Edited by: Marios G. Philiastides, University of Glasgow, United Kingdom
Reviewed by: Elsa Fouragnan, Plymouth University, United Kingdom; Christian Keitel, University of Glasgow, United Kingdom
This article was submitted to Decision Neuroscience, a section of the journal Frontiers in Neuroscience
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 functional role of high beta oscillations (20–35 Hz) during feedback processing has been suggested to reflect unexpected gains. Using a novel gambling task that separates gains and losses across blocks and directly compares reception of monetary rewards to a ‘no-reward/punishment’ condition with equal probability we aimed to further investigate the role of beta oscillations. When contrasting different feedback conditions across rewards, we found that a late low beta component (12–20 Hz) had increased in power during the omission of rewards relative to the reception of rewards, while no differences were observed during the loss domain. These findings may indicate that late low beta oscillations in the context of feedback processing may respond to omission of gains relative to other potential outcomes. We speculate that late low beta oscillations may operate as a learning mechanism that signals the brain to make future adequate decisions. Overall, our study provides new insights for the role of late low beta oscillations in reward processing.
Effective decision-making crucially relies on the ability to improve decisions based on the evaluation of feedback. A learning (e.g., prediction-error) signal is computed after observing the outcome generated by each choice which are used to improve the quality of future decisions (
A further exploration in a follow up experiment demonstrated that beta oscillations tend to respond to rare rewarding events. When comparing cued gain and loss incentives with high or low probability, beta oscillations were stronger in power when the probability of cued rewards had a low relative to a high probability (
Perhaps one possible drawback to the above-mentioned studies is that monetary gains are typically compared directly with monetary losses, neither of which may serve as an adequate baseline (
For this study we aimed to investigate the functional role of beta oscillations by comparing gains and losses separately with a “no gain” and a “no loss” feedback condition. To achieve this goal, we employed a novel risky decision-making task that allows one to compare the reception and omission of monetary incentives separately for gains and losses (
The purpose of this task design was to further explore the functional role of beta oscillations using a risky decision-making paradigm with gains and losses portrayed across blocks. Rather than manipulating probability between trials, this paradigm will allowed us to compare feedback associated with uncertain (risk) outcomes relative to feedback produced from certain (safe) outcomes separately across gains and losses. Based on the prior study that shown how frontal beta oscillations respond to reward valence but not probability (
We aimed to compare feedback display representing the reception and omission of gains and losses with an uncertain outcome (±50 monetary units [MU] with a probability of 50%), relative to a certain outcome (±25 MU with a probability of 100%). Moreover, using trial by trial analysis we aim to test whether power of beta oscillations may predict risky decision making. Together our analyses will allow us to explore whether beta oscillations are responsive to gain or loss events and to assess whether this signal may facilitate decisions in upcoming trials.
Twenty-five healthy participants (23 right-handed; 18 females; mean age 21.61; age range 18–34 years; and
We used a novel risky decision-making task – “rewarded voluntary switch task” (see
Switch-risk task. Risky decision making depends on voluntary switching and repeating task-sets. Safe decisions yield 25 MU with a probability of 100% whereas risky decisions yield 50 MU or 0 MU with a probability of 50%. Figure represents trial in the “Switch = Risk” reward block.
Each trial begins with a centered fixation cross displayed between 500 and 1000 ms followed by a screen containing a single digit (1, 2, 3, 4, 6, 7, 8, or 9). To select between
Gains and losses are separated across blocks. During gain blocks, participants are instructed that safe decisions are defined as “100% probability that you would receive 25 monetary units (MU),” while risky decisions were defined as “50% probability that you would receive 50 MU” (or alternatively 0 MU). In loss blocks, the safe decision are defined as “100% probability that you would lose 25 MU” while risky decisions are defined as “50% probability that you would lose 50 MU” (alternatively 0 MU).
For each response a feedback screen displayed for 2000 ms indicated the amount of MUs rewarded or lost for that particular trial. Positive feedback in gain blocks was 50 and 0 MU for loss blocks. Negative feedback in gain blocks was 0 MU and -50 MU for loss blocks. Neutral feedback was 25 and -25 MU for gain and loss blocks, respectively. For risky choices, a random generator displayed positive or negative feedback such that the distribution of feedback type was not fixed but randomly assigned. If response time exceeded 4000 ms or participants responded erroneously participants viewed negative feedback (e.g., 0 MU for gain block, -50 MU for loss blocks).
