microdialysis sample was diluted with distilled water and the pH was adjusted to 10.5 with 2.7 M Na2CO3. The samples were reacted with (20 µL) 20 mM dansyl chloride for 15 min at 70◦ temperature than the reaction was stopped by 10 µL formic acid. To determine glutamate and GABA content, the volume of 350 µL of the reaction mixture was injected onto the enrichment column.

The levels of serotonin were determined by online column switching separation using Discovery HS C18 50 × 2-mm and 150 × 2-mm columns. The flow rate of the mobile phases ["A" 10 mM potassium phosphate, 0.25 mM EDTA "B" with 0.45 mM octane sulphonyl acid sodium salt, 8% acetonitrile (v/v), 2% methanol (v/v), pH 5.2] was 350 or 450 µl/min, respectively in a step gradient application. The enrichment and stripping flow rate of buffer [10 mM potassium phosphate, pH 5.2] was 4 min. The total runtime was 55 min. The HPLC system used was a Shimadzu LC-20 AD Analytical & Measuring Instruments System, with an Agilent 1100 Series Variable Wavelength Detector set at 253 nm and an electrochemical (EC) amperometric detector BAS 400, Bioanalytical System set at 730 mV potential.

The levels of dansylated amino acids (glutamate and GABA) were separated by the above column system. The flow rate of mobile phases ["A" 10mM ammonium formate, 16.8% acetonitrile (v/v), methanol 4.8% (v/v), "B" 10 mM ammonium formate, 70% acetonitrile (v/v), methanol 20% (v/v), pH 3] was 400 µl/min in a linear gradient mode. The enrichment and stripping flow rate of the buffer [10 mM ammonium formate, 1.9% acetonitrile (v/v), 1.1% methanol (v/v)] was 300 µL/min during 4 min and the total runtime was 55 min. The used analytical system was the above, Shimadzu LC-20 System, with Gilson Model 121 Fluorimeter set at 340 nm excitation and 450 nm emission wavelength.

The recovery of the implanted microdialysis probes was evaluated at the end of experiment. The in vitro extraction efficiency for serotonin, glutamate and GABA were estimated to be 21.1 ± 4.8%, 17.1 ± 2.8%, and 21.9 ± 3.4%, respectively. The concentrations of serotonin, glutamate and GABA were expressed in percentage (mean ± SEM) of baseline concentrations in order to monitor changes from basal levels after optical stimulation.

### Anatomical Analysis

After termination of the behavioral experiments mice were deeply anesthetized (see above) and transcardially perfused with 0.1M phosphate buffered saline (PBS) for 1 min, then with 4% (w/v) paraformaldehyde (PFA) in PBS for 20 min. Optic fibers were carefully removed, brains were taken out, and post-fixed for 24 h in fixative at +4 ◦C. Brains were cryo-protected by 20% glucose-PBS solution for 24 h at +4 ◦C. At the end of the microdialysis optic fiber and microdialysis probe were removed carefully, and brains were postfixed for 24 h in 30% glucose containing PFA at +4 ◦C. To enhance the green fluorescence protein (GFP) signal and to facilitate the identification of the MRR and DR, immunofluorescent staining was carried out on 50-µm-thick coronal sections (prepared on a Vibratome VT1200S, Leica, Wetzlar, Germany). Primary antibodies were diluted in Tris-buffered saline (TBS) (Rabbit-anti-Serotonin, 1:10000, ImmunoStar, Hudson, WI, United States; CatNo: 20080; Chicken-anti-GFP, 1:2000, Life Technologies, Carlsbad, CA, United States; CatNo: A10262) and were incubated for 2 days. After washing, sections were incubated in secondary antibody solution overnight (Cy3-conjugated Donkey-anti-Rabbit, 1:500, Jackson ImmunoResearch West Grove, PA, United States; CodeNo:711-165-152; Alexa488-conjugated Goat-anti-Chicken, 1:1000, Life Technologies, CatNo: A-11039; diluted in TBS). After multiple washes, sections were mounted and were evaluated with a Zeiss Axioplan microscope, and images were taken with an Olympus DP70 camera.

The position of the tip of the optical fiber, microdialysis probe and the size of the virus infected area were determined on micrographs by using on overlay of the stereotaxic atlas images on the series of images of the MRR and DR (Paxinos, 2001) (**Figures 1A,B**). We estimated the laser-illuminated volume based on the measurements by Yizhar et al. (2011). Mice with weak virus expression in the MRR or DR or with the optic fiber outside these regions, or the microdialysis probe outside the PFC were excluded from the analysis.

### Statistics

Data were represented as means ± standard error of the mean. Behavioral differences were evaluated by repeated measure ANOVA when temporal data series were evaluated. Two-way ANOVA was performed when the individual time-points of such temporal data series were averaged (see for instance the righthand panels of **Figure 2**). Factors were indicated in Results. ANOVA was followed by Dunnet post hoc comparisons where main effects were significant. In the case of bite counts, which did not fulfill ANOVA requirements, statistical differences were evaluated by the median test, a subtype of Pearson's chi-squared test. In the in vivo microdialysis study, a two factor repeated measures ANOVA was employed (repeated measures factor 1 was 'time'; factor 2 was 'groups'). P values lower than 0.05 were considered significant; P-values lower than 0.1 but larger than 0.05 were identified as trends.

### RESULTS

## Acute Effects of Intermittent MRR Optic Stimulation

Social interactions were intense during min 0–3 of the 21 min-long encounter in controls, but decreased rapidly and were maintained at low levels throughout the encounter [Ftime(6,90) = 69.14; p < 0.0001] (**Figure 2A**). The optic stimulation of the MRR did not affect social behavior [Ftreatment(1,15) = 0.42; p > 0.6; Finteraction(6,90) = 0.41; p > 0.9]. In sharp contrast, MRR stimulation strongly influenced aggressive interactions. In controls, aggressive interactions showed low levels in the first 3 min of the encounter, reached a peak between min 3 and 6, and gradually decreased thereafter (**Figure 2B**, circles). Changes in MRR-stimulated mice did not follow this pattern (**Figure 2B**, squares). Values comparable with those seen in controls were recorded in the periods when stimulations were not administered, but the duration of aggressive interactions sharply decreased during stimulation periods [Ftreatment(1,15) = 0.03; p > 0.9; Ftime(6,90) = 1.81; p > 0.2; Finteraction(6,90) = 3.51; p < 0.005]. The distribution of biting behavior showed that controls displayed two major bouts of bite delivery, one between min 3 and 6, and another one between min 18 and 21, i.e., toward the end of the

stimulation period (**Figure 2C**). MRR stimulation profoundly altered this distribution: bite delivery was frequent in-between stimulations, but significantly less frequent when stimulations were administered (χ <sup>2</sup> = 31.95; p < 0.005). There were no significant differences between the non-stimulated and shamstimulation periods in controls (χ <sup>2</sup> = 0.28; p > 0.6), but in stimulated mice, non-stimulation and stimulation periods differed significantly (χ <sup>2</sup> = 5.55; p < 0.02) (**Figure 2C**, right-hand panel).

Data suggested a phasic-like effect of MRR stimulation on aggressive behavior. To investigate this issue further we studied the duration of offensive behaviors in bins of 1 min (**Figure 2D**). The time course of offensive behaviors was markedly different in MRR-stimulated mice as compared to controls (**Figure 2D**, lefthand panel) [Ftreatment(1,15) = 0.03; p > 0.9; Ftime(20,300) = 1.64; p < 0.05; Finteraction(20,300) = 1.61; p < 0.05]. For clarity, we illustrated this difference by averaging each non-stimulation min that preceded the stimulation periods, as well as each of the 3 min of the stimulation periods (**Figure 2D**, right-hand panel). Stimulations decreased offensive behaviors rather rapidly, e.g., during the first min of their administration and this effect carried over to the next min. Interestingly, however, offense returned to control levels during the third min when stimulation was still administered.

A similar analysis was not performed for bite counts, as their display was sparse, and a min-by-min analysis would have been meaningless. Other behaviors were not affected by stimulations (**Table 1**).

Taken together, these findings show that the optogenetic stimulation of the MRR specifically inhibits aggressive behaviors in a phasic-like manner, particularly offense and bite delivery. Non-aggressive social interactions remained unaltered.

### Acute Effects of Intermittent DR Optic Stimulation

In controls, the duration of social behavior followed the same temporal evolution as in the first experiment [Ftime(6,72) = 37.56; p < 0.0001] (**Figure 3A**). However, DR stimulation -in contrast to MRR stimulation- significantly increased social behavior throughout the encounter [Ftreatment(1,12) = 4.66; p = 0.05]. The two groups showed small differences before the first stimulation (min 0–3). The two groups showed small differences before the first stimulation (min 0–3). To test whether this affected group differences in later phases of the encounter, we performed a second analysis, in which pre-stimulation and post-stimulation behaviors were evaluated separately. Between min 0–3, differences in social interactions were not significant [F(1,12) = 0.82; p < 0.4]. By contrast, group differences (expressed as % of min 0–3 values) were significant [Ftreatment(1,12) = 6.47; p < 0.03; Ftime(5,60) = 3.57; p < 0.01; Finteraction(5,60) = 0.21; p > 0.9]. There was no interaction between the factors, suggesting that DR stimulation had a toniclike effect. In order to visualize the lack of impact of baseline differences, we showed post-stimulation values as the percentage of pre-stimulation ones in the insert of **Figure 3**. Offensive behavior decreased throughout the encounter [Ftime(6,72) = 5.87; p < 0.0001] (**Figure 3B**). This was slightly different from the pattern seen in the first experiment, where offense was low in min 0–3, increased between min 3–6 and decreased thereafter (**Figure 2B**). However, a min-by-min presentation of the findings suggests that the patterns of change in controls (i.e., non-stimulated mice) were similar of the two experiments (**Figure 3B**, insert). The figure suggests that offensive aggression was decreased by DR stimulation, but due to large variation the change was not significant [Ftreatment(1,12) = 3.09; p > 0.2; Finteraction(6,72) = 0.41; p > 0.9]. When, however, averages were calculated for the duration of this behavior over the whole encounter (**Figure 3B**, right-hand panel), there was a trend toward decreased aggressiveness in stimulated mice [Ftreatment(1,12) = 3.66; p = 0.07]. The temporal evolution of bite delivery was also similar to that observed in the first experiment: in controls, two main bouts of bite delivery were identified particularly between min 6–9 and min 18–21 of the encounter (**Figure 3C**). DR stimulation did affect bite counts (Chi square for all time-points = 11.97; p < 0.05), but this effect was restricted to min 6–9, e.g., to the 3 min block that followed the first stimulation (Chi square for this time-point = 3.89; p < 0.05) (**Figure 3C**).

Non-social behaviors were not affected by DR stimulation (**Table 2**).

Taken together, these findings show that DR stimulation increases non-aggressive social interactions, and decreases offensive behaviors at trend level. DR stimulation also abolished the peak in biting behavior observed in controls between 6 and 9 min. Neither of these effects was restricted to the periods of stimulation, suggesting that the DR exerts tonic effects on behavior.

### Carryover Effects of Repeated MRR and DR Optic Stimulation

Subsequent to the first encounter, mice were submitted to three additional ones at 2-day intervals. Optogenetic stimulations were administered according to a roman square design (see Experimental design). As such, all mice had a history of stimulations by the end of the second trial. To identify carryover effects, the actual stimulation status of mice was not considered, because it was hypothesized that carryover effects, if present, would be independent of the ongoing stimulation. Noteworthy, there was no significant interactions between time and stimulation in trials 2–4.

No carryover effects were observed with MRR stimulation. **Figure 4A** presents behavioral differences as compared to the first trial; no statistically significant changes were observed [social behavior: F(3,48) = 0.43; p > 0.8; offense: F(3,48) = 1.11; p > 0.4]. The same was true for social behaviors in the case of DR stimulation [F(3,24) = 0.38; p > 0.8] (**Figure 4B**, left-hand panel).

By contrast, offensive behavior was dependent on the history of DR stimulation (**Figure 4B**, right-h and panel, and **Figure 4C**). As compared to trial 1, offensive behaviors decreased in trials 3 and 4, when all mice had a stimulation history [F(3,24) = 6.44; p < 0.01]. **Figure 4C** shows that indeed, this effect did not depend on the actual stimulation status of mice. The duration of offensive threats decreased over trials [Ftrial(3,20) = 5.98; p < 0.01], but


TABLE 1 | The effect of MRR optic stimulation on non-social behaviors on day 1.

Data (mean ± SEM) show the duration of behaviors expressed in percent of total test time. Sham and real stimulation periods were indicated in gray and blue, respectively. Values showed a significant temporal evolution without significant impact of MRR stimulation.

no stimulation effects were observed [Fstimulation(1,20) = 0.12; p > 0.8], and there was no interaction between these factors [Finteraction(3,20) = 0.57; p > 0.7]. Behavioral data obtained in trials 2–4 were shown in more detail in **Tables 3**, **4** (for the first trial, see **Figures 1–3**). These tables show that the behavioral effects resulting from stimulation in trial 1 were roughly replicated in subsequent trials, except for the gradual decrease in offense after DR stimulation. No similar decrease was observed after MRR stimulation.

Taken together, these findings show that DR but not MRR stimulations have carryover effects. Particularly, offensive threats decreased in mice with a history of stimulation, and this effect was independent of ongoing stimulations. Considering that offense was affected by DR stimulation only at trend level in trial 1, and that aggression levels decreased in trials 3 and 4 irrespective to current stimulation status, one can hypothesize that the mechanisms underlying this phenomenon are different from those that underlie the acute effects of DR stimulation.

### Neurotransmitter Release in the Prefrontal Cortex After MRR and DR Optic Stimulation

The neurochemical consequences of raphe stimulations were studied in the prefrontal cortex, an area deeply involved in the control of aggression and social behavior in general. Importantly, the particularities of stimulations were similar to those employed in the behavioral studies, albeit mice were anesthetized this time.

The prefrontal release of all three, serotonin, glutamate, and GABA were increased after the optogenetic stimulation of raphe nuclei [serotonin: Ftime(7,119) = 7.06; p < 0.01; glutamate: Ftime(7,119) = 2.56; p = 0.01; GABA: Ftime(7,119) = 4.14; p < 0.01]. Moreover, at the termination of the experiment 100 mM KCl was able to increase neurotransmitter release remarkably in all animals confirming that the cells remained alive and reactive (p < 0.01 comparing the last fraction to all others except stimulated ones) (**Figure 5**, columns). However, the neurochemical consequences of MRR or DR stimulation largely depended on the stimulated brainstem area [serotonin: Ftime∗group(14,119) = 2.44; p < 0.01; glutamate: Ftime∗group(14,119) = 1.82; p < 0.05; GABA: Ftime∗group(14,119) = 1.62; p = 0.082]. As compared to baseline, the extracellular release of serotonin was increased during the optogenetic stimulation of both the MRR and DR (**Figure 5A**). Note that the samples contained a microdialysate of 30 min, whereas stimulations lasted only 3 min. Consequently, release induced by stimulation was considerably diluted, which may explain the relatively low levels of serotonin in the dialysate. The temporal evolution of the release was, however, rather different with the two nuclei [Fgroups(2,17) = 9.68; p < 0.01]. The increase vanished relatively rapidly when the MRR was stimulated., whereas DR stimulation induced a long lasting increase in release: prefrontal serotonin levels were higher 2 h after the last stimulation as compared to controls.

Glutamate release was increased only in mice stimulated in their MRR. DR stimulation had no similar effect (**Figure 5B**). Note that in contrast to serotonin, the increase in glutamate release was observed after a considerable delay, but at the same time the effect was lasting, as it was observed 1h after the first stimulation. GABA release increased immediately after the first stimulation as with serotonin release, but was transient in both groups [Fgroups(2,17) = 2.03; p > 0.1] (**Figure 5C**).

We also investigated the release of dopamine and noradrenaline in the prefrontal cortex; stimulations affected neither (data not shown).

These findings show that the stimulation of the MRR and DR show some similarities as it regards their neurochemical consequences in the prefrontal cortex, but also show important differences. The impact of stimulations on GABA release was



Data (mean ± SEM) show the duration of behaviors expressed in percent of total test time. Sham and real stimulation periods were indicated in gray and blue, respectively. Values showed a significant temporal evolution without a significant impact of DR stimulation.

TABLE 3 | The effects of median raphe region stimulation on social interactions and offense.

(1) Social interactions


(2) Offense


Data (mean ± SEM) show the duration of behaviors expressed in percent of total test time. Real stimulation periods were indicated in blue. Averages were outlined in bold font. For statistics, see text and Figure 4.

