Social threat avoidance depends on action-outcome predictability

Sequestro, Matteo; Serfaty, Jade; Grèzes, Julie; Mennella, Rocco

doi:10.1038/s44271-024-00152-y

Download PDF

Article
Open access
Published: 26 October 2024

Social threat avoidance depends on action-outcome predictability

Communications Psychology volume 2, Article number: 100 (2024) Cite this article

4294 Accesses
2 Citations
14 Altmetric
Metrics details

Subjects

Human behaviour

Abstract

Avoiding threatening individuals is pivotal for adaptation to our social environment. Yet, it remains unclear whether social threat avoidance is subtended by goal-directed processes, in addition to stimulus-response associations. To test this, we manipulated outcome predictability during spontaneous approach/avoidance decisions from avatars displaying angry facial expressions. Across three virtual reality experiments, we showed that participants avoided more often when they could predict the outcome of their actions, indicating goal-directed processes. However, above-chance avoidance rate when facing unpredictable outcomes suggested that stimulus-response associations also played a role. We identified two latent classes of participants: the “goal-directed class” showed above-chance avoidance only in the predictable condition, while the “stimulus-response class” showed no credible difference between conditions but had a higher overall avoidance rate. The goal-directed class exhibited greater cardiac deceleration in the predictable condition, associated with better value integration in decision-making. Computationally, this class had an increased drift-rate in the predictable condition, reflecting increased value estimation of threat avoidance. In contrast, the stimulus-response class showed higher responsiveness to threat, indicated by increased drift-rate for avoidance and increased muscular activity at response time. These results support the central role of goal-directed processes in social threat avoidance and reveal its physiological and computational correlates.

Associative learning shapes visual discrimination in a web-based classical conditioning task

Article Open access 03 August 2021

Evaluation of an integration community project for asylum seekers in Sweden: physical activity adherence and changes in character traits and life satisfaction

Article Open access 13 September 2024

Observational threat learning influences costly avoidance behaviour in healthy humans

Article Open access 28 July 2024

Introduction

Facial emotional displays convey critical information about the sender’s affective state and associated behavioral intentions, but also about potential threats and/or opportunities in the environment¹. Such information helps us to shape adaptive social behavior, such as avoiding threatening individuals and approaching friendly ones^2,3. Although there is a consensus that emotions are intimately linked to action, the nature of such a link is still a matter of debate^1,4.

Responses to socio-emotional stimuli have been thought to be strongly influenced by simple motor tendencies to approach or avoid, selected by evolution and/or development for their adaptive value^5,6,7,8. Such action tendencies are determined by the association between the representation of a specific stimulus and the representation of a specific action⁹, and can either be innate or formed by associative learning (i.e., by habit formation⁷). In the case of angry facial displays, individuals may learn throughout life that such emotional expressions predict the risk of an aggressive confrontation by personal or vicarious experience^10,11. In this sense, habit formation may consolidate stimulus-response (SR) associations thought to be automatically activated when seeing threatening faces, for example, to facilitate fast reactions to threat. Many emotional accounts assume SR associations to be the default determinants of behavior in response to social threat, while cognitively costly and non-automatic goal-directed (GD) processes would only control and refine automatic tendencies^8,12,13. This regulatory process is goal-directed as it is oriented toward one or more goals that may or may not be congruent with the initial SR tendency⁷. This architecture is considered efficient in selecting responses to threat, as automatic SR tendencies allow swift action selection in response to danger, while slower GD processes enhance decision optimality¹⁴.

More recently, however, the dichotomy between SR and GD processes has been criticized as over simplistic^9,15,16. Specifically in relation to defensive behavior, SR and GD processes are nowadays conceived on a continuum, where SR processes, leading to both habitual and innate defensive responses (e.g., fight/flight), prevail under imminent threat, whereas GD processes dominate during safety and pre-encounter states (i.e., before the threat is detected in the environment^17,18,19). Importantly, recent research in humans^{20,21,22,23,24} and rodents^{25,26,27,28,29} converge in assigning a more important role to GD defensive responses than previously thought, both under low and relatively imminent threat^7,9,30. Indeed, not all GD processes are slow and costly³¹, as complex information integration can lead to fast, implicit, and even unconscious estimates of the best action-outcome option, combining automaticity with optimality^7,32,33,34. In line with this, when threat is present but not yet attacking and a fast decision needs to be taken (i.e., post-encounter phase), a mixture of “cognitive” (GD) and “reactive” (SR) defensive modes emerges^18,35,36, raising the question on how the brain arbitrates between the two. Notably, social threats, such as encounters with angry or aggressive individuals, most likely lie in this transition zone.

To test whether GD or SR processes, or a combination of both, are involved in determining behavior, a classical approach consists in manipulating the outcome value of a response (e.g., by outcome devaluation³⁷) or the action-outcome contingency (e.g., by contingency degradation³⁷). Indeed, while GD behavior is sensitive to outcome values and action-outcome contingencies, SR behavior is not affected by learning or environmental feedback^37,38. It has been proposed that the controllability of expected action-outcomes governs the balance between SR and GD processes³⁹. Specifically, when an action reliably leads to a predictable consequence (controllable action-outcome context), GD processes should be privileged as they are more optimal. Conversely, in uncontrollable contexts, when the contingency between actions and consequences is stochastic, the additional computational complexity of GD processes is unnecessary, and SR associations are a more parsimonious policy^39,40.