The experiment was programmed using E-Prime 2.0 software. Stimuli were centered on the screen and remained on the screen until a response was made. The text was displayed in black font on a gray scale background and all participants were instructed to use both hands. Participants received two rounds of training, which consisted of eight blocks of 10 trials, resulting in 80 trials in total. If accuracy was below 95% additional training sessions were provided. This learning phase was reflected in the actual experiment; accuracy for all except one participant (86%) was above 92%.
Initially, nine participants received 16 blocks of 30 trials (480 trials total). Due to the notion that probability of feedback did not vary between trials, choices varied between subjects, and that a substantial amount of trials were removed from the analysis, a potential problem with the analysis may be due to too few trials. Therefore, the number of trials per block was increased to 40 trials (640 trials total) for the remaining sixteen participants. After performing the task, participants were shown the total cumulative feedback on the computer screen. Participants received 500 MU for participation (500 MU ≈ 7 USD) and an additional bonus, between -300 and + 300 MU, based on the feedback outcomes of six randomly selected trials to maintain an equal motivation for risky decision making across blocks (see
The EEG data were recorded with BrainAmp amplifiers and BrainVision Recorder software (Brain Products GmbH, Munich, Germany) using silver ActiCap active scalp electrodes mounted in an elastic cap located at 60 standard positions according to the international 10–20 system. The electrophysiological signals were filtered online using a sampling rate of 500 Hz in the frequency range 0.2–100 kHz. Impedances were kept <10 kΩ. Electrooculogram were recorded with electrodes placed at both lateral canthi and below the left eye. EEG signals were referenced to the mean of the activity at the two mastoid processes.
Data preprocessing of the EEG data was performed using BrainVision Analyzer 2.0. First, signals in bad channels were replaced using nearest-neighbor interpolation. Second, a bandpass filter (0.1–40 Hz) was applied to the data, after which eye-blink- and eye-movement-related activity was suppressed in the data using independent component analysis. Finally, intervals containing non-systematic artifacts produced by electromyographic activity, skin potentials and other sources were manually rejected from the data. Across subjects, 10.1% (σ = 0.090) of trials were excluded from the analysis. For the first group, the mean number of trials excluded from the analysis was 15.6% (σ = 0.123); for the second group 8.0% (σ = 0.061) of trials were excluded. The mean number of valid trials included in EEG analysis across each condition for all subjects was 80.7 (range: 66–101 trials). The range of trials removed from each group was: 0–40 (from 480 trials) and 0–63 (from 640 trials).
EEG analysis for each
Single trial time-frequency analysis was performed on a time window between -1000 and 2000 ms for each condition. For each trial, the segmented EEG data was convolved with a complex Morlet wavelet (from 1 to 40 Hz, linear increase). The frequency and time resolution of were set at the default settings (temporal resolution of 3 s at frequency 1 Hz) in Brainstorm, which uniquely define the temporal and spectral resolution of the wavelet for all other frequencies (
Response times of risky and safe decisions were analyzed across gains and losses using a repeated measures analysis of variance (ANOVA) and a Bonferroni correction procedure. To determine whether the percentage of selected risky gambles was above or below chance level (μ = 50%), a one sample
Mean beta power (12–20 Hz) was calculated for FCz, FC1, FC2, Cz, C1, C2, CPz, CP1, and CP2 electrode positions for each feedback and valence condition within the 700–1000 ms post-response time window and entered into a repeated measures ANOVA test. Greenhouse-Geisser correction was applied.
