All our participants were recruited via advertisements in local newspapers, notices and flyers. Online questionnaires and telephone interviews were used to check eligibility. Exclusion criteria were age under 18 years or over 55 years, any acute or chronic medical condition including neurological disorders, smoking of more than five cigarettes per day, current use of medication, drug or alcohol abuse, a body mass index (BMI) > 30, or shiftwork. Participants could also not participate if they were students of psychology or economics and if they were not naïve to the TSST procedure. Other exclusion criteria for the SAD group were any other current mental disorder except SAD or avoidant personality disorder. For the control group, participants could not have or have had any mental disorder diagnosis. Due to sex hormone influences on the cortisol stress response, we included only male participants (34). We based our sample size calculation on previous studies with similar designs (35, 36) and planned to recruit 30 participants per group including possible drop-outs or missing data. A total of 121 subjects (63 SAD; 58 HC) met these criteria and participated in our experiment. From this sample, three subjects with SAD and six healthy controls had to be excluded for this reasons: current medication use (n = 5), critically elevated global psychiatric symptoms in control sample [t-value of the Brief Symptom Inventory > 70; (37)] (n = 3), one subject was known to one of the experimenters (n = 1). Our final sample consisted of 112 participants, with 60 SADs, age 28.30 ± 8.91 years (mean ± SD), and 52 healthy controls, age 27.85 ± 6.67 years. Participants were randomly assigned to the stress or no-stress condition of the Trier Social Stress Test for groups (TSST-G (38); SAD: 31 in the stress condition, 29 in the no-stress condition; healthy controls: 27 in the stress condition, 25 in the no-stress condition).

The Multifaceted Empathy Test [MET; (39)] was used to measure emotional and cognitive empathic abilities. It consists of 30 pictures showing people in positive and negative emotionally laden situations. To assess cognitive empathy, participants had to choose the correct mental state description out of four labels presented at the bottom of the picture. To measure emotional empathy, participants rated their experienced level of empathic concern for the person in the picture on a 9-point Likert scale. Trial presentation and response registration was controlled by a PC running Presentation (Neurobehavioral Systems Inc., Albany, CA, United States). For statistical analyses, trials within the cognitive or emotional condition were averaged to provide us with one dependent variable. Due to technical problems, the sample is missing MET data from five participants.

For the stress manipulation, we applied the TSST-G, a standardized laboratory protocol inducing psychosocial stress in groups of up to six participants (38). In the TSST-G, subjects are separated by dividing walls to prevent them from interacting. The TSST-G comprises two conditions, namely, a stress and a control (no-stress) condition. In the stress condition, participants have to give a 2-min free speech for a mock job interview and do a mental arithmetic task, both in front of two evaluators and two cameras. In the no-stress condition, participants had to read a text in a low voice and recite number series and are not videotaped or evaluated in any way. This resembles an active control condition with the same physiological features (standing upright, talking) but without the stress-inducing aspects of social evaluation or uncontrollability that are specific to the stress condition. The TSST-G enables a moderate psychosocial stress induction and has been proven to reliably result in activation of the HPA axis with elevated cortisol levels, as well as significant cardiovascular and subjective stress responses with a specific and active control condition that does not lead to psychobiological stress responses (38, 40, 41).

We measured salivary cortisol and heart rate to assess the physiological stress response. To measure the subjective stress reaction, a visual analogue scale (VAS) was used. Cortisol and subjective responses were collected at five time points throughout the experiment (Figure 1).

Saliva samples for cortisol assaying were collected using a commercially available sampling device (Sarstedt^®, Nümbrecht-Rommelsdorf, Germany). Saliva samples were stored at −20°C. For biochemical analyses of free cortisol concentration, saliva samples were thawed and spun at 3.000 revolutions per minute for 10 min to obtain 0.5–1.0 ml of clear saliva of low viscosity. Salivary cortisol concentrations were determined by a commercially available chemiluminescence immunoassay (CLIA; IBL, Hamburg, Germany). Inter- and intra-assay coefficients of variation were below 8%. Because of too little saliva, one participant’s cortisol data were not available (n = 111).