The current study aimed at characterizing the contribution of SR and GD processes to spontaneous avoidance of threatening avatars in Virtual Reality (VR). Participants had to choose to enter one of two elevators, each containing a virtual avatar that displayed an angry or a neutral facial expression. We manipulated the controllability of action-outcomes, by comparing two conditions in which threat-related action-outcome associations were predictable and unpredictable, respectively. For example, choosing to enter the elevator with the neutral avatar in the predictable condition deterministically resulted in the predicted outcome (i.e., avoidance of the angry avatar), guaranteeing full controllability. Conversely, choosing to enter the elevator with the neutral avatar in the unpredictable condition could stochastically result either in anger avoidance or approach, precluding controllability of the outcome. Importantly, participants were instructed to make spontaneous decisions, and the threatening expressions of the avatars were never explicitly mentioned.

In the predictable condition, we expected participants to avoid the angry avatar more than chance, regardless of whether avoidance is supported by GD processes, SR processes, or a combination of the two. In the unpredictable condition, if GD processes are involved in the decision, avoidance rates should decrease, because of the stochasticity in action-outcome contingencies. Conversely, if avoidance is solely driven by SR associations, no reduction in avoidance rates should be observed. Importantly, a simple reduction in avoidance rate under unpredictability would not rule out the involvement of SR processes in the decision, since SR processes might just keep operating in the unpredictable condition as they do in the predictable one. However, an avoidance rate at chance level in this condition would suggest that SR processes are not driving avoidance, either because they are not elicited in response to threatening expressions, or because they are not overtly influencing avoidance. Therefore, by comparing the avoidance patterns between the two conditions, we tested the effect of GD and SR processes over social-threat avoidance.

Methods

The presented studies were not preregistered.

Participants

We conducted three experiments and for each experiment we ran power simulations⁴¹ to determine the sample size required for detecting an effect of the Condition (Predictable vs. Unpredictable) on avoidance choices with 80% power, each time based on the results of the previous experiment (for Experiment 1, see Supplementary Note 1 for a detailed presentation of the pilot experiment on which we based the simulation). The simulation approach for power calculation involves fitting a Generalized Linear Mixed Model (GLMM) to an initial dataset to extract fixed and random effect parameters. These parameters are then used to generate data for a specified number of artificial participants, on which the same GLMM is fitted. This process is repeated 1000 times for each sample size to test. The power is calculated as the proportion of artificial datasets that show a significant effect (p < 0.05) of the desired parameter out of the total 1000 datasets⁴¹. Compared to analytical methods for power calculation, the simulation approach is suitable for GLMMs and accounts for inter-individual variability (i.e., the random structure of the model). Participants in all experiments were aged between 18 and 35 years old, had no history of neurological or psychiatric disorders (a full list of inclusion criteria is provided in Supplementary Methods 1) and were recruited using convenience sampling through flyers on the university campus and online advertisements.

Experiment 1: The power simulation indicated a required sample of 20 participants. Since in the pilot experiment we had data losses due to cybersickness and since the task design was modified to reduce this risk, we decided to aim for 60 participants. Sixty-five volunteers participated in this experiment. We excluded data from 3 participants due to technical issues, 1 due to cybersickness, and 1 because they reported already being familiar with the task. The final sample consisted of 60 participants (n = 30 females, n = 30 males; mean age ± SD = 23.51 ± 3.83). Data for this experiment was collected between July and October 2022.

Experiment 2: The power simulation indicated a required sample of 30 participants. Thirty-two volunteers participated in this experiment. We excluded data from 1 participant due to cybersickness and 1 because they showed lack of understanding of the task instructions. The final sample consisted of 30 participants (n = 15 females, n = 15 males; mean age ± SD = 22.60 ± 3.29). Data for this experiment was collected between December 2022 and January 2023.

Experiment 3: The power simulation indicated a required sample of 60 participants. Sixty-six volunteers participated in this experiment. We excluded data from 3 participants due to cybersickness and 3 due to technical issues. The final sample consisted of 60 participants (n = 32 females, n = 28 males; mean age ± SD = 23.42 ± 4.59). Data for this experiment was collected between April and June 2023.

For all experiments sex and age were self-reported by participants and no data about race/ethnicity was collected as it is illegal in France, except under special circumstances. The experimental protocol was approved by INSERM, licensed by the local research ethics committee (IRB00003888 – Avis 18-544-ter – 25.10.2021). Participants provided written informed consent and were compensated for their participation. See Supplementary Table S47 for a detailed description of the three samples.

Materials (experiments 1, 2 and 3)

Equipment and software

A virtual environment was presented through the Virtual Reality Head Mounted Display (VR HMD) Oculus Quest 2 (Meta Quest). The Oculus HMD has a resolution of 1920*1832 (per eye). Two Oculus Touch Controllers (one per hand) were used to collect participant responses. The VE was developed using the Unity game engine (version 2020.3.19f1, Unity Technologies) and the UXF package (Unity Experiment Framework⁴²). The vertical Field of View of the Oculus device was set to 60 degrees (per eye), and the virtual environment was updated at a constant frequency of 60 Hz.

The software was run on an Alienware Aurora Ryzen Edition PC with a 3.00 GHz 12-core AMD Ryzen 9 5900 processor, a 32GB DDR4 3400 MHz RAM, and an NVIDA GeForce RTX 3080 10GB GDDR6X graphic card.

