Contributions of insula and superior temporal sulcus to interpersonal guilt and responsibility in social decisions

abstract

This study investigated the neural mechanisms involved in feelings of interpersonal guilt and responsibility evoked by social decisions in humans.

In two studies (one during fMRI), participants repeatedly chose between safe and risky monetary outcomes in social contexts.

Across conditions, each participant chose for both themselves and a partner ( Social condition), or the partner chose for both themselves and the participant ( Partner condition), or the participant chose just for themselves ( Solo condition, control).

If the risky option was chosen in the Social or Partner condition, participant and partner could each receive either the high or the low outcome of a lottery with 50% probability, independently of each other.

Participants were shown the outcomes for themselves and for their partner on each trial and reported their momentary happiness every few trials.

As expected, participant happiness decreased following both low lottery outcomes for themselves and for the partner.

Crucially, happiness decreases following low outcomes for the partner were larger when the participant rather than their partner had made the choice, which fits an operational definition of guilt.

This guilt effect was associated with BOLD signal increase in the left anterior insula.

Connectivity between this region and the right inferior frontal gyrus varied depending on choice and experimental condition, suggesting that this part of prefrontal cortex is sensitive to guilt-related information during social choices.

Variations in happiness were well explained by computational models based on participants’ and partners’ rewards and reward prediction errors.

A model-based analysis revealed a left superior temporal sulcus cluster that tracked partner reward prediction errors that followed participant choices.

Our findings identify neural mechanisms of guilt and social responsibility during social decisions under risk.

introduction

Introduction Imagine that you go out to dinner with a friend and it is your turn to choose the restaurant.

You can choose between two restaurants: one that you both know well, with very predictable food of good quality, and a new restaurant that neither you nor your friend has eaten in before.

You decide to try the new one.

Unfortunately, while your dish is nice, your friend’s turns out to be worse than the known restaurant’s dishes.

How would you feel?

Would you feel differently if it had been your friend’s turn to choose the restaurant?

Being responsible for such suboptimal outcomes for others can induce a feeling of interpersonal guilt, formally described as a negative emotional response to harming someone with whom one has a positive social bond ( Baumeister, 1998 ; Baumeister et al., 1994 ; Berndsen et al., 2004 ; Tangney et al., 2007 ; Zeelenberg and Breugelmans, 2008 ).

Guilt influences decisions (e.g., Charness and Dufwenberg, 2006 ), and abnormal sensitivity to guilt is associated with severe social dysfunctions ranging from psychopathy to depression and anxiety, depending on whether sensitivity to guilt is, respectively, reduced or increased ( Tangney et al., 2007 ).

Several brain regions have been associated with the feeling of guilt: the anterior insula (aIns) ( Bastin et al., 2016 ; Lamm and Singer, 2010 ; Piretti et al., 2023 ), the dorsal cingulate cortex ( Bastin et al., 2016 ; Gifuni et al., 2017 ) and the left temporo-parietal junction ( Bastin et al., 2016 ; Piretti et al., 2023 ), the ventromedial prefrontal cortex ( Krajbich et al., 2009 ), and with various other regions and networks involved depending on the method used to induce guilt ( Bastin et al., 2016 ; Gifuni et al., 2017 ).

However, we are interested in the mechanisms involved in a specific, as yet understudied aspect of guilt: the kind that can result from everyday choices that expose others to an uncertain outcome.

To study this, we built on findings from experimental social decision-making tasks.

Several studies have used games from behavioural economics or perceptual tasks linked to punishment of a partner.

For example, participants behaving in accord with an economic definition of guilt aversion during a trust game showed activation in insula, supplementary motor area, dorsolateral prefrontal cortex (dlPFC) and temporal parietal junction ( Chang et al., 2011 ).

In another study, partners received painful stimuli when participants made errors during a difficult perception task.

These errors evoked activations in the left aIns and dlPFC in the participants ( Koban et al., 2013 ).

In a variation of this task, participants could decide to bear a proportion of their partner’s pain ( Yu et al., 2014 ).

The level of pain taken, indicative of guilt, and activations in anterior middle cingulate cortex and aIns were higher when the pain followed errors made only by the participant rather than by both players.

A multivariate reanalysis of these two datasets revealed a neural signature for guilt ( Yu et al., 2020 ), with key regions including the anterior medial cingulate cortex, insula, inferior frontal gyrus (IFG), inferior temporal cortex, thalamus, and cerebellum.

One recent study paired the aforementioned perception-and-pain paradigm with dictator game decisions that allowed the participant to compensate for the partner’s outcome ( Gao et al., 2018 ).

The guilt context increased advantageous-inequity aversion and decreased disadvantageous-inequity aversion and affected the neural correlates of these inequities (respectively, mentalizing-related and emotion- and conflict-related regions).

Finally, another study contrasted guilt and shame: confederates either experienced economic loss due to bad participant advice (participants experienced guilt), or experienced no loss when participants’ bad advice was correctly refused by the confederate (participants experienced shame) ( Zhu et al., 2019 ).

Guilt relative to shame activated supramarginal gyrus and temporo-parietal junction as well as orbitofrontal cortex, ventrolateral, and dorsolateral prefrontal cortex.

Multivariate analyses revealed that guilt could be distinguished from shame based on activation in ventral anterior cingulate cortex and dorsomedial prefrontal cortex.

One landmark study reported on the mechanisms involved in seeking or avoiding responsibility in decisions affecting a group of individuals, and reported involvement of medial prefrontal cortex, aIns and temporo-parietal junction ( Edelson et al., 2018 ).

In this elegant study, participants could delegate their choice between a risky and a safe option to a group or decide themselves; in separate conditions, the choice affected payoff either only for the participant or for all group members.

Interestingly, most participants displayed responsibility aversion, and this effect could not be explained by guilt, suggesting people associate a psychological cost with assuming responsibility for others’ outcomes.

However, in many situations, one does not have the possibility to delegate a decision to others, and choosing a risky option in such a case may evoke guilt when the outcome for others is negative.

The neural mechanisms underlying guilt evoked during such situations of social responsibility are still unknown.

These questions are the subject of the present study.

Similarly to Edelson et al.’s and more recent studies ( Arioli et al., 2023 ; Fareri et al., 2022 ), our paradigm leveraged a risky choice task with social conditions.

In our experiment, participants were paired with a partner and played an ‘ice-breaker’ game that created a positive social bond between them, increasing the likelihood of feeling empathy and guilt for each other ( Baumeister, 1998 ; Julle-Danière et al., 2020 ; Loewenstein et al., 1989 ).

Participants or their partner then chose between risky and safe monetary options in three conditions ( Figure 1 ).⟦>zach claim=no-assertion: @{Participants or their partner then chose between risky and safe monetary options in three conditions ( Figure 1 ).} Narrates the paradigm — three conditions with a partner — rather than stating a result; a component of the paper's scope.⟧

In the Social condition, the outcome of choices affected both participant and partner; participants were and felt responsible for these outcomes because they had agency over their decisions and knew that they could have chosen otherwise ( Frith, 2014 ).

We contrasted this condition with similar choices made by a simple expected-value-maximizing algorithm posing for the partner ( Partner condition), and participant choices only affecting the participant themselves ( Solo condition, acting as control).

Importantly, we assessed the emotional impact of the outcomes of these choices by monitoring participants’ happiness every two trials ( Rutledge et al., 2014 ; Rutledge et al., 2016 ).

This allowed us to fit computational models to happiness data and search for networks sensitive to reward prediction errors resulting from participant or partner choices.

We ran two experiments, Study 1 outside the MRI scanner and Study 2 during fMRI, with separate groups of participants.

Figure 1. Experimental design.⟦>zach claim=no-assertion: @{Figure 1. Experimental design.} A figure title, not a finding.⟧

In every trial, participants were presented with pairs of monetary options (a safe and a risky option; the risky option was a lottery with equally probable high and low outcomes).

There were three conditions: a non-social ‘Solo’ condition, in which the participant’s choice led to an outcome just for themselves (left panel); and two conditions in which choices were made by the participant (‘Social’ condition) or by their partner (‘Partner’ condition) and led to outcomes affecting both players (right panel).

Importantly, selecting the risky option in the social or partner conditions led to the lottery being played out independently for both players; that is, participant and partner could receive the high or low outcome independently from each other (coloured boxes).

Selecting the safe option led to both players receiving an equal outcome.

We analysed our behavioural data using several complementary methods: choices were modelled with mixed-effects regressions serving as manipulation checks; risk preferences expressed in choices were assessed using a comprehensive expected utility model as well as with a simpler, more robust ‘risk premium’ approach; and happiness data were fitted, in addition to the computational models, with several linear mixed models (LMMs) to assess the impact of both the participant’s and their partner’s rewards, the impact of agency and their interactions.

Inspired by findings reported in previous neuroimaging of social emotions, we also used several methods to analyse our fMRI data, including conventional methods (both region-of-interest and mass univariate); mixed-effects regression models; computational model-based analyses (inspired by, e.g., Konovalov et al., 2021a ; Rutledge et al., 2014 ); and functional connectivity (e.g., Edelson et al., 2018 ; Konovalov et al., 2021a ).

