This analysis was part of a larger project (see Ebner et al., 2015, 2016; Lin et al., 2018; Horta et al., 2019; Plasencia et al., 2019; Frazier et al., 2021). The current analysis sample comprised 98 of the originally recruited 105 participants,¹ with 48 young (M = 22.43 years, SD = 2.97, 18–31 years, 48% females) and 50 older (M = 71.08 years, SD = 4.98, 63–81 years, 58% females) adults. Using G*Power, sensitivity analyses showed that with the current sample size and p = 0.05 as type I error threshold, we had 80% power to detect a small effect (Cohen’s f = 0.18; Cohen, 2013) for the interaction between age group and face trustworthiness level on trustworthiness ratings (Hypothesis 1); and 80% power to detect small effects (Cohen’s f² = 0.13; Cohen, 2013) for the linear and quadratic trends of face trustworthiness and their interactions with age group on amygdala activity (Hypothesis 2 and 3). In addition, we had 80% power to detect a small effect (Cohen’s f = 0.22; Cohen, 2013) for the interaction between age group, treatment group, and face trustworthiness level on amygdala activity (i.e., exploratory analysis for oxytocin modulation).

All participants were English-speaking, White, with no history of neurological or psychiatric disorder, and with normal or corrected-to-normal vision. All older participants completed the Telephone Interview for Cognitive Status (TICS; Brandt et al., 1988; M = 35.47, SD = 2.42, Min = 30, Max = 42; cut off < 30) as screening for cognitive impairment. Twenty-six young (46% females) and 26 older (54% females) participants were randomly assigned to self-administer 24 international units (IUs) oxytocin via nasal spray and 22 young (50% females) and 24 older (63% females) participants to self-administer 24 IUs of a placebo nasal spray.

As shown in Table 1, overall, young participants performed better in the cognitive tasks (i.e., higher processing speed measured by the WAIS-R Digital Symbol Substitution Test (Wechsler, 1981) and better short-term memory measured by the Rey Auditory Verbal Learning Test (Rey, 1964) than older participants. The age groups also differed in positive affect with higher positive affect scores in older than young participants measured by the Positive Affect Negative Affect Scale (Watson et al., 1988; Röcke et al., 2009).

The present study used data from a larger project that comprised three test sessions: (1) a phone screen (about 30 min) to determine study eligibility (for drug administration as well as MRI) and to collect background demographic and health information; (2) an in-person screen visit (about 45 min), in which participants gave written informed consent, provided blood and urine samples, completed short questionnaires, and underwent a brief general health evaluation with the study clinician; and (3) an in-person full study visit lasting about 3 h, approximately 2–10 days after the in-person screen visit, during which participants self-administered the study drug and completed the fMRI task reported in this paper (as well as other (f)MRI and questionnaire measures reported elsewhere; Ebner et al., 2015, 2016; Horta et al., 2019; Frazier et al., 2021).

All study sessions were conducted by trained research staff and were completed at the Department of Psychology, the Institute on Aging, and the McKnight Brain Institute at the University of Florida. In-person visits started at about 9 a.m., given the fasting blood draw and diurnal hormone cycles. Participants were instructed to stay hydrated and to avoid caffeine and substance use 24 h before the in-person visits and abstain from food, exercise, and sexual activity at least 2 h before the visits. The study protocol was approved by the University of Florida Institutional Review Board (IRB#39-2013) and registered as a clinical trial with ClinicalTrials.gov (NCT01823146). Only measures considered in the present data analysis are described in detail below.

Short cognitive tasks as well as brief personality and socioemotional questionnaires were completed during the in-person screen visit and are presented in Table 1. Positive and negative affect was assessed at the start of the in-person full study visit (see Table 1 for details). This was followed by saliva sampling and self-administration of 24 IUs (one puff per nostril; IND #100,860) of either oxytocin or a placebo, which contained the same ingredients as the oxytocin spray except for the oxytocin. Treatment assignment was randomized and double-blinded, monitored by the dispensing pharmacy, and followed recommendations for standardized intranasal oxytocin administration (Guastella et al., 2013). Before self-administration, all pre-menopausal women completed a pregnancy test to assure safety.