TABLE 4 | The effects of dorsal raphe stimulation on social interactions and offense.

#### (1) Social interactions

fnbeh-12-00163 July 30, 2018 Time: 20:9 # 12


(2) Offense


Data (mean ± SEM) show the duration of behaviors expressed in percent of total test time. Real stimulation periods were indicated in blue. Averages were outlined in bold font. For statistics, see text and Figure 4.

similar in the groups. By contrast, glutamate release was induced by MRR stimulation only, whereas the release of serotonin -albeit present in both groups- was transient with MRR stimulation, and surprisingly long-lasting with DR stimulations.

### DISCUSSION

### Main Findings

The dorsal and median raphe affected social behavior and aggression differently in our study. MRR stimulations decreased aggression in a phasic-like manner. Effects were restricted to the stimulation periods, and vanished in the non-stimulation periods that separated stimulations. No effects on social behaviors were observed. By contrast, the DR stimulation rapidly promoted social behaviors, but in a tonic fashion. Effects were present during both the stimulation and non-stimulation periods. Aggressive behaviors were marginally diminished by DR stimulation in the first trial, but repeated stimulations administered over 8 days considerably decreased aggression suggesting that repeated DR stimulations have slowly developing effects.

The effects of MRR and DR stimulation on neurotransmitter release were markedly different in the prefrontal cortex, a major site of aggression control. MRR stimulation increased serotonin release relatively rapidly, but transiently, and induced a major and more durable increase in glutamate release. By contrast, DR stimulation had no effect on glutamate release, but persistently increased prefrontal levels of serotonin. Release remained higher than the baseline long after stimulations halted. Effects on GABA release were transient with both nuclei.

### Raphe Nuclei and Serotonin

Ample evidence demonstrates that the neurochemical properties of raphe neurons are heterogenous: about their half or more are non-serotonergic (depending on the study; Moore, 1980). Glutamatergic and GABAergic neurons are significant components of both raphe nuclei, some studies suggesting that they are more numerous than serotonergic ones (DR: (Gamrani et al., 1979; Nanopoulos et al., 1982; Commons, 2009; Jackson et al., 2009); MRR: (Allers and Sharp, 2003; Varga et al., 2009; Sos et al., 2017); moreover, disparate studies suggest that the share of serotonergic neurons is below 10% in the median raphe (Sos et al., 2017). In addition, serotonergic neurons often coexpress (sometimes several) other neurotransmitters, suggesting that even serotonergic neurons release non-serotonergic neurotransmitters (Kachidian et al., 1991; Shutoh et al., 2008;

Gagnon and Parent, 2014; Sos et al., 2017). As such, behavioral effects obtained by the stimulation of raphe nuclei are not necessarily attributable to serotonin.

Although the existence of long-range GABAergic neurones was repeatedly suggested (Melzer et al., 2012; Caputi et al., 2013; Lee et al., 2014), no earlier publication confirmed that the axon terminals of raphe GABA neurons can reach the prefrontal cortex, the GABA response to stimulation was likely secondary to the release of other neurotransmitters, e.g., serotonin or glutamate which responded to raphe stimulation in our study. It is worth to note, however, that a large share of raphe neurons seems to be neither glutamatergic, GABAergic nor serotonergic (Sos et al., 2017). Such neurons may express other neurotransmitters, e.g., dopamine (Jahanshahi et al., 2013). Albeit the connectivity of some non-serotonergic raphe neurons is poorly known, one cannot rule out that they contributed to the behavioral effects observed here, as all three serotonin, glutamate and dopamine contribute to the control of aggression by the prefrontal cortex (Takahashi et al., 2011; Hwa et al., 2015; Tielbeek et al., 2016). The particular roles of these raphe mechanisms can be investigated only by neuron type-specific expression of channelrhodopsin, e.g., by the use of CRE mice. A differential study of such subsystems may be the target of subsequent research.

### Differential Role of Raphe Nuclei in Aggression: Comparisons With Earlier Studies

While the inhibition of aggression by the dorsal raphe is well-established (Pucilowski and Kostowski, 1983; Takahashi and Miczek, 2014; Miczek et al., 2015), the role of the median raphe is more controversial. Early studies provided negative results; e.g., DR lesions lastingly promoted aggressive behavior, whereas MRR lesions were without effect (Jacobs and Cohen, 1976). In a similar fashion, the stimulation of the dorsal raphe did, whereas the stimulation of the median raphe did not inhibit aggression in a study involving muricide (Pucilowski and Kostowski, 1981). It occurs that more subtle manipulations also emphasize the role of the DR over those of the MRR. E.g., the activation of GABA<sup>B</sup> receptors in the DR but not in the MRR promoted the display of escalated aggression (Takahashi et al., 2010). Other studies did find a role for MRR in aggression control; e.g., the 5,7-dihydroxytryptamine-mediated destruction of the MRR decreased submissiveness in rats and elicited behaviors indicative of aggressive arousal albeit not aggression per se (File et al., 1979). In another study, however, counterintuitive effects of MRR downregulation were observed. The 5-HT1A receptor agonist 8-OH-DPAT, an activator of somatodendritic autoreceptors, decreased maternal aggression when microinjected into the MRR of female rats (De Almeida and Lucion, 1997). Thus, the downregulation of MRR serotonin neurotransmission achieved by negative feedback decreased rather than increased aggressiveness.

We suggest that such controversial findings may at least be partly explained by the phasic-like effects of MRR neurotransmission on aggression as revealed by the present study. Such effects may easily be overlooked in studies using different experimental approaches, as the anti-aggressive effects of MRR stimulation seem to vanish rather rapidly. In our study, offense decreased in the first two, but returned to control levels during the third min of stimulation (**Figure 2D**). Earlier findings corroborated with our release studies may even suggest that the anti-aggressive effects of MRR stimulations may be reversed over time. It was shown that a short pulse of serotonin is likely to induce inhibition in the cortex, whereas the prolonged presence of serotonin may result in excitation (Zhou and Hablitz, 1999). In line with these observations, MRR stimulation increased serotonin and GABA release within 15 min in the prefrontal cortex, but these effects disappeared upon repeated stimulations to give raise to a large increase in glutamate release (**Figure 5**). One can tentatively hypothesize that this change in the neurochemical consequences of stimulations may have reversed their behavioral effects if stimulations were more durable. The complex neurochemical effects of MRR stimulation may at least partly explain the controversial findings briefly reviewed above.

### Limitations

As a first attempt to differentiate the roles of the two raphe nuclei in sociability and aggression by optogenetic techniques, our experiments have limitations, which need to be addressed in future studies. We investigated neurotransmitter release only in the prefrontal cortex, and in anesthetized animals. There are several other key regions in the circuitry that controls aggression, and neurotransmitter release may be influenced by anesthesia, albeit controls were also anesthetized. Nevertheless, our microdialysis study revealed two important aspects of raphe function. One was technical: the study showed that the stimulation of the DR and MRR induces serotonin release in areas involved in aggression control. Serotonin was outlined here because in contrast to glutamate and GABA, its release cannot be attributed to local neurons. The second important conclusion of this study was that the stimulation of the DR and MRR elicits substantially different neurochemical responses (at least in the prefrontal cortex). The differential neurochemical consequences of stimulations can be attributed to the particularities of the two raphe nuclei rather than to anesthesia.

The second limitation of the study relates to differences in the temporal resolution of the behavioral and the neurochemical experiment. Behaviors were investigated in bins of 3 min, whereas neurotransmitter release was studied in samples taken at 30 min intervals due to technical reasons. Fraction No. 4 reflected fast responses, because stimulation was started just 15 min before this fraction was collected. Fraction No.5 indicated habituation/serotonin depletion due to repeated stimulations, whereas subsequent fractions indicated prolonged effects. Based on findings, one can confidently assume that the serotonin and GABA response occurred shortly after stimulation, whereas the glutamate response was slower. Yet, the next sample was taken with a delay, thus, it is impossible to evaluate how

slow the glutamate response was. A more rapid sampling and detection methodology can overcome this deficiency in the future. Nevertheless, our findings may provide a preliminary clue on the mechanisms underlying the behavioral effects. For instance, the phasic-like effects of MRR stimulation are unlikely to be mediated by glutamatergic neurotransmission, because the behavioral response appears to be faster than the glutamatergic one. On the other hand, the tonic-like behavioral effects of DR stimulation may be due to serotonin release, as this was the only persistent neurochemical response observed in the microdialysis study. One cannot rule out that blue light per se had effects, as sham stimulations were in fact no stimulations. Yet, our studies performed in parallel showed that blue light per se has no measurable effects on social behavior (Biro et al., 2018). The issue naturally needs further experimentation.

Albeit not necessarily a limitation, we mention here that we failed to observe those rapid behavioral effects of raphe stimulation that were described in several laboratories, including ours (Ohmura et al., 2014; Balazsfi et al., 2017; Correia et al., 2017). In these studies, the stimulation of the DR or the MRR rapidly suppressed locomotion, increased anxiety, or resulted in the emergence of conditioned fear. Effects usually developed within seconds except for conditioned fear, but even in this study (Balazsfi et al., 2017) rapid effects on locomotion were evident when mice were stimulated in the MRR. No locomotion effects were observed in the social interaction test in the present experiments. It is unlikely that the reason was technical, as the aforementioned study of ours was performed under entirely similar conditions than this one. One can tentatively hypothesize that the environment has a decisive impact on the consequences of raphe stimulation. E.g., effects induced in non-social testing environments (Ohmura et al., 2014; Balazsfi et al., 2017; Correia et al., 2017) may be overruled or changed in social contexts (this study).

### CONCLUSION

Our findings suggest that the raphe nuclei provide several ways to control the social behaviors. They inhibit aggression in a phasiclike manner, while increasing amiable interactions tonically. Both effects are rapid, but have a different time-course. The two effects dissociate anatomically: the phasic-like control of aggression can be attributed to the MRR, whereas the tonic control of

### REFERENCES


social behaviors to the DR. The latter also seems to exert a slowly developing anti-aggressive effect, which can be expressed independently of actual DR activation. The differential roles of the two raphe nuclei are likely explained by their differential neurotransmission profiles in target areas.

Understanding the role of serotonin in aggression requires information on both the anatomical source of serotonergic inputs at various release sites, and the elucidation of the interactions between various neurotransmitter systems located within the raphe nuclei. It has been suggested that MRR and DR projections differ in their sensitivity toward pharmacological agents (Hensler, 2006). If true, a better understanding of the separate roles of raphe nuclei in aggression may help understanding controversial findings with the available agents and may also help designing novel treatment strategies.

### AUTHOR CONTRIBUTIONS

DB and DZ microinjected the virus carrier and implanted the optic fibers; they also performed behavioral and microdialysis experiments, and contributed to designing the study, as well as to the analysis and interpretation of findings. KD together with GN and CC checked the optic fiber and microdialysis probe placements and virus infection by immunocytochemistry. CM, ZV, and ÁN contributed to the scoring of behaviors, whereas MB and BS studied neurotransmitters in dialysates obtained from the prefrontal cortex. JH contributed to the designing of the study, the interpretation of findings, and wrote the first draft of the manuscript.

### FUNDING

This study was supported by NKFIH grant no. 112907, ERC-2011-ADG-294313 (SERRACO), and KÖFOP-2.1.2-VEKOP-15- 2016-00001 (all to JH) as well as NKFIH grant no. 120311 to DZ. The agencies had no further role in study design, in the collection, analysis or interpretation of the data.

### ACKNOWLEDGMENTS

The authors wish to thank Éva Dobozi for expert technical assistance (Semmelweis University, Budapest, Hungary).

in adult healthy volunteers. J. R. Soc. Med. 109, 381–392. doi: 10.1177/ 0141076816666805



neurons of the caudal raphe nuclei: a dual labeling in the rat. J. Neurosci. Res. 30, 521–530. doi: 10.1002/jnr.490300309



**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Balázsfi, Zelena, Demeter, Miskolczi, Varga, Nagyváradi, Nyíri, Cserép, Baranyi, Sperlágh and Haller. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# The COMT Val158Met Polymorphism and Reaction to a Transgression: Findings of Genetic Associations in Both Chinese and German Samples

Cornelia Sindermann<sup>1</sup> \*, Ruixue Luo<sup>2</sup> , Yingying Zhang<sup>2</sup> , Keith M. Kendrick <sup>2</sup> , Benjamin Becker <sup>2</sup> and Christian Montag1,2 \*

*<sup>1</sup> Department of Molecular Psychology, Institute of Psychology and Education, Ulm University, Ulm, Germany, <sup>2</sup> MOE Key Laboratory for Neuroinformation, The Clinical Hospital of Chengdu Brain Science Institute, University of Electronic Science and Technology of China, Chengdu, China*

#### Edited by:

*Nelly Alia-Klein, Icahn School of Medicine at Mount Sinai, United States*

#### Reviewed by:

*Silvia Solis-Ortiz, Universidad de Guanajuato, Mexico Francesco Papaleo, Fondazione Istituto Italiano di Technologia, Italy*

#### \*Correspondence:

*Cornelia Sindermann cornelia.sindermann@uni-ulm.de Christian Montag christian.montag@uni-ulm.de*

> Received: *10 April 2018* Accepted: *27 June 2018* Published: *03 August 2018*

#### Citation:

*Sindermann C, Luo R, Zhang Y, Kendrick KM, Becker B and Montag C (2018) The COMT Val158Met Polymorphism and Reaction to a Transgression: Findings of Genetic Associations in Both Chinese and German Samples. Front. Behav. Neurosci. 12:148. doi: 10.3389/fnbeh.2018.00148* After a transgression, people often either tend to avoid the transgressor or seek revenge. These tendencies can be investigated via a trait approach and surprisingly little is known about their biological underpinnings. One promising candidate gene polymorphism, which may influence individual differences in avoidance of a transgressor and vengefulness, is the COMT Val158Met (rs4680) polymorphism known to affect dopaminergic signaling and among others brain activity in situations in which people punish others for their behavior. We therefore investigated the molecular genetics of individual differences in Avoidance Motivation and vengefulness with a focus on this polymorphism. Possible genetic associations were first investigated in a sample of *N* = 730 Chinese participants (*n* = 196 females) using buccal cells to extract the DNA for genotyping. To replicate the findings we carried out a parallelized investigation in a sample of *N* = 585 German participants (*n* = 399 females). Chinese and German versions of the TRIM-12 and the Vengeance Scale were implemented to assess individual differences in tendencies to react to a transgression. Results show that Met allele carriers of the COMT Val158Met polymorphism (Val/Met and Met/Met) score significantly higher on the tendency to avoid a transgressor in the Chinese male and female samples, with an especially pronounced effect in the female subgroup. The same effect could be found in the German sample, again especially in females. Additionally, carrying a Met allele was associated with higher vengefulness in the Chinese sample only, especially in males. The present findings indicate that the COMT Val158Met polymorphism might influence individual differences in the motivation to avoid transgressors across cultures, especially in females. However, its association with vengefulness seems to be more complex and may exhibit some cultural and gender specific effects.

Keywords: vengefulness, Avoidance Motivation, COMT Val158Met, China, Germany

## INTRODUCTION

There are many different potential reactions to a transgression, although the main ones are for the victim to forgive the transgressor, avoid him/her or take revenge. Acts of revenge taking are reported across the globe and also vengefulness as a personality trait has already been investigated in different cultures (e.g., Henrich et al., 2006; Sindermann et al., 2018). Quantitative genetic studies have revealed that individual differences in reactive aggression (heritability estimate: 39%), the reaction to unfair offers (punishing an opponent for an unfair offer; additive genetic effects explained 42% of the variance) and antisocial personality traits such as various types of aggression (heritability estimates: 26–56%), possess a significant heritable component (Vernon et al., 1999; Brendgen et al., 2006; Wallace et al., 2007). This leads to the conclusion that also variance in trait vengefulness might be explained partly by genetic underpinnings. Currently however, surprisingly little is known about the genes/genetic polymorphisms influencing individual differences in trait vengefulness.

