Adaptive behavior, i.e., the optimization of response strategies based on performance-related feedback, is the key to successful survival in dynamic environments. In general, individuals strive to maximize desirable and minimize unfavorable action consequences, as formalized in the Law of Effect by Edward Thorndike (1927
). However, not all contingencies between actions and consequences are learned equally well. A growing body of evidence points to the existence of a specific learning bias due to which reward seeking is particularly coupled with action invigoration, while punishment avoidance is particularly coupled with action inhibition (Gray & MacNaughton, 2003
). In other words, learning to execute a response to obtain a reward is easier than learning to inhibit a response to obtain a reward, and learning to inhibit a response to avoid punishment is easier than learning to execute a response to avoid punishment (e.g., Guitart-Masip et al., 2012a
). This bias has been referred to as “Pavlovian bias” and is thought to originate from a conflict between Pavlovian control of behavior, which favors approach in the prospect of reward and avoidance in the prospect of punishment (Gray & MacNaughton, 2003
), and instrumental control of behavior, in which the behavioral output depends entirely on outcome valence (Guitart-Masip, Duzel, Dolan, & Dayan, 2014a
, Guitart-Masip et al., 2014b
). It has been proposed that prefrontal executive control mechanisms are needed to overcome these biases in the learning process (Cavanagh, Eisenberg, Guitart-Masip, Huys, & Frank, 2013
). In addition, the hippocampus has also been implicated in processing of approach-avoidance conflicts (for a review, see Ito & Lee, 2016
Interestingly, the Pavlovian bias is quite robust and evident not only when performing oneself, but also when learning merely by observing another individual’s actions and their consequences (Peterburs, Frieling, & Bellebaum, 2020
). Aside from contextual factors, such as agency, a growing number of studies have investigated modulation of feedback-based learning by inter-individual factors. For instance, depression, and specifically anhedonia, i.e., a lack of pleasure in response to ordinarily rewarding experiences, is associated with reduced reward sensitivity (e.g.; Huys, Pizzagalli, Bogdan, & Dayan, 2013
; see Must, Horvath, Nemeth, & Janka, 2013
for a mini review). Nevertheless, the Pavlovian bias is preserved in patients with mild to moderate major depressive disorder, suggesting that the motivational deficits typical for this disorder cannot be solely explained in terms of aberrant reward processing, and that an altered Pavlovian bias is unlikely to impair recovery (Moutoussis et al., 2018
). In contrast, a very recent study reported that patients with first-episode psychosis were impaired at overcoming Pavlovian bias (as reflected in lower learning rates and overall performance) and showed blunted sensitivity to both reward and punishment (Montagnese et al., 2020
), suggesting that these alterations in outcome processing may play an important role for the psychopathology of this disorder.
Social anxiety disorder (SAD), which is characterized by disproportional fear in and of (social) performance situations and feedback, has also been linked to alterations in reward processing and feedback-based learning. Cognitive models of the disorder (Clark & Wells, 1995
; Rapee & Heimberg, 1997
) postulate information processing biases towards increased attention to and memory for negative information as well as negative interpretation of ambiguous information. In line with this, empirical studies have found that high socially anxious individuals learned better than non-socially anxious individuals to avoid stimuli that were associated with negative feedback (Abraham & Hermann, 2015
; Voegler, Peterburs, Bellebaum, & Straube, 2019
) or ambiguous stimuli (Stevens, Peters, Abraham, & Hermann, 2014
) in probabilistic learning tasks. These studies only used tasks requiring response invigoration. Social anxiety, however, is characterized particularly by avoidance behavior: high socially anxious individuals often do not act, i.e., they do not go out, do not meet others, they avoid giving speeches, etc., and the persistence of this behavior suggests that it is reinforced because (anticipated) negative consequences of social situations are avoided. Moreover, this avoidance also prevents the opportunity for fear extinction, which also contributes to the persistence of the anxiety. It thus seems conceivable that the Pavlovian learning bias is altered in individuals with high social anxiety, with a strong tendency to avoid negative feedback, especially (but not only) by not acting. The present study aimed to investigate this notion. Importantly, previous investigations tested clinical samples (patients with SAD; e.g., Voegler et al., 2019
) or extreme groups (Abraham & Hermann, 2015
) to characterize the impact of high levels of social anxiety on feedback-based learning. However, it has been proposed that social anxiety is represented on a continuum ranging from subclinical behaviors (e.g., shyness) to clinical manifestation (SAD) based on common underlying dysfunctional mechanisms (Stein, Torgrud, & Walker, 2000
). Of note, applying dimensional rather than categorical approaches when studying psychopathology to better understand the full spectrum of mental health and mental illness is also at the heart of the Research Domaine Criteria Initiative (RDoc) of the National Institute of Mental Health. Along these lines, testing extreme groups or patient samples may not be ideal to map the impact of social anxiety as a continuous variable as it occurs in the healthy population. We therefore recruited a large sample of healthy adults who naturally varied in social anxiety (as determined based on self-report). Of note, this sample also included individuals with particularly low or high social anxiety scores but did not include individuals with a clinical diagnosis of SAD.