Before entering the MRI scanner, participants underwent another MRI safety screening, received instructions on the scanning procedure, and practiced the tasks they were working on inside the scanner. The scanning session started about 45 min after drug self-administration, with brief anatomical scans and a 15-min decision-making paradigm (Frazier et al., 2021), before participants completed the Face Trustworthiness Rating Task analyzed for this study (see detailed description below). This was followed by a facial expression identification task (Horta et al., 2019), an 8-min resting state scan (Ebner et al., 2016), and a series of socioemotional questionnaires (Ebner et al., 2015), post-scan, debriefing, and compensation procedures outside the scanner.

As illustrated in Figure 1, each trial started with a jittered fixation cross (range: 2–12 s, mean = 4 s), followed by an image (either a face or a scrambled face) presented in the center of the screen for 4 s, with the rating scale and respective finger use for the response presented below the image. Participants were asked to rate the trustworthiness of each face on a four-point scale (1 = Not at all trustworthy to 4 = Very trustworthy) using an MRI compatible 4-button box, while the face was on the screen, or press any of the four buttons while the scramble was on the screen. Participants’ response and reaction time for each trial were recorded. The task was presented on Experiment Builder (SR Research), separated into two functional runs (about 7 min/run) to reduce participant burden/fatigue and to account for scanner drift. A total of 202 brain images were acquired on the 3T scanner (see details below) in each run.

Face stimuli were selected from the FACES database (Ebner et al., 2010).² We chose 32 young (age range: 18–31 years) and 32 older (age range: 69–80 years) neutrally expressive faces with equal numbers of male and female faces per age group. Faces were evenly split into two sets based on their age and gender, with eight faces per age-by-gender group in each run. Among each set of 32 face images, 16 faces (four from each age-by-gender group) were selected to create the scrambled face images (Figure 1; also see text footnote 2). In each run, the presentation order of the 48 face images (32 face and 16 scrambled face images) was pseudorandomized so that no more than three faces of the same age-by-gender group or three scrambled face images repeated in a row.

Brain images were acquired on a 3T Philips Achieva MR Scanner (Philips Medical Systems, Best, The Netherlands) with a 32-channel head coil. Whole-brain high-resolution three-dimensional T1-weighted anatomical reference images were acquired using an MP-RAGE sequence (sagittal plane, FOV = 240 × 240 × 170 mm; 1 mm³ isotropic voxels). Functional images were acquired using whole-head gradient-echo-planar imaging and single-shot gradient echo (38 interleaved slices, TR = 2 s, TE = 30 ms, FOV = 252 × 252 × 133 mm, flip angle = 90°, in-plane resolution = 80 × 80, no skip).

We applied standard preprocessing procedures using Statistical Parametric Mapping 12 (SPM 12; Welcome Department of Imaging Neuroscience). This included segmentation and spatial normalization into MNI space for structural images and slice time correction, realignment and unwarp, coregistration with structural data, spatial normalization into MNI space, resampled voxel size of 2 mm³, and smoothing with an 8 mm Gaussian kernel for functional images.

The Mauchly’s test for face trustworthiness level was significant, indicating that the assumption of sphericity was violated. Therefore, we reported any effects involving this variable with Greenhouse-Geisser corrected degrees of freedom. The main effect of age group was not significant [F_{(1, 93)} = 1.03, p = 0.31, η_p² = 0.01], suggesting no age-related differences in overall face trustworthiness ratings. The main effect of face trustworthiness level was significant [F_(3.01, _280.25) = 70.98, p < 0.001, η_p² = 0.43], in that both young and older participants gave higher trustworthiness ratings to faces with higher levels of trustworthiness, indicating high consistency in trustworthiness ratings between the current sample and the independent norm sample (Pehlivanoglu et al., 2022). This main effect of face trustworthiness level was further qualified by a significant age moderation [F_{(3.01, 280.25)} = 3.68, p = 0.013, η_p² = 0.038]. In particular, older (M = 2.66, SE = 0.05) compared to young (M = 2.47, SE = 0.05) participants gave higher trustworthiness ratings to faces from the second level of face trustworthiness [F_{(1, 93)} = 6.46, p = 0.01, η_p² = 0.065], while the age groups did not differ in their ratings of the other four levels of trustworthiness (all ps > 0.10, Figure 2). This result partially supported Hypothesis 1 in that older participants rated somewhat untrustworthy-looking faces as relatively more trustworthy than did young participants. No other main or interaction effects were significant for face trustworthiness ratings (all ps > 0.05; see Supplementary Material for parameter estimates of all variables).