To search for possible candidate gene polymorphisms in the context of trait vengefulness, one can take a closer look at revenge behavior. In comparison with trait vengefulness, there are several studies investigating the neurobiological and genetic basis of revenge taking and associated behaviors (e.g., altruistic punishment) in laboratory settings (e.g., Sanfey et al., 2003; De Quervain et al., 2004; Strobel et al., 2011). In these studies, the dopaminergic reward system appears to be of particular importance since punishing others for unfair behavior (which can be seen as a kind of revenge taking) was associated with activation in brain areas such as for example the (dorsal) striatum, nucleus caudatus (NC), and the (anterior) insula (Sanfey et al., 2003; De Quervain et al., 2004). These brain areas have also been associated with reward processing or its anticipation (e.g., Delgado et al., 2003; Liu et al., 2011; Sescousse et al., 2013) and their activity is known to be modulated by dopaminergic pathways (Gaspar et al., 1989; Arias-Carrión et al., 2010). This suggests that pathways and thereon genes and genetic polymorphisms influencing the dopaminergic reward system are associated with acts of revenge and may therefore also be associated with trait vengefulness.

In this regard, one study in the field is of particular interest. Strobel et al. (2011) observed that brain regions involved in reward processing (cingulate gyrus (CG), insula, dorsolateral prefrontal cortex (DLPFC), nucleus accumbens (NAc), NC) were more strongly activated during trials in which participants punished a person who made an unfair offer compared to trials in which participants did not punish anybody. Additionally, the activation of the NAc and bilateral CG was also higher in trials in which participants punished a person who initially made the unfair offer to the participants themselves (revenge), compared to trials in which participants punished a person, which made an unfair offer to a third person (altruistic punishment). Notably, this study also found that the genotype of the Catechol-O-Methyltransferase (COMT) Val158Met (rs4680) polymorphism modulated the neural responses. In detail, it was found that the activation in reward related brain areas (clusters of the left anterior cingulate cortex (ACC), right posterior insula, right NAc) in the punishment trials (vs. no punishment) was higher in Met allele carriers compared to (homozygote) Val allele carriers. This polymorphism is positioned in the COMT gene on chromosome 22q11 (https:// www.ncbi.nlm.nih.gov/projects/SNP/snp\_ref.cgi?rs=4680) and causes a valine to methionine (G → A) substitution. Moreover, this polymorphism is functional on a biological level as it moderates the activity of the COMT enzyme and thereby dopamine catabolization in the synaptic cleft. This mechanism is of special importance in the prefrontal cortex (PFC) due to a paucity of dopamine transporters (Lewis et al., 2001; for overviews see Mier et al., 2010; Montag et al., 2012). More specifically, it was found that the enzyme activity of Val/Val (GG) carriers was around 38% higher than of Met/Met (AA) carriers in the DLPFC (Chen et al., 2004). Moreover, it was demonstrated that Val/Val (GG) carriers catabolize three to four times more dopamine than Met/Met (AA) carriers. Heterozygote Val/Met (GA) carriers show an intermediate catabolization rate (Lachman et al., 1996). On a psychological level, the COMT Val158Met polymorphism has also been associated with various phenotypes. For example the Met allele of the COMT Val158Met polymorphism is linked to higher reward responsiveness and reward seeking (Lancaster et al., 2012), and higher activity in reward related brain areas (ventral Striatum, lateral PFC) during anticipation of uncertain reward (Dreher et al., 2009). Moreover, it is important to highlight that the different alleles of the COMT Val158Met polymorphism have been associated with individual differences in cognitive abilities and (personality) traits related to stress processing / negative emotionality. In detail, the Met allele seems to be associated with higher abilities in executive cognitive abilities and lower activations in prefrontal brain regions compared to the Val allele while performing cognitive tasks (Mier et al., 2010). In addition, the Met allele of the COMT Val158Met polymorphism has been associated with emotional instability and lower stress resiliency according to the warrior-worrier-hypothesis (but therefore better cognitive functions; Goldman et al., 2005; Montag et al., 2012) and with anxiety and related traits in empirical studies (e.g., Olsson et al., 2005; Stein et al., 2005; Hashimoto et al., 2007; Montag et al., 2008; Lee and Prescott, 2014).

Overall, the results indicate that the COMT Val158Met polymorphism genotype modulates reward processing in association with punishment of unfair behavior and taking revenge. Accordingly, it can be hypothesized that this effect is transferable from revenge-like behavior to trait vengefulness and that Met allele carriers might experience revenge as more rewarding. We therefore hypothesized that carrying the Met allele would be associated with higher scores in trait vengefulness. Since avoiding negative outcomes is also associated with activation in reward related brain areas, we additionally hypothesized that carrying the Met allele would be associated with the tendency to avoid a transgressor (please note that Revenge Motivation and Avoidance Motivation show strong positive correlations and can therefore not be seen as opposites; Kim et al., 2006; Johnson et al., 2010; Szcze´sniak and Soares, 2011). With the present study we aimed at investigating these hypotheses.

### MATERIALS AND METHODS

### Sample

For the present genetic association studies, data collection was conducted in China and in Germany to independently replicate findings (see also for example Montag et al., 2017). Taking into account (i) power analyses but also (ii) sample sizes of other recent genetic association studies using a questionnaire approach to assess traits and (iii) the independent replication as implemented in the present research led to the conclusion that around 500–600 participants per sample (China and Germany) should be investigated. The complete procedure including the software to present the questionnaires as well as the equipment and protocols for genotyping were the same in China and Germany. Participants in China and Germany were recruited at universities. All Chinese participants are part of the Chengdu Gene Brain Behavior Project and all German participants of the Ulm Gene Brain Behavior Project. For a total of N = 730 Chinese participants [n = 534 males, n = 196 females; mean age: 21.62 (SD = 2.35)] complete data were available for the present study. Most of the Chinese participants were Han Chinese (n = 680). Additionally, the data of N = 585 German participants, of which all stated German as their native language, were included in the analyses of the German part of the study [n = 186 males, n = 399 females; mean age: 23.74 (SD = 8.33)]. All participants gave electronic and written informed consent in accordance with the Declaration of Helsinki. The studies/protocols were approved by the local ethics committees at University of Electronic Science and Technology of China, Chengdu, China and Ulm University, Ulm, Germany.

### Self-Report Measures

To measure individual differences in the tendencies to react to a transgression two different self-report measures were assessed via an online platform. Participants completed a Chinese/German version of the Transgression Related Interpersonal Motivations−12 (TRIM-12) inventory (original English version: McCullough et al., 1998; McCullough and van Oyen Witvliet, 2002). This measure consists of 12 items split into two scales named Revenge Motivation [5 items; Cronbach's α = 0.87/0.87 (Chinese sample/German sample)] and Avoidance Motivation [7 items; α = 0.83/0.88 (Chinese sample/German sample)]. Answers are given on a 5-point Likert scale. Participants also completed a Chinese/German version of the Vengeance Scale (original English version: Stuckless and Goranson, 1992). It consists of 20 items forming a single scale and the items are answered on a 7-point Likert scale [α = 0.90/0.93 (Chinese sample/German sample)]. For all scales, higher scores indicate higher tendencies toward vengefulness/avoidance of transgressors. The questionnaires used were the same as in Sindermann et al. (2018), where also full details about the Chinese and German translations of the items, the factorial structure of all scales in the Chinese and German languages as well as associations with prominent personality factors are given. It should be noted that some of the participants from the Sindermann et al. (2018) study are also included in the present study (Chinese sample: n = 234; German sample n = 182).

### Genotyping

DNA was extracted from buccal cells on a MagNA Pure 96 robot (Roche Diagnostics, Mannheim, Germany) using commercial extraction kits. Genotyping of the COMT Val158Met Single Nucleotide Polymorphism (SNP) was conducted on a Cobas Z 480 Light Cycler (Roche Diagnostics, Mannheim, Germany) by means of polymerase chain reaction and subsequent high resolution melting. Simple probe assay designs by TIBMolBiol (Berlin, Germany) were used. Hardy Weinberg Equilibriums (HWEs) were calculated using the Court lab–HW calculator.

### Statistical Analyses

The distributions of all scales under investigation were tested for normal distribution (separately in the Chinese and German sample). The statistical tests as well as histograms are presented in the **Supplementary Material**. As none of the scales was normally distributed according to statistical tests and because especially the histograms of the German sample showed a marked deviation from the normal distribution, it was decided to implement all statistical tests using non-parametric tests.

First, associations of age and gender with all scales were investigated to examine possible confounding variables. Spearman correlations between age and Revenge Motivation, Avoidance Motivation and the Vengeance Scale were calculated and Mann-Whitney U-Tests were used to examine gender differences.

To investigate the relationship between the COMT Val158Met polymorphism and the scales under investigation, groups of Met– (A–: GG) and Met+ (A+: GA/AA) carriers were built. This procedure is justified because in the Chinese sample the distribution of genotypes in the COMT Val158Met polymorphism was skewed with only a few individuals (n=50) carrying the homozygote Met/Met (AA) genotype (see **Table 1**). This is also in line with the expected distribution of the COMT Val158Met polymorphism in the Han Chinese population as indicated by data banks (for further information see also the results section). Therefore, in order to enhance statistical power it is reasonable to group Val/Met (GA) and Met/Met (AA) genotypes into one group called Met+. Despite a similar number of homozygote Val/Val and homozygote Met/Met carriers in the German sample, we implemented the same grouping (Met– vs. Met+) for the data to allow a direct comparison between the Chinese and German samples. Next, we calculated Mann-Whitney U-Tests in the Chinese and German samples to investigate differences between the Met– (A–: GG) and Met+ (A+: GA/AA) groups. Additionally, we calculated the Mann-Whitney U-Tests separately within the samples from each nation and split by gender (this is necessary due to (i) significant effects of gender on the investigated scales and (ii) with regard to the different gender ratios in the Chinese and German samples). Moreover, we also present all the results (using Kruskal-Wallis-Tests) on a genotype level with the grouping Val/Val (GG), Val/Met (GA), Met/Met (AA) in the **Supplementary Material**.

### RESULTS

### Genotype Distribution and Hardy Weinberg Equilibriums

As seen in **Table 1**, the distributions of the genotypes in the COMT Val158Met polymorphism were in the HWE in both samples from China and Germany. The distribution of genotypes also fits with the observations reported at dbSNP in the samples from both nations (https://www.ncbi.nlm.nih.gov/projects/SNP/ snp\_ref.cgi?rs=4680).

### Associations With Age and Gender

Only in the Chinese sample age correlated weakly significantly with Avoidance Motivation (ρ = 0.09, p = 0.011), but no other scale. In the German sample, age did not significantly correlate with any of the scales of interest (all p-values > 0.281). Therefore, it was decided to not include this variable as control variable in further analyses.

Gender differences were found in Avoidance Motivation (U = 43,195.50, Z = −3.63, p < 0.001) in the Chinese sample and in Revenge Motivation (U = 28,436.00, Z = −4.57, p < 0.001) and the Vengeance Scale (U = 28,692.00, Z = −4.42, p < 0.001) in the German sample. Descriptive statistics are presented in **Table 2**.

TABLE 1 | Genotype distributions of the COMT Val158Met polymorphism in the Chinese and German samples.


*HWE = Hardy Weinberg Equilibrium.*

TABLE 2 | Descriptive statistics of all scales under investigation split by nation and gender.


*Mean values and SDs are reported [M (SD)].*

## Effects of the COMT Val158Met Polymorphism

As presented in **Table 3**, significant effects of the COMT Val158Met polymorphism (tested on Met– vs. Met+ allele level) on Revenge Motivation and the Vengeance Scale were observed, but only in the Chinese sample. Met allele carriers (Met+; A+; GA/AA) showed higher scores than the Met– (Val/Val; A–; GG) group. The strongest significant effect of the COMT Val158Met polymorphism was found on Avoidance Motivation in the Chinese sample with Met allele carriers (Met+; A+; GA/AA) showing higher scores than the Met– group (Val/Val; A–; GG). Results with regard to the effect on Avoidance Motivation were significant and in the same direction in the German sample. In the Chinese sample, all effects of the COMT Val158Met polymorphism would survive Bonferroni correction for multiple testing (α = 0.05/3 = 0.0167; accounting for the number of scales under investigation). In the German sample, the association between the COMT Val158Met polymorphism and Avoidance Motivation would also survive Bonferroni correction for multiple testing (α = 0.05/3 = 0.0167; accounting for the number of scales under investigation).

When effects were considered separately for the samples from both nations and split by gender, **Table 4** shows that the direction of the effects was the same for males and females in the Chinese sample, although the associations with Revenge Motivation and the Vengeance Scale were only significant in males. For all the scales, Met allele carriers (Met+; A+; GA/AA) showed higher scores compared to the Met– group (Val/Val; A–; GG). There were no significant effects of the COMT Val158Met polymorphism on any of the scales under investigation in the German male sample (also when testing one-sided based on the hypotheses and the results in the Chinese sample). However, in the German female sample the effect on Avoidance Motivation was significant and in the same direction as in the Chinese male and female samples. Of note, only the associations between the COMT Val158Met polymorphism and Revenge Motivation and Avoidance Motivation found in the Chinese male sample and the association between the COMT Val158Met polymorphism and Avoidance Motivation in Chinese and German females would still be significant after Bonferroni correction for multiple testing (α = 0.05/3 = 0.0167; accounting for the number of scales under investigation). Additionally, when correcting for multiple testing while accounting for the number of scales under investigation as well as gender, only the association between the COMT Val158Met polymorphism and Revenge Motivation found in the Chinese male sample and between the COMT Val158Met polymorphism and Avoidance Motivation found in the German females would still be significant (α = 0.05/6 = 0.0083). The association between the COMT Val158Met polymorphism and Avoidance Motivation in the Chinese female sample just failed to remain significant (with a p-value of 0.009). The strongest effects, namely the ones on Avoidance Motivation are also presented in **Figure 1**.

As a final note, based on the hypotheses derived in the introduction part of the manuscript, also one-sided testing of

#### TABLE 3 | Effects of the COMT Val158Met polymorphism (on allele level) on all scales under investigation split by nation.


*Mean values and SDs are reported [M (SD)] in the first columns. All p-values presented in this table are derived from two-sided testing.*

TABLE 4 | Effects of the COMT Val158Met polymorphism (on allele level) on all scales under investigation split by nation and gender.


*Mean values and SDs are reported [M (SD)] in the first columns. N(Chinese, males, Met*−*)* = *292, n(Chinese, males, Met*+*)* = *242, n(Chinese, females, Met*−*)* = *93, n(Chinese, females, Met*+*)* = *103, n(German, males, Met*−*)* = *49, n(German, males, Met*+*)* = *137, n(German, females, Met*−*)* = *78, n(German, females, Met*+*)* = *321. All p-values presented in this table are derived from two-sided testing.*

the COMT Val158Met effects on all scales would be justified and would lead to halved p-values. Thereon, also the association between the COMT Val158Met polymorphism and Avoidance Motivation in Chinese males and females would survive the strict Bonferroni correction procedure for multiple testing while accounting for the number of scales under investigation and gender (α = 0.05/6 = 0.0083).

### DISCUSSION

Investigating the genetic underpinnings of individual differences in the tendencies to react to a transgression in independent Chinese and German samples, we found that the COMT Val158Met polymorphism is associated with Avoidance Motivation in several (sub)samples. In the Chinese and the German sample Met allele carriers (Met+; A+; GA/AA) reported higher scores in the tendency to avoid a transgressor as compared to homozygote Val/Val (Met–; A–; GG) carriers. In the Chinese sample, this effect was found in both males and females, but stronger in females. In the German sample, the direction of the association was also observed in males and females, although it was stronger (and only significant) in females again; whereas in males only a small descriptive difference in the same direction was found when comparing Met+ and Met– carriers. Of note, when correcting the alpha level for the number of scales under investigation and gender (α = 0.05/6 = 0.0083), the associations between the COMT Val158Met polymorphism and Avoidance Motivation found in Chinese males and females would only survive if testing onesided based on the hypothesis formulated in the introduction part of the manuscript.

In the Chinese male sample, we also found an association between the COMT Val158Met polymorphism and vengefulness

(Revenge Motivation as well as the Vengeance Scale). Again, Met allele carriers (Met+; A+; GA/AA) showed higher vengefulness scores. In the German sample, on the other hand, we could not find such an association (even if testing one-sided based on the hypothesis). Thus, as we hypothesized, both the tendency to avoid transgressors and to react vengefully toward them are associated with the COMT Val158Met polymorphism, although it should be emphasized that the latter finding was only significant in the Chinese (male) sample.