The behavioural modelling is thus complemented by neuroimaging analyses that offer insight about both the activity in regions associated with guilt as well as their place in a wider network, providing an in-depth comprehensive analysis of the mechanisms behind guilt evoked by social responsibility.

results

Results Behaviour Overview We first verified that participants’ choices were reasonable in the Solo and Social conditions, then assessed whether these choices and the risk preferences they revealed changed depending on whether participants chose just for themselves ( solo ) or for themselves and the interaction partner ( social ).

Next, we assessed whether momentary happiness varied with decision outcomes and whether responsibility influenced this relationship.

Lastly, we fitted computational models to the happiness data.

While Study 1 (behaviour only) was run before Study 2 (fMRI), we will report the results of both studies together as their results were highly consistent.

Choices: manipulation check As expected, participants’ probability of choosing the risky option (lottery) increased with the difference between the expected value of the lottery and the value of the safe option (Study 1: Figure 2A , t (4796) = 9.26, p < 3.1e –20 , β = 0.074, 95% CI = [0.059 0.090]; Study 2: Figure 2D , t (3829) = 10.62, p < 5.3e –26 , β = 0.093, 95% CI = [0.075 0.110]; mixed-effects regressions, see Equation 1 , detailed results are reported in Appendix 1—table 1 ).⟦>zach claim=86ccf2d0-0182-4d14-a72d-183c27d73310: @{Choices: manipulation check As expected, participants’ probability of choosing the risky option (lottery) increased with the difference between the expected value of the lottery and the value of the safe option (Study 1: Figure 2A , t (4796) = 9.26, p < 3.1e –20 , β = 0.074, 95% CI = [0.059 0.090]; Study 2: Figure 2D , t (3829) = 10.62, p < 5.3e –26 , β = 0.093, 95% CI = [0.075 0.110]; mixed-effects regressions, see Equation 1 , detailed results are reported in Appendix 1—table 1 ).} participants-probability-choosing-risky-option⟧

Participants chose the lottery more often in the Solo condition than in the Social condition in Study 1 ( t (4796) = 2.54, p = 0.011, β = 0.164, 95% CI = [0.038 0.291]), but this difference was not found in Study 2 ( t (3829) = 0.23, p = 0.82, β = 0.015, 95% CI = [–0.109 0.138]).⟦>zach claim=2952075b-80ca-4849-9ddd-6582a1967dac: @{Participants chose the lottery more often in the Solo condition than in the Social condition in Study 1 ( t (4796) = 2.54, p = 0.011, β = 0.164, 95% CI = [0.038 0.291]), but this difference was not found in Study 2 ( t (3829) = 0.23, p = 0.82, β = 0.015, 95% CI = [–0.109 0.138]).} participants-chose-risky-option-lottery⟧

There was no significant interaction between the difference in expected values and experimental conditions in either study (p > 0.52).⟦>zach claim=630e2396-70a6-46dc-93cd-5e49a3a8b472: @{There was no significant interaction between the difference in expected values and experimental conditions in either study (p > 0.52).} no-significant-interaction-between-difference⟧

Figure 2. Participant choices in Studies 1 (outside fMRI, N = 40) and 2 (inside fMRI, N = 44).⟦>zach claim=37937db1-e635-46d4-ac0a-f482400668e5: @{Figure 2. Participant choices in Studies 1 (outside fMRI, N = 40) and 2 (inside fMRI, N = 44).} findings-rest-two-samples-healthy⟧

( A, D ) The probability of choosing the risky option (lottery) in both Solo and Social conditions is well explained by the difference in expected value of the risky and safe choice options (EV risky – V safe ).⟦>zach claim=2952075b-80ca-4849-9ddd-6582a1967dac: @{( A, D ) The probability of choosing the risky option (lottery) in both Solo and Social conditions is well explained by the difference in expected value of the risky and safe choice options (EV risky – V safe ).} participants-chose-risky-option-lottery⟧

Participants chose the risky option slightly more often in the Solo condition than in the Social condition in Study 1 ( A ) but not in Study 2 ( D ).⟦>zach claim=2952075b-80ca-4849-9ddd-6582a1967dac: @{Participants chose the risky option slightly more often in the Solo condition than in the Social condition in Study 1 ( A ) but not in Study 2 ( D ).} participants-chose-risky-option-lottery⟧

Lines are predicted values of a logit linear mixed regression model fitted to the choice data (see Results).

Error areas indicate 95% pointwise confidence intervals for the predicted values.

Triangles indicate individual average choice proportions binned by EV risky – V safe value; size of triangle reflects the number of participants contributing to a datapoint.

Blue upward-pointing and red downward-pointing triangles are data from the Solo and Social conditions, respectively.

( B, E ) Risk premiums did not differ between Solo and Social conditions.⟦>zach claim=09960051-2473-4a67-b092-89fc6c2b9004: @{( B, E ) Risk premiums did not differ between Solo and Social conditions.} risk-premiums-not-differ-between⟧

( C, F ) Values of the risk aversion parameter ρ in the Solo and Social conditions were broadly consistent with Risk premium values, but showed that participants were slightly more risk averse in the Social than in the Solo condition in Study 1 only (see Results).⟦>zach claim=b4cfe28f-8a72-4aeb-9d52-b405c160af07: @{( C, F ) Values of the risk aversion parameter ρ in the Solo and Social conditions were broadly consistent with Risk premium values, but showed that participants were slightly more risk averse in the Social than in the Solo condition in Study 1 only (see Results).} participants-slightly-more-risk-averse⟧

In panels B, C, E and F, grey lines and markers show individual data, red lines show means with 95% confidence intervals about the means, and grey areas are kernel density plots representing the distribution of the data.

Choices: risk preferences To better assess whether people’s risk preferences varied between the Solo and the Social condition, we evaluated two additional measures.

First, we calculated for each participant a ‘risk premium’, defined as the difference between the expected value of a lottery and its certainty equivalent (see Equation 2 in Methods; positive risk premiums indicate risk aversion).

Second, we used an expected utility theory (EUT) approach to calculate a parameter ρ that describes a decision-maker’s risk attitude under the assumption of constant absolute risk aversion (see Equation 3 in Methods).

We used two measures because fitting the EUT model, the more comprehensive measure that takes into account all choices of a participant, requires many trials to be fitted reliably, a condition that was not satisfied in many participants of both studies; in contrast, the risk premium’s single-point measure of risk aversion could be estimated in all participants.

Risk premiums did not differ between Social and Solo conditions (Study 1: Figure 2B , t (39) = 1.53, p = 0.134, Cohen’s d = 0.24, BF 10 = 0.49; Study 2: Figure 2E , t (43) = –0.21, p = 0.84, d = –0.03, BF 10 = 0.17).⟦>zach claim=09960051-2473-4a67-b092-89fc6c2b9004: @{Risk premiums did not differ between Social and Solo conditions (Study 1: Figure 2B , t (39) = 1.53, p = 0.134, Cohen’s d = 0.24, BF 10 = 0.49; Study 2: Figure 2E , t (43) = –0.21, p = 0.84, d = –0.03, BF 10 = 0.17).} risk-premiums-not-differ-between⟧

Turning to the EUT approach, as differences in risk attitudes for gains and losses are well known ( Kahneman and Tversky, 1979 ), we first estimated ρ separately for gain and loss trials.

As the number of these types of trials varied across participants, we only obtained reliable estimates for both types of trials in some participants (Study 1: 18 of 40; Study 2: 16 of 44).

As ρ did not vary between gain and loss trials (Study 1: t (17) = 0.21, p = 0.84, d = 0.05; Study 2: t (15) = –0.61, p = 0.55, d = 0.15; paired t -test), we then pooled across gain and loss trials, then estimated and compared ρ for the Solo and Social conditions using paired t -tests.⟦>zach claim=6f5c63df-fb60-4e84-b0c6-4094ef525062: @{As ρ did not vary between gain and loss trials (Study 1: t (17) = 0.21, p = 0.84, d = 0.05; Study 2: t (15) = –0.61, p = 0.55, d = 0.15; paired t -test), we then pooled across gain and loss trials, then estimated and compared ρ for the Solo and Social conditions using paired t -tests.} risk-aversion-parameter-not-differ-between⟧

We found that participants were slightly more risk averse in the Social than in the Solo condition in Study 1 ( Figure 2C , t (39) = 2.27, p = 0.03, d = 0.36, BF 10 = 1.69) but not in Study 2Study 2 ( Figure 2F , t (43) = 1.40, p = 0.17, d = 0.21, BF 10 = 0.41).⟦>zach claim=b4cfe28f-8a72-4aeb-9d52-b405c160af07: @{We found that participants were slightly more risk averse in the Social than in the Solo condition in Study 1 ( Figure 2C , t (39) = 2.27, p = 0.03, d = 0.36, BF 10 = 1.69) but not in Study 2Study 2 ( Figure 2F , t (43) = 1.40, p = 0.17, d = 0.21, BF 10 = 0.41).} participants-slightly-more-risk-averse⟧

Risk premium and ρ were highly consistent with each other across participants in both conditions for both studies (Study 1: Solo condition: F (1,38) = 69.9, p < 0.001, R 2 = 0.65; Social condition: F (1,38) = 42.0, p < 0.001, R 2 = 0.55; Study 2: Solo condition: F (1,41) = 57.9, p < 0.001, R 2 = 0.59; Social condition: F (1,41) = 74.2, p < 0.001, R 2 = 0.64; these and all subsequent regressions are robust).⟦>zach claim=db58b24c-22dd-45c9-aaa6-656fe937a3d0: @{Risk premium and ρ were highly consistent with each other across participants in both conditions for both studies (Study 1: Solo condition: F (1,38) = 69.9, p < 0.001, R 2 = 0.65; Social condition: F (1,38) = 42.0, p < 0.001, R 2 = 0.55; Study 2: Solo condition: F (1,41) = 57.9, p < 0.001, R 2 = 0.59; Social condition: F (1,41) = 74.2, p < 0.001, R 2 = 0.64; these and all subsequent regressions are robust).} guilt-effect-occurred-whether-participant⟧

In sum, participants showed very similar risk preferences when making decisions affecting only themselves ( Solo condition) or themselves and their partner ( Social condition), with a tendency towards higher risk aversion in the Social condition in Study 1. Momentary happiness: links to reward Momentary happiness was assessed every two trials in a similar manner to previous studies by Rutledge et al., 2014 ; Rutledge et al., 2016 .