Subjects completed a variant of the orthogonalized go/nogo task first described by Guitart-Masip et al. (2011
) that decoupled action (response execution or inhibition; go/nogo) and outcome valence (win/avoid), thus allowing for direct assessment of the Pavlovian bias. Our main aim was to examine to what extent performance in the orthogonalized go/nogo task was affected by social anxiety. As mentioned above, high levels of social anxiety have previously been associated with better learning from negative than positive feedback (Abraham & Hermann, 2015
; Voegler et al., 2019
). We therefore also applied the so-called probabilistic selection task that enabled us to assess if subjects generally had a propensity to learn better from positive (approach learning) or negative feedback (avoidance learning). Previous work (e.g., Frank, Seeberger, & O'reilly, 2004
, Frank, Woroch, & Curran, 2005
; Stocco et al., 2017
) demonstrated individual differences in learning from positive and negative feedback in healthy participants which was also reflected in their electrophysiological responses to correct and erroneous stimulus choices, suggesting that the individual propensity to learn from positive/negative feedback may be a trait-like phenomenon that can be directly linked to midbrain dopaminergic function (Frank et al., 2005
). We thus included learning propensity as a potential additional predictor that may modulate the Pavlovian learning bias in feedback-based learning.
We expected that higher levels of social anxiety would be associated with reduced Pavlovian bias in a go context due to better learning of go to avoid compared to subjects with lower levels of social anxiety. In contrast, we hypothesized that the Pavlovian bias in the nogo context would be more pronounced in subjects with higher levels of social anxiety, due to better learning of nogo to avoid (compared to nogo to win) relative to subjects with lower levels of social anxiety. These effects of social anxiety on the Pavlovian Bias were expected to interact with learning propensity. Specifically, we expected the hypothesized social anxiety effects (i.e., reduced Pavlovian bias in the go context, more pronounced Pavlovian bias in the nogo context) to be strongest for socially anxious individuals with a strong bias towards better learning from negative than positive feedback.
The present study investigated the impact of social anxiety on the Pavlovian bias in feedback-based learning, with an additional focus on the individual propensity to learn from positive or negative feedback. To this end, a sample of healthy adults who naturally varied in their levels of social anxiety completed an orthogonalized go/nogo task that decoupled action type (go/nogo) and outcome valence (win/avoid) and a probabilistic selection task based upon which the individual propensity to learn from positive and negative feedback was determined. Self-reported social anxiety and learning propensity were used as predictors in LME model analysis of performance accuracy in the go/nogo task. In line with the expectations, we found an interaction between the Pavlovian bias on the one hand, which is reflected in the interplay of the factors action type (go/nogo) and outcome (win/avoid), the additional factor learning progress across blocks, and the predictor variables social anxiety and learning propensity on the other hand. Specifically, high socially anxious subjects with a propensity to learn better from negative feedback showed particularly pronounced learning for nogo to avoid and lacked significant learning for nogo to win as well as go to avoid. This result pattern indicates that high levels of social anxiety in concert with negative learning propensity hamper the overcoming of Pavlovian bias in a win context while facilitating response inhibition in an avoidance context.
