Abstract
Risk is a fundamental factor affecting individual and social economic decisions, but its neural correlates are largely unexplored in the social domain. The amygdala, together with the dorsal anterior cingulate cortex (dACC), is thought to play a central role in risk-taking. Here, we investigated in human volunteers (n = 20; 11 females) how risk (defined as the variance of reward probability distributions) in a social situation affects decisions and concomitant neural activity as measured with fMRI. We found separate variance-risk signals for social and nonsocial outcomes in the amygdala. Specifically, amygdala activity increased parametrically with social reward variance of presented choice options and on separate trials with nonsocial reward variance. Behaviorally, 75% of participants were averse to social risk as estimated in a Becker–DeGroot–Marschak auction-like procedure. The stronger this aversion, the more negative the coupling between risk-related amygdala regions and dACC. This negative relation was significant for social risk attitude but not for the attitude toward variance-risk in juice outcomes. Our results indicate that the amygdala and its coupling with dACC process objective and subjectively evaluated social risk. Moreover, while social risk can be captured with a framework originally established by finance theory for nonsocial risk, the amygdala appears to process social risk largely separately from nonsocial risk.
Significance Statement
Risk—the uncertainty of outcomes—has profound effects on decision-making and behavior. In social situations, uncertainty about others' behavior (“social risk”) similarly guides our decisions and can contribute to social anxiety. Surprisingly little is known about the neural mechanisms processing social risk. Here, we investigated neural activity when humans evaluated social risk derived from the uncertainty of personal compliments received from social partners. Activity in the amygdala, a brain structure implicated in emotion and social behavior, reflected social risk levels. Amygdala functional connectivity to the anterior cingulate cortex reflected individuals' aversion to social risk. Our findings extend risk evaluation into the social domain and pave the way for investigating social risk attitudes in mental health impairments, including social anxiety.
Introduction
Processing risk—the uncertainty associated with outcomes—is crucial for survival and adaptive behavior (Caraco et al., 1980). It allows for an informed prediction of the likely distribution of future outcomes (Knight, 1921). An increase in the spread around the mean of outcomes can serve as an objective measure of an increase in risk (Rothschild and Stiglitz, 1970), operationalized for example as an increase in the variance of outcomes. Decision-makers are often not indifferent to increases in objective risk but subjectively evaluate them, e.g., in the form of prospect (Kahneman and Tversky, 1979) and utility (Samuelson, 1937; Houthakker, 1950; Stauffer et al., 2014). The addition of risk reduces the subjective value of a choice option for a risk averse individual but increases it for a risk seeker, and these preference differences express themselves in value signals in the brain (Tobler et al., 2009; Van Duijvenvoorde et al., 2015).
Many decisions occur in a social context (Crespi, 2001; Bshary et al., 2014; Chen and Hong, 2018) and the feedback provided by others is valuable to decision-makers as it can be a source of acceptance and status (Izuma et al., 2008; Zink et al., 2008). Just like nonsocial rewards, social rewards are distributed and in principle can be captured with summary statistics such as the mean and variance of the distribution. Accordingly, decision-makers may seek or avoid variance-risk in the social domain as such, even without observing the risky choices of others (Chung et al., 2015; Suzuki et al., 2016). Previous research has associated the amygdala and amygdala–dorsal anterior cingulate cortex (dACC) interactions with processing nonsocial decision variables, including risk (De Martino et al., 2010; Li et al., 2011; Zeeb and Winstanley, 2011; Grabenhorst et al., 2012, 2016; Jung et al., 2013; Orsini et al., 2015, 2017; Chen and Stuphorn, 2018; Feldmanhall et al., 2018; Aydogan et al., 2021). The amygdala processes also social decision variables, such as predictions of social partner's choices (Grabenhorst et al., 2019), social value (Chang et al., 2015; Schultz et al., 2019), betrayal aversion (Lauharatanahirun et al., 2012), and information about social stimuli and social hierarchy (Gothard et al., 2007; Rutishauser et al., 2013; Munuera et al., 2018). Moreover, it is more active when decision-makers align themselves with others in risky decision situations (Burke et al., 2010b). However, the precise role of the amygdala in processing objective or subjectively evaluated social variance-risk remains unclear.
To close this gap, we designed a novel risky choice task in which we first measured the individual willingness to pay (WTP) for options with different social and nonsocial variance-risk as well as different expected values (EV). We used photos together with social descriptions addressed to the participant as social rewards and juice varying in quality and quantity as nonsocial rewards, because these two reward types are, or can be made, similar on dimensions such as primacy and duration (Matyjek et al., 2020). For each participant, risk, and EV level, we estimated individual levels of risk proneness with WTP and their relation to choice. Based on the two separate lines of research described above on amygdala involvement in nonsocial risk processing and in general social functions, we hypothesized that objective social variance-risk is encoded in the amygdala. Second, we tested whether functional connectivity between the amygdala and dACC, suggested by anatomical connections (Carmichael and Price, 1995; Beckmann et al., 2009) and functional similarities (Kable and Glimcher, 2007; McClure et al., 2007; Hare et al., 2011; Kolling et al., 2012), correlated with risk attitude.