Across all trials, in both studies, participant momentary happiness correlated with rewards obtained in the current trial by the participant and by the partner (Correlations with participant reward, Study 1: F (1,3598) = 691.5, p < 0.001, R 2 = 0.16; partner reward, Study 1: F (1,2426) = 128.6, p < 0.001, R 2 = 0.05; participant reward, Study 2: F (1,2630) = 650.8, p < 0.001, R 2 = 0.20; partner reward, Study 2: F (1,1735) = 111.8, p < 0.001, R 2 = 0.06; Figure 3A, B, E, F ).⟦>zach claim=db58b24c-22dd-45c9-aaa6-656fe937a3d0: @{Across all trials, in both studies, participant momentary happiness correlated with rewards obtained in the current trial by the participant and by the partner (Correlations with participant reward, Study 1: F (1,3598) = 691.5, p < 0.001, R 2 = 0.16; partner reward, Study 1: F (1,2426) = 128.6, p < 0.001, R 2 = 0.05; participant reward, Study 2: F (1,2630) = 650.8, p < 0.001, R 2 = 0.20; partner reward, Study 2: F (1,1735) = 111.8, p < 0.001, R 2 = 0.06; Figure 3A, B, E, F ).} guilt-effect-occurred-whether-participant⟧

Based on previous findings ( Rutledge et al., 2016 ), we reasoned that participants’ momentary happiness would be influenced not only by rewards obtained in the current trial, but also by expected rewards, reward prediction errors, rewards received in previous trials, and differences between rewards obtained by the participant and the partner.

Crucial for our research question, we aimed to assess whether responsibility for these rewards, that is, taking into account whether the rewards occurred following choices made by the participant or the partner, would also influence variations in happiness.

Figure 3. Participant momentary happiness in Studies 1 ( A–D ) and 2 ( E–H ).⟦>zach claim=a246d73e-44f4-4bee-aff2-96388949b76a: @{Figure 3. Participant momentary happiness in Studies 1 ( A–D ) and 2 ( E–H ).} participant-momentary-happiness-varied-rewards-2⟧

Happiness varied with rewards received by the participant ( A, E ) and by the partner ( B, F ).⟦>zach claim=84c4be17-6129-424a-bfc3-5c04b90608a0: @{Happiness varied with rewards received by the participant ( A, E ) and by the partner ( B, F ).} participant-momentary-happiness-varied-rewards⟧

Each dot is one trial; data are pooled across participants.

Lines are fitted regression lines.

A computational model taking into account expected, previous and current rewards, reward prediction errors for both participant and partner, and decision-maker (Responsibility Redux model, see Results) predicted the variations in participants’ momentary happiness well ( C, G , and Table 1 ).⟦>zach claim=a2e7521f-86cd-472a-9b5e-a23659cce62a: @{A computational model taking into account expected, previous and current rewards, reward prediction errors for both participant and partner, and decision-maker (Responsibility Redux model, see Results) predicted the variations in participants’ momentary happiness well ( C, G , and Table 1 ).} among-computational-models-fitted-momentary⟧

Changes in momentary happiness after lottery choices in Social and Partner conditions varied with lottery outcome and decision-maker ( D, H ).⟦>zach claim=3531ed17-66ac-4b50-8bc8-1bcf3c035e9e: @{Changes in momentary happiness after lottery choices in Social and Partner conditions varied with lottery outcome and decision-maker ( D, H ).} when-partner-received-low-lottery⟧

Data were binned according to outcome for each participant and decision-maker (Social = participant chose the lottery, Partner = partner chose the lottery).

Crucially, responsibility for low lottery outcomes for the partner decreased participant happiness more than the same outcomes following partner choices (see Results), which fits the definition of interpersonal guilt.

In D and H, dots are individual datapoints, the bar indicates the mean, the error bars are 95% confidence intervals about the mean, and the stars indicate the significance of the ‘guilt effect’ (see text): ***p < 0.001; **p < 0.01. Figure 3—figure supplement 1. Parameter recovery for Responsibility Redux model.⟦>zach claim=db58b24c-22dd-45c9-aaa6-656fe937a3d0: @{In D and H, dots are individual datapoints, the bar indicates the mean, the error bars are 95% confidence intervals about the mean, and the stars indicate the significance of the ‘guilt effect’ (see text): ***p < 0.001; **p < 0.01. Figure 3—figure supplement 1. Parameter recovery for Responsibility Redux model.} guilt-effect-occurred-whether-participant⟧

Stability of the estimated parameters of the temporal difference models was evaluated by attempting to recover parameters from synthetic data created using each participant’s real estimated parameters.

After fitting each participant’s momentary happiness data (see above), we synthesized new momentary happiness data based on each participant’s estimated parameters, added 1SD of noise to the happiness data, fitted the model to these synthetic data, and repeated this procedure 10 times, for both studies.

We then compared these new estimated parameters to the actual parameters from which the synthetic data were generated.

For each parameter, we calculated the mean of each participant’s recovered parameters and regressed these means on the participants’ actual parameters (black dots and grey regression line, with 95% confidence interval).

Each coloured dot is one recovered parameter value; different colours represent different participants.

Following Rutledge and colleagues’ methodology, which considers that changes in momentary happiness in response to outcomes of a probabilistic reward task are explained by the combined influence of recent reward expectations and prediction errors arising from those expectations, we fitted computational models to each participant’s happiness data.

In these models, certain rewards, the expected value of chosen lotteries, reward prediction errors and additional outcome parameters are modelled separately with influences that decay exponentially with time (see Methods for equations and details).

We first fitted a ‘ Basic ’ model ( Equation 4 ) containing regressors coding for certain reward, expected value of chosen lotteries and participant reward prediction errors to evaluate whether these fundamental reward variables explained variations in momentary happiness.

This model explained the data reasonably well (see Table 1 ).⟦>zach claim=a2e7521f-86cd-472a-9b5e-a23659cce62a: @{This model explained the data reasonably well (see Table 1 ).} among-computational-models-fitted-momentary⟧

Next, based on the notion that inequality between rewards received by the participant and the partner would influence happiness (e.g., Loewenstein et al., 1989 ), we assessed the fit of an ‘ Inequality ’ model that included one term modelling the difference between the participant and partner outcomes ( Equation 5 ).

We also assessed the fit of Rutledge et al., 2016 ‘ Guilt-envy ’ model in which advantageous and disadvantageous inequality were modelled separately ( Equation 6 ).

Next, we tested a ‘ Responsibility ’ model in which separate terms represented the partner’s reward prediction errors resulting from choices made by the participant and by the partner ( Equation 7 ).

Finally, we optimized the Responsibility model by removing the regressor modelling partner reward prediction errors resulting from choices made by the partner, yielding a ‘ Responsibility Redux ’ model ( Equation 8 ).

Table 1. Fits of computational models to momentary happiness data.⟦>zach claim=a2e7521f-86cd-472a-9b5e-a23659cce62a: @{Table 1. Fits of computational models to momentary happiness data.} among-computational-models-fitted-momentary⟧

Model N param Mean R 2 Mean R 2 adj BIC AIC Study 1 Basic 3 0.328 0.305 –1000 –1399 Inequality 4 0.346 0.316 –916 –1416 Guilt-envy 5 0.354 0.316 –785 –1385 Responsibility 5 0.370 0.333 –866 –1466 Responsibility Redux 4 0.361 0.331 –999 –1499 Study 2 Basic 3 0.374 0.340 –693 –1062 Inequality 4 0.394 0.350 –620 –1080 Guilt-envy 5 0.405 0.350 –491 –1043 Responsibility 5 0.433 0.380 –616 –1168 Responsibility Redux 4 0.422 0.379 –735 –1195 BIC, Bayesian Information Criterion; AIC, Akaike’s Information Criterion.

BIC and AIC values are summed across participants.

As in previous work by Rutledge et al., 2014 ; Rutledge et al., 2016 , model fits were performed with individually Z -scored happiness ratings.

Best values of each variable for each study are highlighted in bold font.