Previous research has confirmed a robust asymmetric interaction of action and outcome type in feedback-based learning: learning to execute a response to obtain a reward or to inhibit a response to avoid punishment is much easier than learning the reverse, a phenomenon referred to as “Pavlovian bias” (Guitart-Masip et al., 2012a
; Peterburs et al., 2020
). This bias is supported by the present findings. In addition, learning was overall better for go than for nogo, which is also consistent with previous observations (e.g., Guitart-Masip et al., 2014b
; Ocklenburg et al., 2017
; Peterburs et al., 2020
) and could be attributed to higher task difficulty and/or cognitive demand in the context of response inhibition (relative to response execution), or a general propensity to respond in experimental sessions. Also in accordance with previous findings (Peterburs et al., 2020
), go to win was an easy condition to learn, as reflected in overall high accuracy rates and a lack of significant learning progress across task blocks. This result pattern strongly suggests a ceiling effect in go to win learning (see also Figs. 3, 4, 5).
Crucially, as expected, we found that both learning propensity and social anxiety together affected performance in the orthogonalized go/nogo task, reflected in an interaction between the five factors block (i.e., learning progress), action type and outcome (reflecting Pavlovian and instrumental conflict), learning propensity (positive/negative learners), and social anxiety. Further analysis revealed that high socially anxious subjects with a propensity to learn better from negative feedback showed a particularly pronounced increase in accuracy across blocks in the nogo to avoid condition, as reflected in a particularly steep slope (see Fig. 3
), while at the same time no learning was seen in the other conditions. High socially anxious subjects with a propensity to learn better from negative feedback also eventually achieved a high accuracy level for nogo to avoid, with > 70% in Block 4. This is partly in line with a negative learning bias/better avoidance learning in social anxiety (Abraham & Hermann, 2015
; Voegler et al., 2019
) and extends these previous findings in showing particularly strong learning of response inhibition to avoid negative feedback. Of note, similar accuracy levels were achieved by low socially anxious subjects with a propensity to learn from negative feedback. Interestingly, high socially anxious subjects with a propensity to learn better from negative feedback also showed a lack of significant learning for nogo to win and also go to avoid. This is somewhat unexpected and somewhat incompatible with the notion of attenuated Pavlovian bias. Rather, the present result pattern appears to suggest that high socially anxious subjects with negative learning bias are impaired in overwriting the Pavlovian bias in a win context. On the other hand, acquisition of response inhibition in context of avoidance is facilitated. In general, these results are consistent with a specific information processing bias in social anxiety that leads to increased attention to and/or memory for negative information (for an overview, see Peschard & Philippot, 2016
). Positive feedback, according to SAD psychopathology and cognitive SAD models (Clark & Wells, 1995
; Rapee & Heimberg, 1997
), is rather unexpected and inconsistent with the self-image of high socially anxious individuals and may thus be processed less efficiently. However, this seems to apply only when a certain degree of task difficulty is reached and (cognitive) resources are taxed, explaining why go to win performance in high socially anxious subjects with negative learning propensity was comparable to that of the other groups in the present study.
It has to be noted that the present sample was subclinical and did not include any individuals with a clinical diagnosis of SAD. Nevertheless, some individuals did report moderate to high social anxiety levels that can typically be observed in clinical samples (LSAS scores > 60; Mennin et al., 2002
; Rytwinski et al., 2009
). In contrast to previous investigations (e.g., Abraham & Hermann, 2015
; Pittig, Pawlikowski, Craske, & Alpers, 2014
), the present study was based on a rather large naturalistic sample and did not include a comparison of extreme groups of high and low socially anxious individuals. LME model analysis allowed inclusion of social anxiety as a continuous predictor and revealed that higher levels of social anxiety were associated with generally decreased learning performance. While this is certainly an intriguing result, it must be stressed that the clinical implications remain rather unclear. Future studies should specifically address the Pavlovian learning bias in patients with SAD. The use of disorder-specific stimulus material, such as faces or otherwise social feedback (rather than abstract stimuli), might be particularly informative in this regard and also increase ecological validity. And even outside of clinical populations, the type of feedback may be an interesting factor to manipulate in future experiments. A recent study involving a card gambling task (Case & Olino, 2020
) found that monetary and social positive and negative feedback both led to comparable learning (i.e., decreases in plays on disadvantageous decks across the task). Importantly, performance on the task with social feedback was associated with fun-seeking and depressive symptoms, indicating that using different types of feedback may help to better characterize social avoidance learning.