Materials and Methods
Participants
After a behavioral session (Day 1, see below), 24 healthy volunteers (age range, 19–29; 14 females) completed the functional magnetic resonance imaging (fMRI) session (Day 2) of this study. Inclusion criteria were normal or corrected-to-normal vision, general liking of dairy products, and normal appetite. Exclusion criteria were lactose intolerance, active avoidance of sugar or fat in the diet, metal implants in the body, being on medication other than contraceptives, psychiatric illness, and pregnancy. Four participants were not analyzed further because they showed excessive head motion or because of technical problems during scanning. Accordingly, we present data from 20 participants (11 females). The study was approved by the Local Research Ethics Committee of the Cambridgeshire Health Authority, and written informed consent was obtained from all participants before the experiment.
Tasks
Risky rewards
As social rewards, we used pictures of human faces and the upper half of the upper body with a neutral to positive expression paired with a compliment (Fig. 1a). Each face and compliment combination was unique, and unrepeated, ensuring that every option element appeared only once. The compliments themselves were carefully designed to vary in perceived value, introducing significant variation in the social rewards. Examples of these compliments include “I think you are terrific,” “Way to go!”, “I am utterly impressed!”, “Your ideas are very clever!”, and “Good job.” In our study, social risk consisted of the equiprobable possibility of receiving a larger or a smaller social reward. With higher risk, the magnitude difference between the two rewards was larger than with lower risk. Thus, we used the formal definition of objective risk as the mathematical variance of a known probability distribution (Rothschild and Stiglitz, 1970), as used previously for neural applications (Tobler et al., 2007, 2009). The subjective values of each social reward were determined in an independent auction task on Day 1. Nonsocial rewards consisted of three differently flavored commercially available smoothies. Participants were asked to not eat or drink for 3 h before the experiment, following approaches in previous studies (Grabenhorst et al., 2010; Zangemeister et al., 2016; Kim et al., 2024). Our goal in designing these stimuli was to achieve meaningful variation in both value and risk. We found in pretests that this was possible by using a relatively small number of different juices offered in different quantities as has been done in previous studies (Grabenhorst et al., 2010; Zangemeister et al., 2016). For the social stimuli, we found that it was possible to achieve sufficient variation in value and risk by using unique faces paired with unique compliments.
Tasks and stimuli. a, Tasks. On Day 1, all three Becker–De Groot–Marschak (BDM) auction tasks followed the same basic procedure. In each trial, a risk-free (Auctions 1 and 2) or a risky option (Auction 3) was presented for 3 s. Subsequently, participants had up to 10 s to indicate their willingness to pay on a vertical mouse-controlled scale of 0–100 pence in steps of 1 penny (1 penny = £1/100). The possible outcomes illustrated by the bars of risky options corresponded to the amounts participants were willing to pay for specific social or nonsocial rewards as determined in the first two auction tasks. After each task, one trial was chosen, and the participant bid in that trial was compared with a computer bid that was randomly drawn from a uniform distribution. If the participant bid in these trials was higher than, or equal to, the computer bid, participants paid the computer bid from their £1 endowment and received the reward at the end of the session. Conversely, if the participant bid was smaller than the computer bid, participants neither paid nor received anything. In the risky choice BDM, participants were informed of the trial type (social/juice) before being presented with the risky option. On Day 2, participants performed the fMRI task, in which they chose between two risky options on every trial, either by eye movement (saccade) or by hand movement (button press). We used different actions to keep participants engaged. Both reward type (juice or social) and action type varied across trials. The trial type information (here “social + eye”; presented for 1 s ± 200 ms) was followed by the sequential presentation of the two options (presented for 1.5 s each). Participants then provided their choice (Option 1 represented by “1” and Option 2 by “2”) with the trial-appropriate action. In the feedback phase, the chosen option remained visible for 1 s. Jittered intervals separated trials by 4–8 s and stimuli by 3–4 s. b, Stimuli. Our stimuli consisted of a rectangular box with two horizontal bars, each representing a reward. These two rewards were equiprobable (p = 0.5). The risky options in the main task were the same as those used in the third BDM auction. Risk levels R1–4 reflect different levels of variance-risk. These were crossed with three levels of expected value, EV1–3, defined as the mean of the two equiprobable rewards. Thus, overall, there were 12 different options (3 EV levels times 4 risk levels), both for the social and the nonsocial domain. Abbreviations: EV, expected value; ISI, interstimulus interval; ITI, intertrial interval.