For model details, see Results and Methods ( Equations 4–8 ).

All models explained variations in happiness reasonably well ( Table 1 ).⟦>zach claim=a2e7521f-86cd-472a-9b5e-a23659cce62a: @{All models explained variations in happiness reasonably well ( Table 1 ).} among-computational-models-fitted-momentary⟧

Overall, the Responsibility and Responsibility Redux ( Figure 3C, G ) models explained the data best (highest R 2 /adjusted R 2 or lowest AIC/BIC); a likelihood ratio test ( Equation 9 ) revealed that the Responsibility model fitted better than all the other models, including the Responsibility Redux model (Study 1: all LR ≥47.36, p < 0.0001; Study 2: all LR ≥77.83, p < 0.0001).⟦>zach claim=982ae364-a8a2-4ee6-8f94-30ebbe71a1e3: @{Overall, the Responsibility and Responsibility Redux ( Figure 3C, G ) models explained the data best (highest R 2 /adjusted R 2 or lowest AIC/BIC); a likelihood ratio test ( Equation 9 ) revealed that the Responsibility model fitted better than all the other models, including the Responsibility Redux model (Study 1: all LR ≥47.36, p < 0.0001; Study 2: all LR ≥77.83, p < 0.0001).} likelihood-ratio-test-showed-responsibility⟧

We also compared the R 2 and adjusted R 2 values obtained for each participant and model using t -tests, for both studies (average values are reported in Table 1 ).⟦>zach claim=a2e7521f-86cd-472a-9b5e-a23659cce62a: @{We also compared the R 2 and adjusted R 2 values obtained for each participant and model using t -tests, for both studies (average values are reported in Table 1 ).} among-computational-models-fitted-momentary⟧

The Responsibility model yielded higher R 2 values than all the other models (Study 1: all t > 3.6, p < 0.007; Study 2: all t > 2.9, p < 0.034; Bonferroni-corrected t -tests) except for the Guilt-envy model in the data of Study 1 ( t = 2.19, p = 0.17).⟦>zach claim=59ea6ccb-75b3-4756-9a5d-d19542972a93: @{The Responsibility model yielded higher R 2 values than all the other models (Study 1: all t > 3.6, p < 0.007; Study 2: all t > 2.9, p < 0.034; Bonferroni-corrected t -tests) except for the Guilt-envy model in the data of Study 1 ( t = 2.19, p = 0.17).} responsibility-model-yielded-higher-values⟧

The Responsibility and Responsibility Redux models yielded higher adjusted R 2 than the Basic model (Study 1: all t > 3.6, p < 0.01; Study 2: all t > 3.6, p < 0.01).⟦>zach claim=59ea6ccb-75b3-4756-9a5d-d19542972a93: @{The Responsibility and Responsibility Redux models yielded higher adjusted R 2 than the Basic model (Study 1: all t > 3.6, p < 0.01; Study 2: all t > 3.6, p < 0.01).} responsibility-model-yielded-higher-values⟧

Weights for certain rewards, expected value of the lotteries, participant reward prediction errors and the forgetting factor γ were overall positive across participants, models and studies (Study 1: all Z > 5.4, p < 0.0001; Study 2: all Z > 4.9, p < 0.0001, Wilcoxon sign-rank tests were used because data were not normally distributed).⟦>zach claim=982ae364-a8a2-4ee6-8f94-30ebbe71a1e3: @{Weights for certain rewards, expected value of the lotteries, participant reward prediction errors and the forgetting factor γ were overall positive across participants, models and studies (Study 1: all Z > 5.4, p < 0.0001; Study 2: all Z > 4.9, p < 0.0001, Wilcoxon sign-rank tests were used because data were not normally distributed).} likelihood-ratio-test-showed-responsibility⟧

Medians of the forgetting factor γ varied little across models, ranging from 0.39 to 0.44 in Study 1 and from 0.42 to 0.46 in Study 2, indicating stable influences of previous trials on happiness across models.

Participants’ reward prediction errors (sRPE) influenced happiness more than partner’s reward prediction errors, whether the latter resulted from the participant or the partner’s choices (respectively, social_pRPE and partner_pRPE ): weights for sRPE were higher than for social_pRPE or partner_pRPE (Study 1: all Z > 6.0, p < 0.001; Study 2: all Z > 3.7, p < 0.003).⟦>zach claim=d2189d9d-2aea-4cff-bff6-d1ace9b9fa33: @{Participants’ reward prediction errors (sRPE) influenced happiness more than partner’s reward prediction errors, whether the latter resulted from the participant or the partner’s choices (respectively, social_pRPE and partner_pRPE ): weights for sRPE were higher than for social_pRPE or partner_pRPE (Study 1: all Z > 6.0, p < 0.001; Study 2: all Z > 3.7, p < 0.003).} participants-own-reward-prediction-errors⟧

The stability of these estimated parameters was verified using a parameter recovery procedure (see Methods and Figure 3—figure supplement 1 ).⟦>zach claim=744191f0-c636-4388-887c-cb552e458d9f: @{The stability of these estimated parameters was verified using a parameter recovery procedure (see Methods and Figure 3—figure supplement 1 ).} parameter-recovery-procedure-synthetic-data-generated — The parameter-recovery claim states exactly this verification of the fitted happiness-model parameters.⟧

These results thus replicate the finding that outcomes of risky social decisions influence momentary happiness ( Rutledge et al., 2016 ).

Crucially, we find here that the partner’s reward prediction errors ( social_pRPE and partner_pRPE ) contributed to explaining changes in participants’ momentary happiness: the Responsibility and ResponsibilityRedux models explained the data better than the models without these parameters (see Table 1 ).⟦>zach claim=a2e7521f-86cd-472a-9b5e-a23659cce62a: @{Crucially, we find here that the partner’s reward prediction errors ( social_pRPE and partner_pRPE ) contributed to explaining changes in participants’ momentary happiness: the Responsibility and ResponsibilityRedux models explained the data better than the models without these parameters (see Table 1 ).} among-computational-models-fitted-momentary⟧

In particular, the partner’s reward prediction errors resulting from the participants’ decisions ( social_pRPE ), that is, those pRPE for which participants were responsible contributed to explaining our data (weights for social_pRPE were greater than 0: Responsibility model : Study 1: Z = 2.85, p = 0.004, Study 2: Z = 3.26, p = 0.001; ResponsibilityRedux model : Study 1: Z = 2.93, p = 0.003, Study 2: Z = 3.30, p = 0.001; weights for social_pRPE tended to be higher than weights for partner_pRPE: Responsibility model : Study 1: Z = 2.14, p = 0.033; Study 2: Z = 1.41, p = 0.16).⟦>zach claim=31dbdf89-fbc1-47e0-88ff-e01183873dbd: @{In particular, the partner’s reward prediction errors resulting from the participants’ decisions ( social_pRPE ), that is, those pRPE for which participants were responsible contributed to explaining our data (weights for social_pRPE were greater than 0: Responsibility model : Study 1: Z = 2.85, p = 0.004, Study 2: Z = 3.26, p = 0.001; ResponsibilityRedux model : Study 1: Z = 2.93, p = 0.003, Study 2: Z = 3.30, p = 0.001; weights for social_pRPE tended to be higher than weights for partner_pRPE: Responsibility model : Study 1: Z = 2.14, p = 0.033; Study 2: Z = 1.41, p = 0.16).} partner-reward-prediction-errors-resulting⟧

Momentary happiness: effects of agency, responsibility, and guilt Next, we assessed whether happiness varied depending on the participant’s agency ( Social + Solo vs. Partner ), and found happiness to be lower when the participant chose, independent of the outcome (Study 1: t (3600) = –3.92, p < 0.0001, β = –0.14, 95% CI = [−0.20 to 0.07]; Study 2: t (2870) = –6.07, p < 0.0001, β = –0.24, 95% CI = [−0.31 to 0.16]).⟦>zach claim=8a2fcf5d-5473-4f45-b060-506f4d3d95d9: @{Momentary happiness: effects of agency, responsibility, and guilt Next, we assessed whether happiness varied depending on the participant’s agency ( Social + Solo vs. Partner ), and found happiness to be lower when the participant chose, independent of the outcome (Study 1: t (3600) = –3.92, p < 0.0001, β = –0.14, 95% CI = [−0.20 to 0.07]; Study 2: t (2870) = –6.07, p < 0.0001, β = –0.24, 95% CI = [−0.31 to 0.16]).} participant-happiness-lower-when-participant⟧

This is interesting in itself and may reflect the drive behind responsibility aversion reported by Edelson et al.’s 2018 study: being assigned the role of the decider in a social setting may make people slightly unhappy, perhaps due to ‘weight of the responsibility’.

To specifically search for a sign of interpersonal guilt, we analysed happiness values reported after outcomes of lottery choices in the Social and Partner conditions using conventional LMMs ( Equation 10 ).