Recent findings have linked both pronounced positive and negative learning biases to better access to subordinate word meanings in a lexical ambiguity priming task (Ceballos et al., 2020
). This result suggests that basal ganglia function, which is reflected in these learning biases, directly impacts behavioral flexibility. In the context of the present task, increased behavioral flexibility could be expected to be linked with attenuated Pavlovian bias which would be reflected in more efficient learning of go to avoid as well as nogo to win. However, contrary to this notion, learning propensity only influenced task performance as a function of social anxiety. Of note, there was no population level bias towards positive or negative learning in the present sample and most subjects presented with mild to moderate biases in either direction. This is consistent with previous findings in healthy adults (Frank et al., 2005
). There also was no correlation between learning bias and social anxiety, indicating that there is no linear relationship between the severity of social anxiety symptoms and increased learning from negative feedback.
Interestingly, individual differences in learning from positive and negative feedback have been linked to basal ganglia dopaminergic function and underlying genetic differences. Consistent with findings showing that go learning in the context of positive feedback relies particularly on striatal dopamine D1 receptors, a polymorphism in the DARPP-32 gene, which has been associated with D1 receptor effects in synaptic plasticity, was linked to a positive learning bias (Frank, Moustafa, Haughey, Curran, & Hutchison, 2007
). Similarly, nogo learning in the context of negative feedback was linked to the C957T polymorphism of the DRD2 gene, a gene associated with striatal D2 receptor function (Frank et al., 2007
). Unfortunately, the current study was purely behavioral and thus cannot directly inform about gene or brain level effects. However, based on the present results we could speculate that high levels of social anxiety might be associated with altered basal ganglia function in response to negative feedback primarily affecting the nogo pathway and thus D2 receptor function. More research is needed to directly investigate this notion.
It has to be noted that the present study tested a rather homogenous sample both with regard to age and educational attainment, so findings may not generalize to a (more diverse) community sample. Moreover, subjects were not recruited to balance sex. Sex differences in learning from positive and negative feedback have been reported, with females showing better learning from positive feedback than males (Evans & Hampson, 2015
), and may thus present a possible confound in the present study. To our knowledge, potential sex differences with regard to learning to execute or inhibit a response to obtain a reward or avoid punishment have not been investigated yet. A substantial body of literature has explored sex differences in anxiety and mood disorders. With regard to social anxiety, women are more likely than men to have SAD and present with more severe symptoms (e.g., Asher & Aderka, 2018
), although findings regarding functional impairment are inconclusive (Asher, Asnaani, & Aderka, 2017
). Since our sample included only 17 men, meaningful subgroup analysis elucidating possible sex differences in the impact of social anxiety and learning propensity cannot be performed. However, this may be an interesting avenue for further research, as might be potential effects of acute stress, which has been shown to decrease learning from negative (but not from positive) feedback (Petzold, Plessow, Goschke, & Kirschbaum, 2010
To conclude, the present study investigated the interplay between individual levels of social anxiety and biases in feedback-based learning. The results confirmed a robust Pavlovian bias: learning to execute a response to obtain positive feedback was easier than learning to inhibit a response to obtain positive feedback, and learning to inhibit a response to avoid negative feedback was easier than learning to execute a response to avoid negative feedback. Importantly, as expected, this asymmetric coupling of action and outcome valence was modulated as a function of social anxiety and individual learning propensity. High socially anxious subjects with a propensity to learn better from negative feedback showed particularly pronounced learning for nogo to avoid and a lack of significant learning for nogo to win as well as for go learning. Thus, high levels of social anxiety in concert with negative learning propensity appear to interfere with the overcoming of Pavlovian bias in a win context while facilitating response inhibition in an avoidance context. In general, these findings add to a growing body of evidence for altered outcome processing and adaptive behavior in social anxiety. It could also be speculated that they might yield interesting clinical implications although the present study did not test a clinical sample. Possibly, determining an individual’s learning propensity and taking into account differences in learning as a function of outcome and action type might help to develop therapeutic interventions targeting SAD.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.