Stimuli

The virtual environment contained dyads of full-body virtual avatars presenting neutral and angry face expressions selected from the Radboud Face Database⁴³. The faces of each dyad of avatars to be presented together were matched for trustworthiness and threat traits⁴⁴. The 2D faces were converted into 3D head models and merged into 3D body models. We selected images for 10 individuals for each sex, resulting in 10 individuals × 2 sexes (male and female) × 2 facial expressions (neutral and angry), for a total of 40 avatars. Later, we discarded 4 avatars (a dyad for each sex) due to their poor quality in virtual reality, leaving us with a total of 36 stimuli. Detailed information about the creation of the stimuli used in this study can be found in Supplementary Methods 2.

VR Task

In the virtual environment, participants faced a wall with two elevators and a frame with two arrows (one pointing up and one pointing down) between them (Fig. 1A). Each trial began as soon as participants fixated the central arrows for 2.5 s. At the elevators’ arrival the doors opened, revealing a 3D avatar inside each elevator. Two-thirds of the trials presented an avatar with an angry face and an avatar with a neutral face (Threat trials), while one-third of the trials presented the two avatars with only neutral faces (Neutral trials). While most analyses were focused on participants’ behavior in Threat trials, the addition of Neutral trials was motivated by the goal of dissimulating as much as possible the presence of threatening expression. Furthermore, this allowed us to run control analyses by testing differences between Threat and Neutral trials in terms of response times and physiological responses (Experiment 3).

**Fig. 1: Task structure and avoidance rates.**

Participants were asked to freely choose which elevator to enter as quickly as possible after the doors were fully opened. Notably, in Threat trials, participants could choose to enter the elevator with the angry avatar (approach response) or the one with the neutral avatar (avoidance response). The avatars were presented in the periphery of the participants’ visual field at a visual angle of 20°. Participants were instructed to maintain fixation on the central arrows during choice. They held in each hand a VR response controller to provide left/right responses and had limited time to make a choice (i.e., 1.5 s). If the response was not given in time, the trial ended and a red text prompt “Plus rapide!” (“Faster!”) was displayed in the virtual environment.

Importantly, the arrow at the center of the scene served as the cue associated with the Condition (Predictable or Unpredictable). Participants were informed that when the arrow pointing up lit (Predictable condition), the avatars would remain in their respective elevators after choice. For example, when participants clicked on the left button, an automatic camera movement after the choice led them next to the avatar present in the left elevator at the moment of choice. Therefore, participants could predict, and thus control in a deterministic fashion, the avatar to be next to (or far from). Conversely, participants were informed that when the down arrow lit (Unpredictable condition), the avatars in both elevators would walk outside after the choice and leave the virtual environment, while two new avatars (not previously visible) would enter the elevators. Importantly, the face identities of the new avatars were always different from those of the avatars present at the moment of choice. Participants were unaware that the new avatars could display an angry or a neutral expression with a 50% probability. Therefore, they had no control over whether they would ultimately approach or avoid angry avatars in Threat trials. Participants were led next to the avatar in the chosen elevator only after the new avatar entered. At the end of each trial, participants were “teleported” back to the starting position and the scene reset within 0.5 s between trials.

It must be stressed that participants were only instructed to freely choose to enter the elevator they preferred. They were further informed of the contingency between the cue (the arrow between the elevators) and the fact that the avatars would stay or leave the elevators after participants’ response. No other explicit instructions were given, and the presence of threatening expressions was never mentioned. Overall, we chose this experimental setting to investigate spontaneous avoidance behavior. Furthermore, note that the terms “Predictable” and “Unpredictable” do not refer to the avatars’ movements, which were known and explicit, but to the outcome of the choice in terms of anger approach/avoidance in Threat trials.

The task consisted of one to three practice blocks of 12 trials each, followed by four main blocks of 24 trials each. Between each block, participants were given a 1.5-minute break during which they could remove the VR HMD. To prevent cybersickness, these breaks could be extended for as long as participants wished. All participants completed a first practice block to familiarize with the virtual environment and the task and completed one or two additional practice blocks if their response accuracy was below 80% (if responses were not made within 1.5 s). After each block, participants received feedback indicating the percentage of responses they had provided in the correct timing. If this percentage was lower than 80%, they were asked to make faster decisions.

Avatars with angry expressions were presented only in the main blocks and not in the practice blocks. In each block, half of the trials were from the Unpredictable condition and half from the Predictable condition. Each trial presented only avatars of the same sex, including the new avatars that entered the scene after the choice in the Unpredictable condition. The side of the elevator with the angry avatar (left, right) and the sex of the avatars in the elevators were randomized and counterbalanced throughout the task.

Finally, there was a slight modification of the task between Experiment 1 and Experiments 2 and 3. In Experiment 1, immediately after trial initiation, one of the two arrows lit signaling the arrival of the elevators and, after a jittered time interval (mean interval ± SD = 0.9 s ± 0.3), the doors opened revealing a 3D avatar inside each elevator (Supplementary Fig. S12). In Experiments 2 and 3, the arrow lit simultaneously with the opening of the doors (Fig. 1A; videos of both versions of the task are provided in the OSF repository associated with this study⁴⁵). To prevent participants from anticipating responses, we maintained the jittered interval but shifted it before the onset of the stimulus. The duration of the interval remained the same as in Experiment 1 (mean interval ± SD = 0.9 s ± 0.3).

This choice was motivated by results from Experiment 1, which suggested that participants initiated their decision at the onset of the cue, rather than at door opening. Accordingly, the trial-by-trial duration of this jittered interval in Experiment 1 was negatively correlated with Response Times (RTs; Gamma GLMM with identity link function: β = −0.048; p = 0.005). We reasoned that this was problematic, as a longer time to decide may favor GD processes, by facilitating implementation intentions (i.e., goal-directed plans that automatically link anticipated situational cues to behavioral responses^46,47,48). Furthermore, since RTs recorded from the moment the doors opened may not capture the entirety of the decision process, this could bias models that include response times, such as the drift-diffusion models (see below).