The selection of the specific frequency band, latency interval, and electrode positions was purely data-driven and based on statistical analysis of the ERSP data averaged over the experimental conditions (the analysis was orthogonal to our main analysis). The ERSP data was tested against zero using permutational statistics on t-score maps transformed with the TFCE (threshold-free cluster enhancement) algorithm (see
To assess whether the spectral power density of beta frequency influenced risky decisions in the following trial, we included several generalized linear models (GLMs) with a logit link function, performed separately for gains and losses. Spectral power density is characterized by the distribution of power for each frequency range within a specified time series (
Generalized Logistic Model (GLM) predicting risk decision making in the following trial for rewards
β | |||||
---|---|---|---|---|---|
Theta PSD | -0.155 | 0.058 | -2.672 | 0.007 | 0.063 |
Beta PSD | 0.094 | 0.046 | 2.025 | 0.042 | 0.378 |
Theta∗Beta PSD | 0.019 | 0.025 | 0.756 | 0.449 | >0.999 |
Theta PSD∗Fb (+50) | 0.112 | 0.069 | 1.616 | 0.106 | 0.954 |
Theta PSD∗Fb (+0) | 0.008 | 0.069 | 0.124 | 0.901 | >0.999 |
- |
- |
||||
Beta PSD∗Fb (+0) | -0.130 | 0.074 | -1.751 | 0.079 | 0.711 |
Theta PSD | -0.050 | 0.056 | -0.891 | 0.372 | >0.999 |
Beta PSD | 0.061 | 0.052 | 1.184 | 0.236 | >0.999 |
Theta∗Beta PSD | 0.024 | 0.025 | 0.977 | 0.328 | >0.999 |
Theta PSD∗Fb (-0) | 0.123 | 0.074 | 1.654 | 0.098 | 0.882 |
Theta PSD∗Fb (-50) | -0.079 | 0.078 | -1.020 | 0.307 | >0.999 |
Beta PSD∗Fb (-0) | -0.197 | 0.080 | -2.453 | 0.014 | 0.126 |
Beta PSD∗Fb (-50) | -0.116 | 0.072 | -1.601 | 0.109 | 0.981 |
Generalized Logistic Model (GLM) predicting risk decision making in the following trial for rewards
β | |||||
---|---|---|---|---|---|
Theta PSD | -0.146 | 0.061 | -2.365 | 0.018 | 0.162 |
Beta PSD | -0.035 | 0.067 | -0.519 | 0.603 | >0.999 |
Fb (+50) | -0.025 | 0.071 | -0.357 | 0.721 | >0.999 |
- |
- |
||||
Theta∗Beta PSD | 0.019 | 0.025 | 0.756 | 0.449 | >0.999 |
Theta PSD∗Fb (+50) | 0.103 | 0.072 | 1.426 | 0.154 | >0.999 |
Theta PSD∗Fb (+25) | -0.008 | 0.069 | -0.124 | 0.901 | >0.999 |
- |
- |
||||
Beta PSD∗Fb (+25) | 0.130 | 0.074 | 1.751 | 0.079 | 0.711 |
Theta PSD | -0.130 | 0.065 | -1.979 | 0.047 | 0.423 |
Beta PSD | -0.054 | 0.070 | -0.776 | 0.437 | >0.999 |
Fb (-0) | 0.111 | 0.071 | 1.555 | 0.120 | >0.999 |
- |
- |
||||
Theta∗Beta PSD | 0.024 | 0.025 | 0.977 | 0.328 | >0.999 |
Theta PSD∗Fb (-0) | 0.203 | 0.083 | 2.439 | 0.014 | 0.126 |
Theta PSD∗Fb (-25) | 0.079 | 0.078 | 1.020 | 0.307 | >0.999 |
Beta PSD∗Fb (-0) | -0.081 | 0.089 | -0.915 | 0.360 | >0.999 |
Beta PSD∗Fb (-25) | 0.116 | 0.072 | 1.601 | 0.109 | 0.981 |
For the beta frequency component, source localization for each feedback condition across gains and losses were performed on single trials between 12 and 20 Hz between the 700–1000 ms time window. A default anatomy of the standard MNI brain was used to compute a head model using OpenMEEG software (
Participants performed the task correctly: mean accuracy was 96.7% (σ = 0.029). Overall subjects preferred risky decisions (58.2%, σ = 0.121) more often than safe decisions (
Boxplots representing
Unexpectedly, a late low beta (12–20 Hz) frequency component during the feedback display between 700 and 1000 ms was shown. Individual scores for beta power are shown in Figure
Time-frequency power (total) across negative (+0 MU), neutral (+25 MU) and positive (+50 MU) feedback for gain blocks.
Time-frequency power (total) across negative (–0 MU), neutral (–25 MU) and positive (–50 MU) feedback for loss blocks.
The repeated measures ANOVA also revealed a statistical significant moderate effect of
To further support the claim that beta power oscillations were specific to the omission of gains, we tested the differences in beta frequency power across gains and losses producing no monetary value (i.e., +0 MU for gains versus -0 MU for losses). This contrast allowed us to deduce whether beta oscillations were sensitive to gain omission, and not necessarily to the monetary value. A direct comparison between negative feedback during gain blocks (+0 MU) and positive feedback during loss blocks (-0 MU) relieved a two-way interaction effect between
We next determined whether there were differences between negative feedback across gain and loss contexts (i.e., +0 MU for gains versus -50 MU for losses). No differences were observed for this contrast. Finally, we compared neutral feedback across domains (i.e., +25 MU for gains versus -25 MU for losses). This contrast also revealed no significant effects.
