A corticostriatal pathway mediating self-efficacy enhancement

Forming positive beliefs about one’s ability to perform challenging tasks, often termed self-efficacy, is fundamental to motivation and emotional well-being. Self-efficacy crucially depends on positive social feedback, yet people differ in the degree to which they integrate such feedback into self-beliefs (i.e., positive bias). While diminished positive bias of this sort is linked to mood and anxiety, the neural processes by which positive feedback on public performance enhances self-efficacy remain unclear. To address this, we conducted a behavioral and fMRI study wherein participants delivered a public speech and received fictitious positive and neutral feedback on their performance in the MRI scanner. Before and after receiving feedback, participants evaluated their actual and expected performance. We found that reduced positive bias in updating self-efficacy based on positive social feedback associated with a psychopathological dimension reflecting symptoms of anxiety, depression, and low self-esteem. Analysis of brain encoding of social feedback showed that a positive self-efficacy update bias associated with a stronger reward-related response in the ventral striatum (VS) and stronger coupling of the VS with a temporoparietal region involved in self-processing. Together, our findings demarcate a corticostriatal circuit that promotes positive bias in self-efficacy updating based on social feedback, and highlight the centrality of such bias to emotional well-being.


Validation of perceived valence of the rating scale
In order to characterize the perceived valence of different values on the 11-pt scale that we used for the speech evaluation, we conducted an online experiment on undergraduate psychology students from Tel-Aviv University (n=54; Meanage±SD: 22.65 ± 3.16 years, 37 females).
Throughout the experiment, participants were first instructed to imagine they have just delivered a public speech, and that professional judges were about to evaluate their performance.Moreover, they were told that they would be evaluated on a 0-10 scale and in comparison to a sample of 100 people that are similar to them in terms of age and education level.Furthermore, they were instructed that on the 0-10 scale, 0 marks the lowest possible score and is thus labeled "much below average"; 5 marks the average performance and is labeled "average" accordingly; and 10 is the maximal score and is therefore labeled "far above average".Next, the rating scale was presented to participants with numerical indications ranging from 0-10 and the 3 labels mentioned above were positioned in their matching locations.Participants were instructed to mark the range of scores that they believe would be experienced by them as either positive, neutral or negative.Supplementary Figure 1 depicts the association of each score with either positive, neutral or negative valence.Note that more than 50% of the participants associated each of the scores 7-10 with positive valence and scores 5-6 with neutral valence, in accordance with our definition of the positive and neutral feedback conditions.However, the score 4 was associated more often with negative valence than with neutral valence.Regarding this matter, note that the scale used in this experiment differed from the one we actually used in the social feedback experiment in several terms.First, the scale in the online experiment referred to a hypothetical sample of similar others, whereas the scale in the feedback experiment lacked such social reference.Second, the scale in the online experiment included numerical indications and different labels, which may have influenced participants' ratings negatively (e.g., the score "4" or the label "average", which did not appear in the feedback experiment, may have enhanced the negativity of intermediate scores).Nonetheless, these results confirm that the two conditions differed in terms of their perceived valence.

Posterior predictive checks
To validate the accurate performance of the winning model, we implemented posterior predictive checks assessing the correspondence between the actual data and simulated data.To this end, we first generated posterior distributions of data that were simulated based on the estimated model parameters.Thus, for each of the five rating types this procedure resulted in 10,000 posterior samples for each of the 2,000 ratings (i.e.50 participants x 40 trials).Since our main parameters of interest (Ω  ) reflected the correlation between pairs of ratings for each subject, we computed the Pearson correlation between pre-to-post feedback simulated ratings for each subject on each posterior draw.This was done specifically for correlations between the post-speech self-evaluation and post feedback self-evaluation; as well as for pre-speech self-efficacy and post-feedback selfefficacy.This resulted in a posterior distribution of 10,000 correlation coefficients for each subject, from which we extracted the median.Next, we correlated these simulated correlation coefficients with the actual correlation coefficients describing the coupling between the original ratings across participants, using Pearson correlations (see Supplementary Figure 3).

Descriptive statistics of positive and neutral components of the update bias indices
The update bias indices of interest were computed as the difference between the feedbackrating correlations given positive and neutral feedback.However, as range restriction is known affect the covariance between variables, it might be the case that the covariance on positive vs. neutral feedback trials varied due to the differential ranges we identified with each of these conditions (i.e.4-6 for neutral vs. 7-10 for positive).To address this issue, we computed descriptive statistics (mean and standard deviation) for the positive and neutral components for each update bias indices.Descriptive statistic were similar for both the self-evaluation update bias (positive component [M±SD]: 0.43±0.15;neutral component: 0.41±0.16)and the self-efficacy update bias (positive component [M±SD]: 0.39±0.14;neutral component: 0.39±0.18).

Association of self-efficacy update bias with feedback memory
To assess whether self-efficacy update bias was related do differential memory of judges' positive vs. neutral feedback, we performed a control analysis where we tested the correlation between update bias and memory errors.We first computed the Mean Squared Error (MSE) of the difference between the actual and recollected feedback scores under each of the two conditions for each participant.We then subtracted between the memory MSE of each condition, such that   Δ =    −    .We tested the correlation between   Δ and self-efficacy update bias using partial Pearson correlations, where we controlled for individual differences in between-conditions' mean prediction error difference.The correlation between self-efficacy update bias and   Δ was not significant (r(47)=-.21,p=.15).

Questionnaire-based components of psychopathology related symptoms
Association of self-negativity with positive vs. neutral components of self-efficacy update bias One of our main findings was that higher scores on the self-negativity component associated with less positive self-efficacy update bias.The latter was computed as the difference between updates made for positive vs. neutral feedback (see Equation (4) in main text), and this raises the question of whether this correlation was driven more strongly by either positive or neutral feedback.To test this, we examined the separate correlations of self-negativity with the distinct positive and neutral components of self-efficacy update bias (i.e.Ω   ( 3 , 5 ) and Ω   ( 3,  5 ) , respectively).We used partial Pearson correlations, as we wanted to control for individual differences in the corresponding mean deviance between expected and actual feedback scores (i.e.prediction error) in each of the conditions.None of these correlations were significant (correlation with positive component of self-efficacy update bias: r(47)=.07,p=.65; correlation with neutral component of self-efficacy update bias: r(47)=.15,p=.31).Thus, it seems that the association between self-negativity and self-efficacy update bias was mainly related to the difference between updates made for positive vs. neutral feedback conditions.

Association of self-negativity with speech-induced distress
To test whether self-negativity also associated with changes in subjective distress that were induced by the public speaking task, we correlated the self-negativity scores with the delta in the reported distress between each pair of adjacent ratings.Since we collected five distress ratings, we had four indices of speech-induced distress changes: (1) Difference between distress experienced at baseline (i.e.prior to speech announcement) and just before the speech performance.We assumed that this difference represented anticipatory related distress.(2) Difference between distress experienced just before the speech performance and during the speech performance.We assumed that this difference represented distressful reactivity to the speech.(3) Difference between distress experienced during the speech performance and immediately after its termination.We assumed that this difference represented immediate recovery from the speech-induced distress.(4)   Difference between distress experienced immediately after speech termination and subjective distress reported ~35 min later.We assumed that this difference represented long-term sustainment of-, or recovery from-, speech-induced distress.We found that higher self-negativity was associated with less immediate recovery from speech-induced distress (i.e.difference (3); Spearman correlation: (48)=.401,p=.004, pFDR<.05).This suggests that participants who scored higher on this psychopathology related domain, were affected more negatively by the speech performance.None of the correlations between self-negativity and the remaining rating changes were statistically significant (with (1): (48)=.07,p=.62; with (2): (48)=.02,p=.92; with (4): (48)=-.15,p=.29).

Brain activity in the auxiliary fMRI tasks
Prior to conducting the similarity analysis between the social feedback task and the auxiliary tasks, we assured that the latter activated our ROIs.Indeed, VS activation was evident in response to winning vs. losing money in the monetary reward task (Supplementary Figure 5).In the guided self-evaluation task, judging whether traits were descriptive of oneself vs. performing a lexical control task on the same traits activated extensive portions of the default-mode network, including the VMPFC (Supplementary Figure 5).

Association of self-efficacy update bias with VS encoding of judges' feedback positivity: complementary analyses
Association of ventral striatum activity with positive vs. neutral components of self-efficacy update bias A central result reported in the main text is the association between self-efficacy update bias and VS activity during encoding of judges' positivity.One question surrounding this result is whether this correlation was driven specifically by the delta between update parameters obtained for positive vs. neutral feedback conditions; or whether the positive and neutral components of the update bias (i.e.Ω  ( 3 , 5 ) and Ω   ( 3,  5 ) , respectively) contributed separately to this brain-behavior correlation as well.To inspect this, we computed the covariance between the positive and neutral components of self-efficacy update bias and brain encoding of social feedback positivity in two separate multiple regression analysis, and restricted this analysis to a mask covering the bilateral VS.To control for the potential influence of individual differences in prediction error on both bias parameters and brain activity, we regressed out the mean prediction error during positive or neutral conditions from the corresponding model parameters; and also entered them as betweensubject covariates in the relevant regression analysis.To detect significant covariance, we set the statistical threshold at voxel-level p<.001 and small-volume corrected family-wise error (SVC FWE) p<.05, as we did in the analysis in the main text.This analysis did not reveal any significant covariance of either the positive or neutral components of self-efficacy update bias with VS activity.Nonetheless, we explored the association of the positive or neutral components of selfefficacy update bias with VS activity in a 4mm sphere centered around coordinates in the right VS wherein significant covariance with self-efficacy update bias was observed (x=14, y=10, z=-7).
We controlled for individual differences in mean prediction error during positive or neutral conditions here as well, by using partial Pearson correlations.These analyses showed only nonsignificant trends, suggesting that right VS activity correlated positively with the positive component of self-efficacy update bias (r(45)=.2,p=.18) and negatively with the neutral component of self-efficacy update bias (r(45)=-.13,p=.39).Thus, the association we found between self-efficacy update bias and right VS activity encoding judges' positivity was related primarily to the difference between updates made for positive vs. neutral feedback conditions.
Controlling for expected feedback score at the trial-level A central confound in our experimental design was the deviance between expected and actual social feedback scores (i.e.prediction error) on each trial in the fMRI social feedback task.This factor varied both across participants and across the positive and neutral conditions.Thus, we controlled for the effect of individual differences in prediction error delta between the positive and neutral conditions, on the different brain functionality indices we examined.Yet, in an additional attempt to control this issue, we modeled the expected feedback scores (i.e. the post-speech selfevaluation scores) at the trial-level, and tested whether brain activations in our ROIs and their association with self-efficacy update bias were affected by this factor.To this end, we executed an additional GLM on the social feedback task fMRI data, which took into account the expected feedback score on each trial.This GLM was similar to the one that is detailed in the Methods section in the main text, except for two main differences.First, the feedback reception phase was modeled with two parametric modulatorsone capturing the absolute valence of feedback on each trial, which was our main effect of interest; and another capturing the expected feedback score on each trial.Second, in order to examine the unique effects of the two parametric modulators, their regressors were mean-centered and were not serially orthogonalized 1 .
We first questioned if within this design, the absolute valence parametric regressor explained more variance in the VS than the expected score parametric regressor, during the feedback reception phase.Analysis of the contrast between these regressors confirmed that this was the case in the right VS (Supplementary Figure 6, left; voxel-level p<.001 SVC FWE<.05).Thus, right VS activity upon feedback reception was mainly driven by the absolute valence of feedback scores.
Next, we tested if the positive association between self-efficacy update bias and right VS activity during encoding of judges' feedback positivity remained significant also when controlling for the expected score at the trial-level.We repeated the regression analysis that is described in the main text.This analysis affirmed that voxels in the right VS showed significant positive covariance with self-efficacy update bias also when controlling for the expected feedback score on a trial-by-trial basis (Supplementary Figure 6, right; SVC<.05), but this was evident in fewer voxels (1 voxel at voxel-level p<.001 and 5 voxels at p<.005; peak t-value=3.6at x=11, y=7, z=-7).

Exploration of the neural correlates of self-evaluation and self-efficacy updates
A caveat of the current experiment is that we did not record brain activity during the selfevaluation and self-efficacy assessments.Therefor, this study is limited in terms of differentiating between the neural correlates of these self-beliefs and their updating following social feedback (see further discussion in main text).Nonetheless, in an attempt to address this issue, we examined the neural correlates of self-evaluation and self-efficacy during feedback reception.Specifically, we assumed that if feedback is used to update self-efficacy and self-evaluation, then it is possible that the pre-level of both of them is negatively encoded during feedback, while the feedback itself is positively encoded.This is because the prediction errors that drive the updates are 'feedback minus pre-feedback self-efficacy' and 'feedback minus pre-feedback (aka post-speech) selfevaluation'.To test this option, we conducted two additional GLMs.These GLMs were similar to the one used for modelling social feedback in the main text, except that each of them included two parametric regressors during feedbackone for the feedback itself and one for the pre-feedback self-efficacy or self-evaluation ratings, which were modelled on a trial-by-trial basis.In these GLMs the parametric regressors were mean-centered and were not serially orthogonalized.
As mentioned above, we focused this analysis on the negative contrast of the parametric regressors of pre-feedback self-evaluation and self-efficacy against baseline.We first examined effects within our key ROIs -the VS and VMPFC.Activity in the right VS negatively encoded both pre-feedback self-evaluation and self-efficacy ratings (Supplementary Figure 7; voxel-level p≤.001 SVC FWE<.05; self-evaluation: 1 voxel at p≤.001 and 6 voxels at p<.005, peak t-value=3.6at x=7, y=12, z=-7; self-efficacy: 1 voxel at p≤.001 and 6 voxels at p<.005, peak t-value=3.16at x=7, y=10, z=-7).No significant correlation with brain activity was found in the left VS or VMPFC ROIs for both ratings.Next, we explored the neural correlates of self-evaluation and self-efficacy at the whole-brain level.We found that activity in the occipital cortex (lingual gyrus) negatively encoded self-evaluation ratings during feedback (Supplementary Figure 7; voxel-level p<.001 and cluster-level pFDR<.05;107 voxels, peak t-value=4.96at x=9, y=-89, z=-1).We did not find any significant clusters of brain activity that correlated negatively with the selfefficacy ratings during feedback.However, it is noteworthy that at a statistical threshold of voxellevel p<.001 and minimal cluster threshold of k=10 (i.e.without cluster-level correction), a cluster of activation in the right hippocampus was evident (Supplementary Figure 7; additional clusters at this threshold included the supplementary motor area and the cuneus).A recent perspective has highlighted the critical role of the hippocampus, a region that is well-known for its role in forming long-term memories, in encoding reward-relevant information that guides future behavior and decision-making 2 .This preliminary result might suggest that updating self-efficacy at the time of feedback may have engaged long-term mnemonic processes, but this option requires further investigation given the lenient statistical threshold of the results.Lastly, it is also noteworthy that although the above-mentioned ROI analyses in the VMPFC were not significant, activity in this area did negatively encode pre-feedback self-evaluation and self-efficacy ratings during feedback at an uncorrected statistical threshold of voxel-level p<.005 (Supplementary Figure 7).