Self-Report Measures

Numerous studies have shown that social interactions and decision-making can be affected by changes in positive affect⁴⁹, depression⁵⁰, anxiety and social anxiety⁵¹, as well as reward/punishment sensitivity⁵². Therefore, participants completed a series of self-administered questionnaires to account for these variables in our sample. Of note, characterizing the effect of these affective and personality traits on behavior is beyond the scope of the present article, which focuses on the balance between SR and GD processes in social threat avoidance. We nonetheless include descriptive results from self-report measures including the observed Cronbach alphas in Supplementary Table S47.

The administered questionnaires were:

The Positive and Negative Affect Schedule (PANAS⁵³), to measure affect;
The Patient Health Questionnaire depression scale (PHQ-8⁵⁴) to measure depressive symptoms;
the State Trait Anxiety Scales (STAI⁵⁵) and the Liebowitz Social Anxiety Scale (LSAS⁵⁶), to measure anxiety and social anxiety;
The Behavioral Inhibition/Behavioral Activation scales (BIS-BAS⁵⁷).

Finally, to track possible cybersickness symptoms due to VR, participants filled in the Simulator Sickness Questionnaire⁵⁸.

Materials specific to Experiment 3: physiological and subjective measures

Previous studies have shown that phasic variations in heart rate (HR) are associated with distinct stages of approach and avoidance decisions. Specifically, the magnitude of cardiac deceleration around the presentation of a threatening stimulus has been associated with the efficiency of information sampling and action preparation, while subsequent cardiac acceleration has been associated with action execution^59,60,61. Rather than being a passive reaction to threatening situations, bradycardia has also been considered as an active state that facilitates the preparation of value-based actions and the selection between different action possibilities^60,62. To test for the association between cardiac deceleration and value integration during the decision, we recorded ECG during the main task, and subsequently ran a subjective evaluation task, asking for an explicit subjective evaluation of each possible outcome (approach/avoidance) encountered in the main task. We predicted that subjects would subjectively prefer avoidance vs. approach scenarios and that heart rate deceleration would be more pronounced under threat in predictable contexts, facilitating the integration of subjective values into the decision⁶⁰.

Also, as an exploratory analysis, in Experiment 3, we measured EMG activity of response effector muscles during the main task. Muscle activity, such as involuntary hand muscles contraction following transcranial magnetic stimulation of the primary motor cortex, measured with EMG^63,64, or sustained force exerted due to voluntary isometric wrist extension⁶⁵, has been used as an objective marker of action readiness during viewing of emotional images. Similarly, we measured muscle contraction of the flexor pollicis brevis with EMG, at the moment of response production, as a measure of action readiness. Here, we wanted to explore whether action readiness is modulated either by the simple presence of threat in the scene, or by the predictability of the outcome, which would clarify whether both SR and GD mechanisms can influence motor preparation.

All physiological data were collected using ADInstruments hardware (ADInstrument, Amsterdam, NL) and amplified prior to digitalization (Dual BioAmp, ADInstrument). All the processing steps and analyses were conducted using custom scripts on MATLAB⁶⁶ (R2022b).

Electrocardiography (ECG) recording and preprocessing

ECG data were recorded continuously at a sampling rate of 1KHz using three disposable adhesive Ag/AgCl electrodes placed under the left collar bone (ground electrode), right collar bone (negative electrode), and below the left ribs (positive electrode). Raw data were band-pass filtered between 1 and 100 Hz, and R peaks were automatically detected using a custom amplitude-threshold method individually calibrated for each participant. The detected peaks were visually inspected to check for errors in the algorithm and manually corrected when necessary. Inter-Beat-Intervals (IBIs) were computed as the time difference in milliseconds between two consecutive R-peaks. IBIs that differed in value more than 30% from the previous one were considered artifacts and replaced with a cubic spline interpolated value. Finally, IBIs were converted to instantaneous Heart Rate (iHR) by dividing each of them by 60,000.

To extract event-related cardiac responses, we constructed stimulus-locked iHR time series in each trial by linear interpolation to achieve a sampling rate of 10Hz⁶⁷. The time series thus obtained were band-pass filtered between 0.01 and 2 Hz with a second‐order zero-phase forward and reverse Butterworth filter to remove slow drifts and smooth the angles introduced by the interpolation⁶⁷. Finally, the resulting iHR series were epoched from −4 to 10 s around the stimulus onset and baseline corrected by subtracting the mean value of a segment from −2 to −1 seconds before stimulus onset. Note that we did not use a baseline right before the stimulus onset to avoid correcting for a time segment already active due to anticipatory cardiac deceleration. Data from 5 participants was excluded from ECG analysis due to poor signal quality (remaining sample: N = 55, n = 30 females, n = 25 males; mean age ± SD = 23.25 ± 4.58).

Electromyography (EMG) recording and processing

EMG activity of the flexor pollicis brevis muscles was recorded continuously at 2 kHz using two 4 mm circular Ag-AgCl shielded electrodes attached to each hand as recommended⁶⁸. Prior to electrode fixation, the skin was lightly abraded with an exfoliant gel (Nuprep - Weaver and co.) and then wiped with a damp cloth. Two electrodes per hand were placed in a bipolar montage parallel to the muscle fibers, along an axis connecting the ulnar aspect of the metacarpophalangeal joint and the pisiform, with one electrode placed midway along this axis and the other electrode placed 2 cm (center to center) away from the first electrode⁶⁹. A ground electrode was placed on the processus spinosus of the C7 vertebra⁷⁰.