Heart rate was measured via a wireless chest heart rate transmitter and wrist monitor recorder (Polar RS800TM, Polar Electro^®, Oy, Kempele, Finland) and recorded continuously as an indicator of the sympathetic adrenal medullary system (SAM) (see experimental course, Figure 1). We assessed an aggregated baseline mean over 5 min in a standing position. For the stress induction period, means from 35 1-min intervals were entered into the analyses. Due to technical problems, heart rate measures were not available for n = 11 at the baseline and n = 8 participants during the stress induction period. Heart rate analyses were thus conducted with n = 101 participants at baseline and n = 104 during the stress induction period.

We assessed social performance via an everyday conversation task, based on Voncken and Bögels (17). This paradigm enables an experimentally controlled investigation of social interaction with high ecological validity. Social performance comprises social behavior and anxious appearance (see below). Participants had to start a 5-min conversation with a confederate. Instructions were as follows: “We would like you to have a conversation with another person. The purpose is to get to know each other. It is up to you to start the conversation and to keep it going.”

Confederates were female undergraduate students. All had undergone a 3 h training session in the experimental conversation paradigm and rating social interaction with the Social Behavior and Anxious Appearance Rating Scale [SBA-rating scale; (17)]. They had been trained to remain in a neutral but friendly posture during the conversation and to answer with up to three pieces of information. They were trained to only pose a question after 7 s of silence. That way, the responsibility for keeping the conversation going remained with the participant. We based these instructions on Voncken and Bögels (17). Confederates were blinded for the condition (stress/no-stress) and group (SAD/HC). After the conversation, confederates rated the participants’ social performance via a modified version of the SBA-rating scale (17).The SBA-rating scale consists of 27 items and two scales: anxious appearance and social behavior. Anxious appearance covers signs of nervousness such as trembling, blushing or stuttering. Examples for the questions are: “Was the participant shaking/trembling (e.g., hands, legs or parts of the face)?” or “Was the participant laughing nervously?” Social behavior comprises items relating to formal aspects of interaction behavior, such as maintaining eye contact or formulating full sentences, as well as more complex aspects, such as the degree of self-disclosure or showing interest in the conversation partner. Examples for questions are: “Did the participant seem confident?” or “Did the participant make eye contact with you?” The two-factor structure was confirmed by Voncken and Bögels (17) and Bögels et al. (42). The original scale refers to a conversation involving three persons and was adapted for a two-party conversation in our study. Therefore, one item of the SBA scale had to be removed (“Could the participant divide attention between you both?”) which resulted in 26 items for the current investigation. Social behavior and anxious appearance represent the two main dependent behavioral variables of social performance in the conversation paradigm.

To assess social anxiety as a continuous variable, we applied the Liebowitz Social Anxiety Scale [LSAS; (43)]. Depressive symptoms were rated with the Beck Depression Inventory [BDI; (44)]. To assess the general burden of psychiatric symptoms, we used the Brief Symptom Inventory [BSI; (45)]. We relied on the Perceived Stress Scale to measure chronic stress levels [PSS; (46)].

Participants came to the laboratory twice. The first appointment was scheduled for individual diagnostics including SKID I and SKID II [German version of the Structural Clinical Interview for DSM IV; (47)], and the Multifaceted Empathy Test [MET; (39)] for measuring cognitive and emotional empathy abilities.

After providing informed consent, all participants were interviewed by a trained clinician and performed the MET individually on a PC. The diagnostic session lasted approximately 2 h.