In sum, the association between the Met allele of the COMT Val158Met polymorphism and the scale Avoidance Motivation was the most stable and robust finding, especially across the two female samples with completely different cultural backgrounds. This indicates a general cross-cultural effect, and maybe gender-specificity. We also want to mention that an independent replication of associations, such as the one reported here, especially in ethnically and cultural different samples, is difficult to observe. Thus, in the present study the independent replication of the effect, mainly in the female subsamples, increases confidence in the overall robustness of our finding. Hence, we argue for the importance of these independently derived effects in the same direction, but also mention that results in the German male sample were particularly weak and not significant. Effects of single genetic markers are often weak [around 1% of explained variance in a certain phenotype (often lower)] and therefore the non-significant findings in the German male sample could be due to a lack of power (with n = 186, this sample was the smallest subsample in this study; in detail, in the German male sample n = 49 participants were in the Met– group and n = 137 in the Met+ group, whereas for example in the Chinese female sample (n = 196), we tested n = 93 (Met–) vs. n = 103 (Met+) participants). However, in line with our findings the effect of the COMT Val158Met polymorphism on Avoidance Motivation might be particularly strong in females. Support for this interpretation comes from previous studies demonstrating gender specific effects of the COMT Val158Met polymorphism on various traits, cognitive abilities and cortical thickness (e.g., Stein et al., 2005; Lang et al., 2007; Chen et al., 2011; Sannino et al., 2014). To further illuminate this: the COMT enzyme does not only metabolize catecholamines but also methylates catecholestrogens (Harrison and Tunbridge, 2008). Additionally, it has been shown that estradiol can inhibit COMT gene transcription and mRNA expression and thereon also COMT enzyme activity in certain cells (Xie et al., 1999; Jiang et al., 2003). In line with this, studies showed that COMT enzyme activity in the liver and erythrocytes was lower in healthy females compared to healthy males (Fähndrich et al., 1980; Philippu et al., 1981; Boudíková et al., 1990). This indicates potentially higher baseline dopamine levels in females compared to males as COMT catabolizes dopamine [But it needs to be mentioned that this will most likely only apply in COMT active brain regions. Additionally, also this argumentation is limited to the COMT related mechanisms to catabolize dopamine. However, there are several more mechanisms to clear the synaptic cleft from dopamine except the enzymatic mechanism by COMT (e.g. MAO-A,...)]. Hence, in females, which are carrying the Met allele, the reduced COMT enzyme activity by carrying the Met allele together with the (generally) reduced COMT enzyme activity by higher estrogen levels (compared to males) might lead to especially pronounced effects on psychological phenotypes in females compared to males (Lachman et al., 1996; Chen et al., 2004). In detail, the potentially higher dopamine levels in female Met allele carriers especially in COMT active regions of the brain might lead to higher Avoidance Motivation compared to female Val/Val carriers. Support for the influence of estrogens (and perhaps other sex hormones) on the (gender specific) effects of the COMT Val158Met polymorphism also comes from a study, which showed that genetic variation in the COMT gene associated with extreme COMT enzyme reduction [22q11DS patients (only one copy of the COMT gene) carrying the Met allele in the COMTVal158Met polymorphism], was associated with cortical thinning only in females and only after puberty (similar effect also found in genetically modified mice; Sannino et al., 2017). Moreover, another study showed that the Met allele of the COMT Val158Met polymorphism was associated with better performance in a working memory task in males and postmenopausal (but not pre-menopausal) women (Papaleo et al., 2015). These studies strengthen the idea that the hormonal status (with regard to sex hormones) is important for the gender specific effects of the COMT Val158Met polymorphism. However, it needs to be noted that the importance of estrogens in inhibiting COMT enzyme activity with a focus on mechanisms in the brain has been challenged by studies showing that estradiol does not affect (i) COMT activity in the rat brain and (ii) in a glioblastoma cell line (Cohn and Axelrod, 1971; Jiang et al., 2003; see also Harrison and Tunbridge, 2008 for an overview). In conclusion, the exact biochemical underpinnings of the here found results of the COMT Val158Met polymorphism cannot be tested with the present dataset as no further biological marker of interest except the genotype in the COMT Val158Met polymorphism was assessed. In so far this discussion part of our work is speculative.

To assume possible reasons and mechanisms on a psychological level, which underlie the association between the COMT Val158Met polymorphism and Avoidance Motivation, it must be noted that there was a substantial correlation between Avoidance and Revenge Motivation in the present study (China: ρ = 0.36, p < 0.001; Germany: ρ = 0.41, p < 0.001). Hence, one could conclude that the tendency to avoid a transgressor could reflect the tendency toward seeking to punish transgressors by ending a relationship (e.g., friendship) and thus incorporates a component of vengefulness (e.g., a typical item of Avoidance Motivation is "I cut off the relationship with him/her"; McCullough and van Oyen Witvliet, 2002). In this case, the association between the Met allele of the COMT Val158Met polymorphism and higher scores in Avoidance Motivation might be explained by higher experience of reward during punishment of the transgressor (see experimental studies e.g., Sanfey et al., 2003; De Quervain et al., 2004; Strobel et al., 2011; and also the following studies: Gaspar et al., 1989; Delgado et al., 2003; Arias-Carrión et al., 2010; Liu et al., 2011; Sescousse et al., 2013). On the other hand, at least in the German sample, the association of the COMT Val158Met polymorphism with Avoidance Motivation differed from that with vengefulness. This could indicate different underlying mechanisms for the associations of the COMT Val158Met polymorphism with vengefulness compared with its association with Avoidance Motivation. As such, avoiding an aversive stimulus/outcome can also be understood as rewarding by itself (Kim et al., 2006) and higher reward responsiveness has previously been associated with the Met allele of the COMT Val158Met polymorphism (Dreher et al., 2009; Lancaster et al., 2012). Hence, also in this way one could explain the association found in the present study. Next, in accordance with the warrior-worrier hypotheses mentioned above (Goldman et al., 2005) and as the Met allele has also been associated with anxiety and related traits previously (e.g., Olsson et al., 2005; Stein et al., 2005; Hashimoto et al., 2007; Lee and Prescott, 2014), we shortly wanted to test post-hoc whether neuroticism (as an indicator of emotional instability) would mediate the effects of the Met allele of the COMT Val158Met polymorphism on Avoidance Motivation. More detailed information and results are reported in the **Supplementary Material**. In short, at least neuroticism as measured with the NEO-FFI (Costa and McCrae, 1992) does not seem to explain the present findings. But in this regard it is still important to note that for several anxiety related traits, gender specific effects of the COMT Val158Met polymorphism with especially pronounced results in females (Met allele associated with higher anxiety related traits) have been observed (e.g., Olsson et al., 2005; Stein et al., 2005). Hence, the hypothesis that the association between the Met allele of the COMT Val158Met polymorphism is associated with Avoidance Motivation via an anxiety related trait seems likely in the light of the present gender specific effects already discussed above. In conclusion, there are different ways to explain the present results. Thereon, future large-scale studies investigating the relationship between the COMT Val158Met polymorphism and Avoidance Motivation are needed. In this regard it would be of great interest to investigate the exact motivation/motives for why participants tend to avoid transgressors and how this might relate to reward responsiveness and/or specific anxiety-related traits.

In terms of the relationship between the Met allele of the COMT Val158Met polymorphism and vengefulness, which could only be observed in the Chinese (male) sample, it would be of great interest to investigate possible confounding variables contributing to this apparent cultural (and gender) specificity. Next to catecholestrogens, which might explain gender differences, for example, also cross-cultural differences in societal norms and manners might contribute to observed effects of the COMT Val158Met polymorphism as well as differences between the Chinese and German samples. As such, the works by Hofsede are of potential importance, revealing differences in power distance and individualism/collectivism between both countries, with China scoring higher on collectivism and power distance (https://www.hofstede-insights.com/ country/china/; https://www.hofstede-insights.com/countrycomparison/germany/). In this regard, it may be important to consider both normative and evaluative assessments of tendencies toward these constructs (as outlined by Montag et al., 2016; Sindermann et al., 2018). Next, other objective measures might be additionally used in future studies to expand the knowledge gained by the present results. As such, it could be considered to assess tendencies to react to a transgression on the behavioral and neural level to overcome some shortcomings associated with selfreport measures. Lastly, another very interesting research question would be if / to what extent other dispositions to react to a transgression, for example forgivingness, are influenced by genetic markers such as the COMT Val158Met polymorphism.

In conclusion, the present results show for the first time that there seems to be an association between the Met allele of the COMT Val158Met polymorphism and the personality trait of the tendency to avoid a transgressor, especially in females. Future research will need to explore the underlying mechanisms (e.g., motives for revenge, anxiety and cultural influences) explaining this association, particularly regarding the importance of specific anxiety-related factors.

### AUTHOR CONTRIBUTIONS

CS and CM planned the design of the present study. CS and RL collected the data from the Chinese part of the study. CS, RL, and YZ conducted the genetic analyses of the Chinese samples. CS supported data collection for the German part

### REFERENCES


of the study and analyzed around half of the genetic samples from the German participants. CS wrote the manuscript and carried out the statistical analyses. YZ checked the statistical analyses. CM worked over the first draft of the manuscript. BB and KMK provided helpful comments and worked over the complete manuscript. All authors approved the final version of the manuscript.

### FUNDING

CS is stipend of the German Academic Scholarship Foundation (Studienstiftung des deutschen Volkes). KMK is supported by the National Natural Science Foundation of China (NSFC) (grant 31530032). CM is funded by a Heisenberg grant (DFG, MO2363/3-2) from the German Research Foundation (Deutsche Forschungsgemeinschaft).

### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fnbeh. 2018.00148/full#supplementary-material


genotype with sensation seeking personality trait. Neuropsychopharmacology 32, 1950–1955. doi: 10.1038/sj.npp.1301335


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Sindermann, Luo, Zhang, Kendrick, Becker and Montag. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# The Urge to Fight: Persistent Escalation by Alcohol and Role of NMDA Receptors in Mice

Herbert E. Covington III<sup>1</sup> \*, Emily L. Newman<sup>1</sup> , Steven Tran<sup>1</sup> , Lena Walton<sup>1</sup> , Walae Hayek<sup>1</sup> , Michael Z. Leonard<sup>1</sup> , Joseph F. DeBold<sup>1</sup> and Klaus A. Miczek1,2,3,4 \*

<sup>1</sup> Department of Psychology, Tufts University, Medford, MA, United States, <sup>2</sup> Neuroscience, Sackler School of Biomedical Sciences, Tufts University, Boston, MA, United States, <sup>3</sup> Pharmacology, Sackler School of Biomedical Sciences, Tufts University, Boston, MA, United States, <sup>4</sup> Psychiatry, Sackler School of Biomedical Sciences, Tufts University, Boston, MA, United States

Alcohol drinking, in some individuals, culminates in pathologically aggressive and violent behaviors. Alcohol can escalate the urge to fight, despite causing disruptions in fighting performance. When orally administered under several dosing conditions the current study examined in a mouse model if repeated alcohol escalates the motivation to fight, the execution of fighting performance, or both. Specifically, seven daily administrations of alcohol (0, 1.8, or 2.2 g/kg) determined if changes in the motivation to initiate aggressive acts occur with, or without, shifts in the severity of fighting behavior. Responding under the control of a fixed interval (FI) schedule for aggression reinforcements across the initial daily sessions indicated the development of tolerance to alcohol's sedative effect. By day 7, alcohol augmented FI response rates for aggression rewards. While alcohol escalated the motivation to fight, fighting performance remained suppressed across the entire 7 days. Augmented FI responding for aggression rewards in response to a low dose of alcohol (1.0 g/kg) proved to be persistent, as we observed sensitized rates of responding for more than a month after alcohol pretreatment. In addition, this sensitization of motivated aggression did not occur with a general enhancement of motor activity. Antagonism of NMDA or AMPA receptors with ketamine, dizocilpine, or NBQX during later challenges with alcohol were largely serenic without having any notable impact on the expression of alcohol-escalated rates of FI responding. The current dissociation of appetitive and performance measures indicates that discrete neural mechanisms controlling aggressive arousal can be distinctly sensitized by alcohol.

Keywords: alcohol, aggressive behavior, motivation, glutamate receptors, NMDA/AMPA, tolerance, sensitization, neuroplasticity

## INTRODUCTION

Alcohol-escalated violence inflicts serious harm and suffering on a global scale as documented over many decades (Pernanen, 1993; Bye and Rossow, 2010). More than half of violent criminal acts are associated with alcohol in the perpetrator or victim or both (Beck et al., 2014). In such cases, alcohol consumption prompts a motivational state that culminates in attempts to act violently, which is distinct from impaired and uncoordinated behavior during intoxication (Mayfield, 1976; Jaffe et al., 1988). Cognitive models attempt to explain alcohol-instigated aggression through processes such as fear reduction, cortical disinhibition, anticipation of expected outcomes, or selectively attending to

#### Edited by:

Etsuro Ito, Waseda University, Japan

#### Reviewed by:

John J. Woodward, Medical University of South Carolina, United States Carla Cannizzaro, Università degli Studi di Palermo, Italy Christian P. Müller, Friedrich-Alexander-Universität Erlangen-Nürnberg, Germany

\*Correspondence:

Herbert E. Covington III herbert.covington@tufts.edu Klaus A. Miczek Klaus.miczek@tufts.edu

Received: 17 May 2018 Accepted: 20 August 2018 Published: 13 September 2018

#### Citation:

Covington HE III, Newman EL, Tran S, Walton L, Hayek W, Leonard MZ, DeBold JF and Miczek KA (2018) The Urge to Fight: Persistent Escalation by Alcohol and Role of NMDA Receptors in Mice. Front. Behav. Neurosci. 12:206. doi: 10.3389/fnbeh.2018.00206

provocative cues (i.e., alcohol-induced myopia) (Schmutte et al., 1979; Pernanen, 1993; Pihl et al., 1993; Sayette et al., 1993; Zhang et al., 2002; Giancola et al., 2011). Here, we focus on the motivation to engage in aggressive behavior when it is ostensibly rewarding and outcomes are predictable in a mouse model (Ginsberg and Allee, 1942; Fish et al., 2005). The currently selected experimental conditions aim to systematically dissect how alcohol over repeated exposures alters the appetitive and performance (i.e., consummatory) components of aggressive behavior (Miczek et al., 2015; Golden et al., 2017; Hashikawa et al., 2017).

The neural circuitry of appetitive and consummatory behaviors overlap considerably (Wise, 2013). Quantification of appetitive behaviors, particularly when maintained by fixed interval (FI) schedules, indicates the state of "arousal" immediately prior to reward receipt (Wenger and Dews, 1976). This method of schedule-controlled behavior allows for dissociating motivational processes that precede performance measures (Gonzalez and Goldberg, 1977). Several types of aggression can be highly arousing and represent evolutionarily conserved, natural rewards (Scott, 1958). Reactive "hot" acts of violence are often produced by repeated cycles of alcohol intoxication (see Beck et al., 2014 for a review of clinical data). The neural architecture supporting such maladaptive aggression remains unknown, but key epidemiological findings provide some insight into their origins, including the predictably high rate of reoccurrence and their progressive escalation in alcohol use disorders (Fergusson and Horwood, 2000; Coid et al., 2006). Preclinical data corroborate these trends, such that the proportion of alcohol-heightened aggressors in a sample increases with a history of intermittent voluntary drinking (Fish et al., 2002; Hwa et al., 2015). We hypothesize that repeated exposures to alcohol - in certain contexts and when winning a confrontation is expected – can potentially trigger an intense motivation to engage in future aggressive acts.

The dose-dependent biphasic modulation of aggressive performance (Miczek et al., 1992, 1998) is clearly characterized by lower alcohol doses, which reliably increase threats and attacks; yet, the motivational indices prior to fighting require more evaluation (Fish et al., 2008). In addition to its acute effects on behavior, repeated EtOH administrations increase the propensity for the later expression of an alcohol-heightened aggressive phenotype (Lessov et al., 2001; Fish et al., 2002; Didone et al., 2016). In line with a dopamine-dependent theory of behavioral plasticity, increases in synaptic strength, particularly excitatory synapses on dopamine (DA) neurons in the ventral tegmental area (VTA), occurs after an in vivo administration of alcohol, like DA-positive modulators (Saal et al., 2003). Interestingly, levels of operant responding that are reinforced by aggression require intact DA receptor activation in the ventral striatum (Couppis and Kennedy, 2008). Moreover, persistently augmented behavioral responses to alcohol rely on the activation of N-methyl-d-aspartate (NMDA) and alpha-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid (AMPA) receptors in the VTA (Phillips and Shen, 1996; Broadbent et al., 2003).