The EMG signal was band-pass filtered between 20 and 500 Hz with a zero-phase forward and reverse 4th order Butterworth filter, while spectral interpolation was used to remove power line noise⁶⁸. The signal was then full-wave rectified and EMG envelopes were created by low-pass filtering the signal at 10 Hz with a zero-phase forward and reverse 4th order Butterworth filter. We used a single-threshold algorithm⁷¹ to detect onsets and offsets of EMG bursts. We identified a baseline as the segment with the lowest amplitude in a moving average with a window of 500 msec from 1 s before the stimulus onset to the response time⁷². We defined EMG bursts as the series of contiguous sample points 3 baseline standard deviations above the baseline mean, and the RT generating burst as the EMG burst containing the response time. We defined burst onset and offset as the first and last samples above this threshold. Burst windows closer than 20 ms were considered as the same burst and merged. Isolated bursts shorter than 25 ms were not considered. Trial-by-trial windows of EMG activity with the automatically detected bursts were visually inspected by the experimenter to check for misidentifications by the algorithm. Onsets and offsets of activity were manually corrected when necessary. Furthermore, noisy trials were manually rejected after visual inspection. The peak EMG response of a trial was defined as the peak amplitude within this window. Peak amplitudes lower/greater than 1.5 interquartile ranges over the first/third quartile within individual participant distributions were considered artifacts, and the corresponding trial was rejected. Time series of within-trial EMG activity were time-locked to peak responses and epoched at ± 500 ms around this point. Note that data from nine participants was excluded from the EMG analysis due to poor signal quality (remaining sample: N = 51, n = 27 females, n = 24 males; mean age ± SD = 23.88 ± 4.79).

Subjective evaluation task

2D snapshots of the scenes presented in the virtual environment were used for this task. We created images representing each pair of angry-neutral avatars (32 dyads). The task was implemented in MATLAB⁶⁶ (R2022b) using Psychtoolbox⁷³ (v3.0.18). A typical trial (Supplementary Fig. S13) began with a 0.5 s long gray screen followed by a superimposed white fixation cross with a duration randomly sampled from a uniform distribution between 0.8 and 1.2 s. Once the fixation cross disappeared, the scene was presented for 1 s. The vertical frame containing the arrows in the original task was rotated 90 degrees in a horizontal position, with the arrows pointing to the left or right elevator. After the scene, a screen with a visual analog scale asked participants how much they would like to be in the elevator pointed by the horizontal arrow, ranging from “Pas du tout” (“Not at all”) to “Vraiment beaucoup” (“Very much”). To respond, participants had to click and drag the cursor on the scale before confirming the response by clicking on the “Continue” button, which ended the trial. Based on the position on the scale, responses were attributed a score between 0 and 100. Each possible pair of avatars was presented twice, once with an avoidance situation (i.e., the arrow pointing to the elevator with the neutral face) and once with an approach situation (i.e., the arrow pointing to the elevator with the angry face). This resulted in 64 trials. The order of stimulus presentation and the direction of the arrow (i.e., approaching or avoiding the elevator with the angry avatar) were randomized across trials. The task was preceded by a brief training session of six neutral trials (three pairs of females and three pairs of male avatars) to familiarize participants with the task. Due to a technical issue, data from this task is available for a sub-sample of participants (46 individuals, 28 females; mean age ± SD = 23.24 ± 4.62).

Procedure

The day before the experiment, participants were asked to complete the LSAS, the STAI-Trait, the BIS-BAS, the PHQ-8 and the PANAS scales via Qualtrics (Qualtrics, Provo, UT). On the day of the experiment, participants were asked to sign an informed consent upon arrival and to complete the STAI-State scale. Next, and only in Experiment 2 and 3, an experimenter attached the ECG electrodes, and a five-minute resting baseline was collected. Note that ECG data were collected in Experiment 2 only to assess the feasibility of such recordings in preparation for Experiment 3. This process helped us identify a technical issue with trigger alignment which we solved for the next experiment. In Experiment 3, in addition to the ECG electrodes, EMG electrodes were attached to participants’ hands. Participants were then introduced to the equipment (VR HMD and controllers) and received verbal instructions on the task from the experimenter, who was not blind to the experimental conditions or the study hypothesis. Finally, the experimenter assisted participants in putting on the equipment and allowed them to adjust the headset to maximize comfort. Participants sat on a chair with their arms resting on a table throughout the task. Once they were ready to start, the experimenter left the room. After the experiment, participants completed the Simulator Sickness Questionnaire and answered some debriefing questions. In Experiment 3 only, participants also performed the Subjective Evaluation task (see below). The entire procedure took approximately 75 minutes in Experiment 1, 90 in Experiment 2 and 100 in Experiment 3.

Statistical analysis (Experiments 1, 2, and 3)

Model-free behavioral analyses

Model-free statistical analyses were conducted using R⁷⁴ (v 4.1.1) and RStudio⁷⁵ (v 2023.6.2.561). Most analyses were performed on Threat trials (i.e. 64 trials, 32 per Condition). We considered trials with RTs below 150 ms as anticipations⁷⁶ and excluded them from the analyses. Following this rule, in Experiment 1 we dropped seven trials from the full dataset (0.18% of the total), one trial in Experiment 2 (0.05% of the total), and four trials in Experiment 3 (0.1% of the total).

Dichotomous choices (Avoid the angry avatar = 1, Approach the angry avatar = 0) were modeled trial-by-trial through a Bayesian hierarchical logistic generalized linear model (or Generalized Linear Mixed Effect Model or GLMM) using the R package brms⁷⁷ (v2.16.1). This method estimates regression parameters through Markov Chain Monte Carlo (MCMC) sampling. Three Markov chains were run in parallel with 10,000 iterations each, 2000 of which were used as a warm-up period. We used as priors for parameter estimation a normal distribution with µ=0 and SD = 2.5 (i.e., skeptical priors) in the case of the intercept and other regression coefficients, an exponential distribution with λ = 2 for the standard deviation parameter, and a Lewandowski-Kurowicka-Joe (lkj) distribution with η = 2 for the correlation matrix. Similar priors have been previously suggested in the case of GLMMs⁷⁸. We considered the Condition (treatment coded: Unpredictable = 0 vs. Predictable = 1) as predictor. Furthermore, because SR and GD processes are supposed to influence behavior at different response latencies^79,80, we also added RTs (grand average centered) and their interaction with the Condition as predictors. For all models we considered the individual subject as a random effect, and we kept a maximum random slope structure⁸¹.

Stepwise model selection was performed by comparing models with different numbers of predictors through their WAIC values. Only the best-fitting model was considered for further analyses. We based inference on group-level parameters’ 95% Credible Intervals (95%CrI) defined as the interval between the 2.5th and 97.5th quantiles of the posterior parameter distributions. We considered an effect to be credible if the relative 95%CrI did not contain zero. In addition, we report for each parameter the proportion of posterior samples lower and greater than zero (p_<0 or p_>0) and Bayes Factors for parameters different from zero (BF₁₀) for credible effects, or Bayes Factors for parameters not different from zero (BF₀₁) for non-credible effects. Detailed information about model fitting and selection, complete regression tables and prior sensitivity analyses are provided in Supplementary Methods 3. For analyses with RTs as dependent variables see Supplementary Note 2.

Drift-diffusion models

DDMs computationally characterize decision between two options as an evidence accumulation process, which terminates with the response. A simple version of a DDM fits response rate and RTs distributions to estimate three parameters of interest: (1) the drift-rate (v), which represents the rate at which evidence accumulates in favor of a response and is positively affected by the value assigned to the choice^82,83; (2) the boundary separation (a), which represents the amount of evidence that needs to be accumulated before executing a response, and thus reflects response caution⁸⁴; (3) the non-decision time (t0), which represents the time needed to encode the stimuli and to execute the motor response. In our task: (a) an effect on the drift-rate would suggest that contextual information about the predictability of the action-outcome is sampled in order to guide the decision, consistent with a GD process; (b) an effect of boundary separation would suggest that participants choose more carefully in one of the two conditions, e.g., because they assign more salience to an action that is contingent with the desired outcome; (c) an effect on non-decision time would suggest that participants in one condition spend more time in encoding the stimuli or in implementing a motor response, when they know that their action consistently achieves the desired outcome. To disentangle among these possible effects (or combinations of them), we fitted several DDMs and then compared them to identify which best described participants’ behavior.

We used the HDDM package⁸⁵ (v0.9.7) in Python (v3.7.13) to fit Bayesian hierarchical DDMs. This method generates estimates of DDMs individual- and group-level parameters through MCMC sampling. For each model three Markov chains were run in parallel with 20,000 iterations each (5000 of which as warm-up period) and a thinning factor of 3. We used the population-informed priors by Matzke and Wagenmakers⁸⁶ to fit the models. In all models the starting point bias was kept fixed and equidistant between boundaries. Consistent with our model-free analyses, we modeled only Threat trials. We set avoidance responses from the angry avatar as the upper boundary and approach responses as the lower boundary. We fitted several models, varying the drift-rate, the boundary separation, the non-decision time, a combination of these, or none of them, by Condition, and then compared their DIC and BPIC values⁸⁷. Further analyses were based on the best fitting model (i.e., lowest DIC and BPIC). Due to the limited number of trials in our task, we only estimated regressors (i.e., the Condition effect) at a group level. We based inference on the group-level parameters’ 95% Credible Intervals (95%CrI), defined as the interval between the 2.5^th and 97.5^th quantiles of the estimates’ posterior distributions. We also provide the proportion of posterior samples for each parameter lower and greater than zero (p_<0 or p_>0). Detailed information about DDM model fitting and selection can be found in Supplementary Methods 4.

Statistical analysis specific to Experiment 3: physiological and subjective measures

Physiological data

Physiological time series were analyzed using a one-dimensional statistical parametric mapping approach. Through this method, individual time points of physiological signals (iHR or EMG amplitude) from each participant were fitted separately trial-by-trial using a Generalized Linear Model (GLM) with the following formulas (1–3):

$$\,{{{{\boldsymbol{iHR}}}}}_{(t)}\,{or}\,{{{{\boldsymbol{EMG}}}}}_{(t)} \sim {{{{\boldsymbol{\beta }}}}}_{0(t)}+{{{{\boldsymbol{\beta }}}}}_{{condition}(t)}+{{{{\boldsymbol{\beta }}}}}_{{threat}(t)}+{{{{\boldsymbol{\beta }}}}}_{{condition}^{*} {threat}(t)}+{{{{\boldsymbol{\beta }}}}}_{{RT}}$$