Fitting several models revealed that lottery outcomes influenced happiness, and that the identity of the decision-maker played a role ( Appendix 1—table 2 ).⟦>zach claim=no-assertion: @{Fitting several models revealed that lottery outcomes influenced happiness, and that the identity of the decision-maker played a role ( Appendix 1—table 2 ).} Analysis narration introducing the guilt LMM; the effects it foreshadows are stated as their own claims below.⟧

Crucially, the interaction between partner outcome and decision-maker was significant (Study 1: t (1180) = 3.52, p = 0.0004, β = 0.37, 95% CI = [0.16 0.58]; Study 2: t (937) = 2.85, p = 0.0045, β = 0.33, 95% CI = [0.10 0.56]).⟦>zach claim=3531ed17-66ac-4b50-8bc8-1bcf3c035e9e: @{Crucially, the interaction between partner outcome and decision-maker was significant (Study 1: t (1180) = 3.52, p = 0.0004, β = 0.37, 95% CI = [0.16 0.58]; Study 2: t (937) = 2.85, p = 0.0045, β = 0.33, 95% CI = [0.10 0.56]).} when-partner-received-low-lottery⟧

When the partner received the low lottery outcome, participant happiness was lower when they rather than the partner had chosen the lottery ( Figure 3D, H ).⟦>zach claim=3531ed17-66ac-4b50-8bc8-1bcf3c035e9e: @{When the partner received the low lottery outcome, participant happiness was lower when they rather than the partner had chosen the lottery ( Figure 3D, H ).} when-partner-received-low-lottery⟧

When the partner received the low lottery outcome of a participant-chosen lottery, they would presumably feel let down because the participant could have chosen the better safe option.

As participants knew this, they were likely to feel ‘simple guilt’ ( Battigalli and Dufwenberg, 2007 ).

This behavioural effect (difference in happiness obtained when the partner received low lottery outcomes after participant rather than partner choices) is thus compatible with ‘simple guilt’, and we will thus refer to it as ‘guilt effect’.

The ‘guilt effect’ occurred whether the participant received the high lottery outcome (Study 1: t (39) = –3.58, p < 0.001, d = 0.56, BF 10 = 32; Study 2: t (43) = –2.68, p = 0.01, d = 0.4, BF 10 = 3.8) or the low lottery outcome (Study 1: t (39) = –3.39, p = 0.002, d = 0.54, BF 10 = 19; Study 2: t (43) = –3.58, p < 0.001, d = 0.54, BF 10 = 33.5).⟦>zach claim=db58b24c-22dd-45c9-aaa6-656fe937a3d0: @{The ‘guilt effect’ occurred whether the participant received the high lottery outcome (Study 1: t (39) = –3.58, p < 0.001, d = 0.56, BF 10 = 32; Study 2: t (43) = –2.68, p = 0.01, d = 0.4, BF 10 = 3.8) or the low lottery outcome (Study 1: t (39) = –3.39, p = 0.002, d = 0.54, BF 10 = 19; Study 2: t (43) = –3.58, p < 0.001, d = 0.54, BF 10 = 33.5).} guilt-effect-occurred-whether-participant⟧

In Study 2, we will compare individual ‘guilt effect’ values to individual guilt-related brain activation patterns (see end of the section on BOLD signal results).

Responsibility for choices did not influence happiness following positive lottery outcomes for the partner (both studies, all | t| < 1.3, p > 0.2, BF 10 < 0.2).⟦>zach claim=a22c0bd0-79d3-4ae6-b9d5-8227716b52f4: @{Responsibility for choices did not influence happiness following positive lottery outcomes for the partner (both studies, all | t| < 1.3, p > 0.2, BF 10 < 0.2).} responsibility-choices-not-influence-happiness⟧

In sum, in both studies, we found evidence that participants’ happiness was influenced by the outcomes of choices for their partner, especially by outcomes resulting from participant decisions, that is, partner outcomes for which participants were responsible.

Within these outcomes, participants felt worse following low lottery outcomes for the partner if those outcomes were consequences of their own choice rather than the partner’s, which we interpret as interpersonal guilt.

BOLD signal Next, we sought to uncover the neural mechanisms associated with our guilt effect and those involved in tracking consequences of participants’ decisions on their partner.

We analysed the BOLD responses of brain regions engaged during decision-making and at the time of receiving the outcomes of the choice using conventional as well as computational model-based analyses, using the fMRI data collected in Study 2. Brain regions engaged during social decision-making Given that our task involved several conditions and successive trial components, we aimed to first replicate previous results related to the neural correlates of decisions under economic risk.

We searched for brain regions engaged more when participants chose the risky instead of the safe option and found such responses in the bilateral ventral striatum (Cohen’s d = 0.72 and 0.85 in the left and right clusters, respectively; Figure 4A and Appendix 1—table 3 ), which replicates previous findings ( Cui et al., 2022 ; Preuschoff et al., 2006 ).⟦>zach claim=8eccf358-8cda-4fec-b608-34581e8afbe7: @{We searched for brain regions engaged more when participants chose the risky instead of the safe option and found such responses in the bilateral ventral striatum (Cohen’s d = 0.72 and 0.85 in the left and right clusters, respectively; Figure 4A and Appendix 1—table 3 ), which replicates previous findings ( Cui et al., 2022 ; Preuschoff et al., 2006 ).} bilateral-ventral-striatum-more-active⟧

Next, we aimed to identify the brain regions associated with social decision-making under risk and searched for regions more engaged during decisions in the Social condition compared to the Solo condition.

Three significant clusters of voxels were identified ( Figure 4B and Appendix 1—table 3 ), in the precuneus ( d = 0.79), the left temporo-parietal junction (TPJ; d = 0.59) and the medial prefrontal cortex (mPFC; d = 0.54).⟦>zach claim=ee017229-e215-452e-a719-932af2d716f2: @{Three significant clusters of voxels were identified ( Figure 4B and Appendix 1—table 3 ), in the precuneus ( d = 0.79), the left temporo-parietal junction (TPJ; d = 0.59) and the medial prefrontal cortex (mPFC; d = 0.54).} decisions-social-compared-solo-condition⟧

This also replicates previous findings associating this region with social decisions (e.g., Fareri et al., 2012 ; Jung et al., 2013 ; Nicolle et al., 2012 ; Ogawa et al., 2018 ; Piva et al., 2019 ).

Figure 4. BOLD responses.⟦>zach claim=8eccf358-8cda-4fec-b608-34581e8afbe7: @{Figure 4. BOLD responses.} bilateral-ventral-striatum-more-active⟧

( A ) Regions showing a greater response when participants chose the risky (lottery) rather than the safe option, irrespective of Social or Solo condition.⟦>zach claim=8eccf358-8cda-4fec-b608-34581e8afbe7: @{( A ) Regions showing a greater response when participants chose the risky (lottery) rather than the safe option, irrespective of Social or Solo condition.} bilateral-ventral-striatum-more-active⟧

( B ) Regions showing a greater response when participants chose for both themselves and their partner rather than just for themselves (Social > Solo).⟦>zach claim=ee017229-e215-452e-a719-932af2d716f2: @{( B ) Regions showing a greater response when participants chose for both themselves and their partner rather than just for themselves (Social > Solo).} decisions-social-compared-solo-condition⟧

( C ) Coefficients of linear mixed models (LMMs) indicate that two of these regions, precuneus and TPJ, were most active when participants chose the lottery in the Social condition.⟦>zach claim=1ba55a5a-b2c9-4b3d-94ff-48c91db9541f: @{( C ) Coefficients of linear mixed models (LMMs) indicate that two of these regions, precuneus and TPJ, were most active when participants chose the lottery in the Social condition.} only-precuneus-tpj-showed-positive⟧

R–S indicates results of LMMs based on the Risky–Safe response difference.

All coefficients and differences reported are significantly different from 0 (see Appendix 1—table 4 ).⟦>zach claim=no-assertion: @{All coefficients and differences reported are significantly different from 0 (see Appendix 1—table 4 ).} A pointer to Appendix 1—table 4 that the coefficients differ from zero, not a result stated here.⟧

( D ) Brain regions more active during receipt of the outcomes of lotteries than safe choices (all conditions).⟦>zach claim=0248969b-e2f4-4d2b-bddb-6536f7de95b9: @{( D ) Brain regions more active during receipt of the outcomes of lotteries than safe choices (all conditions).} during-receipt-lottery-versus-safe⟧

( E ) Coefficients of LMMs indicated that insula ROIs (see D) mirrored the guilt effect observed in our behavioural data: voxels here responded more to low lottery outcomes (L) for the partner when these resulted from participant’s rather than the partner’s choices, even when responses to high outcomes were subtracted (L–H).⟦>zach claim=3d98c046-68f2-491d-97c0-2b75571e838c: @{( E ) Coefficients of LMMs indicated that insula ROIs (see D) mirrored the guilt effect observed in our behavioural data: voxels here responded more to low lottery outcomes (L) for the partner when these resulted from participant’s rather than the partner’s choices, even when responses to high outcomes were subtracted (L–H).} insula-rois-responded-more-low⟧

All coefficients and differences reported are significantly different from 0 (see Appendix 1—table 6 ).⟦>zach claim=no-assertion: @{All coefficients and differences reported are significantly different from 0 (see Appendix 1—table 6 ).} A pointer to Appendix 1—table 6 that the coefficients differ from zero, not a result stated here.⟧