In the auction task on Day 1 and the choice task on Day 2, we presented social and individual (i.e., nonsocial) options with a specific risk and EV. As usual, the expected value of a gamble corresponded to the sum of the probability-weighted reward amounts that could be won in that gamble (Knutson et al., 2005; Tobler et al., 2005, 2009). Because our gambles all consisted of two equiprobable rewards, the expected value of a gamble was simply the mean of the two possible reward amounts. We used four levels of risk and three levels of EV, resulting in 12 unique risk–EV combinations. Each option consisted of a rectangular box and two red horizontal lines at different heights (Fig. 1b), corresponding to different reward amounts (O’Neill and Schultz, 2010). Participants learned the social and individual reward levels associated with line height on Day 1. The same stimuli informed participants about social and individual risk in the fMRI task on Day 2. Therefore, differential brain activity to social and individual risk cannot be due to visual differences in the stimuli.
Becker–DeGroot–Marschak auction-like task (Day 1)
On Day 1, participants evaluated social and nonsocial (juice) rewards in a social and a juice valuation task. Specifically, they used a computer mouse to provide a bid on how much money they were willing to pay in order to obtain a given social or juice reward. Before each reward task, participants received an endowment of £1. Each trial consisted of a reward presentation phase of 3 s followed by a bidding phase of up to 10 s. Rewards were kept constant across participants. Juice rewards were cued by a rectangle in one of three different colors, indicating the type of juice (Matyjek et al., 2020), and an amount, indicated in milliliters. These amounts ranged from 15 to 330 ml in steps of 15 ml.
For the valuation of social rewards, each face and each compliment appeared only once. For each reward type, there were 66 unique rewards, yielding 66 unique reward–willingness-to-pay combinations. Following a standard Becker–DeGroot–Marschak (BDM) auction (Becker et al., 1964), the actual price of the reward in one randomly selected trial was determined by a uniformly distributed random draw from 0 to 100 pence (1 penny = £1/100). If the bid of the participant was larger than, or equal to the actual price, the participant received the reward and paid the actual price. In contrast, if the bid of the participant was smaller than the price, the participant did not receive the reward but did not pay anything either. One trial of each auction task was implemented at the end of the session. Participants first completed the BDM auction for social rewards, where they viewed actual images of faces paired with compliments. This was followed by a BDM auction for juice rewards, serving as a nonsocial comparison.
Next, to measure the subjective value of variance-risk, participants provided bids to play various risky options with two equiprobable rewards in a third BDM auction task. Before each trial, participants were notified about the reward type (social or juice) they were bidding for in that trial. Participants evaluated all twelve options, both for social and juice reward. Again, the computer randomly selected one trial at the end of the task, and this trial was implemented according to the randomly determined actual price and the bid of the participant. Finally, participants practiced the main task to be used on Day 2 (see next section).
fMRI task: risky binary choice (Day 2)
On Day 2, participants first reacquainted themselves with the main task inside the MRI scanner. Participants performed three blocks of 48 trials (17.73 ± 0.34 min per block, mean ± SEM). In every trial of this task, participants decided between two risky options (Fig. 1a). These options were the same as those used in the risky gamble BDM task on Day 1. Each trial started with information (1 ± 0.2 s) about its reward (social or juice) and required action type (eye or hand movement). Subjects then saw two consecutively presented risky options (1.5 s each). The two options were separated by a jittered interstimulus interval (ISI; fixation cross, 3–4 s). All choices were between a low-risk (Fig. 1b, Level 1) and a higher-risk (Levels 2–4) option, presented in randomized order, with an equal number of trials in which the low-risk option appeared first or second. The EV of the low-risk option varied from trial to trial, which prevented participants from fully predicting the properties of the second option even if it was low risk and thereby precluded decision-making already during the presentation of the first option. The sequential presentation of options allowed us to look for neural correlates of single-option risk and EV. After viewing the options, participants chose either the option they saw first (“1”) or the option they saw second (“2”). The “1” and the “2” were randomly presented to the left and right of a fixation cross, preventing movement preparation already at the time of the second option. The placeholder of the chosen option (“1” or “2”) remained on the screen for 1 s after the choice. Thus, participants did not receive rewards immediately or see faces but made choices between options that differed in risk and reward denoted by two lines for each option (as in pretest Task 4 on Day 1). Social and nonsocial options thereby were visually as similar as possible. Trials were separated by jittered intertrial intervals (ITIs) of 4–8 s. Both ISIs and ITIs were jittered according to a Poisson distribution to increase design efficiency.