(1)

$${{{{\boldsymbol{iHR}}}}}_{(t)} \sim {{{{\boldsymbol{\beta }}}}}_{0(t)}+{{{{\boldsymbol{\beta }}}}}_{{condition}(t)}+{{{{\boldsymbol{\beta }}}}}_{{outcome}(t)}+{{{{\boldsymbol{\beta }}}}}_{{condition}^{*} {outcome}(t)}+{{{{\boldsymbol{\beta }}}}}_{{RT}}$$

(2)

$${{{{\boldsymbol{EMG}}}}}_{(t)} \sim {{{{\boldsymbol{\beta }}}}}_{0(t)}+{{{{\boldsymbol{\beta }}}}}_{{condition}(t)}+{{{{\boldsymbol{\beta }}}}}_{{response}(t)}+{{{{\boldsymbol{\beta }}}}}_{{condition}^{*} {response}(t)}+{{{{\boldsymbol{\beta }}}}}_{{RT}}$$

(3)

with t being the time-point, representing 100 ms in the case of iHR (interpolated at 10 Hz) and 1 ms in the case of EMG (sampled at 1 kHz). In the models: β_0(t) represented the intercept, β_condition(t) represented the effect of the Condition (Unpredictable = −0.5; Predictable=0.5) on the iHR or EMG amplitude time point, β_outcome(t) the effect of the Outcome (Threat = −0.5; Safe=0.5), β_response(t) the effect of the Response (Approach the angry avatar = −0.5; Avoid=0.5), β_threat(t) the effect of Threat (Neutral trial = −0.5; Threat trial=0.5), β_RT the effect of Response Times (grand average centered), β_{condition*outcome(t)} the Condition by Outcome interaction, and β_{condition*response(t)} the Condition by Response interaction, β_{condition*threat(t)} the Condition by Threat interaction. Note that when testing for the effect of Threat, we could not simultaneously test for the effect of Response, as there were no approach or avoidance responses in Neutral trials. Therefore, for each physiological signal, we fitted different GLMs with different datasets, namely the GLM including the Response effect only on Threat trials, and the GLM including the Threat effect on the full datasets. Before fitting the models, both iHR and EMG data were z-scored within participants using the mean and standard deviation of the concatenated trial data.

We tested the significance of the effects obtained using a cluster-based permutation approach⁸⁸. First, we computed t-values against zero for the beta coefficients for each regressor and each time point. Second, we applied two thresholds to the resulting t-values series at the 2.5^th and 97.5^th most extreme values of a t-distribution (p_thres = 0.05). Clusters of effects were defined as contiguous points exceeding these thresholds for more than 100 ms for ECG and 20 ms for EMG. The significance of these clusters was tested by permuting the series of beta coefficients 10,000 times with vectors of zeros (i.e., no effect) for a random number of participants. For each permutation, we extracted the sum of the t-values (t_mass) within each detected cluster to create a ‘null’ distribution of t-masses. Finally, the t-masses for each observed cluster were compared to this null distribution to compute non-parametric cluster-corrected p-values (p_corr) as the proportion of the absolute value of the null distribution that exceeded the absolute value of the observed t-mass. Clusters with a p_corr < 0.05 were considered significant.

Data from subjective evaluation task

First, we conducted a two-sided paired t-test on the mean reported value between the approach and the avoidance situations (i.e., with the arrow pointing toward or away from the angry avatar). The assumption of normality was formally verified through a Shapiro-Wilk test, and the assumption of homogeneity of variance was formally verified through a Levene’s test.

Second, since the subjective evaluation task presented the same dyads of stimuli as in the VR task, we were able to compute a per-dyad Subjective Value (SV) score as the difference between the avoidance and the approach value from the angry avatar presented in a given trial of the VR task (originally from −100 to 100, then scaled between −1 and 1 to facilitate model fitting). Thus, negative values indicated a preference for approaching the angry avatar over avoidance, positive values indicated a preference for avoidance, and zero indicated indifference. These per-dyad values were then fitted on a trial-by-trial basis to a Bayesian hierarchical logistic model predicting avoidance decisions (see below Effect of Heart Rate and Subjective Value on Avoidance Decisions).

Effect of heart rate and subjective value on avoidance decisions

To test the hypothesis that heart rate deceleration sustains avoidance decision-making by facilitating the integration of subjective values into the decision⁶⁰, we computed a trial-by-trial measure of cardiac deceleration by taking the maximum iHR change from baseline value in the cluster of beta parameters significantly lower than zero on the intercept of our GLM (see Results - Electrocardiography). Heart rate change values lower/greater than 1.5 interquartile ranges over the first/third quartile within individual participants’ distributions were considered artifacts, and the corresponding trial rejected. iHR change values were then z-scored within participants. We then fitted a Bayesian hierarchical logistic model predicting avoidance decisions and including as predictors the Condition, RTs (as in the best fitting model in the previous analyses, see Results - Avoidance Rate), iHR change, Subjective Value (SV; see above for a description) and their interactions. Importantly, because the Subjective Evaluation task was only administered to a sub-sample of participants, and since the ECG data for one of these participants was rejected due to low quality of the recording, this model was fitted on 45 individuals (n = 28 females, n = 17 males; mean age ± SD = 23.31 ± 4.65).