The current objective was to determine how alcohol, when orally administered under several dosing conditions increases both (1) the motivation to fight, and (2) the execution of fighting. We sought to confirm that lower acute doses of alcohol will increase the intensity of offensive aggression (e.g., escalated number of bites and threats), without affecting the motivation to engage in fighting. The impairing effects of higher alcohol doses were expected to reduce both anticipatory responding and performance measures (Fish et al., 2008). We hypothesized that during repeated exposures to alcohol, tolerance first develops to the sedative effects, and eventually sensitized responding emerges, which may serve as an index of motivation. We determined whether or not changes in the motivation to initiate aggressive acts occur with, or without, shifts in the intensity of fighting behavior (Newman et al., 2018). Responding under the control of the FI schedule of reinforcement is a highly sensitive measure of appetitive behavior, indicating that the pattern of responding (i.e., the "scallop") during successive administrations of alcohol may therefore be attributed to underlying changes in incentivemotivation. Thus, our final hypothesis explored to what extent the activation of ionotropic glutamate receptors (iGluRs) is necessary for maintaining any lasting changes in aggressive reinforcement resulting from repeated oral administrations of alcohol.

## MATERIALS AND METHODS

### Subjects

Eight-week-old male C57BL/6J mice (C57; Jackson Labs, Bar Harbor, ME, United States) were housed in polycarbonate cages (28 cm × 17 cm × 14 cm) lined with pine chip bedding. Food and water were available at all times. For Experiments 1–4, "resident" male mice (n = 80) were housed with a female of the same age and strain for at least 1 month to facilitate aggressive behaviors and avoid social isolation (Crawley et al., 1975; Miczek and O'Donnell, 1978). All pups were culled at 3 weeks of age. Female partners were removed following the display of consistent aggression by the resident male, at which point these aggressive resident males were singly housed for the remainder of the experiment. For Experiment 5, 8-week-old male C57 mice (n = 22) were singly housed in polycarbonate cages for the duration of the experiment.

Additional male C57 mice (8 weeks) were group-housed as "intruders" in large polycarbonate cages (46 cm × 24 cm × 15 cm), with unrestricted food and water available. These intruder mice were used for daily tests of aggression by residents for approximately 1 week, and then replaced with a new cohort of intruders. The vivarium was maintained at 21 ± 1 ◦C, 30– 40% humidity, and 12-h reverse light/dark cycles (lights on at 17:30 h). Experimental procedures were approved by the Tufts Institutional Animal Care and Use Committee following the Guide for the Care and Use of Laboratory Animals (National Research Council 2011).

### Procedures

### Measurements of Aggression and Aggressive Motivation

After at least 4 weeks of cohabitation and the birth of one litter of pups to confirm successful mating behavior, each resident male

was quantitatively screened for consistent display of aggressive behavior (See **Figure 1**). During this phase of screening for aggression, the female and pups were removed from the resident's home cage, before an intruder was introduced (Miczek and O'Donnell, 1978). This social confrontation lasted for 5 min following an initial attack bite by the resident, and confrontations were permitted until a total of 30 bites accumulated, or 5 min elapsed if no fight occurred. An experimenter tallied the frequency of attack bites, and recorded the duration of the confrontation. Daily confrontations were conducted at 24-h intervals until each resident displayed a stable level of aggression toward an intruder (30 bites within 1 min, over seven sessions). By the final screening session >90% of residents were highly aggressive, displaying vigorous attacks (i.e., 30 bite limits reached within 1 min) with a short latency (<10 s) to the initial attack. Intruders were systematically rotated to ensure that the resident did not habituate to a specific intruder (Winslow and Miczek, 1983).

Once stable and reliable aggression was established, residents (n = 74, six mice never established stable levels of aggression during screening trials) were then conditioned to perform a nose poke task according to a FI schedule reinforced by the opportunity to fight (Fish et al., 2002). A panel with two nose-poke operanda was inserted into the resident's home cage and affixed to the walls. The first nose poke in the assigned "active" hole after the interval

FIGURE 1 | Resident male C57BL/6J (C57) mice were housed with breeding C57 females for at least 1 month. In daily resident–intruder confrontations, each resident male encountered a novel, male C57 intruder for 5 min in the resident home cage. After establishing an aggressive phenotype, each resident was trained during a fixed interval (FI) schedule that was reinforced by the presentation of an intruder. The FI was progressively increased from 1 s to 10 min over the course of 1 month. Mice were divided into experimental groups upon establishing consistent patterns of FI responding. Experiment 1 revealed the effects of acutely administered water or EtOH (0.5, 1.0, or 1.8 g/kg, PO) on FI responding and subsequent aggressive behavior. Experiment 2 evaluated the effects of repeated daily administrations of EtOH (1.8 or 2.2 g/kg, PO) on FI and aggression trials. Experiment 3 confirmed and extended these findings by measuring the persistence of alcohol-escalated motivation to fight. Experiment 4 examined the role of iGluRs during the expression of sensitized FI responding and aggressive performance in response to a 1.0 g/kg EtOH challenge that occurred at least 10 days after repeated administrations of water or EtOH (2.2 g/kg, PO).

had elapsed was reinforced by the presentation of an intruder which was promptly attacked. Specifically, after completing the behavioral requirement of the FI schedule, a house light was illuminated and an intruder was simultaneously introduced into the resident's home cage. After 1 min of aggressive interactions, the intruder was removed, the house light turned off, and the session terminated. The response panel was removed from the resident's home cage immediately after each daily session. During the first 2 weeks of conditioning, the female partner of the male resident was returned to the home cage at the completion of each FI session. After the first 2 weeks of FI training, the female was permanently removed and the male resident was singly housed under the same housing conditions. FI sessions were conducted daily for all experimental mice. On the first day of FI conditioning, the FI interval was 30 s. Over the next 30 daily FI sessions the interval was gradually increased to 10 min.

Once the FI reached 10 min, five to seven sessions were conducted per week until the mice demonstrated stable rates of responding in the "active" nose poke hole. Over successive daily sessions, the pattern of responses reliably increased in frequency toward the end of the FI (i.e., demonstrating an FItypical "scalloped" pattern of responding). The scalloped pattern of nose pokes allows for assessments of the rate of responding and the index of curvature (Fry et al., 1960). This index of curvature ranges from a value of −0.75, indicating that all responses are made during the first quarter of the interval, to a value of +0.75, indicating that all responses are made during the last quarter of the interval. If responses are evenly distributed across the interval, the index of curvature is 0. Consistent with previous uses of this procedure the curvature values of all experimental mice for the current series of experiments approximated +0.30 (Fish et al., 2002).

### Experiment 1: Effects of Acutely Administered Alcohol on Motivation to Fight and Fighting Performance

After stable rates of FI responding were observed (i.e., <20% variation in FI responding over 3 days), male residents (n = 10) were habituated to oral administrations of tap water via gavage (per os, PO) 10 min before each daily FI session for 1 week. After habituation to these handling procedures, residents were given either tap water or various doses of EtOH (0.5, 1.0, and 1.8 g/kg, PO) 10 min before FI sessions in an unsystematic sequence at 72 h intervals.

### Experiment 2: Effect of Repeated Alcohol Administrations on the Motivation to Fight and Fighting Performance

After stable rates of FI responding were observed, male residents (n = 30) were habituated to oral administrations of tap water 10 min before each daily FI session for 1 week. These residents were given seven consecutive days of either water (n = 10) or one of two different doses of EtOH (1.8 or 2.2 g/kg/day, PO, n = 10/dose) 10 min before their daily FI session. Ten days after their last daily dose of water or EtOH, mice were orally administered 1.0 g/kg EtOH to assess the potentially sensitizing

effects of repeated EtOH administration on FI responding and fighting.

### Experiment 3: The Long-Term Consequences of Repeated EtOH on the Motivation to Fight and Fighting Performance

Once stable rates of FI responding were observed, male residents (n = 14) were habituated to oral administrations of tap water via gavage prior to each daily FI session. Each of these residents was then given EtOH (1.8 g/kg/day, PO) 10 min prior to each daily FI session for seven consecutive days, adhering to a within-subjects design. To carefully observe the long-term effects of these repeated, intermittent EtOH administrations on FI responding and fighting performance, these mice were subsequently challenged with EtOH (1.0 g/kg, PO) 14, 40, and 60 days after the last 1.8 g/kg dose. Five days prior to each EtOH challenge, residents were re-evaluated for baseline FI responding after PO water treatments before each daily session. Three of these trained resident mice lost weight and failed to respond during the day 40 EtOH challenge and were excluded from analyses of these later time points.

Upon completion of Experiment 3, blood was collected 10 min after EtOH 1.0 g/kg administration on the last EtOH challenge day from the submandibular vein. Blood samples were centrifuged at 4◦C for 10 min at 3,000 rpm, and plasma (5 µL) was extracted for blood ethanol concentration (mg/dL) analysis (AM1 Alcohol Analyzer, Analox Instruments Inc., Lunenburg, MA, United States).

### Experiment 4: Role of iGluRs During the Expression of Alcohol-Escalated Motivation to Fight

After establishing stable rates of FI responding, male residents (n = 20) were habituated to an oral administration of tap water via gavage 10 min prior to each daily FI session, as described above. These mice were then given either water (n = 10) or EtOH (2.2 g/kg/day, PO, n = 10) 10 min prior to their daily FI session for the next 7 days. Ten days later, the role of iGluRs on FI responding and fighting performance was assessed 10 min after an EtOH administration (1.0 g/kg, PO). Specifically, every 72 h these male residents were administered an IP injection of the AMPA receptor antagonist NBQX (0, 10, 17, and 30 mg/kg), the NMDA receptor antagonist ketamine (0, 5.6, 7.5, and 10 mg/kg) or the NMDA receptor antagonist dizocilpine (0.01, 0.1, and 0.3 mg/kg). Each iGluR antagonist dose was administered 15 min before EtOH (1.0 g/kg, PO). Ten minutes after receiving EtOH, FI responding for the opportunity to fight and fighting performance were measured.

### Experiment 5: Effect of Repeated Alcohol Administrations on Locomotor Activity in an Open Field

A separate cohort of mice were habituated to oral administrations of tap water via gavage for 1 week. These mice were given either water (n = 8) or EtOH (1.8 or 2.2 g/kg/day, PO, n = 7/EtOH treatment) immediately prior to locomotor assessments for the next 7 days. The locomotor behavior of each mouse was observed in a 51 cm× 36 cm × 31 cm plastic enclosure (Rubbermaid) that served as an open field. The total distance traveled (cm) was measured using video tracking software (EthoVision, Noldus, Wageningen, Netherlands). Mouse images were captured under red illumination at a rate of three samples/second through a 0.5-lux camera (Cohu, Model 4815–2100/AL09), which was positioned 165 cm above each open field. Three days after their seventh oral EtOH or water administration, locomotor activity was again assessed for 1 h for three consecutive days (experimental days 10–13) in response to a gavage administration of either water, 1.0 or 2.0 g/kg EtOH, in a semi-randomized order.

### Video Analysis

Agonistic behavior was recorded using a digital webcam (Logitech <sup>R</sup> HD Pro Webcam C920, Newark, CA, United States). A trained observer (intra-observer reliability: r > 0.95) analyzed video recordings during the fixed-interval and the aggressive encounter of the male residents using Observer XT software (Noldus). The first 60 s of the FI, the last 60 s of the FI, and the 60 s aggressive confrontation were analyzed. Key presses on a custommade keyboard coded the frequency, duration, and latency of each operationally defined behavior (**Table 1**). Aggressive behaviors quantified during social confrontations included attack bites and sideways threat. Non-aggressive behaviors included anogenital and nasal contact, pursuit, self-grooming, rearing, and walking (Miczek and O'Donnell, 1978). Arousal behaviors included tail rattle, digging, and jumping (Krsiak and Steinberg, 1969).

### Drugs

NBQX, ketamine and dizocilpine were obtained from Tocris Bioscience (Minneapolis, MN, United States). All compounds were dissolved in 0.9% NaCl. Each drug dose was injected intraperitoneally (IP) in a volume of 1 ml/100 g of body weight. For EtOH procedures, 95% ethyl alcohol was purchased from Pharmco-AAPER Products, Inc (Brookfield, CT, United States) and diluted with tap water to obtain 5%, 10%, 18%, or 22% EtOH concentrations (w/v). It was administered via gavage (PO) in a volume of 1 ml/100 g of body weight.

### Statistics

To observe the acute effects of EtOH on motivated responding for aggression reward, time-stamps of each nose-poke during a 10 min FI were carefully examined. All mice were administered either water or EtOH (0.5, 1.0, or 1.8 g/kg, PO) 10 min prior to the start of the FI schedule. The average rate of FI responding over the FI and the number of attack bites following the FI schedule was compared for each gavage treatment using a one-way repeated measures ANOVA. Post hoc comparisons for each dose of EtOH to water were made using Dunnett's test.

Three separate groups of mice trained to respond under the demands of the FI10 schedule were next examined to determine the effects of EtOH dose (0, 1.8, and 2.2 g/kg) on the induction and expression of alcohol-escalated motivation to seek aggression. For this experiment, two-way repeated measures ANOVA were used to assess the impact of EtOH doses (0, 1.8,


and 2.2 g/kg) administered over 7 days on the average rate of responding during the FI, and on attack behavior during reward receipt. After a 10-day EtOH free interval, the effect of water or EtOH (1.0 g/kg) on FI responding and attack behavior were again assessed using a one-way ANOVA. Dunnett's tests were used to make post hoc comparisons between water and EtOH treatments for both ANOVA.

To examine how long the effect of alcohol-escalated responding persists, the lower dose of alcohol (1.8 g/kg) was again examined (using a within-subjects design) on the rate of nosepoke responding during a 10 min FI schedule for an aggressive reward across seven daily administrations. Daily rates of nosepoking during each FI session were compared using a one-way repeated measures ANOVA. In addition, the average Index of Curvature for each daily FI session was also compared over each of the seven daily sessions using a one-way repeated measures ANOVA. Both of these indices of motivated responding during an FI schedule for aggression reinforcement were again assessed at much later time-points (i.e., after increasingly extended EtOHfree intervals) after being challenged with either water or 1.0 g/kg EtOH. Comparisons between water and EtOH on days 13 and 14, 39 and 40, and 59 and 60 (respectively) were made using paired t-tests for the average of both response rate and the Index of Curvature. In the case of significance, pairwise comparisons of behavioral elements were made using the Holm–Sidak method. In addition, the frequency and duration of behavioral elements collected and scored by a trained observer over 1 min bins at the beginning and end of the FI, and at the start of the aggressive encounter, were compared for water and EtOH on the 1st, 3rd, and 5th daily oral administrations using a oneway repeated measures ANOVA. These same behavioral elements scored during the FI and aggressive encounters during the three later challenge tests (i.e., day 14, 40, and 60) were compared by paired t-tests between temporally complimentary water and EtOH (1.0 g/kg) days.

To examine the neuropharmacology of the persistent expression of alcohol-escalated motivation for aggression, iGluR antagonists were administered prior to EtOH (1.0 g/kg) challenges. For this iGluR antagonism study, one-way ANOVA were initially performed on cumulative FI responding and attack bite frequency data, comparing repeatedly water-treated versus repeatedly EtOH-treated groups (i.e., control or repeated EtOH groups) after acute water gavage. There were no significant differences in measures of motivation or aggressive behavior, so a water baseline was calculated from averaging data across control and repeated EtOH groups. Two-way repeated measures ANOVA were conducted on baseline data after water or 1.0 g/kg EtOH and IP vehicle treatment to detect interactions between acute fluid treatment and history of repeated water or 2.2 g/kg EtOH. Additional two-way RM ANOVA were performed to detect interactions between repeated EtOH treatment and doses of MK-801, ketamine or NBQX administered prior to acute 1.0 g/kg EtOH. All pairwise comparisons were made using the Holm–Sidak method.

Finally, three separate groups of mice were examined to determine the effects of EtOH dose (0, 1.8, and 2.2 g/kg) on the induction and expression of locomotor sensitization to EtOH

fnbeh-12-00206 September 11, 2018 Time: 18:51 # 5

under the same oral administration conditions used above for studies on schedule-controlled aggression. For this experiment, two-way repeated measures ANOVA were used to assess the impact of EtOH doses (0, 1.8, and 2.2 g/kg, PO) administered on days 1, 3, 5, and 7 on the average distance traveled (cm). After a 3-day EtOH free interval, the locomotor response of these mice to EtOH (0, 1.0, or 2.0 g/kg) was again assessed over three consecutive days (one PO dose condition/day). Specifically, locomotor activity (cm traveled) on each challenge day was totaled across 5 min bins and analyzed using a two-way ANOVA (EtOH treatment × minute). Dunnett's tests were used to make post hoc comparisons between EtOH treatment groups (0, 1.8, and 2.2 g/kg) across 5 min time intervals for the first 30 min of each challenge day.