If participants fulfilled our inclusion criteria, they were randomly assigned to the stress or control condition. They were scheduled for the experimental session in groups of four to six participants. The groups were always mixed with participants in the SAD and HC group. Participants were asked not to do any physical exercise and to abstain from caffeine, nicotine, alcohol or any medication for 24 h prior to the experiment. They were instructed to eat their usual standard breakfast and lunch and to stop food consumption at 4 P.M. Sessions started between 5 and 6 P.M. to control for diurnal variation in cortisol secretion. After arriving at the lab, participants again provided informed consent. They were not allowed to communicate with each other, were provided with a heart rate device (Polar, RS800TM, Polar Electro^®, Oy, Kempele, Finland) and were seated individually in a cubicle computer lab. They were given information about saliva collection and how to set markers for the heart rate recording. After filling out the first visual analogue scales, they underwent the TSST-G (38). Each of the two parts of the TSST-G (speech/reading and mental arithmetic/counting, see next section), was followed by a set of decisions on a behavioral economic paradigm [adapted from von Dawans et al. (35)]. Due to space shortage, these results are being prepared for a second manuscript. At the end of the TSST and the second decision paradigm, participants were informed that the next task would be a conversation with a stranger. They were guided individually into separate rooms where a trained female confederate was waiting and the conversation task took place for 5 min. Participants then came back into the computer lab for further questionnaires (BDI, BSI, LSAS, and PSS), saliva sampling and VAS (see Figure 1). All participants were debriefed and reimbursed at the end of the study, which was approved by the local ethics committee of the University of Freiburg, Germany. Participants gave written informed consent on both study days. Subjects received a flat fee of 40 Euros and could earn additional money in the decision paradigm (depending on their own decisions and the decisions of another group of participants not subjected to the TSST-G procedure) with a mean of 5.69 ± 0.90 Euros (SD).

To test for differences in groups with respect to psychometric data or baseline measures we conducted two factorial ANOVAs with the factors stress/control and SAD/HC. We also ran two MANOVAS, one for empathy (cognitive and emotional) and one for social performance (anxious appearance and social behavior). With respect to the moderation of the effects of stress on social behavior by empathic abilities we conducted two moderation analyses separately for each HC and SAD (one for anxious appearance, one for social behavior). The PROCESS macro for SPSS by Hayes (48) was used (model 2). Moderation analyses were conducted for both SAD and HC with condition as the independent variable, and anxious appearance and social behavior, respectively, as dependent variable; cognitive and emotional empathy as two moderators.

Participants with SAD and HC did not differ in age [F(1, 108) = 0.07, p = 0.786, η_p² = 0.001] or education [χ² (df = 2) = 0.248, p = 0.883]. Participants in the stress group tended to be older [F(1, 108) = 3.31, p = 0.072, η_p² = 0.030]. As expected participants with SAD exhibited significantly higher levels of psychopathological symptoms, and they reported higher levels of social anxiety [LSAS: F(1, 104) = 138.56, p < 0.001, η_p² = 0.571], depressive symptoms [BDI: F(1, 107) = 60.70, p < 0.001, η_p² = 0.362] and general psychiatric symptoms [BSI global severity index: F(1, 108) = 137.79, p < 0.001, η_p² = 0.561] as well as higher levels of chronic stress [PSS: F(1, 108) = 77.84, p < 0.001, η_p² = 0.419]. There were neither significant differences between stress and control condition nor any significant interactions between the two factors (all p > 0.10) (Table 1).

Participants with SAD and healthy controls did not differ in terms of empathic abilities, in their level of cognitive or emotional empathy. We also found that randomization to the experimental conditions succeeded: there were no differences in empathic abilities between the stress and control condition and no interaction effects (all p > 0.10) (Table 2).

We noted a divergence in baseline measures between psychological stress levels and biological stress markers. Participants with SAD showed significantly higher levels on the VAS stress at baseline [F(1, 108) = 12.45, p = 0.001, η_p² = 0.103] whereas they did not differ from HC in cortisol [F(1, 107) = 1.35, p = 0.247, η_p² = 0.012] or heart rate [F(1, 97) = 1.81, p = 0.181, η_p² = 0.018] at baseline.

Our within-subjects results showed that inducing stress via the TSST-G was successful. Besides an overall effect of time [F(3.44, 358.18) = 43.11, p < 0.001, η_p² = 0.293], participants in the stress condition displayed a significantly higher subjective stress response measured with the VAS stress [time × stress: F(3.44, 358.18) = 9.92, p < 0.001, η_p² = 0.087] compared to the control condition. Moreover, there was a significant time × SAD interaction documenting higher subjective responses in SAD than HC [F(3.44, 358.18) = 4.93, p = 0.001, η_p² = 0.045]. The three way interaction [time × stress × SAD: F(3.44, 358.18) = 2.12, p = 0.077, η_p² = 0.021] reached significance on the trend level.

With respect to the between-subject effects we detected significant main effects for the stress condition [F(1, 104) = 6.09, p = 0.015, η_p² = 0.055], SAD [F(1, 104) = 42.80, p < 0.001, η_p² = 0.292], and a significant interaction between stress × SAD—yet more evidence of higher overall subjective stress levels in SAD than HC [F(1, 104) = 4.28, p = 0.041, η_p² = 0.039; Figure 2A].