For all feedback conditions beta oscillations were localized to the right frontal cortex, left parietal cortex, and medial frontal structures, possibly overlapping with the medial frontal cortex and the striatum. These source estimations seem to correspond with prior lesion studies (
Since beta oscillations were specific to gain blocks, corresponding to previous studies showing an increase in beta power during gains compared to losses (
First of all, our results reveal main effects of positive and negative feedback for both GLMs reflecting gain and loss blocks. Within the gain blocks, positive compared to neutral feedback (β = 0.271;
Furthermore, GLMs representing gain blocks (Tables
In summary, power of beta oscillations increased during gain omission and predicted a decrease in risky taking in next trials within the gain blocks. Taken together, beta oscillations may signify a reward learning mechanism which modulates future decisions. Perhaps this learning mechanism plays a specific role in risky decision making in the context of uncertain gains.
Previous studies have revealed a mid-frontal beta oscillatory activity between 20 and 35 Hz elicited by gain compared to loss outcomes (
We recorded EEG while participants performed a task that yielded reception and omission of monetary incentives separately for gains and losses. Rather than demonstrating a high beta component, the results demonstrated a significant moderate effect of late low beta band (12–20 Hz) for negative feedback in the gain context, but not for the loss context. Specifically, when participants selected risky gambles a significant increase in beta power during the omission of gains (negative feedback) compared to the reception of gains (positive feedback) was found. This increase in beta power during the omission of gains was also significant when compared to reception of gains after selecting the safe option (neutral feedback).
Additional analysis was performed to test whether beta oscillations during the omission of gains differed from the omission of losses; i.e., we compared neural responses to omission of gain (+0 MU) and loss blocks (-0 MU; see
Supplementary to the time-frequency analysis, we calculated source estimates of the low late beta component. Beta oscillations across all conditions were localized within the right prefrontal cortex, medial frontal cortex, left parietal cortex, and the striatum, corresponding to previous studies (
An important distinction between the current results and prior studies relate to the spectral and temporal counterparts of beta oscillations. In the current study, beta oscillations were relatively low in frequency (12–20 Hz) and late in time (700–1000 ms) compared to previous studies (
Although no explanation has been provided to explain the functional role of this late beta frequency component, others have offered the possibility that multiple beta frequency components may co-occur during feedback processing (
Secondly, the source localization of the current study showing activity within the right lateralized frontal area corresponds with high-beta oscillations in an earlier study (
Importantly, probability, expected value, and magnitude of outcomes remained constant throughout the entire experiment and thus cannot account for the increase of beta oscillatory activity observed during the omissions of gains. This finding corresponds with the previous study demonstrating no change in beta power synchrony under manipulations of probability and expected value (see
To explore the functional role of beta oscillations on risky decision making, we also investigated whether beta power density on each trial would predict the tendency to select risky decisions on the following trial. The GLM predicting risky decision making in the following trial demonstrated an interaction effect between beta oscillatory power and positive feedback, yet specifically for gain blocks. The relationship between the interaction (beta PSD × positive feedback) and risky decision making was negative (i.e., β = -0.390), suggesting that during positive feedback a
To interpret this result, we propose a reward learning mechanism marked by changes in beta oscillations between trials. When receiving positive feedback, an increase in beta power reinforces the decision maker to continue to select risky gambles. However, during the absence of gains, a violation of rewards occurs in which the gain omission relative to alternative prospective outcomes results to an increase in beta oscillatory power as the result of perceiving gain omission as a “loss” (see
Perhaps this proposed reward learning mechanism may also explain the observed results in a prior experiment in which induced 20 Hz transcranial electric current stimulation increased risky decision making (
In the current study we showed that late low beta oscillations between 12 and 20 Hz are functional sensitive to gain omission relative to other potential gains. Furthermore, beta oscillations elicited by positive feedback in the gain domain were negatively associated with risky decision making in the following trial. From these two novel findings, we propose a reward learning mechanism by which the power of beta oscillations manifested by outcome violation, motivates responders to change subsequent choices as a means to compensate for reward omission on the current trial. We further contend that due to the novelty of this finding, further work is necessary to determine whether late low beta oscillations reflect a similar or alternative feedback-related beta component reported in the high beta range.
ZY, AS, and VK designed the research and wrote the paper. ZY performed the research. ZY, MM-S, NN, and DA analyzed the data.
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 CK and handling Editor declared their shared affiliation.
The Supplementary Material for this article can be found online at: