Human beings strive to be accepted by others and to maintain a positive social image1. Thus, social evaluation of our behavior can pose a threat to our social image, eliciting a stress response in our body2,3,4. This initiates various physiological processes5 and is associated with negative affective consequences, like anxiety or embarrassment6,7,8. Social evaluation, however, is fundamental to self-related learning processes, as it gives one the opportunity to integrate the feedback we receive from others and update the beliefs about ourselves accordingly9,10. Biases in how we process self-related feedback on our behaviors, i.e. whether we focus more on negative or positive feedback, impact our affective reactions11,12 and, in the case of self-serving processing, may function as a coping strategy11. While (social) stress is a risk factor for many psychiatric conditions13, successful coping is an important factor in maintaining mental health14. In the current study we implemented a computational modeling approach to investigate the coping mechanism of self-beneficial belief updating after social-evaluative stress and tested whether shifted information processing after stress predicts recovery from stress-induced negative affect.

When we receive feedback regarding our behaviors, information processing and belief updating is shaped by self-relevant motivations15, especially the motivation to maintain optimistic beliefs about the self16. Many studies have demonstrated that the process of self-related belief updating is biased in favor of positive information, i.e. self-related beliefs are updated more strongly when feedback is better than expected17,18,19,20. However, updating biases towards negative feedback have been reported in performance contexts21,22, which indicates that the context of learning (i.e. learning about own abilities or learning about one’s personality), type of feedback and prior assumptions are important factors when explaining self-related belief updating biases.

While there are only relatively few studies on the effects of stress on self-related belief updating, various studies on reward processing and non-self-related feedback processing have shown that stress is an influencing factor in this regard. One key mechanism for feedback-based learning is the prediction error signal, indicating the difference between a predicted and an actual outcome23,24, which is being minimized by updating beliefs during learning. This signal is generated by dopaminergic neurons of the ventral striatum25, which might be particularly important for the stress-induced modulation of prediction error signals as the dopamine system is sensitive to stress26,27. However, these effects depend on the type, intensity and schedule of the stress exposure28, which might also explain heterogeneous effects of stress on reward processing and feedback-based learning. Research on declarative memory has shown that timing of stress matters. In the acute stress phase, mainly characterized by a rapid sympathetic response, catecholamines and non-genomic glucocorticoid actions lead to increased memory formation of the stressful event. Cortisol is released with a delay and inhibits memory consolidation later on to avoid interference with non-stress-related information29,30. Besides this inhibition of hippocampus-dependent declarative memory, neuro-imaging research on classification learning also found a shift towards striatum-based procedural learning after stress, i.e. also non-declarative learning processes are modulated by stress31. Acute stress is associated with an increased extinction resistance in fear conditioning32. When learning takes place with a delay to stress, trace conditioning33, and updating in reversal fear conditioning34 are attenuated, cortisol is associated with reduced fear conditioning32 and working memory is reduced35,36. Timing of stress seems to be important for feedback-based or reward-based learning as well37. Initially, acute stress (e.g. a threat of a shock during learning) impairs feedback-based learning of reward38. Neurally, acute stress attenuates the response to reward in the striatum and orbitofrontal cortex39,40 and enhances the striatal response to aversive feedback41. Accordingly, under acute stress self-related belief updating is more strongly driven by unfavorable feedback, i.e. the learning bias in favor of positive information (optimism bias) usually found in self-related belief updating is absent42. The opposite effects are reported when learning takes place with a delay to stress (e.g. after a public speech), a phase mainly characterized by an increase of cortisol29. Here, feedback processing is more strongly driven by stimuli signaling reward and possibly associated with stress-induced cortisol change43 while learning from negative feedback is decreased, potentially linked to cortisol levels before learning44. On the neural systems level, stress recovery is associated with increased striatal responses to rewarding feedback at 50 min after stress37,45. Moreover, specifically individuals with low striatal reward reactivity showed an association of recent life stress with lower positive affect, which makes striatal reactivity a potential factor of successful stress coping46.

According to classic appraisal theories of stress47, different strategies such as seeking social support, positive revaluation or acceptance are helpful in coping with stress-induced negative affect47,48,49. In the context of social-evaluative stress a self-protection strategy is to view oneself in a positive light, i.e. emphasizing the own desirability, focusing on own successes and attributing failure externally50. This strategy has also been successful in alleviating stress-induced negative affect following a performance situation11,51. Generally, an optimistic way of processing self-related feedback has been associated with better mental health52,53. On the contrary, processing self-related feedback in a more negative way may result in negative beliefs about the self54 and ultimately lead to lower self-esteem or depressive symptomatology. Studies on self-related belief updating in individuals with depression suggest that information processing is distorted in a negative direction55 and that coping strategies for situations of social-evaluative stress are less readily available in these patients56.

In the present study, we aim to investigate the effects of social-evaluative stress on the updating of self-related ability beliefs and the propensity to engage into self-beneficial learning after social-evaluative stress. By means of two well validated and highly reliable paradigms, the Trier Social Stress Test3 (public speech) and the Cold Pressor Test57, as well as a no stress control condition, we directly manipulated levels of social-evaluative stress in a between-groups design. After stress manipulation we used computational modeling to describe participants’ self-related belief updating behavior using the learning of own performance (LOOP) task21. In this task participants continuously update beliefs about their abilities in epistemologically novel behavioral domains. We then used participants’ learning bias from positive and negative feedback to predict their recovery from stress-induced negative affect. We found that social but not physical stress shifted subsequent self-related belief updating in a more self-beneficial direction which predicted better recovery from negative affect. We elaborate on the relationship between stress (specifically the components of negative affect and cortisol response), self-related belief updating and affect regulation in healthy participants and discuss the potential of our findings for a better understanding of maladaptive self-related belief systems in psychiatric conditions such as depression.


After exposure to social-evaluative stress (SOC, Trier Social Stress Test), non-social, physical stress (PHY, Cold Pressor Test) or a no stress control condition (CON, reading) participants performed the LOOP task21, which was covered as a measure of cognitive estimation skills (see Fig. 1). The central idea of the LOOP task is to create a performance context and provide manipulated positive or negative feedback in comparatively neutral domains in which people have only vague prior assumptions. By this means, individuals form a concept about their own abilities over the course of the experiment. In a previous study, we showed that this process of self-related belief updating can be described best by a computational prediction error learning model (adapted from Rescorla and Wagner58) with two separate learning parameters for positive and negative prediction errors21. During the LOOP task, participants were asked to answer estimation questions in two different estimation domains (e.g. estimating the weight of animals and the height of buildings) and received manipulated performance feedback implying a rather good performance in one category and a rather bad performance in the other one (high vs. low ability condition). In the beginning of each trial participants saw a cue indicating the estimation category and had to rate their expected performance for the upcoming estimation question in this category. A manipulated feedback on their estimation performance in relation to an alleged reference group was presented afterwards. Saliva cortisol as well as negative affect, including perceived stress, embarrassment, anger, and frustration, were assessed several times during the experiment. Pre-stress baseline measures (T1AFF/CORT) were taken after a 10-min-period of rest in the beginning of the session. Post-stress negative affect was rated immediately after the stress exposure or control task (T2AFF) to calculate the mean change of negative affect (ΔAFF). Post-stress cortisol samples were taken after another 10-min period of rest (T2CORT) to calculate the mean cortisol change (ΔCORT). After performing the LOOP task, saliva samples and negative affect were again obtained (T3AFF/CORT, for a detailed description see methods).

Figure 1
figure 1

(a) Experimental timeline and procedure. SOC: social-evaluative stress group (public speech [audience icon], n = 29), PHY: physical stress group (Cold Pressor Test [ice cubes icon], n = 30), CON: no stress control group (reading task [paper icon], n = 30), salivette icon: saliva collection for cortisol determination; paper pencil icon: rating of negative affect including perceived stress, embarrassment, anger, and frustration. (b) Sequence of one trial. 1. Cue: display of the upcoming estimation category associated with a high or low ability condition, 2. Performance expectation rating, 3. Estimation question, 4. Performance feedback. Figure adapted from Müller-Pinzler et al.21.

Cortisol response and negative affect

Cortisol change

The stress manipulation was effective and social-evaluative stress, as well as physical stress, led to a stronger increases of cortisol levels from baseline T1CORT to post-stress T2CORT than in the no stress control group (Scheirer-Ray-Hare test on ΔCORT controlled for time of the day [TIME]: main effect factor Stress group H2 = 18.9, p < 0.001, post-hoc Dunn-Bonferroni-Tests for factor Stress group: SOC vs. CON: z =  − 4.29, p < 0.001; PHY vs. CON: z =  − 2.76, p = 0.018; see Table S1). There was no statistically significant difference between the two stress groups (SOC vs. PHY: z = 1.56, p = 0.355; baseline cortisol levels did not significantly differ between groups H2 = 1.74, p = 0.419 controlled for TIME, see Fig. 2a,c).

Figure 2
figure 2

(a) Cortisol levels over the course of the experiment separately for the three stress groups (social-evaluative stress [n = 29] vs. physical stress [n = 30] vs. no stress [n = 30]). Lines connect group medians. Group means are depicted as gray ovals within the box. (b) Mean negative affect ratings (embarrassment, anger, frustration and perceived stress) over the course of the experiment separately for the three stress groups depicted as in (a) (c) Change in saliva cortisol levels after stress induction (post-stress T2CORT − baseline T1CORT), (d) Change in negative affect (post-stress T2AFF − baseline T1AFF), SOC = Social-evaluative stress group, PHY = Physical stress group, CON = no stress control group. Line inside box: median, lower/upper box hinges: 25th and 75th percentile, lower/upper box whiskers: smallest/largest value within 1.5 × inter-quartile range from hinges, *p < 0.050, ***p < 0.001.

Change in negative affect

Mean negative affect increased significantly after social-evaluative stress but not after physical stress compared to the control group (Kruskal Wallis test on ΔAFF: H2 = 43.9, p < 0.001, post-hoc Dunn-Bonferroni-tests: SOC vs. CON: z =  − 6.45, p < 0.001, PHY vs. CON: z =  − 1.88, p = 0.182, SOC vs. PHY: z =  − 4.59, p < 0.001; baseline negative affect did not significantly differ between groups H2 = 3.2, p = 0.201; see Fig. 2b,d and Supplementary Fig. S1).

Forming self-related beliefs over time

In a model free behavior analysis we replicated previous findings regarding the LOOP task which indicates self-related belief updating in response to the feedback21. Over the time of 30 trials, participants adapted their performance expectation ratings (EXP) towards the positive and negative feedback of the two ability conditions, i.e. they updated their self-related beliefs (Fig. 3c, significant factor Ability condition high vs. low t86 = 8.52, p < 0.001, significant Trial x Ability condition interaction t5156 = 32.72, p < 0.001). Social-evaluative stress modulated self-related belief updating over time, i.e. performance expectation ratings became increasingly higher compared to physical stress or no stress (Trial x Ability condition x Stress group split into the contrasts social [SOC] vs. non-social [PHY, CON] and the orthogonal contrast PHY vs. CON: interaction for contrast SOC vs. [PHY, CON]; t5156 = 4.01, p < 0.001). In the physical stress group performance expectation ratings were even more negative over time than in the no stress control condition (Trial x Ability condition x Contrast PHY vs. CON t5156 =  − 2.15, p = 0.031; mixed-effects model with the within-group factor Ability condition, the between-group factor Stress group, and the continuous variable trial, plus interactions, see Supplementary Table S2).

Figure 3
figure 3

(a) Structure of the model space. αUni = one learning rate for the whole time course; αHigh ability/ αLow ability = two separate learning rates for the two ability conditions; αPE+PE- = two separate learning rates for positive and negative prediction errors; adapted from Müller-Pinzler et al.21. (b) Protected exceedance probabilities resulting from the Bayesian Model Selection procedure including the prediction error learning models depicted in (a) and a mean model (M0) assuming stable means for each ability condition instead of continuous learning (c) Performance expectation ratings (EXP, solid line) and performance expectations predicted by the winning model (EXP − pred., dashed line) over the time course of 30 trials. Ratings and predicted values were averaged across participants separately for the two ability conditions and the three experimental groups. Shaded areas represent the standard errors of the expectation ratings for each trial. (d) Learning rates derived from the Valence Model (winning model). A significant interaction effect (*) of PE-Valence x Stress group (SOC = social-evaluative stress, PHY = physical stress, CON = no stress control) indicates that a bias towards increased updating in response to negative prediction errors (αPE-) in contrast to positive prediction errors (αPE+) is absent in the social-evaluative stress group. (d) Rank-based regression plot of valence bias score predicting the recovery from negative affect (REC, ratings T2AFF − T3AFF) in the subsample of the social-evaluative stress group (n = 29) controlled for the stress-induced change in negative affect (ΔAFF, ratings T2AFF − T1AFF), i.e., residuals of the valence bias score and the recovery predicted by ΔAFF are plotted. More self-beneficial belief updating (higher valence bias score) is associated with a better recovery from stress-induced negative affect.

Model selection for computational models of learning behavior

To capture the updating of the performance expectation ratings over time in a learning model, a similar model comparison to that of Müller-Pinzler et al.21 was performed. All three main models of the model space followed the idea of a Rescorla-Wagner model58 with one or two learning rates for each participant reflecting the degree to which people weighted prediction errors (PE = Feedbackt − EXPt) to update their expectation rating (see Fig. 3a and for model descriptions see method section).

In line with Müller-Pinzler et al.21, the Valence Model outperformed all other models in all three groups according to Bayesian Model Selection59 (see Fig. 3b; protected exceedance probability for the whole sample pxptotal > 0.999, Bayesian omnibus risk BORtotal < 0.001 as well as separately for the three groups pxpSOC = 0.985, BORSOC = 0.019, pxpPHY > 0.999 , BORPHY < 0.001, pxpControl > 0.999 , BORControll < 0.001; see Table 1 and Supplementary Table S3 for more details on model comparisons). This model, with two separate learning rates for positive PEs (αPE+) and negative PEs (αPE-) across ability conditions, assumes that learning differs depending on the valence of prediction errors. Learning parameters from the Valence Model were used for further analysis.

Table 1 PSIS-LOO scores for the whole sample.

The modeled performance expectations of our winning model predicted the performance expectation ratings on the individual subject level within each ability condition with R2= 0.33 ± 0.24 (M ± SD). Repeating the model free analysis with the modeled performance expectations confirmed the results from the original analysis (see Supplementary Table S4).

Stress and learning parameters

In line with Müller-Pinzler et al.21, the physical stress and no stress control group showed a negativity bias in their learning behavior, i.e. a stronger self-related belief updating after negative than positive prediction errors (αPE+ vs. αPE- within group comparison for PHY: W = 100, Z =  − 2.73, p = 0.005 and CON: W = 84, Z =  − 2.89, p = 0.003, Wilcoxon test). This negativity bias was absent after social-evaluative stress (αPE+ vs. αPE- within group comparison for SOC: W = 193, Z =  − 0.53, p = 0.609; significant PE-Valence x Contrast SOC vs. [PHY, CON] interaction bVALxSOC = 0.114, t85 = 2.30, p = 0.024, PE-Valence x Contrast PHY vs. CON: bVALxPHY =  − 0.036, t85 =  − 0.72, p = 0.471; betas standardized, see Fig. 3d and Supplementary Table S5).

To better capture biased learning behavior, a valence bias score was computed (valence bias score = (αPE+—αPE−)/(αPE+  + αPE−))21,61,62, which represents updating after positive compared to negative prediction errors. More positive valence bias scores indicate more self-beneficial belief updating, while negative valence bias scores speak for stronger self-related belief updating after negative feedback.

Negative affect and cortisol change predict subsequent self-beneficial belief updating

To further assess which aspect of the stress response is associated with self-beneficial belief updating, we correlated negative affect and cortisol with the valence bias scores across all three experimental groups. While both stress groups (SOC and PHY) only differ significantly in terms of an increase in negative affect but not cortisol levels, both measures show large variance within and across groups (see Fig. 2) that could explain differences in self-related belief updating that is only partially captured by the group effect. We found that a stronger increase in negative affect (ΔAFF = T2AFF − T1AFF) predicted more self-beneficial belief updating (bΔAFF = 0.100, t86 = 2.066, p = 0.042, rank regression63 of the valence bias score predicted by ΔAFF for the whole sample). Also, a higher increase in cortisol levels (ΔCORT = T2CORT − T1CORT) predicted more self-beneficial belief updating (bΔCORT = 0.020, t85 = 2.377, p = 0.020, rank regression63 controlled for TIME for the whole sample, for further information see Supplementary Results and Supplementary Table S6).

Learning bias and affective recovery

A more positive valence bias score predicted better recovery from stress-induced negative affect during learning in the social-evaluative stress group, the only group with significantly increased levels of negative affect after stress (Fig. 3e; REC, change in negative affect post-stress T2AFF − post-learning T3AFF, bBIAS = 0.584, t26 = 2.131, p = 0.043, rank regression63 controlled for the increase in negative affect [ΔAFF = T2AFF − T1AFF]). This supports the idea of self-beneficial belief updating as a coping strategy. Analysis across the whole sample trend-wise confirmed this effect (bBIAS = 0.238, t85 = 1.922, p = 0.058). However, regression coefficients of the valence bias score predicting the affective recovery within the two other experimental groups alone, which exhibit no substantial increase in negative affect in the first place, were not significant (PHY: bBIAS = 0.147, t27 = 0.999, p = 0.327, CON: bBIAS = 0.000, p > 0.999, rank regression). A stronger relationship in the social stress group compared to the other groups, i.e. a modulation of the factor Stress group, showed a trend-wise effect (BIAS x Contrast SOC vs. [PHY, CON] interaction bBIASxSOC = 0.322, t85 = 1.70, p = 0.093, rank regression63 of the affective recovery predicted by valence bias score, Stress group split into the contrasts social [SOC] vs. non-social [PHY, CON] and PHY vs. CON, the increase in negative affect, plus the bias x Stress group interactions for the two contrasts, Supplementary Table S7).


After being devalued for example at work or school we need to empower ourselves in order to uphold or boost our self-image. Research has shown that the ability to adopt a positive attitude towards oneself after receiving criticism is central to positive affect and good mental health outcomes in the long run11,53,64. In the current study we investigated how people apply self-beneficial belief updating during a performance feedback situation as a means to counter their negative affect. Using computational modelling, we provide a mechanistic explanation on how individuals engage in more self-beneficial updating of ability beliefs after experiencing a threat to their social image and how this shift in social learning of self-related information predicts recovery from stress-induced negative affect.

The positive shift of self-related updating of ability beliefs after social-evaluative stress, going along with a better recovery from negative affect, fits nicely to the notion of a belief’s own value as recently posited by Bromberg-Martin and Sharot15. In their revised framework, general belief updating is not solely driven by external outcomes like rewards or punishments but also by the agent’s motivation to optimize internal states like positive affect15,65. In the present study we show this direct link between self-related belief updating and a change in the affective state indicating that self-related belief updating might be motivated by the wish to uphold or even recover a positive affective state. This is in line with the idea of motivated cognition, i.e. the assumption that cognitive processes like attention, information processing and decision making are not neutral on their own, but are always shaped by needs, feelings and desires of the individual66. Especially when processing information that challenges one’s self-image, self-related belief updating is not only informed by the history of previous feedback, as it has often been assumed in classical reinforcement learning tasks, but also by various self-relevant needs and goals67. Transferred to the present study, this implies that the motivation to restore an endangered self-image and to regulate one’s affect back to a set point directly impacts self-related information processing. The pattern of an active counter-regulation of negative affect by self-beneficial belief updating can be described as a striving for homeostasis11. To better capture the fluctuation of the affective state and its involvement in the trial-by-trial self-related belief updating loop, following the framework by Bromberg-Martin and Sharot15, future studies should consider repeated assessments of affective states during the task to predict the empowering potential of shifts in learning on the single trial level.

Since negative self-related beliefs are at the core of psychiatric conditions like depression54, this study targets clinically highly relevant processes. Depression is associated with seeking negative feedback which confirms negative self-related beliefs68 and seeking negative feedback in combination with a stressful life event can further increase depressive symptoms69. Furthermore, depression is associated with a weaker stress recovery mediated by an attentional bias towards negative feedback. Understanding the mechanisms of how people form self-related beliefs in a context mimicking everyday performance settings and linking these to the regulation of negative affect after stress has important implications for understanding the etiology of depressive symptoms. The present study set-up, including a social-evaluative stress induction followed by a social-evaluative performance situation also addresses one of the fundamental fears of individuals with social anxiety: being devalued by others. Since both depression and social anxiety are associated with negatively biased updating behavior in response to self-related feedback12,21,55,70, we assume that the affect-regulating and empowering potential of self-beneficial belief updating after social-evaluative stress would be less pronounced in depression or social anxiety and would thereby possibly exacerbate the symptomatology in a self-fulfilling way. Future studies with similar experimental set-ups and clinical samples could examine the relationship between self-beneficial belief updating and affect regulation in more detail and develop potential intervention strategies based on empowering individuals on their way to processing newly incoming information.

Replicating a previous study of ours21, self-related belief updating was negatively biased in the control condition in which participants were not exposed to any stress. In the prior study, this negativity bias has been shown to be specific for self-related belief updating in comparison to belief updating about another person21. In the present study, we found that after physical stress participants also exhibited a negativity bias in forming self-related beliefs, i.e. participants tended to make greater updates in response to negative prediction errors in contrast to positive prediction errors. The negativity bias stands in contrast to other studies reporting a positivity or optimism bias in feedback-based learning e.g. when receiving feedback about the chance to encounter negative life events20,71, about one’s intelligence17 or about one’s personality18,19 (for a review see16). There are several possible explanations for the motivation behind the negativity bias in context of the LOOP task in contrast to the reported positivity biases of other studies which was, however, not the focus of the present study (for a discussion on the negativity bias see21). In order to test for the specificity of self-beneficial belief updating after social-evaluative stress, it would be interesting to test if this effect also accounts for experiments that typically yield a positivity bias (e.g. for life events, IQ or personality) in feedback-based learning tasks.

Here, we demonstrated that both, negative affect and cortisol stress responses, go along with a shift in self-related belief updating. It has been shown before that experiencing social emotions (e.g. embarrassment or shame) is related to increased cortisol levels in situations which threaten one's social image, like the social-evaluative stress induction6. Cortisol has been linked to reward processing and feedback-based learning in the stress triggers additional reward salience—STARS—model which proposes that stress and the associated release of cortisol modulates the dopamine system, resulting in an increased salience of rewards, thus biasing learning towards rewarding feedback43,72. The current results, however, suggest that the quality of stress (here, social vs. physical) might make a difference, and the STARS model, based on a rather unspecifically triggered cortisol response, cannot fully explain the present stress effect on self-related belief updating after social but not physical stress.

Although both groups showed a significantly greater increase in cortisol compared to control, this effect was less pronounced in the physical stress group. While the Cold Pressor Test is known to elicit a strong sympathetic activity, studies reported only low to moderate cortisol effects73, that were weaker than after the Trier Social Stress Test74,75. Our alteration of the Cold Pressor Test, to remove the social element of the conductor, might have even further reduced stress effects as compared to the original Cold Pressor Test protocol. We did find an association of the negative affect response (i.e. self-evaluative emotions like embarrassment), which is typically specific for social-evaluative stress and rather absent during physical stress, with shifts in learning behavior in our study. But also cortisol as a rather unspecific stress component was associated with shifts in learning behavior. Thus, we cannot rule out that a more intense physical stress protocol with higher cortisol responses would have led to similar effects on learning rates. A more detailed recording of negative affect and a comparison between different negative affective states as well as more detailed recording of the physiological stress response might help in future studies to better differentiate between different stress qualities and understand specific effects of social-evaluative stress on self-beneficial belief updating.

To summarize, our results indicate a shift towards more self-beneficial belief updating after social-evaluative but not physical stress. This shift goes along with a better recovery from stress-induced negative affect. Linking self-related belief updating to affect is an important step in understanding biases in self-related learning and its relation to affect regulation. The special feature of the present study was the study-set that allowed to examine a link between negative affect and self-related belief updating. By introducing a performance context with consecutive self-related feedback, corresponding to real-life school or work-related performance situations, individuals can form beliefs about their own abilities over time and potentially use this formation process as a means to regulate their affect. With this approach we aimed to increase the ecological validity of the study in order to trigger and investigate motivational processes that might be less relevant in more abstract study settings. Since social evaluation represents a constant stressor in every-day life, the question of an appropriate coping strategy to regulate negative affect is of great importance when handling everyday social situations.

Materials and methods


Eighty-nine participants recruited at the University of Lübeck Campus were included in the study. Upon appearance, participants were assigned to either a social-evaluative stress group (SOC; n = 29, 21 female, aged 18–28 years; M = 22.9; SD = 2.76), a physical stress group (PHY; n = 30, 20 female, aged 19–27 years; M = 22.5; SD = 1.94) or the control group (CON; n = 30, 20 female, aged 18–32 years; M = 22.3; SD = 3.00, data of the control group were published before21). From the initially recruited N = 96 subjects, seven had to be excluded—five because they did not believe the cover story and two due to technical problems. All included participants were fluent in German, non-smokers with a body-mass index between 18.5 and 30. They were not diagnosed with acute or chronic psychiatric conditions or diseases affecting the hormone system and did not take psychiatric drugs or medication affecting the hormone system (except hormonal contraceptives). Participants had normal or corrected-to-normal vision and did not study psychology to avoid previous experience with experiments using cover stories. Additional exclusion criteria for participants who underwent the physical stress protocol were cardiovascular diseases, frequent fainting or seizures and current hand injuries. For more details on the sample characteristics see Supplementary Table S9a. All participants gave written informed consent prior to the participation and received monetary compensation for their participation. They were naive to the background of the study during the session and debriefed about the cover story afterwards. The study was conducted in compliance with the ethical guidelines of the American Psychological Association (APA) and was approved by the ethics committee of the University of Lübeck.

Manipulation procedure

Social-evaluative stress

Social-evaluative stress was induced by a public speech similarly to the Trier Social Stress Test3. Participants were instructed to prepare a short self-presentation for an application for a scholarship, which had to be presented in front of a selection committee who would allegedly assess the participant’s verbal skills and body language. The selection committee consisted of the experimenter, who was passive during the speech, a second experimenter, who was allegedly responsible for measuring verbal skills, and a passive camera assistant, who pretended to videotape the speech. Before starting the 10-min preparation period, participants briefly visited the room with the selection committee. After the preparation time was over, participants were asked to come back to this room and present their speech. Talking time was 5 min (M = 4.9 min, SD = 0.16) with a minimum of 3 min of uninterrupted speech. If the participant finished the speech before the time was over, the second experimenter waited for at least 15 s with a motionless face and then asked the participant to continue. If the participant stopped speaking again and the 3 min of free speech had passed, the second experimenter asked standardized questions until the 5 min of talking time were over (“Explain why it is important for you to achieve a good performance.”, “Do you think it is important to improve yourself throughout your life?”, “Do you consider yourself a person who values his/her independence?”). Average social-evaluative stress duration (start subsequent rest period − start speech preparation) was M = 16.4 min, SD = 1.2.

Physical stress

Physical stress was induced by an exposure to ice water according to the Cold Pressor Test protocol57,76. Participants were asked to dip their non-dominant hand in cold water (water temperature 3–5.5 °C = 37.4–41.9°F, M = 4.26 °C, SD = 0.50) for as long as possible up to 3 min (duration 48 s–3 min, M = 2.7 min, SD = 0.7). The water was kept in motion with a small electrical pump to prevent the water temperature from rising around the participant’s hand. To control for the procedure of the social-evaluative stress condition, participants visited the room with the cold pressor apparatus first, had a 10-min preparation period and came back into the room for the stress exposure. During the preparation time, participants were asked to imagine dipping their hands in a freezing cold environment and write down their associations. To make the stress exposure less social, the experimenter was not present in the room but waited in an adjacent room. If the participant took out their hand before the 3 min were over, they had to signal this immediately by ringing a bell. The experimenter could roughly observe the participant in the reflection of the glass door, thus ensuring that she/he dipped the hand into the water. Average physical stress duration (including preparation period) was M = 16.2 min, SD = 1.5.

No stress control condition

In the control condition, participants performed a reading task that was described to them as measuring reading speed. They had 10 min to rehearse two different texts about applying for a scholarship. Afterwards, they were guided to the other room with nobody present and were asked to measure their reading time, while reading the two texts aloud at a natural speed. Average control duration was M = 15.3 min, SD = 1.3.

Manipulation checks


Three saliva samples were collected during the experiment for cortisol analysis (see Fig. 1a). The first sample (baseline T1CORT) was taken after a 10 min period of rest immediately before starting the instruction for the stress manipulation (mean time between T1CORT and start of the SOC, PHY or CON preparation phase: M = 3.7 min, SD = 1.4). The post-stress cortisol sample T2CORT was collected after another 10 min resting period following the stress manipulation and the last sample (T3CORT) was collected after the learning task (M = 45.6 min (SD = 3.3) post stress). The stress-induced cortisol change (ΔCORT) was determined by subtracting the cortisol levels of T2CORT − T1CORT. Saliva was collected with Salivettes (Sarstedt, Nümbrecht, Germany), stored at − 30 °C and sent to the bio-psychological lab at TU Dresden, Dresden, Germany for analysis (here stored at − 20 °C until analysis). Salivary free cortisol levels were determined using a chemoluminescence immunoassay (IBL International, Hamburg, Germany).

Negative affect

We assessed negative affect by means of a short pen and paper questionnaire, covering the emotions embarrassment, anger, frustration, as well as the perceived stress with one rating each. The questionnaires were handed out at baseline (T1AFF) as well as at the very end of the experiment (T3AFF). The post-stress negative affect was measured immediately after the stress manipulation (T2AFF; see Fig. 1). Ratings were averaged for each measurement point to get a composite measure of negative affect (see Supplementary Fig. S1 for separate scores). The change in negative affect after stress (ΔAFF) was determined by subtracting T1 negative affect from T2 (T2AFF − T1AFF). The recovery from negative affect (REC) was determined by subtracting T3 negative affect from T2 (T2AFF − T3AFF).

Behavioral task

Learning of own performance task

The Learning of own performance (LOOP) task21 (Fig. 1b) allows to measure self-related belief updating through trial-by-trial performance expectation ratings and subsequent performance feedback. The task included estimation questions in two different estimation categories (heights of houses and weights of animals) and was presented to the participants as a measure of estimation abilities. To make participants learn about their estimation ability the two estimation categories were paired with manipulated performance feedback implying high ability for one category and low ability for the other (e.g. heights of houses = high ability and weights of animals = low ability, estimation categories were counterbalanced between ability conditions). The assignment of the categories to the ability conditions was independent of the participants’ actual performance and their performance expectation ratings. Thus, participants could learn over the course of the experiment that they were good in one estimation category and rather bad in the other one. Each trial began with a cue displaying the category of the next estimation question followed by a performance expectation rating for this question. Afterwards, the estimation question was presented together with a picture for 10 s. Continuous response scales below the pictures determined a range of plausible answers for each question, and participants indicated their responses by navigating a pointer on the response scale with a computer mouse. Subsequently, feedback indicating the estimation accuracy as percentiles compared to an alleged reference group of 350 university students was presented for 5 s (e.g. “You are better than 72% of the reference participants.”). The order of the two estimation categories/ability conditions was intermixed with a maximum of two consecutive trials of the same condition and 30 trials per condition in total. The estimation questions were randomized within the estimation category/ability conditions. A fixed sequence of ability conditions and feedback was presented for all participants. In the low ability condition, feedback was approximately normally distributed around the 35th percentile (SD ≈ 16; range 1–60%) and in the high ability condition around the 65th percentile (SD ≈ 16; range 40–99%). The task started with detailed instructions and three test trials. All stimuli were presented using MATLAB Release 2015b (The MathWorks, Inc.) and the Psychophysics Toolbox77.


To minimize noise in the cortisol saliva samples, participants were asked to follow behavioral rules prior to the experimental session. These were in detail: no alcohol on the evening before the experiment and bed rest at about 10 p.m. (ideal case eight hours of sleep); one hour before the session: no sport, no smoking, no drinks containing caffeine or theine, no food (including bonbons and chewing gums) and no juices. Upon arrival at the laboratory, participants read the participant information including the cover story regarding the stress manipulation and the LOOP task. After signing the consent form, they were asked to fill out a questionnaire checking the adherence to the behavioral rules. Participants rested for 10 min before the baseline measurement, including saliva cortisol and negative affect, was obtained (T1AFF/CORT). During the resting period, they filled out a short personality questionnaire (not included in this study). Subsequently, participants of the social and physical stress groups were challenged with a stress protocol while participants of the control group did the control reading task. Directly afterwards, participants rated their affective state (T2AFF) followed by another 10 min resting period, which was terminated with a saliva sampling (T2CORT). In the second part of the experiment participants performed the LOOP task. Finally, another cortisol sample and affective ratings were collected (T3AFF/CORT). After completing a post-experimental interview, including additional questionnaires, participants were debriefed about the cover story. The experimental sessions were run between 10.00 a.m.–12.00 p.m., 1.00–3.00 p.m. or 3.45–5.45 p.m. The allocation to the time slots did not differ between the experimental groups (Pearson's Chi-squared test p = 0.867, see Supplementary Table S9b). See Fig. 1a for a graphical illustration of the procedure.

Statistical analysis

Stress manipulation

To test whether the stress manipulation was effective, the stress-induced changes in cortisol as well as affect were compared between the three experimental groups. Due to the stress manipulation, the variance of the cortisol and negative affect responses were unequal between the three experimental groups (Levene test ps < 0.05). Since the distributions of the cortisol and affective stress response were skewed in some groups (Lilliefors-corrected Kolmogorov–Smirnov normality test ps < 0.05 for the cortisol change in the control group and for the change in negative affect in all groups) non-parametric tests were used. Since cortisol levels are known to underlie circadian fluctuations78 all cortisol analysis were controlled for time of the day (morning vs. noon vs. afternoon, see Procedure). Responses in negative affect were compared with the Kruskal–Wallis test, the cortisol response was compared with the Scheirer-Ray-Hare test, an extension of the Kruskal–Wallis test that allows to control for time of the day. Post-hoc comparisons between the groups were performed with Dunn’s test.

Model free analysis of performance expectation ratings

The analysis of the expectation ratings including computational modeling was adapted from Müller-Pinzler et al.21. To illustrate basic effects of the expectation ratings, a linear mixed model with the factors Ability condition (high ability vs. low ability), the continuous variable Trial (30 Trials), and Stress group (with the two contrasts SOC vs. [PHY, CON] and PHY vs. CON) as a between subject factor was performed.

Computational modeling of learning behavior

The dynamic changes in self-related beliefs, which were measured by the performance expectation ratings in response to the provided performance feedback, were modeled using prediction error delta-rule update equations (adapted from Rescorla-Wagner model58). There were three main models of the model space with one or two learning rates modeled separately for each participant (see Fig. 3a). The first model (Unity Model) included a single learning rate for the whole time course (EXPt+1 = EXPt + αUni PEt). The second model (Ability Model) contained two separate learning rates for the two ability conditions allowing to capture a difference in expectation updating when receiving feedback in a high ability context (αHigh ability) or low ability context (αLow ability). The third model (Valence Model) with two separate learning rates for positive PEs (αPE+) and negative PEs (αPE-) across ability conditions allows to model learning that differs depending on the valence of prediction errors rather than different ability conditions. The three models were compared to a Mean Model with two performance expectations means reflecting the assumption of stable expectations for each ability condition without learning over time. In addition to the learning rates, we fitted two parameters for the initial belief about the participant’s performance, separately for both ability conditions (see Table 1).

Model fitting

For model fitting we used the RStan package79, which uses Markov chain Monte Carlo (MCMC) sampling algorithms. All learning models of the model space were fitted separately for each subject. To sample posterior parameter distributions, a total of 2400 samples were drawn after 1000 burn-in samples (overall 3400 samples; thinned with a factor of 3) in three MCMC chains. Convergence of the MCMC chains to the target distributions was assessed by \(\widehat{R}\) values80 for all model parameters. One subject was excluded due to implausible model parameters, i.e. mean learning rate of almost 1, as well as \(\widehat{R}\) values of 1.1 and low effective sample sizes (neff, estimates of the effective number of independent draws from the posterior distribution) for some model parameters of the valence model. Otherwise the effective sample sizes were greater than 1000 (> 1400 for most parameters). Posterior distributions for all parameters for each of the participants were summarized by their mean resulting in a single parameter value per subject that we used to calculate group statistics.

Bayesian model selection and family inference

To select the model that describes the participants’ updating behavior best, we estimated pointwise out-of-sample prediction accuracy for all fitted models separately for each participant by approximating leave-one-out cross-validation (LOO)60. To this end, we applied Pareto-smoothed importance sampling (PSIS) using the log-likelihood calculated from the posterior simulations of the parameter values as implemented by Vehtari et al.60 (loo R package81). Sum PSIS-LOO scores for each model as well as information about \(\widehat{k}\) values, the estimated shape parameters of the generalized Pareto distribution, indicating the reliability of the PSIS-LOO estimate, are depicted in Table 1. As summarized in Table 1 very few trials resulted in insufficient parameter values for \(\widehat{k}\) and thus potentially unreliable PSIS-LOO scores (on average 0.20% of trials per subject with \(\widehat{k}\) > 0.7). Bayesian model selection on PSIS-LOO scores was performed on the group level accounting for group heterogeneity as described by Stephan et al.59,82. This procedure provides the protected exceedance probability for each model (pxp), indicating how likely a given model has a higher probability explaining the data than all other models, as well as the Bayesian omnibus risk (BOR), the posterior probability that model frequencies for all models are all equal to each other82. Additionally, difference scores of PSIS-LOO for all models in contrast to the winning model were computed, which can be interpreted as a simple ‘fixed-effect’ model comparison60 (see Table 1).

Posterior predictive checks

To test whether the predicted values of the winning model could capture the variance in the performance expectation ratings a regression analysis (EXP ~ pred. values) was performed for each subject separately for the two ability conditions. R-squared statistic was determined and averaged. In addition, the model free analysis of the expectation ratings was repeated with the predicted values of the winning model to assess if the predicted data captured the effects that were present in the data of the expectation ratings.

Analysis of learning parameters

Learning rates for positive (αPE+) and negative prediction errors (αPE-, factor PE-Valence) were compared between the three groups in a linear mixed model with the factors PE-Valence and group (split into the contrasts SOC vs. [PHY, CON] and PHY vs. CON). Additional post-hoc tests for the PE-Valence within each stress group were performed with the Wilcoxon test. To test whether the variance in affective response and the cortisol response created by our stress manipulation is related to a bias in the updating behavior, we calculated a normalized learning rate valence bias score (valence bias score = (αPE+  − αPE−)/(αPE+  + αPE−))21,53,54 and tested whether the affective response and the cortisol response predicted this bias. Since the distribution of the cortisol and affective response is left-skewed due to the experimental manipulation and thus absence of the response in parts of the subjects (valence bias score is normally distributed in the whole sample as well as all subgroups) we used rank regressions following Kloke et al.63. In case of the cortisol response, time of the day was additionally included in the regression analysis as a control variable to take into account circadian fluctuations of cortisol levels. To test whether the learning bias is associated with the recovery from negative affect elicited by stress (change in affective ratings post-stress T2AFF − post-learning T3AFF), rank regressions with the valence bias score predicting the recovery were computed with the stress-induced increase in negative affect as an additional control variable to take into account regression to the mean. This was computed within the three experimental groups as well as for the whole sample, here with the additional variable group (split into the contrasts SOC vs. [PHY, CON] and PHY vs. CON) as well as the interaction bias x group to test whether the correlation differs between the three experimental groups. Data was analyzed in with the software R version 3.6.083 and plots were made with the R package gglpot284.