To keep participants engaged, we asked them to provide their choices either by button press on the button box or by saccade. Using the word “eye” or “hand,” participants were informed of the required action type at the beginning of each trial, together with the juice/social information. There was a fixation requirement in all trials such that if participants made a saccade in a hand trial or a button press in an eye trial, the trial was counted as an error and repeated. Before the MRI task, we informed participants that one randomly selected decision per reward category would be played out with the risk and EV of the selected option. Participants received the rewards they won after the MRI session in the form of a glass of chosen juice reward and a printed card with the social reward.
Neuroimaging
Data acquisition
Participants were scanned with a Siemens 3 T Trio Scanner at the Cognition and Brain Sciences Unit, Cambridge, United Kingdom. We acquired between 310 and 420 volumes of T2*-weighted echo-planar images (EPIs) in three runs. For each volume, we acquired 56 slices in ascending order and with the following parameters: in-plane resolution, 3 × 3 mm; slice thickness, 2 mm; repetition time (TR), 3 s; echo time (TE), 30 ms; flip angle, 90°; slice gap, 0.5 mm; FOV, 192 × 192 × 140 mm3; matrix, 64 × 64. Four dummy volumes before each scanning run served to achieve steady-state magnetization. We acquired a high-resolution T1 structural scans using an MPRAGE sequence: 192 slices; slice thickness, 1 mm; no gap; in-plane resolution, 1 × 1 mm2; FOV, 256 × 256 × 192 mm3; TR, 2.3 s; TE, 2.98 ms; inversion time, 900 ms; flip angle, 9°.
Image preprocessing
We used Statistical Parametric Mapping (SPM12) to preprocess the fMRI data, including slice time correction and motion correction. All preprocessed fMRI data were aligned to the high-resolution anatomical image. After segmentation of the anatomical images into the gray and white matter, they and the coregistered EPIs were DARTEL-normalized to MNI space. Finally, the EPIs were spatially smoothed using a three-dimensional Gaussian filter (6 mm full-width at half-maximum).
Behavioral data analysis
BDM auction tasks
To assess the influence of objective variance-risk, expected value, and their interaction on individual willingness to pay (i.e., subjective value), we performed the following multiple linear regression for each participant, separately for the social and nonsocial domain:
Risky binary choice task
To assess the influence of the willingness to pay (derived from the separate risky option BDM auction task; WTP) for each option, we performed the following logistic regression:
Risk proneness and EV learning across tasks
To quantify individual proneness to take social or nonsocial risk, we used the regression coefficients
Neuroimaging data analysis
To detect neural activity related to variance-risk, EV, and WTP, we estimated two general linear models (GLM) in SPM12 at the participant (first) level and then took individual contrast images to the group (second) level.
Participant level analysis (first level)
In the first GLM (GLM1), for each participant and block of trials, we modeled five main time periods within each trial: (1) information screen onset (juice/social reward, eye/hand action); (2) social Option 1 onset; (3) social Option 2 onset; (4) nonsocial Option 1 onset; (5) nonsocial Option 2 onset; (6) time of decision; and (7) feedback screen onset. Periods 2–5 each had two parametric modulators, one for social or nonsocial variance-risk and the other for social or nonsocial EV. To ensure that the regressors competed for explaining independent components of variance, serial orthogonalization of parametric regressors was turned off (Mumford et al., 2015). We modeled each period as an event with a duration of zero. Each of these regressors was convolved with the canonical basis functions. Furthermore, we added the six deconvolved motion regressors and a regressor for each block to account for potential confounds. Simple contrasts on parametric modulators for social or nonsocial risk (averaged across Options 1 and 2) served to construct individual contrast images.
In the second GLM (GLM2), we modeled both social and nonsocial as one trial type and included a parametric modulator, reflecting WTP as determined via a BDM auction-like task. We conducted again multiple linear regression for each participant, analyzing the domain-general WTP coding. Similar to GLM1 above, we modeled each period as an event (duration of zero), convolving all regressors with canonical basis functions. Motion regressors and a block regressor were added to address potential confounds. Individual contrast images were constructed using simple contrasts on parametric modulators for WTP.
Group analysis (second level)
To identify regions processing social or nonsocial variance-risk at the group level, we used flexible factorial designs that included the contrast images on the parametric modulators from the individual GLM. We report whole-brain results (p < 0.05, peak-level FWE-corrected) as well as activations in the amygdala, our a priori region of interest (p < 0.05, peak-level FWE-corrected). To define the region of interest, we used the third version of the automated anatomical labeling atlas (AAL3; Rolls et al., 2020). To extract activity from amygdala regions we used the right amygdala from AAL3.