Our finding that older compared to young adults evaluated somewhat untrustworthy-looking faces as more trustworthy is in line with emerging evidence of decreased sensitivity to deceptive cues in aging (Ruffman et al., 2012; Frazier et al., 2021; see Ebner et al., 2022, for an overview). This finding also suggests that the age-related positivity effect (Reed et al., 2014) applies to face trustworthiness evaluation (see also Zebrowitz et al., 2013, 2017), in that older adults rated untrustworthy-looking faces, as faces depicting negative facial cues, as less negative/more positive than young adults did.

Previous studies have reported comparable findings (i.e., relatively more favorable ratings for untrustworthy-looking faces from older than young adults; Castle et al., 2012; Zebrowitz et al., 2017; Cassidy et al., 2019; see also Bailey et al., 2016). Our study importantly qualifies this previous work by demonstrating that higher trustworthiness ratings were only given for somewhat untrustworthy-looking faces, but not for very untrustworthy-looking faces. While both signal negative facial cues, somewhat untrustworthy-looking faces display ambiguous (negative) facial cues. Thus, it is possible that the positivity effect in aging for face trustworthiness is selective; that is, older adults may be as sensitive as young adults to negative/deceptive cues in non-ambivalent faces, but not in ambivalent faces.

In fact, supporting this interpretation, ambivalent information requires rather complex cognitive operations such as interference resolution (e.g., attending to specific aspects of a stimulus while ignoring (salient) other aspects; Stanley and Blanchard-Fields, 2008; O’Connor et al., 2019). Given age-related decline in working memory (Pehlivanoglu et al., 2014; for a meta-analysis see, Bopp and Verhaeghen, 2007) combined with age-related reduction in sensitivity to deceptive cues (Castle et al., 2012) and an increased positivity effect with age (Carstensen and Mikels, 2005), older participants may have had difficulty in focusing on cues of untrustworthiness while filtering out cues of trustworthiness in ambiguously untrustworthy-looking faces. In line with this explanation, other studies showed that older compared to young adults used fewer negative words to describe ambiguous scenarios (Juang and Knight, 2016; Mikels and Shuster, 2016) and gave more positive evaluations to ambiguous facial expressions (i.e., surprise; Neta and Tong, 2016; Neta et al., 2018).

This age-related reduced differentiation when confronted with ambiguously negative (untrustworthy) cues may be reflective of older adults’ perceptual dedifferentiation (Park et al., 2012) and/or may reflect a shift in decision criteria in trust-related decision making. That is, when cues of untrustworthiness are not distinct, older adults may not pick up on them, and perhaps particularly so for negative cues. In support of this interpretation, age-related dedifferentiation in face emotion recognition is greater/or only occurs for negative (e.g., anger, disgust) but not positive or neutral expressions (Franklin and Zebrowitz, 2017).

Consistent with previous findings in young adults (Said et al., 2009; Freeman et al., 2014; see Santos et al., 2016, for a meta-analysis), we identified a complex pattern of amygdala response to face trustworthiness in our age-heterogeneous sample, which qualifies prior work (Said et al., 2009). In particular, we observed a lateralization in amygdala response to facial trustworthiness that was further moderated by age, as discussed next.

Comparable across both age groups, right amygdala response was greater to more and less trustworthy-looking than to moderately trustworthy-looking faces (i.e., U-shaped association). In left amygdala, in contrast, we observed a negative linear association; in that left amygdala activity decreased with increasing face trustworthiness. Right amygdala has been shown to mediate unconscious processing of emotional information, while left amygdala has been associated with conscious processing of emotional stimuli (Morris et al., 1998, 1999; Allen et al., 2021). Also, a U-shaped amygdala response to face trustworthiness has been observed in implicit tasks (Mattavelli et al., 2012; Rule et al., 2013) or when stimuli were presented so briefly (e.g., 200 ms) that elaborate processing was not possible (Freeman et al., 2014). Mapping onto these previous findings, the U-shaped association between right amygdala response and face trustworthiness ratings in our study may reflect more implicit processing, while the negative linear association between left amygdala response and face trustworthiness ratings may reflect more explicit processing.

Further, supporting an age-differential response pattern in right amygdala, we found that young but not older participants showed a negative linear trend in addition to the U-shaped association. As a result, young but not older participants showed enhanced right amygdala activity to less untrustworthy-looking faces. As amygdala response toward face trustworthiness may reflect the level of social salience (Sander et al., 2003; Vuilleumier, 2009), which further signal approach/avoidance behavior (Todorov, 2008; Todorov and Engell, 2008), age-differential linear response in right amygdala could be interpreted as suggesting greater salience of untrustworthy-looking faces among young but not older adults (or at least, that salience of untrustworthy-looking faces is reduced among older relative to young adults). Thus, our amygdala findings are in accord with brain data on the positivity effect suggesting reduced response to negative but not positive information (Mather et al., 2004; see also Reed and Carstensen, 2012, for an overview).

Visual inspection of the two clusters identified in left and right amygdala suggested their location in the lateral nuclei, known as the gateway of amygdala, that receive perceptual input (LeDoux, 2007). Stronger magnetic field imaging such as with a 7T scanner in future studies will allow enhanced localization (Vu et al., 2017) and facilitate interpretation of the processes underlying face trustworthiness evaluation in young and older adults.

In this study we found no evidence of a moderation of face trustworthiness evaluation by intranasal oxytocin, neither in behavior nor in brain response. Oxytocin has been promoted as crucial modulator of social cognition and behavior (Harari-Dahan and Bernstein, 2014; Quintana et al., 2018), including in aging (Horta et al., 2020), and has been discussed in the context of trustworthiness perception (Van IJzendoorn and Bakermans-Kranenburg, 2012). However, some empirical work does not support such effects (Lambert et al., 2014; Grainger et al., 2018; see also Declerck et al., 2020). The present study findings align with studies that do not support an effect of a single dose of intranasal oxytocin on face trustworthiness evaluation and extend this null effect to older adults (see Grainger et al., 2019).

There are various possible explanations for the observed non-significant effects of oxytocin on face trustworthiness in brain and behavior. The present study used faces with neutral expressions, which may have reduced emotional reaction/approach-avoidance tendencies to the images. Some of the previous studies that found social-cognitive effects of oxytocin used strong social (e.g., in-group vs. out-group; De Dreu et al., 2010) and emotional (e.g., emotional expression; Horta et al., 2019) manipulations in their experimental designs. It is possible that stimuli with affluent social and emotional content facilitate and/or enhance oxytocin’s social-cognitive effects.

Also, the majority of faces used in the present study fell somewhere in the middle of the trustworthiness spectrum, as they were retrieved from a naturalistic set of images taken from real people (see details about the FACES database in Ebner et al., 2010). The present study’s 4-point rating scale with two extreme, opposite ends (i.e., very untrustworthy to very trustworthy) may have had limited sensitivity to differentiate along trustworthiness among this set of naturalistic images and thus also limited our ability to model the BOLD signal along this dimension. For example, with a comparable sample size of 96 participants who underwent a single dose of either intranasal oxytocin or placebo and using neutral faces as stimuli but a 7-point trustworthiness rating scale, Theodoridou et al. (2009) found higher face trustworthiness ratings in the oxytocin than the placebo group. Future studies will be able to test if such methodological differences underlie variability in detecting oxytocin effects on face trustworthiness evaluation.

Moving forward, studies with a larger sample size are warranted to follow up on the oxytocin null effects observed in the present study. In fact, a recent meta-analysis (Leppanen et al., 2017), which was published after completion of the present data collection and suggested that for a study with oxytocin vs. placebo administration as between-subject factor a sample size of 64 participants per treatment group is necessary to obtain sufficient power (80%) to detect a moderate effect of oxytocin on social cognition (see Supplementary Material for details regarding the interaction we observed between treatment group and face trustworthiness level, which were not statistical significant as per our alpha level cut off, but approached significance). While randomized for treatment group assignment, the present study used a between-subject design and all measures were collected after treatment administration. Future extension of this work would benefit from a cross-over design to account for possible individual differences at baseline. In fact, following recent recommendations (Naselaris et al., 2021), within-subject designs are better suited as their higher sampling of a more limited number of subjects compared to the smaller sampling of a larger number of individuals in a between-subject design are more powerful to detect significant effects. This emphasis on depth of assessment in individual brains over breadth of assessment among multiple brains is also beneficial for enhancing resolution and the signal-to-noise ratio in neuroimaging analysis, and thus a fruitful path forward.