### RESULTS

### EtOH Dose-Dependently Reduced FI Responding for Aggressive Reinforcement With Bi-phasic Effects on Aggressive Behavior

Acute, oral administration of alcohol dose dependently reduced FI responding for aggression reward [F(3, 39) = 14.21, p < 0.001; **Figure 2**, left]. The number of attack bites emitted by each resident was increased by EtOH (0.5 g/kg), and decreased by the highest dose (1.8 g/kg) of EtOH [F(3, 39) = 21.40, p < 0.001; **Figure 2**, right]. Additional behavioral elements, including a longer attack latency [F(3, 13) = 4.23, p = 0.011], a shorter duration of physical contact [F(3, 13) = 7.11, p = 0.001] and decrease in tail rattles [F(3, 13) = 9.01, p = 0.001] were also observed after the administration of the highest dose of EtOH during the aggressive encounter subsequent to FI performance.

### Daily Administrations of Alcohol Dose Dependently Reduced, and Thereafter Escalated, the Motivation to Fight

The rate of FI responding for an aggressive reward was dose dependently attenuated by 1.8 and 2.2 g/kg EtOH over the first few days after administration [F(14, 203) = 5.47, p < 0.001; **Figure 3**, top]. The amount of aggressive behavior at the completion of each FI was significantly reduced in both groups of EtOH treated mice (1.8 and 2.2 g/kg) following each oral administration [F(14, 203) = 4.32, p < 0.001; **Figure 3**, bottom]. When challenged 10 days later with EtOH (1.0 g/kg), both groups of EtOH treated mice produced significantly more FI responding for an aggression reward [F(2, 29) = 20.29, p < 0.001; **Figure 3**], with no notable changes in aggressive performance.

### Repeated Daily Administrations of EtOH Persistently Intensify EtOH-Motivated Responding for an Aggressive Reward

The rate of nose-poke responding during a 10 min FI over 7 days was significantly affected by the administration of 1.8 g/kg EtOH within a large cohort of mice [F(7, 91) = 18.89, p < 0.001; **Figure 4**, top]. Specifically, the first administration of EtOH reduced responding when compared to baseline (water), and this disruptive effect dissipated over the next 3 days, until an increase in motivated responding emerged after the 5th daily administration of EtOH. A sensitization of FI responding for aggression reward was revealed after seven subsequent EtOH free days [day 14, t(13) = 3.6, p = 0.003], and again on experimental days 40 [t(10) = 3.09, p = 0.01] and 60 [t(10) = 3.5, p = 0.005],

frequency of attack bites (Right) is significantly increased by 0.5 g/kg EtOH and decreased by 1.8 g/kg EtOH. Significant post hoc comparisons to water administration are denoted as <sup>∗</sup>p < 0.05 or #p < 0.05.

(Bottom). Dashed light gray lines represent the average baseline water values. Significant effects are depicted as: <sup>∗</sup>p < 0.05 compared to water

baseline; #p < 0.05 compared to the water-treated group.

when all mice were challenged with 1.0 g/kg EtOH as compared to when water was administered the day before (**Figure 4**, top). According to the Index of Curvature, the pattern of FI responding over the course of seven daily EtOH administrations also changed significantly, as responses were found to be more evenly distributed (i.e., IC < 0.3) across the 10 min FI [F(7, 85) = 5.02, p < 0.001, **Figure 4**, bottom]. Following the FI schedule on days 1, 3, and 5 of EtOH administrations, behavioral elements recorded during the aggressive encounters revealed a longer attack latency [F(3, 13) = 15.60, p = 0.001], a shorter duration of physical contact [F(3, 13) = 6.156, p = 0.002] and a decrease in tail rattles directed toward the opponent [F(3, 13) = 9.907, p = 0.001; **Table 1**]. Interestingly, no differences between EtOH (1.0 g/kg) and water were detected on the amount of motor activation during the FI or on subsequent aggressive behaviors during any of the later challenge days (i.e., see day 14, **Table 2**). Average blood EtOH concentrations were 101.6 ± 5.9 mg/dL 10 min after the last 1.0 g/kg gavage administration.

### Dizocilpine (MK-801) Recovered FI Responding for Aggression in Controls Given 1.0 g/kg EtOH but Suppressed Aggressive Performance

Acutely administered EtOH (1.0 g/kg) significantly reduced FI responding for aggression reinforcement in control mice with

a history of water administrations, while mice with a history of repeated EtOH responded significantly more during the FI than water controls upon receiving the same acute dose of EtOH [F(1, 16) = 12.78, p = 0.003; **Figure 5A**]. Twoway repeated measures ANOVA also detected a significant interaction between repeated fluid treatment group and acutely administered 1.0 g/kg EtOH and dizocilpine [F(3, 48) = 4.13, p = 0.011]. Specifically, after receiving 1.0 g/kg EtOH the lowest dose of dizocilpine (0.01 mg/kg) recovered FI responding to baseline in control mice without having any detectable effect on FI responding in mice with a history of repeated 2.2 g/kg EtOH treatments (**Figure 5A**). A main effect of dizocilpine [F(3, 48) = 48.82, p < 0.001] was driven by suppressed FI responding in both control and EtOH groups treated with 1.0 g/kg EtOH and the highest dose of dizocilpine (0.3 mg/kg). While FI responding for aggression was significantly affected by the historical, repeated administration of water or EtOH, performance during aggressive interactions did not differ between controls and repeatedly EtOH-treated mice that received 1.0 g/kg EtOH. However, two-way RM ANOVA detected a main effect of dizocilpine on attack bite frequency, with all three doses (0.01, 0.1, and 0.3) significantly reducing aggressive behavior [F(3, 48) = 29.17, p < 0.001; **Figure 5B**].

TABLE 2 | A week after repeated EtOH (1.8 g/kg/day, PO) the effects of an acute water or EtOH (1 g/kg) challenge on motor behaviors during fixed interval (FI) responding and aggressive performance.


### Ketamine Suppressed FI Responding for Aggression and Aggressive Performance in Water-Treated Controls and EtOH-Treated Mice Given an Acute Dose of EtOH

Mice with a history of repeated EtOH increased their cumulative FI responding for an aggressive encounter while control mice showed decreased responding upon receiving a 1.0 g/kg EtOH challenge (**Figure 5C**). Two-way RM ANOVA detected this interaction between treatment history (control vs. EtOH) and acute fluid administration [water vs. 1.0 g/kg EtOH; F(1, 16) = 12.94, p = 0.002] along with a main effect of increased FI responding by historically EtOH-treated mice [F(1, 16) = 8.83, p = 0.009]. Another two-way RM ANOVA revealed an interaction between EtOH treatment history and ketamine, administered after mice received a 1.0 g/kg EtOH challenge [F(3, 48) = 4.79, p = 0.005]. EtOH-treated mice responded more during the FI for an aggressive encounter after receiving acute EtOH and the vehicle injection compared to controls. However, these animals were more sensitive to the effects of ketamine and showed a significant reduction in their responding after receiving 7.5 or 10.0 mg/kg ketamine; in contrast, water control animals only reduced their responding when given the highest, 10.0 mg/kg dose of ketamine. Like MK-801, ketamine (7.5 and 10.0 mg/kg) suppressed aggressive performance in both control mice and in historically EtOH-treated animals (**Figure 5D**).

### NBQX Reduced Aggressive Performance Without Affecting Responding for Aggression

Mice given 2.2 g/kg EtOH repeatedly showed an increase in cumulative responding for aggression after receiving an acute dose of 1.0 g/kg EtOH and IP vehicle compared to the water baseline and compared to water-treated controls (**Figure 5E**). In addition to this interaction between repeated EtOH treatment and responding after acute EtOH [F(1, 16) = 15.01, p = 0.001], an additional two-way RM ANOVA detected a significant main effect of treatment group, indicating increased responding in mice with a history of EtOH treatment compared to water

revealed by distance traveled in an open field (Top, Left). A later challenge with an EtOH (1.0 g/kg, PO; light gray) did not prompt the expression of locomotor sensitization in these mice (Top, Right; Bottom, Middle). No conditioned locomotor effects of daily EtOH administrations were apparent when these groups when later challenged with a water gavage (Bottom, Left). A challenge administration of 2.0 g/kg EtOH, however, did significantly increase locomotor activity in mice that previously received seven oral administrations of 2.2 g/kg EtOH (Bottom, Right). Significant effects are depicted as: <sup>∗</sup>p < 0.05 compared to corresponding water treatment.

controls when data were collapsed across NBQX dose [F(1, 16) = 10.07, p = 0.006]. NBQX, unlike MK-801 and ketamine, did not suppress responding for aggression (**Figure 5E**). However, two-way RM ANOVA revealed a significant interaction between EtOH treatment history and aggression after acute EtOH and NBQX [F(3, 48) = 4.92, p = 0.005], as well as main effects of both historic exposure to EtOH [F(1, 16) = 5.47, p = 0.033] and of NBQX [F(3, 48) = 4.01, p = 0.013; **Figure 5F**]. While moderate doses of NBQX (10, 17 mg/kg) diminished aggression in control animals, only the highest dose reduced the number of attack bites inflicted by animals with a history of repeated 2.2 g/kg EtOH treatments.

### Repeated Oral Administrations of EtOH Does Not Engender Locomotor Sensitization to a Later 1.0 g/kg EtOH Challenge

Seven daily oral EtOH administrations progressively enhanced locomotor activity, when measured consistently in the same open field, as determined by a main effect for treatment day [F(2, 54) = 14.4, p < 0.001], and a main effect for treatment group [F(2, 54) = 11.1, p < 0.001; **Figure 6**, top]. However, in response to a later water or EtOH (1.0 g/kg) challenge in the same context, locomotor activity was not significantly altered between these three groups of mice. In fact, a significant main effect for treatment group [F(8, 162) = 9.3, p < 0.001] and time-bin [F(2, 162) = 7.4, p < 0.001] for locomotor activity was only detected during the 2.0 g/kg challenge day. Post hoc analyses revealed a selective increase in locomotion within the first 10 min after receiving 2.0 g/kg in those mice previously treated with the highest dose of 2.2 g/kg EtOH (**Figure 6**, bottom).

### DISCUSSION

Appetitive and performance measures in the context of aggression are clearly dissected with the implementation of an FI schedule of reinforcement (Skinner and Morse, 1957). In the present study, FI response curves, with a characteristic scallop shape, were reliably established and stable for more than a month of successive daily sessions. In confirmation of previous observations, aggressive behavior that reinforces FI responding was more intense than species-typical forms of aggression (Fish et al., 2002). A single administration of a low dose of alcohol (i.e., 0.5 g/kg) significantly increased fighting performance, without affecting FI responding. As the dose of alcohol increased, its sedative effect emerged that resulted in the suppression of both behavioral measures. Specifically, alcohol, at both 1.8 and 2.2 g/kg/day, initially disrupted both responding during the FI and subsequent fighting performance. With repeated daily administrations, however, the disruptive effect of these alcohol doses on FI responding quickly recovered to the original response rates, and eventually a sensitization of FI responding emerged. The lasting expression of intensified aggressive arousal emerged only when alcohol was administered, such that on days when alcohol was not delivered, both FI responding and fighting behavior were comparable to water-treated control mice. This biphasic action of repeated alcohol on the curvature of FI responding supports previous studies suggesting that this schedule of reinforcement is a sensitive measure of anticipatory arousal prior to aggression reward (Skinner and Morse, 1957; Sanger, 1988; Fish et al., 2002).

An acute dose of alcohol (1.0 g/kg) significantly increased FI responding for aggression reward in animals with a history of repeated daily alcohol administrations (1.8 or 2.2 g/kg), whereas the same acute dose of alcohol significantly reduced FI responding for aggression in alcohol-naive mice. The 1.0 g/kg dose of alcohol was selective for FI response rate (i.e., aggressive motivation), and did not differentially affect aggressive performance, or locomotor behavior, in mice with or without a history of receiving daily alcohol treatments (**Table 2** and **Figure 6**). Excitatory and inhibitory amino acid regulatory elements in somatic and terminal regions of the DA motive system (Volkow et al., 2017) are promising targets for escalated alcohol drinking and alcohol-heightened aggression (Gourley et al., 2005; Takahashi et al., 2010, 2015; Newman et al., 2012, 2018). Thus, we examined the general role of iGluRs during later challenges with alcohol in mice that were historically treated with alcohol (2.2 g/kg) or water for 7 days.

Under the present conditions, a low dose of the NMDA receptor antagonist dizocilpine (0.01 mg/kg) increased FI responding for aggression in alcohol-naive animals that were treated acutely with 1.0 g/kg alcohol. These same mice significantly reduced their aggressive performance. Because the effects of alcohol and dizocilpine on these behavioral measures were diametrically opposed, it is likely that alcohol suppresses fighting in alcohol-naive animals through a non-glutamatergic mechanism, perhaps by positively modulating GABA<sup>A</sup> receptor activity (Ticku and Kulkarni, 1988) rather than through a direct and synergistic inhibition of NMDA receptors – the later would be expected to increase fighting (Newman et al., 2018). Unlike dizocilpine, ketamine did not increase FI responding in alcoholnaive mice, which may result from differences in drug kinetics or its interactions with alcohol (Petrakis et al., 2004; Wai et al., 2013).

In contrast with alcohol-naive animals, mice with a history of repeated alcohol exposures were only sensitive to the serenic effects of dizocilpine. Repeated alcohol exposures may upregulate NMDA receptor expression (Haugbol et al., 2005; Wang et al., 2010), thereby preventing dizocilpine from increasing responding in mice that previously received daily alcohol treatments. While the present behavioral findings are suggestive of altered sensitivity to NMDAR antagonists in EtOH-sensitized mice, detailed evaluations of NMDA receptor subtype expression patterns in animals with a history of repeated EtOH exposure and aggression are required to further address this hypothesis. Interestingly, the AMPA receptor antagonist, NBQX, selectively reduced aggressive behavior without affecting FI responding in historically water- or alcohol-treated mice. These data point to a specific role for AMPA receptors in the regulation of schedule-induced aggressive behaviors, but not during the arousal associated with an impending

social confrontation. Together, these findings suggest that the behavioral plasticity associated with the long-term expression of escalated motivation for aggression may not be strictly tied to iGluR-dependent mechanisms and are likely to be centered around changes in homeostatic elements of the DA motive circuit. It remains to be determined whether iGluRs are necessary for the induction of this lasting change in behavior, like other types of mesocorticolimbic-dependent behavioral plasticity (Vanderschuren and Kalivas, 2000; Wolf and Ferrario, 2010; Camarini and Pautassi, 2016). It is, however, noteworthy that the current interrogation of iGluRs during scheduleinduced "anticipatory" aggression dampens this abnormally intense display of attack behavior unlike the pro-aggressive effects of iGluR antagonists observed specifically in "alcoholheightened" aggressors, although more detailed analyses of physical confrontations after schedule-induced aggression are required to fully elucidate these findings (Newman et al., 2018).

### Do Fixed Interval Schedules Capture the Motivation to Fight?

The FI schedule of reinforcement effectively separates the behavioral output into appetitive and consummatory components, allowing for the quantification of effort exerted during the interval before having the opportunity to fight and the evaluation of fight performance. Using a 10-min FI schedule results in scalloped response patterns, that can be illustrated mathematically by their index of curvature (Fry et al., 1960). More positive indices of curvature indicate an increasingly greater number of responses generated toward the anticipated end of the interval. In the current study, alcohol increased or decreased the rate of anticipatory FI responses, depending on the frequency of alcohol administrations. While an initial alcohol administration lowered the index of curvature, repeated alcohol administrations induced an alcohol-escalated pattern of FI responding. The "upward shift" in alcohol-escalated FI responding persisted when mice were challenged with a low unit dose of alcohol for more than a month. The selective increase in FI responding, as compared to fighting performance, indicates a specificity toward the motivational aspects of aggressive behavior. Taken together, these results suggest that repeated alcohol exposures in a certain context (i.e., associated with winning a social confrontation) may increase the incentive salience of aggression rewards (Ginsberg and Allee, 1942; Segal et al., 1974; Robinson and Berridge, 1993). While the current study examined oral "gavage" administrations of EtOH, future experimental approaches that allow for self-administered alcohol will provide even more translational value. It will also be interesting to learn if the escalation of FI responding for aggression by EtOH can be generalized to other natural rewards.

Progressive ratio (PR) schedules of reinforcement, from a historical perspective, are more often used to characterize the motivation to achieve rewards (Hodos, 1961), including aggression (Golden et al., 2017). PR schedules are considered extinction trials (i.e., the animal ceases to respond at their "breaking point") and inherently requires multiple reward presentations and consumptions in a single session. Throughout PR schedules, successive reinforcements arguably influence the reward value of each future reinforcement delivery (Stafford et al., 1998). FI schedules measure the acceleration of responding rather than cessation and rely on a single aggression reward per daily session. Employing complex chain schedules of reinforcement may also carefully allow for detailed behavioral analyses of the motivation to acquire social rewards like aggression or sex (Everitt, 1990). Nonetheless, the FI schedule used currently was sensitive to both increases and decreases in anticipatory arousal.

### Neural Contributions to the Emergence of Alcohol-Escalated Motivation for Aggression

Alcohol, and its metabolites, directly and indirectly activate VTA DA neurons, ultimately enhancing DA release in terminal areas that are important for reward processing (Di Chiara and Imperato, 1988; van Erp and Miczek, 2007; Plescia et al., 2014; Bassareo et al., 2017). Many DA, and serotonin (5-HT), receptor populations (e.g., DA D1 and D2; 5HT1A, 5HT1B, and 5HT2C) in midbrain, cortical and limbic areas are also critical for the execution of offensive aggression (Nikulina and Kapralova, 1992; Bondar and Kudryavtseva, 2005; de Boer and Koolhaas, 2005; Takahashi et al., 2012). Neural adaptations in response to the repeated actions of alcohol are likely centered around an augmentation of the DA motive circuitry (Volkow et al., 2017). This form of augmented neural plasticity may well be linked to a sensitization of incentive salience for the positively reinforcing effects of winning aggressive confrontations (Robinson and Berridge, 1993), and the potential to form an adaptive bias toward hyperexcitability during aggressive arousal. Glutamatedependent forms of plasticity are necessary for the long-term behavioral consequences of alcohol to occur (Brodie, 2002; Broadbent et al., 2003), perhaps altering the perceived outcome of aggression rewards. Indeed, appetitive responses are particularly sensitive to synaptic changes in posterior VTA DA neurons (Wise and McDevitt, 2018), and it is tempting to hypothesize that biogenic amines and amino acids play a concerted role in the development of alcohol-escalated aggressive motivation. In an apparent dissociation from other forms of behavioral sensitization, the expression of alcohol-escalated motivation to fight is not readily attenuated by iGluR antagonists at dose ranges that are free from disrupting performance measures, as observed herein (**Figure 5**).

Circuits involving anticipatory arousal and the neurobiology of EtOH's actions also overlap considerably. Hypothalamic and extra-hypothalamic nuclei that are rich in neuropeptides establish such a link (Cannizzaro et al., 2010; Plescia et al., 2014; Pleil et al., 2015; Rinker et al., 2017). It is reasonable to hypothesize that EtOH and its metabolites alter the regulation of sympathetic responses to exacerbate "hot" acts of aggression. Ongoing studies continue to focus on direct and indirect modulation of mesocorticolimbic DA during the expression of alcohol-escalated aggressive motivation. It appears promising to differentiate the neural circuits mediating the urge to fight vs. those responsible for the performance of aggressive acts.

### AUTHOR CONTRIBUTIONS

fnbeh-12-00206 September 11, 2018 Time: 18:51 # 12

HC and KM contributed to designing and conducting the experiments, analyzing the data, interpretation of the results, and writing the manuscript. EN contributed to analyzing the data, interpreting the results, and writing the manuscript. JD and ML assisted with the design of experiments. ST, LW, and WH assisted with experimental procedures.

### REFERENCES


### FUNDING

This work was supported by funding provided by NIH grants R01 AA013983 (KM) and F31 AA025827 (EN).

### ACKNOWLEDGMENTS

The authors wish to thank J. Thomas Sopko and Kurtis Chien-Young for their technical assistance as well as their artistic and scholarly contributions during the preparation of this manuscript.


withdrawal from intermittent alcohol in outbred mice. Psychopharmacology (Berl.) 232, 2889–2902. doi: 10.1007/s00213-015-3925-y


Zhang, L., Welte, J. W., and Wieczorek, W. W. (2002). The role of aggressionrelated alcohol expectancies in explaining the link between alcohol and violent behavior. Subst Use Misuse 37, 457–471. doi: 10.1081/JA-120002805

**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Covington, Newman, Tran, Walton, Hayek, Leonard, DeBold and Miczek. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# The Role of Estrogen Receptor β (ERβ) in the Establishment of Hierarchical Social Relationships in Male Mice

#### Mariko Nakata1,2 , Anders Ågmo<sup>3</sup> , Shoko Sagoshi <sup>1</sup> and Sonoko Ogawa<sup>1</sup> \*

<sup>1</sup>Laboratory of Behavioral Neuroendocrinology, University of Tsukuba, Tsukuba, Japan, <sup>2</sup>Research Fellow, Japan Society for Promotion of Science (JSPS), Tokyo, Japan, <sup>3</sup>Department of Psychology, University of Tromsø, Tromsø, Norway

Acquisition of social dominance is important for social species including mice, for preferential access to foods and mates. Male mice establish social rank through agonistic behaviors, which are regulated by gonadal steroid hormone, testosterone, as its original form and aromatized form. It is well known that estrogen receptors (ERs), particularly ER α (ERα), mediate effects of aromatized testosterone, i.e., 17β-estradiol, but precise role played by ER β (ERβ) is still unclear. In the present study, we investigated effects of ERβ gene disruption on social rank establishment in male mice. Adult male ERβ knockout (βERKO) mice and their wild type (WT) littermates were paired based on genotypeand weight-matched manner and tested against each other repeatedly during 7 days experimental period. They underwent 4 trials of social interaction test in neutral cage (homogeneous set test) every other day. Along repeated trials, WT but not βERKO pairs showed a gradual increase of agonistic behaviors including aggression and tail rattling, and a gradual decrease of latency to social rank determination in tube test conducted after each trial of the social interaction test. Analysis of behavioral transition further suggested that WT winners in the tube test showed one-sided aggression during social interaction test suggesting WT pairs went through a process of social rank establishment. On the other hand, a dominant-subordinate relationship in βERKO pairs was not as apparent as that in WT pairs. Moreover, βERKO mice showed lower levels of aggressive behavior than WT mice in social interaction tests. These findings collectively suggest that ERβ may play a significant role in the establishment and maintenance of hierarchical social relationships among male mice.

Keywords: gonadal steroid hormone, testosterone, aggressive behavior, dominance hierarchy, rank determination, agonistic behavior, social interaction, tube test

### INTRODUCTION

Individuals of social species often establish hierarchical social relationships with their conspecifics. Once their social rank is determined, dominant (higher rank) individuals can get preferential access to resources including food, territory and mates. It is known that group-housed male mice establish a social hierarchy both in the wild and in laboratory housing conditions. Their social order is determined through agonistic interactions, which include not only active aggressive behavior but also various behavioral responses to opponents' aggression, such as fleeing, immobility and upright submissive posture (Grant and Mackintosh, 1963). Once hierarchy is

#### Edited by:

Rosa Maria Martins De Almeida, Universidade Federal do Rio Grande do Sul (UFRGS), Brazil

#### Reviewed by:

Paola Palanza, Università degli Studi di Parma, Italy James P. Curley, University of Texas at Austin, United States

\*Correspondence:

Sonoko Ogawa ogawa@kansei.tsukuba.ac.jp

Received: 08 July 2018 Accepted: 01 October 2018 Published: 22 October 2018

#### Citation:

Nakata M, Ågmo A, Sagoshi S and Ogawa S (2018) The Role of Estrogen Receptor β (ERβ) in the Establishment of Hierarchical Social Relationships in Male Mice. Front. Behav. Neurosci. 12:245. doi: 10.3389/fnbeh.2018.00245 settled, a highest rank male mouse (α-dominant male) consistently attacks subdominant, subordinate and intruder males and successfully defends his territory from rivals (Singleton and Hay, 1983; Palanza et al., 1996; Miczek et al., 2001; Wang et al., 2011).

For assessment of social rank among mice, various testing paradigms, such as direct observation of agonistic behaviors (Miczek et al., 2001) and tube test (Wang et al., 2011), have been developed. Previous studies on neural mechanisms underlying the establishment of social hierarchy in male mice revealed that the gonadal steroid hormone, testosterone, plays an essential role (Machida et al., 1981; Giammanco et al., 2005). Testosterone is mainly secreted from the testes into the blood stream and binds not only to androgen receptors but also to estrogen receptors (ERs), after conversion to 17β-estradiol by aromatase in the brain. Two subtypes of nuclear ERs, ERα and ERβ, are known to mediate intracellular actions by 17β-estradiol, aromatized testosterone.

It is well established that ERα is necessary for the induction of male sexual and aggressive behaviors in mice (Ogawa et al., 1997, 1998, 2000; Rissman et al., 1997; Wersinger et al., 1997). In contrast, the role of ERβ in the regulation of male social behaviors is still not fully understood. ERβ has been thought to modulate male social behaviors in a complex manner, rather than simply induce a stereotyped behavioral pattern. Ogawa et al. (1999) initially reported altered aggressive behavior in male ERβ knockout (βERKO) mice. In aggression tests using a residentintruder paradigm, wild-type (WT) mice showed a gradual increase of aggression levels over three consecutive tests. On the other hand, βERKO males showed high levels of aggression (longer duration of aggressive bouts) starting on the first trial and kept steady levels of aggression throughout the repeated tests. Moreover, βERKO males were much more aggressive particularly during pubertal period compared to WT mice (Nomura et al., 2002; Handa et al., 2012; Tsuda et al., 2014). These experienceand age-dependent influences of ERβ gene disruption suggested that ERβ might regulate male social behaviors in a specific context such as establishment of social hierarchy.

In the present study, we investigated the influence of ERβ gene disruption on the process of establishment of a hierarchical social relationship among socially naïve mice. βERKO and WT male mice were paired with same-sex and same-genotype individuals. Agonistic and prosocial behaviors were analyzed in social interaction tests performed repeatedly (4 trials) over 7 days. Social rank was also assessed with the use of the tube test, right after each trial of the social interaction test. In order to identify critical behavioral acts between paired mice for the determination of social rank, we also analyzed behavioral transition patterns and compared these between WT pairs and βERKO pairs.

### MATERIALS AND METHODS

### Subjects

Gonadally intact and sexually naïve male βERKO and WT littermate mice (βERKO: 12 pairs, n = 24, WT: 10 pairs, n = 20) were used as experimental animals. They were obtained from a breeding colony maintained at the University of Tsukuba. Original breeding pairs were provided by Dr. KS Korach at the National Institute of Environmental Health Sciences (Research Triangle Park, NC, USA) and completely backcrossed to C57BL/6J mice (Krege et al., 1998). Mice were weaned at 3 weeks of age and then group housed with same-sex littermates in genotype-mixed manner. They were kept in polypropylene clear plastic cages (19 × 29 × 12 cm) until the experiment started. They were kept under standard housing conditions (23 ± 2 ◦C, 12:12 light/dark cycle with lights off at 12:00). Food and water were provided ad libitum. All procedures were conducted in accordance with the National Institutes of Health guidelines and were approved by the Animal Care and Use Committee and the Recombinant DNA Use Committee at the University of Tsukuba. All efforts were made to minimize the number of animals and their suffering.

### Experimental Procedures

Starting at 17 ± 4.5 weeks old and throughout the experiment, all mice were individually housed in small transparent plastic home-cages (12.5 × 20 × 11 cm). Non-littermate mice from the same genotype and matched body weight (±3.5 g) were paired (homogeneous pair) and tested against each other throughout the experiment. After 1 week of individual housing, mice were trained for the tube test on two consecutive days. Starting on the next day, each pair underwent the social interaction test followed by the tube test on every other day (days 1, 3, 5 and 7) for a total of 4 trials (trials 1, 2, 3 and 4). All behavioral tests were recorded using digital video cameras and scored by an experimenter unaware of the animals' experimental group, using a digital event recorder program (Recordia 1.0b, O'Hara & Co., Ltd.).

### Social Interaction Test

Social interaction behaviors between the paired mice were assessed in a neutral testing cage (19 × 29 × 12 cm) for 15 min. All tests were done under red light illumination during the dark phase of the light/dark cycle. At first, the testing cage was divided into two compartments by inserting a black Plexiglas board (divider) at the middle of the cage and mice were habituated in each compartment for 5 min. At the beginning of the test, the divider was removed and social interaction behaviors were observed. The cumulative number and duration of aggression, fleeing, immobility, upright submissive posture, approach, sniffing, huddling and grooming were recorded. Aggression was defined as a series of behavioral interactions consisting of at least one of the following: chasing, boxing, wrestling, biting and offensive lateral attack, often accompanied by biting. The cumulative number of tail rattling was also recorded. These nine behavioral acts were classified into two groups for further analysis: agonistic behaviors (aggression, fleeing, immobility, submissive posture and tail rattling) and prosocial behaviors (sniffing, grooming, approaching and huddling). Sniffing and grooming were further categorized into face-targeted or body-targeted. Thereafter, ratio of the face-targeted was calculated and compared between the ranks determined by the tube test. Sniffing and grooming were combined for this analysis (see **Supplementary Table S1**).

### Tube Test

Right after the completion of the social interaction test, the tube test was conducted to assess social rank between the paired mice. All tests were done under red light illumination during the dark phase of the light/dark cycle. A clear plexiglass tube (3 cm inner diameter and 45 cm long) was placed at the center of the testing arena (70 × 50 cm) surrounded by black wall (20 cm). Starting from 2 days before the first social interaction test, all mice were trained individually to run through the tube from one end to the other eight times per day for two consecutive days, as previously described (Wang et al., 2011). A black plastic escape box (13 × 14 × 13 cm) was placed at the end of the tube during these training sessions.

On each testing day, all mice were individually given two pre-test trials to run through the tube without an escape box. In test trials, mice in each pair were released simultaneously from one of two ends of the tube. Each test trial lasted until one mouse forced the other to retreat from the tube. The former mouse remaining in the tube was judged as a ''winner'' and the latter mouse ejected from the tube was judged as a ''loser''. The winner animal ID and latency to loser ejection were recorded in each test trial. Winner shift, defined as the winner being different between two consecutive test trials (Oakeshott, 1974; Wang et al., 2011), was also analyzed. Since video recordings of three pairs (two WT and one βERKO pairs) on Day 1 failed, all data from these pairs were excluded from the analysis.

### Analysis of Behavioral Transition During Social Interaction Tests

Behavioral transitions of two consecutive behavioral events occurring with an interval of less than 6 s were analyzed. They were classified as monad or dyad transitions depending on actor(s) of the behavioral events. Among nine behavioral acts, upright submissive posture was excluded from this analysis because only limited mice showed this behavior. In monad transitions, actors of the two consecutive behavioral events were the same mouse. In dyad transitions, an actor of the first behavioral event (initiator) and that of the second behavioral event (responder) were different mice. For monad transitions, the probabilities of transitions were calculated and 8 kinetograms were constructed for each trial and genotype. For analysis of dyad transitions, eight behavioral events recorded in social interaction test was partially combined as follows; subordinate behaviors including fleeing and immobility, and prosocial behaviors including sniffing, grooming and huddling. Probabilities of dyad transitions were then calculated and 8 kinetograms were constructed for each test, genotype and rank (winner or loser in the tube test). Differences between tests, genotypes and ranks were analyzed qualitatively based on the diagrams. For dyad transitions, the number of all transitions, transitions initiated with approach, and transitions responded with subordinate behaviors were also counted and statistically analyzed.

### Statistics

Agonistic and prosocial behaviors in social interaction tests, and latency to loser ejection in tube tests were analyzed by a two-way analysis of variance (ANOVA), repeated measurements of the main effects for genotype, trials and their interaction. Post hoc power analyses for the main effects and their interaction of ANOVAs (Cohen, 1992) were conducted with G∗Power version 3.1.9.2 (Faul et al., 2007). Genotype differences in percentage of animals showing aggression in social interaction tests, winner shift frequency of tube tests, and winner/loser ratio of initiator of dyad transitions were analyzed by a Fischer's exact test, with stratified analysis of Benjamini and Hochberg method. Genotype differences in total number of all dyad transitions were analyzed by a Chi-squared test with stratified analysis of Benjamini and Hochberg method. Rank differences in approachor subordinate-transitions were analyzed by a Binomial test. ANOVAs were conducted using the SPSS version 21 (SPSS Inc., Chicago, IL, USA). Fischer's exact test, Chi-squared test and Binomial test were conducted with js-STAR (v. 8.0.0 j) software. Statistically significant differences were considered at p < 0.05.