Functional connectivity between the amygdala and medial prefrontal cortex
To assess risk level-dependent functional interactions between the amygdala and medial prefrontal cortex during risky decision-making, we performed a psychophysiological interaction (PPI) analysis (Friston et al., 1997; Gitelman et al., 2003). For each participant, we first extracted the eigenvariate time series from the activations related to social and juice variance-risk in the right amygdala as the seed region. The signals were deconvolved to construct a time series capturing neural activity in the seed area, which served as the physiological regressor for the PPI analyses. As psychological regressors, we used (1) social variance-risk and (2) juice variance-risk. The interaction regressors corresponded to the multiplication of the physiological and psychological regressors. Again, we added the six motion regressors and block regressors of no interest. Next, we correlated the social or juice risk-related interaction contrast images with the individual proneness for social or juice risk (
Results
To investigate whether and how the amygdala, in possible interaction with cortical regions, processes social variance-risk, we asked participants to choose between two sequentially presented options with different variance-risk and expected values. We determined the subjective value of the available social and nonsocial rewards, as well as the willingness of participants to take social and nonsocial variance-risk, in separate bidding tasks.
Behavioral results
Our bidding tasks (Becker–DeGroot–Marschak auctions) indicated that our participants were willing to pay for social (one-sample t test, p < 0.001 for all EV and risk levels) or juice (p < 0.001 for all EV and risk levels) rewards, indicating that despite their subtle nature, both types of rewards are revealed preferred rewarding (only 2.08% of responses had WTP < 5). Moreover, participants showed considerable individual variation in their valuation of variance-risk (Eq. 1 in BDM auction tasks), both in the social (mean range ± SEM range, 48.13 ± 3.46) and nonsocial (50.23 ± 3.54) options. When we averaged risk attitudes across social and nonsocial domains (Fig. 2a; Table 1), 12 out of our 20 participants were risk averse [i.e.,
Risk proneness across and within domains. a, Risk proneness averaged across social and nonsocial domains. Twelve participants were risk averse (β2 < 0; light blue) and eight risk seeking (β2 > 0; orange). Participants are ordered from least risk-prone to most risk-prone. b, Risk proneness in the social domain. c, Risk proneness for juice variance-risk. In b and c, the ordering of participants was the same as in a.
Risk proneness
In a supplementary analysis, we examined whether the average willingness to pay (WTP) differed between the two domains. Our results indicated that they did not for seven out of twelve conditions [i.e., across three mean EV levels (1, 2, 3) and four risk levels (1, 2, 3, 4)]. However, in five cases, WTP for nonsocial rewards was significantly higher than for social rewards, particularly in conditions with higher EV or higher risk [paired sample t test, p < 0.05; (4, 1), (4, 2), (4, 3), (2, 3), (1, 3) for (risk level, EV level)]. Additionally, we tested whether willingness to pay was a predictor of binary choices (Eq. 2 in Materials and Methods, Risky binary choice task). Indeed, we found that WTP for risky options predicted binary choices between risky options (logistic regression; group-level t test; p = 0.05, t(19) = 2.08). A more detailed analysis of single-participant regression coefficients revealed that this was only the case for 15 out of 20 participants. This finding suggested that WTP bids for risky options can predict binary choices, but that congruence does not occur in all participants. The lack of congruence in these participants was likely due to additional value comparisons between binary choice options beyond the simple WTP assigned to individual options. Nevertheless, we included all 20 participants for the fMRI analyses of risk, expected value, and WTP.
Neural results
Social variance-risk coding in the amygdala
To test whether amygdala activity correlated with social variance-risk, we regressed BOLD responses against trial-by-trial parametric modulators capturing variance-risk when the two options were presented (see Materials and Methods, Neuroimaging data analysis). Amygdala activity increased with variance-risk for both social (small-volume FWE peak-corrected, p < 0.05; peak MNI coordinate x = −18, y = −4, z = −12 for the left amygdala, t = 4.97; x = 32, y = 0, z = −16 for the right amygdala, t = 4.01; Fig. 3a, yellow) and juice reward (small-volume FWE peak-corrected, p < 0.05; x = 24, y = 2, z = −22 for the right amygdala, t = 4.18; Fig. 3a, green. For other regions, see Table 2). The direct comparisons between social and juice risk showed no significant effects after correction. Furthermore, to assess common activations related to both social and nonsocial risk, we used conjunction analysis and found domain-general risk signals in the amygdala (Fig. 3a, red), as well as other regions (Table 2). The correlation with aversion to social risk at this location was not significant (t = 1.16, p = 0.5). Moreover, we found no relation to risk proneness or risk aversion in any other voxel of the amygdala, even at lenient thresholds (p < 0.005, uncorrected). Thus, the amygdala processed social and juice variance-risk.