For the within-subjects effects, the repeated measures ANOVA for cortisol revealed an effect of time [F(2.21, 235.99) = 28.33, p < 0.001, η_p² = 0.209]. Participants in the stress condition exhibited significantly stronger cortisol stress responses than those in the control condition [time × stress: F(2.21, 235.99) = 39.86, p < 0.001, η_p² = 0.271]. The time × SAD and three way interaction time × stress × SAD were not significant. The between-subjects effects show a main effect for stress that reflects higher overall cortisol levels in the stress than in the control condition [F(1, 107) = 23.54, p < 0.001, η_p² = 0.181; Figure 2B].

The same pattern appears again concerning the areas under the curve: significant effects are revealed for stress induction for the AUC_G [F(1, 107) = 24.85, p < 0.001, η_p² = 0.188] as well as the AUC_I [F(1, 107) = 64.48, p < 0.001, η_p² = 0.376]. There were no differences in SAD vs. HC, nor any significant stress × SAD interactions (all p > 0.10, Figure 2B bar charts).

For heart rate responses we again found successful stress induction revealed through a significant effect of time [F(8.24, 824.14) = 35.64, p < 0.001, η_p² = 0.263] and time × stress interaction [F(8.24, 824.14) = 2.34, p = 0.016, η_p² = 0.023]. There were no other significant effects in heart rate responses to stress (Figure 2C).

Both our results on cortisol and heart rate demonstrate a divergence between subjective and biological variables that is reflected in stronger subjective responses, but similar biological responses in SAD compared to HC (Figure 2).

In a multivariate analysis of variance, we tested for differences in social behavior and anxious appearance in SAD compared to healthy controls, as well as for the effects of the acute stress induction. The latter failed to show a significant main effect [F(2, 107) = 0.161, p = 0.852, η_p² = 0.003, Wilk’sΛ = 0.997]. As expected, we observed a significantly better performance in HC than SAD [F(2, 107) = 8.884, p < 0.001, η_p² = 0.142, Wilk’sΛ = 0.858]. Moreover, there was a significant stress × SAD interaction revealing a different pattern for the stress induction in SAD patients vs. HC [F(2, 107) = 3.550, p = 0.032, η_p² = 0.062, Wilk’sΛ = 0.938; Figure 3].

With our final step, we wanted to clarify whether empathy modulates the significant stress × SAD interaction of on social performance and used a moderation model conducted with PROCESS (Model 2) for SAD and HC separately (Figure 4).

There were no significant effects for HC. Neither stress induction nor empathic abilities nor the interaction explained anxious appearance [F(5, 44) = 1.18, R² = 0.118, p = 0.336] or social behavior in HC [F(5, 44) = 1.09, R² = 0.110, p = 0.382].

In SAD, the model summary for anxious appearance was not significant [F(5, 51) = 1.91, R² = 0.158, p = 0.109]. When looking at the model, we noted a significant stress by cognitive empathy interaction [stress × cognitive empathy: F(1, 51) = 5.07, change R² = 0.084, p = 0.029]. Moreover, stress alone predicted anxious appearance on the trend level [b = 5.97, t(51) = 1.91, p = 0.062; Figure 4A].

Our moderation model for social behavior was significant [F(5, 51) = 5.45, R² = 0.348, p < 0.001]. We identified a significant negative effect of stress on social behavior [b = −7.12, t(51) = −3.02, p = 0.004]. With respect to the direct effects of empathic abilities, cognitive empathy shows significance on the trend level [b = 0.12, t(51) = 1.97, p = 0.055] while emotional empathy had no direct effect on social behavior (p = 0.803). However, both facets of empathy moderate the effects of stress on social behavior [stress × cognitive empathy: F(1, 51) = 5.23, change R² = 0.067, p = 0.026; stress × emotional empathy: F(1, 51) = 4.05, change R² = 0.052, p = 0.050; Figures 4B, 5]. All results of the moderation models are found in the Supplementary Material (SOM). Moreover, the figure shows the level of significance for the conditional effects of the focal predictor at values of the moderators (all parameters can be found in the SOM).