( F ) A mass-univariate, voxel-wise analysis showed a compatible result: A cluster of voxels within the left insula ROI showed higher responses to low lottery outcomes for the partner if these resulted from participant rather than partner choices.⟦>zach claim=071373be-51ef-4d9f-8101-1379e5fae233: @{( F ) A mass-univariate, voxel-wise analysis showed a compatible result: A cluster of voxels within the left insula ROI showed higher responses to low lottery outcomes for the partner if these resulted from participant rather than partner choices.} mass-univariate-voxel-wise-analysis-found-small⟧

( G ) Activation in bilateral ventral striatum explained by a computational model-based regressor coding participant rewards.⟦>zach claim=eff3fbf5-c0c4-46e8-b43e-7b5e6208a969: @{( G ) Activation in bilateral ventral striatum explained by a computational model-based regressor coding participant rewards.} manipulation-check-bilateral-ventral-striatum⟧

( H ) Within brain regions sensitive to outcomes of risky choices, one cluster in the left superior temporal sulcus region showed a higher response to partner reward prediction errors resulting from participant rather than partner choices.⟦>zach claim=6d82bb56-0a43-4200-b2cd-15d288f5e004: @{( H ) Within brain regions sensitive to outcomes of risky choices, one cluster in the left superior temporal sulcus region showed a higher response to partner reward prediction errors resulting from participant rather than partner choices.} one-cluster-left-sts-responded⟧

( I ) Response in this cluster to the computational-model-based regressors coding participant reward prediction resulting from participant and partner choices, for both sessions of the experiment.

All results shown survive thresholding at p < 0.05 corrected for multiple comparisons at the cluster level, based on a voxel-wise uncorrected threshold of p < 0.001. Colours in panels A–I indicate T values.⟦>zach claim=db58b24c-22dd-45c9-aaa6-656fe937a3d0: @{All results shown survive thresholding at p < 0.05 corrected for multiple comparisons at the cluster level, based on a voxel-wise uncorrected threshold of p < 0.001. Colours in panels A–I indicate T values.} guilt-effect-occurred-whether-participant⟧

In C and E, bars indicate the estimated coefficients.

In C, E, and I, error bars are 95% confidence intervals.

To assess whether activations in these five regions were sensitive to the interaction between choice and condition using a sensitive method, we analysed the responses in all the voxels in these regions using LMMs.

Our model included factors Choice ( Safe or Risky ), Condition ( Social , Partner or Solo ), Run (run one or two), and ROI (left and right ventral striatum, mPFC, precuneus, and TPJ), with all interactions.

We also tested the same model without the factor Run , but this fitted less well (higher BIC) and was thus discarded.

As all factors and interactions were significant (data not shown for sake of brevity), we applied the same models to each region separately ( Appendix 1—table 4 ).⟦>zach claim=no-assertion: @{As all factors and interactions were significant (data not shown for sake of brevity), we applied the same models to each region separately ( Appendix 1—table 4 ).} Narration of how the region-wise models were applied, with data not shown; asserts no result.⟧

All regions showed a significant interaction between Condition and Choice , prompting us to examine these responses in detail.

As we were particularly interested in the response to Risky decisions, we subtracted from it the responses to Safe decisions (labelled as ‘(R–S)’ in Figure 4C ) and compared this difference between (1) Social and Solo conditions, and (2) between the Social and Partner conditions using LMMs.⟦>zach claim=1ba55a5a-b2c9-4b3d-94ff-48c91db9541f: @{As we were particularly interested in the response to Risky decisions, we subtracted from it the responses to Safe decisions (labelled as ‘(R–S)’ in Figure 4C ) and compared this difference between (1) Social and Solo conditions, and (2) between the Social and Partner conditions using LMMs.} only-precuneus-tpj-showed-positive⟧

As above, we tested these models both with and without the factor Run and associated interaction, and we report the best-fitting model (see Appendix 1—table 5 and Appendix 1—table 6 ).⟦>zach claim=no-assertion: @{As above, we tested these models both with and without the factor Run and associated interaction, and we report the best-fitting model (see Appendix 1—table 5 and Appendix 1—table 6 ).} Narration of model selection with and without the Run factor, pointing to two appendix tables.⟧

Only the precuneus and TPJ showed positive differences in both comparisons ( Figure 4C ), indicating that these regions were most active when participants chose the lottery in the Social condition, the critical situation in which participants assume responsibility over others.⟦>zach claim=1ba55a5a-b2c9-4b3d-94ff-48c91db9541f: @{Only the precuneus and TPJ showed positive differences in both comparisons ( Figure 4C ), indicating that these regions were most active when participants chose the lottery in the Social condition, the critical situation in which participants assume responsibility over others.} only-precuneus-tpj-showed-positive⟧

Brain regions engaged during receipt of outcomes Next, we focused on responses during choice outcomes.

A cluster of voxels more active during receipt of lottery outcomes than outcomes of safe choices was identified in the bilateral anterior insula, dorsal mPFC (dmPFC), right superior temporal sulcus (STS), bilateral ventral striatum, right dorsolateral prefrontal cortex, and bilateral inferior parietal lobe ( Figure 4D ; for details including effect sizes, see Appendix 1—table 7 ).⟦>zach claim=0248969b-e2f4-4d2b-bddb-6536f7de95b9: @{A cluster of voxels more active during receipt of lottery outcomes than outcomes of safe choices was identified in the bilateral anterior insula, dorsal mPFC (dmPFC), right superior temporal sulcus (STS), bilateral ventral striatum, right dorsolateral prefrontal cortex, and bilateral inferior parietal lobe ( Figure 4D ; for details including effect sizes, see Appendix 1—table 7 ).} during-receipt-lottery-versus-safe⟧

Except for the STS, all these regions have been previously associated with processing of risk and/or ambiguity ( Wu et al., 2021 ).

In our definition, guilt occurs due to responsibility for low lottery outcomes for the partner.

To search for regions involved in this situation, we again analysed voxel responses using LMMs.

Our first model included factors Lottery outcome ( High or Low ), Condition ( Social or Partner ), Run (run one or two), and ROI (left and right ventral striatum, left and right insula, left and right parietal cortex, dmPFC, right prefrontal, and right middle temporal regions), with all interactions.

We also tested the same model without the factor Run , which fitted less well and was discarded.

As in the analysis of responses to decisions above, all factors and interactions were significant (data not shown for sake of brevity), and we applied the same models to each region separately ( Appendix 1—table 8 ).⟦>zach claim=no-assertion: @{As in the analysis of responses to decisions above, all factors and interactions were significant (data not shown for sake of brevity), and we applied the same models to each region separately ( Appendix 1—table 8 ).} Narration that all factors and interactions were significant (data not shown), then how the per-region models were fitted.⟧

To identify regions likely to be involved in the guilt effect, we selected those satisfying two conditions: higher activity in the Social compared to the Partner condition, and a significant Social:LowOutcome interaction.

This procedure revealed the insulae ( Figure 4E ) and the right middle temporal cortex (trend significant Social vs. Partner difference in the latter region).⟦>zach claim=3d98c046-68f2-491d-97c0-2b75571e838c: @{This procedure revealed the insulae ( Figure 4E ) and the right middle temporal cortex (trend significant Social vs. Partner difference in the latter region).} insula-rois-responded-more-low⟧

To test if these regions responded more to LowOutcomes in the Social condition, we ran additional models on the responses to Low lottery outcomes only (labelled as ‘(L) Social > Partner’ in Figure 4E ), and on this response minus the response to High lottery outcomes (labelled as ‘(L–H) Social > Partner’ in Figure 4E ); and indeed, the insulae showed the sought-after effect (see Appendix 1—table 9 , Appendix 1—table 10 , and Figure 4E ).⟦>zach claim=3d98c046-68f2-491d-97c0-2b75571e838c: @{To test if these regions responded more to LowOutcomes in the Social condition, we ran additional models on the responses to Low lottery outcomes only (labelled as ‘(L) Social > Partner’ in Figure 4E ), and on this response minus the response to High lottery outcomes (labelled as ‘(L–H) Social > Partner’ in Figure 4E ); and indeed, the insulae showed the sought-after effect (see Appendix 1—table 9 , Appendix 1—table 10 , and Figure 4E ).} insula-rois-responded-more-low⟧

Thus, activation in our insula ROIs increased in situations during which participants experienced guilt for low outcomes impacting their partner, compared to similar outcomes resulting from the partner’s choices.

To attempt to confirm these results with a classic fine-grained mass-univariate voxel-wise analysis, we searched, within regions responding more to outcomes of risky compared to safe choices, for higher responses to low lottery outcomes for the partner following participant choices compared to the same outcomes resulting from partner choices.

We found a weak response in a small cluster within the left anterior insula (peak T = 3.95, d = 0.59, 22 voxels, peak intensity at [–28 24 –4]; Figure 4F ).⟦>zach claim=ee017229-e215-452e-a719-932af2d716f2: @{We found a weak response in a small cluster within the left anterior insula (peak T = 3.95, d = 0.59, 22 voxels, peak intensity at [–28 24 –4]; Figure 4F ).} decisions-social-compared-solo-condition⟧

Given the documented association between anterior insula and guilt (see Introduction), we proceeded to test whether this result survived correction for family-wise errors due to multiple comparisons restricted to the left anterior insula grey matter [defined anatomically and thus independently from our findings, as the anterior short gyrus, middle short gyrus, and anterior inferior cortex in an anatomical maximum probability map ( Faillenot et al., 2017 )].