Variance-risk signals in the amygdala. a, Location of variance-risk signals. Social variance-risk signals (left amygdala, t = 4.97; right amygdala, t = 4.01; yellow) located bilaterally whereas juice variance-risk signals (right amygdala, t = 4.18; green) were unilateral. Furthermore, domain-general risk signals in the right amygdala from the conjunction analysis (right amygdala, t = 4.04; red). Neural activations are small volume FWE peak-level corrected with p < 0.05, cluster-forming threshold: p < 0.001 (for visualization purposes; cluster-forming threshold, p < 0.005). b, Parameter estimates for the entire right amygdala (magenta in a). Responses to the risky options increased with both social and juice variance-risk. Error bars represent ±1 standard error of the mean.
Brain regions processing social or nonsocial variance-risk
In exploratory analyses, we investigated social and nonsocial variance-risk coding in the anterior insular cortex, striatum, and lateral orbitofrontal cortex (OFC), as these regions have been associated with risk processing previously (Preuschoff et al., 2006; Tobler et al., 2007; Cooper and Knutson, 2008; Burke et al., 2010b; O’Neill and Schultz, 2010, 2013). We observed significant social variance-risk signals in the striatum (small-volume FWE-corrected, p < 0.05), specifically in the nucleus accumbens and putamen, as well as in the OFC (small-volume FWE-corrected, p < 0.05 in combination of the medial, anterior, posterior, and lateral OFC). The insular cortex showed a relation to social risk that did not survive correction (small-volume FWE-corrected, p = 0.088). Relations to nonsocial risk occurred in the striatum (small-volume FWE-corrected, p < 0.05), but did not reach significance in the insula and OFC (both p < 0.005, uncorrected).
WTP coding in the mPFC and middle and posterior cingulate cortex
We found whole-brain FWE cluster-corrected WTP signals (Fig. 4) in the medial prefrontal cortex (mPFC; t = 6.87), middle cingulate cortex (MCC; t = 7.44), and posterior cingulate cortex (PCC; t = 6.25). Other regions showing WTP signals included the occipital regions (Occipital_Inf_R, t = 7.85; Lingual_R, t = 6.69; Cuneus_L, t = 7.55) and SMA (t = 6.38; Table 3). Furthermore, we found WTP signals in the right amygdala (small-volume FWE peak-corrected, p < 0.05; x = 24, y = −4, z = −18 for the right amygdala). Thus, consistent with frequently identified value-processing regions (Bartra et al., 2013), we found WTP signals in parts of the medial prefrontal and cingulate cortex and amygdala.
Neural representation of willingness to pay (WTP). a, WTP parametrically activated MCC (t = 7.44), PCC (t = 6.25), mPFC (t = 6.87), SMA (t = 6.38), and occipital cortex (Occipital_Inf_R, t = 7.85; Lingual_R, t = 6.69; Cuneus_L, t = 7.55). Neural activations are whole-brain FWE cluster-corrected with p < 0.05, cluster-inducing voxel threshold p < 0.001. b, The dorsal anterior cingulate cortex (dACC, t = 4.50) also encodes WTP. Neural activations are small volume FWE peak-level corrected with p < 0.05 (cluster-forming threshold, p < 0.001).
Brain regions processing WTP
Negative functional coupling between the amygdala–dACC, amygdala–TPJ, and subjective risk proneness
Our activation results relate the right amygdala to variance-risk for both juice and social reward. Previous studies have shown that the amygdala is anatomically (Carmichael and Price, 1995; Beckmann et al., 2009) and functionally (Apps et al., 2016) connected to the dACC. Moreover, resting-state connectivity between the amygdala to the other regions (i.e., node strength) and particularly dACC seems to be related to attitudes concerning nonsocial risk positively (Jung et al., 2018). However, the relation of task-induced activity to taking either social or nonsocial risk remains unknown. We therefore investigated whether amygdala–dACC connectivity during the presentation of risky choice options changed as a function of risk proneness.
In our psychophysiological interaction (PPI) connectivity analysis, we used right amygdala activity as the physiological regressor and parametric variance-risk as the psychological regressor. This analysis revealed that a lower proneness to take social variance-risk correlated with stronger connectivity to the dACC [p < 0.05, small-volume FWE peak-corrected within the dACC as defined by the anterior rostral cingulate zone of the connectivity-based ACC/OFC atlas (Neubert et al., 2015), with a cluster-forming threshold of p < 0.001].
We also observed negative functional coupling between the amygdala and TPJ area and social risk proneness regression coefficients
Negative correlation of amygdala–dACC connectivity with subjective risk attitude. For the PPI analysis, the region of the right amygdala (yellow for social risk; green for juice risk) served as a seed. a, Negative correlation of connectivity with social risk attitude. For social variance-risk, risk proneness correlated negatively with amygdala–dACC connectivity. Neural correlation is small volume FWE peak-level corrected with p < 0.05. The dACC small volume was defined by the connectivity-based atlas of Neubert and colleagues (2015). b, Similarly, social risk proneness is negatively correlated to amygdala–TPJ connectivity. The neural correlation is whole-brain familywise error (FWE) cluster-level corrected with p < 0.05, cluster-forming threshold, p < 0.001. TPJ was defined by the connectivity-based bilateral anterior TPJ atlas (Mars et al., 2012). c, In contrast, for juice risk proneness, there were no significant relations to amygdala–dACC and amygdala–TPJ connectivity (d). e, Estimated PPI betas in dACC and (f) TPJ were derived from social and juice trials, with a 90% confidence interval. Parameter estimates were extracted from peak coordinates (dACC, x = −8, y = 12, z = 38; right TPJ, x = 66, y = −32, z = 12) using a 4 mm radius.
Discussion
We estimated the proneness of human participants to take social and nonsocial variance-risk using the Becker–DeGroot–Marschak auction-like procedure. We found that the amygdala processed both social variance-risk and nonsocial variance-risk independently of risk proneness. Furthermore, the functional connectivity with social risk-coding areas dACC and TPJ during the presentation of risky choice options was negatively correlated to subjective risk proneness for social, but not for juice variance-risk.
Subjective risk attitude measured by willingness to pay for risky options
Although explicit valuations of options as measured with willingness to pay do not always predict choices (Abdellaoui et al., 2007), there typically is a relationship with choice, even in less consequential rating tasks (Lopez-Persem et al., 2017). Because willingness to pay is not derived from choices, it avoids any potential circularity of defining risk attitudes through choice. It is therefore noteworthy that in line with previous choice-based research (Holper et al., 2014), our participants were predominantly risk averse for nonsocial rewards. We extend these previous findings by showing that risk aversion is more common than risk proneness also for social rewards. Moreover, although the risk attitudes of our participants varied in both social and nonsocial domains, we found a significant correlation between these variations across domains. These findings converge with those factor analytic approaches (Frey et al., 2017) that identified a common factor explaining general, domain-independent risk-taking tendencies. Note though that these correlations were not well-powered and should therefore be interpreted with caution. Another point worth mentioning is that in more than half of the cases, the average WTP did not significantly differ between juice and social rewards. However, in conditions with higher EV or risk levels, the WTP of juice rewards was higher than for social rewards. This may be due to the 3 h fasting period, which may have amplified reward seeking behavior, particularly for the juice reward. The difference underscores the importance of carefully determining the subjective value of rewards when measuring risk aversion and comparing it between reward domains. Furthermore, we note that during the WTP measurements, nonsocial juice rewards were cued by abstract cues whereas social rewards were displayed directly on the screen. Despite this difference, the ranges of measured WTP bids were comparable between social and nonsocial rewards. These points underscore the importance of carefully determining the subjective value of rewards when measuring risk aversion and comparing it between social and nonsocial reward domains. Our approach to deliver juice rewards with flavor and sugar components follows previous neuroimaging studies which showed these stimuli effectively activate neural reward systems and guide WTP bids and economic choices (Grabenhorst et al., 2010; Zangemeister et al., 2016; Kim et al., 2024). We adapted this approach to the social domain by presenting photographs and compliments designed to vary in attractiveness. The measured range of WTP bids, economic choices, and activation of neural valuation structures (amygdala, mPFC) suggests these stimuli elicited subjective valuations in participants. Future studies could extend our approach to investigating social risk using more complex, naturalistic social stimuli (e.g., movies and interactions with virtual partners).
Our approach suggests that not only the intrinsic value of social reward but also the value of the risk associated with such rewards can be measured with precise methods from behavioral economics. The intrinsic value of social reward appears to be reduced (and the amygdala response to social situations increased) in individuals with social anxiety (Schultz et al., 2019). An open question worthy of future research may be whether these individuals also show increased aversion to social variance-risk as measured by willingness to pay methods (and/or stronger amygdala responses to social risk).
From a theoretical perspective, our approach is in the finance tradition, which captures the subjective value of a choice option as a linear combination of the moments (mean, variance, etc.) that characterize the distribution of outcomes associated with that option (Markowitz, 1991). However, it is worth noting that this is not the only theory capturing value-based decision-making. Indeed, both the behavior of individuals and the activity of value-related brain regions can be captured also by other theories, such as prospect theory or expected utility theory (Grabenhorst and Rolls, 2011; Konova et al., 2020; Williams et al., 2021). More importantly, our study shows that the formal approach to risk evaluation is not limited to the nonsocial domain but can also be applied to the social domain.
The amygdala encodes social and nonsocial variance-risk
We found social variance-risk signals in the amygdala. Although previous studies (Chang et al., 2015; Grabenhorst et al., 2019) suggest that the amygdala plays an important role in social decision-making, these studies did not investigate social variance-risk. Interestingly, in our study, amygdala activity specifically correlated with objectively defined social risk but not with the subjective valuation of that risk. These findings are compatible with the notion that subcortical processes are closer to the objective inputs, while cortical regions more strongly reflect the idiosyncrasies with which individuals process these inputs (Tobler et al., 2008; Genevsky et al., 2017). We note that although the amygdala has traditionally been associated with processing negative outcomes, substantial evidence across species implicates the amygdala in processing reward (Paton et al., 2006; Grabenhorst and Schultz, 2021). The present results suggest that the amygdala's functions in reward processing extend to processing the variance-risk associated with the distribution of social reward outcomes. Future studies could extend our approach to investigate the processing of value and risk for social stimuli that also include punishments, in addition to rewards and neutral stimuli used here. Together, our findings shed new light on amygdala functions in risk processing and subcortical contributions to risky decision-making.
Subjective risk attitudes correlate with amygdala–dACC and amygdala–TPJ coupling
The social variance-risk–related coupling between the amygdala and dACC decreased with individual proneness to take social risks. Parts of dACC have been associated with processing social reward and social prediction errors (Burke et al., 2010a; Apps and Ramnani, 2014; Apps et al., 2015; Lockwood et al., 2015; Sul et al., 2015; Kumaran et al., 2016; Wittmann et al., 2016; Fariña et al., 2021; Westhoff et al., 2021). Specifically, the processing of social value appears to rely on the interplay of the dACC with the amygdala (Pujara et al., 2022). Similarly, our findings indicate that the TPJ, which is associated with processing social cognition (Mars et al., 2012; Hill et al., 2019; Konovalov et al., 2021), follows the same pattern of interaction. The data suggest that this interplay is relatively specific to the assessment of social, as opposed to nonsocial, risks.
Furthermore, we find an increasingly negative interplay between the amygdala and dACC with increasing aversion to social risk during the presentation of risky choice options. This negative coupling (Fig. 4b, Table 3) in the social domain extends and refines the previously reported increasingly positive amygdala–dACC coupling associated with aversion to nonsocial risk in resting state—that is, without engaging in any task (Jung et al., 2018). Our study also identified a negative functional coupling between the amygdala–TPJ and social risk proneness. Taking the two studies together, it seems that the influence of risk attitude on the interaction between these brain regions is contingent upon the active engagement in processing risk, specifically in relation to social variance-risk.
In primates, neurons in both the amygdala and ACC process rewards and choices in social situations (Chang et al., 2013, 2015; Grabenhorst et al., 2019), and amygdala–ACC frequency-specific coupling is implicated in making choices that affect the reward outcomes of social others (Dal Monte et al., 2020). In rodents, amygdala–ACC connections are implicated in the transfer of information about aversive outcomes during a social, observational learning situation (Allsop et al., 2018). We previously showed that distinct groups of amygdala neurons during observational learning process valuations and decision computations for self and social others (Grabenhorst et al., 2019). Thus, the current findings suggest that related neural processing and amygdala–ACC connectivity in humans play a role in the evaluation of risk derived from social others. Compared to interactions and functional similarities between the amygdala and ACC, much less is known in relation to the amygdala and the TPJ. Our finding of social risk-dependent coupling between the amygdala and TPJ could stimulate further investigation of how these two areas interact functionally.
Conclusions
By combining functional neuroimaging with a Becker–deGroot–Marschak auction task for social and nonsocial outcomes, we showed that the amygdala processes both social and nonsocial variance-risk. Moreover, social risk-related coupling of the amygdala with social risk-coding areas dACC and TPJ changes as a function of risk attitude (more negative coupling with stronger risk aversion), and this relation is relatively specific for attitude to social rather than nonsocial risk. Our findings extend formal approaches to risk evaluation into the social domain and pave the way for investigating social risk attitudes in mental health impairment with dysfunctional processing of social uncertainty, including social anxiety.
Footnotes
This work was supported by the Wellcome Trust and the Royal Society (Wellcome/Royal Society Sir Henry Dale Fellowship Grants 206207/Z/17/Z and 206207/Z/17/A to F.G.; Wellcome Trust Principal Research Fellowship and Programme Grant 095495 to W.S.). J-C.K. received a Doc.Mobility Fellowship (P1ZHP1_184166) from the Swiss National Science Foundation. P.N.T received support from the Swiss National Science Foundation (Grants 10001C_188878, 100019_176016, and 100014_165884). We thank Alaa Al-Mohammad, Philipe Bujold, and Konstantin Volkmann for the helpful discussions. This work was supported in whole, or in part, by the Wellcome Trust. For the purpose of open access, the author has applied a CC BY public copyright license to any Author Accepted Manuscript version arising from this submission.
↵*J-C.K. and L.Z. contributed equally to this work.
The authors declare no competing financial interest.
- Correspondence should be addressed to Fabian Grabenhorst at fabian.grabenhorst{at}psy.ox.ac.uk.