### RESULTS

### Agonistic and Prosocial Behaviors in the Social Interaction Test

The cumulative number of agonistic behaviors gradually increased over the four trials in WT mice, but did not change in βERKO mice (**Figure 1A**, left panel; genotype: F(1,42) = 14.741, p < 0.001, d = 0.592, power (1-β) = 0.997; trial: F(3,126) = 6.333, p < 0.001, d = 0.388, power (1-β) = 0.999; genotype × trial: F(3,126) = 3.617, p = 0.015, d = 0.293, power (1-β) = 0.998). In WT mice, the number of agonistic behaviors in trials 3 and 4 were significantly higher than in trial 1 (p < 0.01), whereas no significant difference was observed between trials in βERKO mice. Moreover, βERKO mice showed a significantly lower number of agonistic behaviors, compared to WT pairs in trial 2 (p < 0.05), and trials 3 and 4 (p < 0.01), although there was no genotype difference in trial 1. βERKO mice also showed significantly a shorter overall cumulative duration of agonistic behaviors, compared to WT mice (**Figure 1A**, right panel; genotype: F(1,42) = 8.629, p = 0.005, d = 0.453, power (1-β) = 0.903). Significant genotype difference was observed in trial 2 (p < 0.05), and trials 3 and 4 (p < 0.01), although there was no significant main effect of trial and interaction of genotype and trial (trial and genotype × trial, n.s.). In contrast, both number and duration of prosocial behaviors were not different between genotypes and did not change over the four trials (**Figure 1B**; main effects of genotype, trial, and interaction, n.s.). Additionally, detailed analysis of sniffing and grooming revealed that in both WT and βERKO pairs, there was no difference in the probability of face-targeted sniffing and grooming between the winner and loser in the tube test conducted in the same experimental day following the social interaction test (**Supplementary Table S1**).

wild type (WT; •), ER<sup>β</sup> knockout (βERKO; ) mice did not show an increase in the number (left panel) of agonistic behaviors over trials. Moreover, <sup>β</sup>ERKO mice showed shorter duration of agonistic behaviors (right panel). (B) There was no difference between βERKO and WT groups in the number (left panel) and duration (right panel) of prosocial behaviors. All data are presented as mean ± SEM. <sup>a</sup>p < 0.01 compared with trial 1 of the same genotype; <sup>∗</sup>p < 0.05; ∗∗p < 0.01 compared with WT in the same trial.

### Tube Test

The latency to loser ejection in tube tests decreased significantly during the four test trials in WT but not in βERKO pairs (**Figure 2**; trial: F(3,51) = 9.143, p < 0.001, d = 0.733, power (1-β) = 0.999; genotype × day: F(3,51) = 8.551, p = 0.007, d = 0.709, power (1-β) = 1.000; genotype: n.s.). As for the winner shift, there was a trend of a gradual decrease from trial 2 to trial 4 only in WT, but not in βERKO mice, although there were no statistically significant genotype differences (**Table 1**; trials 2, 3 and 4; n.s.).

### Monad Behavioral Transition Patterns During Social Interaction Tests

Monad-type behavioral transition patterns during social interaction tests, in which two consecutive behavioral

FIGURE 2 | Influence of ERβ gene disruption on the latency to loser ejection in the tube test. Unlike WT (•), <sup>β</sup>ERKO () mice did not show a decrease in the latency to loser ejection over trials. Data are presented as mean ± SEM. <sup>a</sup>p < 0.05; <sup>b</sup>p < 0.01 compared with trial 1 of the same genotype; <sup>∗</sup>p < 0.05 compared with WT in the same trial.

TABLE 1 | Number of pairs with winner shift in each trial.


There was no significant genotype difference in frequency of the winner shift in each trial.

events were acted by the same mouse, were visualized using kinetograms for each genotype and trial. To construct each kinetogram, data from all mice were combined. Kinetograms for WT mice (**Figure 3**) indicated that WT mice mainly showed investigative behavior, particularly transitions from approach to sniffing, in trial 1. Along repeated trials, behavioral transition patterns of WT mice shifted from investigation to threatening which includes tail rattling. Significant increases of aggression and tail rattling in trials 3 and/or 4 (**Supplementary Figure S1A**) were consistent with these behavioral changes in WT mice. In contrast, βERKO mice did not show any obvious changes of their behavioral patterns throughout the four trials and mainly exhibited investigative behavior (**Figure 4**).

### Analysis of Dyad Social Interaction Patterns Between Winners and Losers

To examine social interaction between winners and losers, dyad behavioral transitions, in which one mouse (responder) responded to a preceding behavioral event acted by the other mouse (initiator), were analyzed.

The total number of dyad transitions, as an index of richness of social interaction, was first examined and compared between WT and βERKO pairs (**Table 2**). WT pairs showed a gradual increase of the number of dyad transitions along repeated trials whereas βERKO pairs did not show such changes. Statistical analysis revealed that the total number of all dyad transitions of βERKO pairs was not different from that of WT pairs on trial 1, but significantly fewer in trials 3 and 4 (**Table 2**; trial 2: X 2 (1) = 2.976, 0.050 < p < 0.100; trial 3: X<sup>2</sup> (1) = 25.638, p < 0.010; trial 4: X<sup>2</sup> (1) = 23.554, p < 0.010). It should be noted that throughout the 4 trials, the numbers of dyad transitions initiated by winners and losers were roughly equal in both genotypes (**Table 2**; trials 1, 2, 3 and 4; n.s.).

Further analysis using kinetograms revealed that in WT pairs approach responded by approach (approach—approach) was a predominant type of dyad transitions in trial 1 (**Figure 5**). However, WT winners showed one-sided aggression thereafter. In trial 2, a strong asymmetry pattern of dyad transitions, initiated by winners' aggression or approach and followed by losers' subordinate behaviors, became obvious. Consistent with the findings in monad transition analysis, transitions as tail rattling—tail rattling and approach—tail rattling became predominant in trials 3 and 4.

In contrast to WT pairs, the most predominant transition was approach—approach in βERKO pairs in all four trials (**Figure 6**). In trial 1, an asymmetry transition pattern in which winners' approach was followed by losers' subordinate behavior, was observed. Starting with trial 2 and thereafter, most of the transitions were symmetrical between winners and losers.

The number of transitions initiated by approach (approachtransitions; **Table 3**) and those responded by subordinate behaviors (subordinate-transitions; **Table 4**) were counted separately for winners and losers in each trial. Statistical analysis revealed that WT winners initiated more approach-transitions

transition.

Arrow width is proportional to (Total number of each transition)/(Total number of all transitions within each genotype and trial). Arrowhead indicates direction of each

TABLE 2 | Total numbers of dyad transitions in each trial and genotype group.


Estrogen receptorβ knockout (βERKO) mice showed a significantly lower number of dyad transitions in trials 3 and 4 than wild type (WT) mice, when winners and losers were combined. Data are presented as total of each trial and group. ∗∗p < 0.01 compared with WT in the same trial.

than WT losers in trial 2 whereas βERKO winners did so in trial 1 (**Table 3**; WT, trial 2; p = 0.016; βERKO, trial 1; p < 0.001, Binomial test). Consistent with these findings, losers of WT pairs showed a significantly higher number of subordinate behavior than winners in trial 2 (**Table 4**; WT, trial 1; p = 0.087, trial 2; p < 0.001, trial 3; p = 0.080 in Binomial tests). On the other hand, losers of βERKO pairs showed a significantly higher number of subordinate behaviors than winners in trial 1 (**Table 4**; βERKO, trial 1; p = 0.002, Binomial tests).

### DISCUSSION

In the present study, we investigated the effects of ERβ gene disruption on the establishment of hierarchical social relationships among male mice. We assessed behavioral changes during repeated trials in social interaction tests conducted in neutral cages (homogeneous set test) followed by tube tests for social rank evaluation.

by the winners. Gray arrows indicate dyad transitions initiated by the losers.

Over four trials, WT pairs showed a gradual increase in agonistic behaviors, such as aggression and tail rattling in social interaction tests. In tube tests, a corresponding decrease of latency to loser ejection was observed. After an initial investigative period in trial 1, WT winners defeated the losers and acquired their dominance. Detailed analysis of behavioral transitions revealed that an asymmetry in behaviors of the winners and losers appeared in trial 2 in WT pairs—i.e., losers responded to winners' one-sided attack with subordinate behaviors. Summarized kinetograms (**Supplementary Figure S2**) clearly demonstrate one-sided dyad transitions from winners' aggression and/or tail rattling to losers' subordinate behaviors in trial 1 and 2 (indicated by thickness of black lines in the left top kinetogram). In trials 3 and 4, winners tried to defend their dominance status through continuous agonistic interactions including frequent tail rattling (indicated with large size circles and thick transition lines in the left bottom kinetogram of **Supplementary Figure S2**). Occurrence of aggressive behavior and winner shifts, even in trial 4, suggested that not all WT pairs successfully established stable dominant-subordinate relationships by the end of four trials. Observation of upright submissive posture also supports these hypotheses (**Supplementary Figure S3**). In WT group, four mice in three pairs showed submissive posture during the social interaction tests and three out of these four mice were losers in the tube test. As one exception, the winner of the pair W15 showed upright submissive posture in trial 4. However, in this trial, both winner and loser showed aggression as well. Additionally, in WT pairs, the results of the tube test reflect dominant-subordinate relationships in the social interaction test after trial 2, consistent with previously reported findings (Wang et al., 2011). These results collectively suggest that WT pairs went through a process of social rank establishment.

In contrast to the findings in WT pairs, βERKO pairs showed little behavioral changes throughout the four trials. Although rank asymmetry in behavioral patterns was observed in trial 1, apparent one-sided aggressive behavior was not observed in βERKO pairs. They kept investigating each other intensively without any increase of aggression or tail rattling throughout all four trials (**Supplementary Figures S1**, **S2**, right kinetograms). The total number of dyad transitions was significantly lower in βERKO pairs, compared to WT pairs in trials 3 and 4. Thus, a dominant-subordinate relationship in βERKO pairs was not as apparent as observed in WT pairs. Although some of

βERKO mice showed submissive posture, it was not necessarily observed in losers in the tube test (**Supplementary Figure S3**). These results suggest that disruption of ERβ gene suppressed the expression of typical behavioral interactions necessary for progression of hierarchical social relationship establishment among male mice.

It has been reported previously that βERKO mice show a tendency of prolonged investigation of social stimuli (Handa et al., 2012; Tsuda et al., 2014). Consistent with these findings, investigatory behavior, compared to WT mice throughout the four trials. These behavioral characteristics may be partly due to changes in the level of anxiety in βERKO mice. Several lines of evidence have suggested that ERβ plays a role in the estrogenic regulation of anxiety-related behaviors (Walf and Frye, 2007; Weiser et al., 2008; Tomihara et al., 2009). In male mice, treatment with specific agonist to ERβ decreased anxietyrelated behavior in a non-social context (Frye et al., 2008).

βERKO pairs in the present study showed higher levels of social



In WT pairs, winners initiated significantly more approach-transition than losers in trial 2. On the other hand, this rank asymmetry in βERKO pairs was observed in trial 1. Data are presented for each trial, genotype and rank. <sup>∗</sup>p < 0.05; ∗∗p < 0.01 compared with losers in the same genotype.

TABLE 4 | Total number of dyad transitions ended by subordinate behaviors (subordinate-transition).


In WT pairs, losers responded with subordinate behaviors more frequently than winners in trial 2. On the other hand, losers in βERKO pairs did so in trial 1. Data are presented for each trial, genotype and rank. <sup>∗</sup>p < 0.05; ∗∗p < 0.01 compared with winners in the same genotype.

Increased anxiety levels may also underlie the reduced aggressive behavior in βERKO mice during social interaction tests in the present study. βERKO mice showed a lower level of aggression and tail rattling than WT mice, although they were not different from WT in terms of other components of agonistic behaviors and prosocial behaviors (**Supplementary Figure S1**). Previous studies reported a consistent tendency of increased aggressive behavior in βERKO male mice in residentintruder paradigm tests (Ogawa et al., 1999; Nomura et al., 2002; Handa et al., 2012; Tsuda et al., 2014). Additionally, although the levels of aggressive behavior were not affected, treatment with selective agonist of ERβ increased dominant behaviors of resident male mice (Clipperton Allen et al., 2010). It should be noted that in those experiments, test mice were presented with a stimulus mouse in their own territory. In addition, olfactory-bulbectomized or gonadectomized males, which rarely counterattack, were used as stimulus animals. Therefore, it is assumed that ERβ activation suppress aggressive behavior in male mice in their own territory toward a stimulus animal which is less likely to counterattack—i.e., in case of an α-dominant male mouse. On the other hand, the homogeneous set test used in the present study provides a completely different ecological situation from the resident-intruder paradigm test. Since both mice in the pairs had intact gonads and olfactory bulbs, they were more likely to counterattack compared to stimulus animals in the resident-intruder paradigm test. Moreover, the social interaction test was conducted in a neutral cage, which was the territory of neither mouse. Thus, when the experimental animal is not in an advantageous situation—i.e. not an α-dominant male in his own territory—ERβ activation may be necessary to induce aggressive behavior and tail rattling for the establishment of a dominantsubordinate relationship.

In the present study, βERKO pairs showed little enhancement of aggression through repeated encounters. This behavioral phenotype is consistent with the previous studies with the resident-intruder paradigm tests (Ogawa et al., 1999). Decreased responsivity to repeated aggressive encounter may alternatively explain disruption of rank establishment in βERKO pairs. However, lack of behavioral change throughout the trials and active behavioral interactions toward rank establishment in βERKO mice may not be due to disruption in social memory and/or social recognition induced by ERβ gene disruption. To establish a social relationship through repeated behavioral interaction, mice need to recognize their opponents and keep social memory until the next trial. It was reported that βERKO male mice possess long-term social memory and are able to discriminate two male stimulus mice in social recognition tests (Sánchez-Andrade and Kendrick, 2011). It should be noted, however, in the present study, mice were group-housed with same-sex littermates

### REFERENCES

Clipperton Allen, A. E., Cragg, C. L., Wood, A. J., Pfaff, D. W., and Choleris, E. (2010). Agonistic behavior in males and females: effects of an estrogen receptor beta agonist in gonadectomized and gonadally intact mice. Psychoneuroendocrinology 35, 1008–1022. doi: 10.1016/j.psyneuen.2010.01.002 until the experiments started. Thus, we cannot exclude the possibility that experience such as social defeat by a cage-mate during group-housed period may be different between genotypes and strengthen the effects of ERβ gene disruption.

In summary, we found that ERβ gene disruption may prevent social rank establishment among male mice. Unlike previously reported findings with the resident-intruder paradigm tests, βERKO male mice showed reduced levels of aggressive behavior in a neutral testing situation. It is hypothesized that ERβ activation may promote aggressive behavior in male mice to acquire social dominance until they establish the status as an α-dominant male. After that, ERβ activation may inhibit excess aggressive behavior by an α-dominant male to avoid further unnecessary injuries of subordinates. In addition, these behavioral phenotypes of βERKO male mice are observed in dyad tests (interaction between two mice). In future studies, it is needed to investigate further whether the βERKO mice are able to establish hierarchical social relationships in larger groups since being in a pair and in a larger group induce different hormonal status in male mice (Williamson et al., 2017). Taken together, we believe that ERβ may be involved in facilitating both the establishment and maintenance of hierarchical social relationships among male mice by regulating aggressive behavior in a social status-depending manner. Further studies are necessary to determine possible underlying neural mechanisms of ERβ-mediated regulation of social behavior.

### AUTHOR CONTRIBUTIONS

MN, AÅ, and SO designed research. MN, AÅ and SS performed research and analyzed data; MN and SO wrote the article.

### FUNDING

The work was supported by the Japan Society for the Promotion of Science (JSPS) Grant-in-Aid for Scientific Research 15H05724 to SO and 17J09810 to MN.

### ACKNOWLEDGMENTS

We thank Professor C. Pavlides for reviewing the manuscript, and Dr. A. Takahashi for valuable discussion.

### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fnbeh. 2018.00245/full#supplementary-material

Faul, F., Erdfelder, E., Lang, A.-G., and Buchner, A. (2007). G∗Power 3: A flexible statistical power analysis program for the social, behavioral and biomedical sciences. Behav. Res. Methods 39, 175–191. doi: 10.3758/ bf03193146

Cohen, J. (1992). A power primer. Psychol. Bull. 112, 155–159. doi: 10.1037/0033- 2909.112.1.155


**Conflict of Interest Statement**: The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Nakata, Ågmo, Sagoshi and Ogawa. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

digital media

of impactful research

article's readership