This correction resulted in a p value of 0.024. This result, although it is only a small effect in a small cluster, is consistent with the mixed model analysis reported earlier.

Neural correlates of responsibility for partner reward prediction errors revealed by computational model-based analysis Next, we attempted to explain BOLD responses using predictions of the ‘Responsibility ’ computational model (see Behaviour/Computational modelling of happiness data, above).

We used the model to create expected BOLD responses for each participant (see Methods) and as a manipulation check searched for responses in ventral striatum evoked by participant rewards ( O’Doherty et al., 2004 ; O’Doherty et al., 2007 ).

We found that activation in bilateral ventral striatum indeed increased with the amount of expected certain rewards and the expected values of chosen lotteries (left: p FWE = 0.002, T = 5.63, d = 0.75, Z = 5.41, 110 voxels, peak at MNI [–14 8 –8], right: p FWE = 0.005, T = 5.46, d = 0.70, Z = 5.26, 80 voxels, peak at MNI [10 10 −4], correction for multiple tests applied across the whole brain; Figure 4G ).⟦>zach claim=eff3fbf5-c0c4-46e8-b43e-7b5e6208a969: @{We found that activation in bilateral ventral striatum indeed increased with the amount of expected certain rewards and the expected values of chosen lotteries (left: p FWE = 0.002, T = 5.63, d = 0.75, Z = 5.41, 110 voxels, peak at MNI [–14 8 –8], right: p FWE = 0.005, T = 5.46, d = 0.70, Z = 5.26, 80 voxels, peak at MNI [10 10 −4], correction for multiple tests applied across the whole brain; Figure 4G ).} manipulation-check-bilateral-ventral-striatum⟧

We thus used this model to search for voxels responding more to partner reward prediction errors resulting from participant rather than partner choices, within the regions sensitive to outcomes of risky choices.

We found this effect in one cluster within the left STS (p FWE = 0.022, T = 4.70, d = 0.53, Z = 4.57, 100 voxels, peak at MNI [−52 –32 0]; Figure 4H ).⟦>zach claim=6d82bb56-0a43-4200-b2cd-15d288f5e004: @{We found this effect in one cluster within the left STS (p FWE = 0.022, T = 4.70, d = 0.53, Z = 4.57, 100 voxels, peak at MNI [−52 –32 0]; Figure 4H ).} one-cluster-left-sts-responded⟧

Responses to the computational-model-based regressors coding for partner reward prediction errors resulting from participant and from partner choices in this cluster are shown in Figure 4I .

This finding suggests that this region of the left STS tracks a partner’s unexpected outcomes less when they do not follow from the participant’s decisions.

Functional connectivity Functional connectivity analyses have revealed differences in networks engaged by social and self-only choices ( Jung et al., 2013 ; Ogawa et al., 2018 ), interactions between midbrain and anterior cingulate during compensation for guilt ( Yu et al., 2014 ), and links between insula connectivity and responsibility aversion ( Edelson et al., 2018 ).

We hypothesized that connectivity with regions that showed guilt- and responsibility-related responses during the outcome phase (see previous paragraph) might change depending on whether participants made decisions for themselves only or for themselves and their partner, and depending on the type of choice ( Safe or Risky ).

To this end, we used the left insula and left STS as seed regions for whole-brain seed-to-voxel psychophysiological interaction (PPI) analyses (see Methods), to search for connectivity changes, during the choice phase of the trial, as a function of Condition and Choice (i.e., voxels with a significant interaction to Condition by Choice ).

The first analysis revealed a cluster in the right IFG whose connectivity to the insula (the seed region) was highest when participants made Risky choices for themselves and Safe choices for both players (p FWE = 0.020, T = 4.34, d = 0.80, Z = 4.21, 115 voxels, peak at MNI [46 16 22]; Figure 5 ).⟦>zach claim=5d2ef557-c6ec-4052-82c2-0b8e4c297e01: @{The first analysis revealed a cluster in the right IFG whose connectivity to the insula (the seed region) was highest when participants made Risky choices for themselves and Safe choices for both players (p FWE = 0.020, T = 4.34, d = 0.80, Z = 4.21, 115 voxels, peak at MNI [46 16 22]; Figure 5 ).} functional-connectivity-between-left-anterior⟧

A smaller cluster in the left IFG showed the same effect but did not survive correction for multiple tests (p uncorrected = 0.001, T = 4.08, Z = 3.97, 38 voxels, peak at MNI [−38 –2 26]).⟦>zach claim=1b9b2289-3ccd-461b-92c8-4f41b5632178: @{A smaller cluster in the left IFG showed the same effect but did not survive correction for multiple tests (p uncorrected = 0.001, T = 4.08, Z = 3.97, 38 voxels, peak at MNI [−38 –2 26]).} left-ifg-insula-connectivity-subthreshold — A left-IFG cluster with the insula seed showed the same connectivity effect but did not survive correction; no claim records this sub-threshold cluster.⟧

The second analysis revealed a smaller cluster in the left IFG that did not survive corrections for multiple tests, where connectivity with the left STS (the seed region) showed the opposite pattern: connectivity was highest when participants made Safe choices for themselves and Risky choices for both players (p uncorrected = 0.001, T = 4.44, Z = 4.30, 35 voxels, peak at MNI [–48 14 6]; Figure 5—figure supplement 1 ).⟦>zach claim=c6a31726-78b5-45f6-861c-b2a3b8ac3249: @{The second analysis revealed a smaller cluster in the left IFG that did not survive corrections for multiple tests, where connectivity with the left STS (the seed region) showed the opposite pattern: connectivity was highest when participants made Safe choices for themselves and Risky choices for both players (p uncorrected = 0.001, T = 4.44, Z = 4.30, 35 voxels, peak at MNI [–48 14 6]; Figure 5—figure supplement 1 ).} left-ifg-cluster-showed-opposite — The claim states this STS-seed left-IFG cluster with the opposite (Safe-self/Risky-both) pattern that did not survive correction.⟧

Figure 5. Changes in functional connectivity between the left anterior insula (seed) and a cluster in the right inferior frontal gyrus at the time of the choice as a function of condition (Social vs. Solo) and choice (Risky or Safe).⟦>zach claim=5d2ef557-c6ec-4052-82c2-0b8e4c297e01: @{Figure 5. Changes in functional connectivity between the left anterior insula (seed) and a cluster in the right inferior frontal gyrus at the time of the choice as a function of condition (Social vs. Solo) and choice (Risky or Safe).} functional-connectivity-between-left-anterior⟧

In the righthand panel, dots are data of individual participants, the markers represent means, and error bars indicate 95 confidence intervals about the mean.

Figure 5—figure supplement 1. Functional connectivity with the left TD-model-defined superior temporal sulcus (STS; seed) during choices in Solo and Social conditions.⟦>zach claim=no-assertion: @{Figure 5—figure supplement 1. Functional connectivity with the left TD-model-defined superior temporal sulcus (STS; seed) during choices in Solo and Social conditions.} A figure-supplement title naming what is plotted rather than stating a finding.⟧

Connectivity between the left STS (seed) and a cluster in the left inferior frontal gyrus that did not survive corrections for multiple tests, where connectivity with the left STS (the seed region) showed the opposite pattern: connectivity was highest when participants made Safe choices for themselves and Risky choices for both players (p uncorrected = 0.001, T = 4.44, Z = 4.30, 35 voxels, peak at MNI [–48 14 6]).⟦>zach claim=c6a31726-78b5-45f6-861c-b2a3b8ac3249: @{Connectivity between the left STS (seed) and a cluster in the left inferior frontal gyrus that did not survive corrections for multiple tests, where connectivity with the left STS (the seed region) showed the opposite pattern: connectivity was highest when participants made Safe choices for themselves and Risky choices for both players (p uncorrected = 0.001, T = 4.44, Z = 4.30, 35 voxels, peak at MNI [–48 14 6]).} left-ifg-cluster-showed-opposite — Restates the STS-seed left-IFG opposite-pattern connectivity that the claim records as not surviving correction.⟧

In the righthand panel, dots are data of individual participants, the markers represent means, and error bars indicate 95 confidence intervals about the mean.

Comparison with multivariate neural guilt signature ( Yu et al., 2020 ) A recent study by Yu and colleagues re-analysed the results of two previous neuroimaging studies of guilt and obtained a neural multivariate guilt-related brain signature (GRBS) ( Yu et al., 2020 ).

As the GRBS and code to compare neural responses to it is freely available at GitHub ( Wager, 2019 ), we compared this brain signature to the neural responses obtained in our task (response to low partner outcomes resulting from participant vs. partner responses).

The dot products between individual responses and the GRBS varied between –40.1 and 36.7, but overall these values were positive (mean = 5.22; median = 6.97; sign test: p = 0.017; Cliff’s Delta = 0.4 = medium effect size; data are not normally distributed).⟦>zach claim=079ebf00-e69a-40ab-863a-03d6a7d34129: @{The dot products between individual responses and the GRBS varied between –40.1 and 36.7, but overall these values were positive (mean = 5.22; median = 6.97; sign test: p = 0.017; Cliff’s Delta = 0.4 = medium effect size; data are not normally distributed).} dot-products-between-individual-neural⟧

We assessed whether inter-individual differences in these dot product values correlated with the behavioural guilt responses, but did not find a significant association [ Spearman’s Rho = –0.058, p = 0.725].⟦>zach claim=4dc9ae47-d845-475f-931c-7428487c66de: @{We assessed whether inter-individual differences in these dot product values correlated with the behavioural guilt responses, but did not find a significant association [ Spearman’s Rho = –0.058, p = 0.725].} individual-grbs-dot-product-values-not⟧

discussion

Discussion We report findings from an experiment on social responsibility and guilt in risky economic decisions, and their neural correlates.

Being responsible for choosing a lottery that yielded a low outcome for a partner made our participants feel worse than witnessing the same outcome resulting from their partner’s choice, which we interpret as interpersonal guilt; although we note that we have not asked participants specifically about which emotion they felt in these situations.

Activation in the left anterior insula (aIns) reflected this effect, replicating previous associations between this region and feelings of guilt.

Whole-brain activation patterns also resembled a neurometric marker of guilt ( Yu et al., 2020 ).

Connectivity between aIns and the right IFG varied depending on whether participants chose the risky or safe option and whether only the participant or both themselves and the partner were affected by the outcome of this choice, suggesting that this part of prefrontal cortex is sensitive to guilt-related information during social choices.

Computational models explained trial-by-trial variations in momentary happiness during the task; the best-fitting model differentiated between partner reward prediction errors resulting from participant and partner choices, indicating that the impact of outcomes for the partner on participants’ happiness varied depending on who chose.

This confirms the importance of responsibility in determining the emotional consequences of risky social choices. fMRI analyses based on this computational model identified a left STS region responding more to partner reward prediction errors resulting from participant rather than partner choices.

This suggests a critical role of the STS in monitoring the consequences of one’s risky decisions on others, an essential social cognitive function.

A last analysis showed that connectivity between this STS region and the right IFG varied depending on condition and choice, suggesting that the information processed in the STS is relayed to the IFG during social decisions.

Our findings add to current understanding of the neural mechanisms underlying responsibility for, and guilt evoked by, the outcomes on others of social decisions under economic risk.

We used several approaches to compare choices made for self only or for both participant and partner.

Participants made slightly more risk-seeking choices when deciding for themselves than for both themselves and the partner in Study 1, but this difference disappeared in Study 2. The ρ parameter on which this finding in Study 1 is based could only be estimated in a minority of participants due to a relatively low number of trials, which suggests that this finding may not be very reliable.

The simpler and more robust method (evaluation of a risk premium) showed no difference in risk aversion across conditions in either study.

Overall, we believe that we do not have strong evidence of differences in risk preferences across conditions.

Our findings are not unexpected given previous work.

At least two studies reported that being responsible for somebody else’s payoffs increases risk aversion ( Fareri et al., 2022 ; Pahlke et al., 2015 ).

However, participants in a recent neuroimaging study were more loss averse when choosing for themselves rather than known other people ( Arioli et al., 2023 ).

Indeed, recent large meta-analyses comparing risky choices for oneself vs. for others report either no difference in risk preferences ( Batteux et al., 2019 ) or a small shift towards more risky decisions for others ( Polman and Wu, 2020 ), with large variations across studies.

Personal closeness to the others for which we decide seems to reduce differences in risk preferences ( Fareri et al., 2022 ; Zhang et al., 2017 ), as does making decisions for self before making decisions for others ( Ifcher and Zarghamee, 2020 ).

In our study, decisions never only affected the partner, which most likely ‘watered down’ any differences in risk attitude in decisions for oneself vs. for someone else.

We also created a friendly prosocial environment in which people felt closer to each other than two strangers would (see Experiment Partner in Methods), and interleaved decisions for self only vs. for self and other.

Similar risk preferences in decisions for self vs. self and other are thus unsurprising.

The fact that participants made similarly risky decisions for themselves or for both during the fMRI study was to our advantage, because it made the neural signals evoked in these situations more comparable.

Anterior insula (aIns) response, particularly in the left hemisphere, was highest when partners received low lottery outcomes resulting from participants’ risky choices.

This replicates a large body of evidence associating aIns with feelings of guilt evoked during social decisions (see Introduction).

Because we have neither asked our participants specifically what they felt in these situations, nor specifically whether they experienced guilt, we cannot exclude the possibility that they have instead or in addition felt empathy for their partner, a feeling of failure or bad luck, or some other emotion.

The aIns region has also been associated with empathy for negative emotions such as disgust ( Wicker et al., 2003 ) or pain ( Gu et al., 2012 ; Lamm et al., 2011 ), with affective empathy during charitable giving ( Tusche et al., 2016 ), and more generally with emotion awareness ( Bird et al., 2010 ; Gu et al., 2013 ).

These functions might be supported by interoception ( Craig, 2002 ): posterior insula is thought to receive interoceptive information from the body and to pass it on to the anterior insula for integration with sensory, emotional, cognitive, and motivational signals from other regions ( Bid Craig, 2009 ; Rogers-Carter and Christianson, 2019 ).

This integration would allow one to represent one’s own, as well as estimates of other people’s, feelings and bodily states and allow error-based learning based on these ( Lamm and Singer, 2010 ; Rogers-Carter and Christianson, 2019 ; Singer et al., 2009 ).

Our finding of changing functional connectivity between aIns and prefrontal cortex as a function of decision taken and experimental condition during choice may reflect changing emotional information transfer between these structures depending on choice and social context.

How we feel when we witness our decisions’ consequences on others is an important signal to consider when attempting to make good social decisions.

Unsurprisingly, lesions of aIns are associated with reduced altruistic attitudes ( Chau et al., 2018 ), individuals with higher levels of psychopathic traits show reduced modulation of aIns response to anticipated guilt ( Seara-Cardoso et al., 2016 ), and show reduced guilt aversion ( Gong et al., 2019 ).

Deciding for both self and partner rather than just for oneself evoked increased activations in precuneus, left TPJ, and dmPFC, areas classically associated with social cognition ( Frith and Frith, 2006 ; Schurz et al., 2014 ; Van Overwalle, 2009 ), and also with feelings of guilt (for meta-analyses, see Gifuni et al., 2017 ; Piretti et al., 2023 ).

This replicates previous findings of engagement of dmPFC, STS, and/or TPJ when making decisions involving others ( Jung et al., 2013 ; Nicolle et al., 2012 ; Ogawa et al., 2018 ; Piva et al., 2019 ).

Interestingly, our computational model-based analysis revealed a cluster of voxels in the left STS that responded positively to partner prediction errors, but only when these resulted from decisions made by the participant instead of the partner.

This is congruent with associations between STS and cognitive perspective taking during charitable giving ( Tusche et al., 2016 ), representation of other people’s interests during altruistic choice ( Hutcherson et al., 2015 ), or more generally with mentalizing computations during strategic social choice ( Carter et al., 2012 ; Hampton et al., 2008 ; Hill et al., 2017 ).

The features of our experimental design (direct contrast of similar decisions made by the participant and partner; independent lottery outcomes for self and other; quantification of the consequences of decisions through variations of momentary happiness) allowed us to identify this very specific neural signal indicative of a neural sensitivity to the consequences of one’s own actions on others, whether positive or negative.

A neural signal coding partner reward prediction errors resulting from one’s decisions seems essential to guide social decisions and as a basis for empathic concern for the people influenced by our actions.

Anecdotally, we found a weak functional connection between this left STS cluster and left aIns that varied as a function of choice and experimental condition; this is interesting because connections between a region in the left TPJ and aIns have been shown to vary depending on people’s tendency to seek or avoid responsibility for others ( Edelson et al., 2018 ).

Given the roles of the left STS/TPJ and aIns in guilt and social decisions, it is not surprising to find information exchanges between these structures.

More experiments investigating specific computations in which these regions are involved (as in, e.g., Charpentier and O’Doherty, 2018 ; Konovalov et al., 2021a ; Konovalov and Ruff, 2021b ) are likely to help understand the neural mechanisms by which guilt and responsibility influence social decision-making.

There are several limitations to our study.

The first limitation is that the partner’s decisions were in fact taken by an algorithm.

This was not communicated to the participants and was thus a case of deception, which is inadmissible in behavioural economics.

To our defence, this study was planned as a social neuroscience study and was executed before we had taken into account the practices of behavioural economists.

However, this approach did have the advantage of eliminating potentially complex iterative reciprocal influences of decisions and outcomes between the players, which would have led to much more complicated emotional states and decisions that would have been difficult to understand and model.

Thus, from an analytical point of view, our approach might have actually allowed a cleaner comparison between decisions and outcomes than if the partner had really made the decisions.

The fact that partner outcomes also influenced participants’ momentary happiness demonstrates that participants were not emotionally detached from the consequences of their actions on their partner.

This finding is, of course, essential for the validity of our results and suggests that the effects could get stronger with more direct interactions.

Therefore, although we will abstain from using deceptive practices when pursuing this research, we believe our findings to be valid.

Another limitation is that we have not asked participants t