Finite mixture modeling (FMM)

A limitation of common statistical models is that they assume homogeneity of behavior, with inter-individual differences typically interpreted as variance around a central tendency. Nonetheless, heterogeneity in free choice tasks^34,89, and especially in avoidance behavior, is the norm in both animal and human populations^90,91,92,93. This seems relevant in the current study’s case and especially in the interpretation of our behavioral results. Specifically, some participants may avoid above-chance in both conditions and with no difference between them, irrespective of action-outcome predictability, in an SR fashion. Conversely, other participants may learn from outcomes and avoided at chance level in the unpredictable condition and significantly more in the predictable condition, in a GD fashion. To address this possibility and explore possible latent classes in our sample, we used Finite Mixture Modeling (FMM). Mixture models assume that data are sampled from a mixture of generating processes, or different populations. This method allows for data-driven clustering within datasets without any a-priori knowledge.

To improve the power of the finite mixture modeling algorithm, we performed this clustering by combining the samples from Experiments 2 and 3, which employed the same task. The combined sample of Experiments 2 and 3 consisted of 90 participants (n = 47 females, n = 43 males; mean age ± SD = 23.14 ± 4.2). We implemented FMM using the R package flexmix⁹⁴ (v.2.3-19). We performed a stepwise mixture model fitting of the previously identified best fitting model in Experiments 2 and 3 (i.e., response predicted by Condition, RTs, and their interaction, see Results – Model-Free results on avoidance rate) considering an increasing number of clustering components (i.e., latent classes) from 1 to 5. The best-fitting clustering for each number of components is identified by an Expectation-Maximization (EM) algorithm that optimizes parameters’ maximum likelihood over 1000 iterations. We compared models with the best fitting clustering for each number of components using Bayes Information Criterion (BIC), which has been shown to be optimal for comparing mixture models⁹⁵. Finite mixture modeling revealed that two latent classes best explained our data (see Supplementary Fig. S14). Based on the behavioral pattern presented by these classes (see Results – Comparing GD and SR latent classes – Model-Free results on avoidance rate), we will refer to them as ‘GD Class’ (n = 61, 17 from Experiment 2 and 44 from Experiment 3) and ‘SR Class’ (n = 29, 13 from Experiment 2 and 16 from Experiment 3).

When considering the effect of latent classes estimated by FMM, Bayesian hierarchical logistic models on approach and avoidance choices were first fit including the Class as a between-group fixed effect (treatment coded: SR Class = 0, GD Class =1). As a robustness check, we also fit models separately for each latent class, obtaining overall the same results (see Supplementary Table S25).

Concerning DDM analyses by class, we allowed the drift-rate (v), boundary separation (a), and non-decision time (t0) to vary by Class. Furthermore, to improve the interpretability of the results, we adopted a cell design matrix, thereby estimating a parameter (v, a, or t0) for each Condition and Class (instead of effects on an intercept as in the previous analyses).

For ECG and EMG data, we rerun the previously described GLM models, separately for each FMM class. Because ECG and EMG data were only available for participants in the third experiment, this analysis included only that sample but retained the classification from the algorithm run on Experiments 2 and 3 combined.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Results

Results (Experiments 1, 2, and 3)

Model-free results on avoidance rate

Across experiments, model comparison revealed the best fitting model to be the one including the main effects of Condition and RT. As expected, participants avoided the angry avatar credibly more in the Predictable than in the Unpredictable condition [Exp1: p(Avoid|Pred.)= 0.726, p(Avoid|Unpred.)= 0.647; Exp2: p(Avoid|Pred.)= 0.659, p(Avoid|Unpred.)= 0.599; Exp3: p(Avoid|Pred.)= 0.613; p(Avoid|Unpred.)= 0.574, see Supplementary Table S48 for avoidance rates and median RTs across Conditions and Experiments] as reflected in a credible model Condition effect (Exp1: β = 0.475, p_<0 = 0, 95%CrI = [0.253, 0.699], BF₁₀ = 128.297; Exp2: β = 0.303, p_<0 = 0.007, 95%CrI = [0.065, 0.539], BF₁₀ = 0.927; Exp3: β = 0.188; p_<0 = 0.015; 95%CrI = [0.019, 0.360]; BF₁₀ = 0.372). However, participants still avoided credibly more than chance in the Unpredictable condition as revealed by a credible model Intercept (Exp1: β = 0.703, p_<0 = 0, 95%CrI = [0.519, 0.896], BF₁₀ = 5.645e + 14; Exp2: β = 0.416, p_<0 = 0, BF₁₀ = 29.969, 95%CrI = [0.206, 629]; Exp3: β = 0.301; p_<0 = 0; 95%CrI = [0.152, 0.450]; BF₁₀ = 81.860, Fig. 1B), and confirmed by a two-sided Wilcoxon test against chance (Exp1: V = 1450.5, p < 0.001, d = 0.893, 95%CI_d = [0.640, 1.210], BF₁₀ = 3.1428e + 06; Exp2: V = 340.5, p = 0.002, d = 0.678, 95%CI_d = [0.330, 1.130], BF₁₀ = 37.446; Exp3: V = 1166.5, p < 0.001, d = 0.502, 95%CI_d = [0.280, 0.740], BF₁₀ = 90.432). Of note, although the 95% credible interval for the Condition effect did not include zero in Experiments 2 and 3, the evidence for this effect represented by the Bayes Factor was still low to null. This was likely because the difference between Conditions depended on RTs, with a credible Condition by RTs interaction (Exp2: β = 1.476, p_<0 = 0, 95%CrI = [0.726, 2.195], BF₁₀ = 182.234; Exp3: β = 1.115; p_<0 = 0.001; 95%CrI = [0.458, 1.756]; BF₁₀ = 35.638, Fig. 1D, E). Interestingly, the main effect of RTs was non-credible in Experiment 2 (95%CrI = [−0.159, 1.019], BF₀₁ = 3.044), but credible in Experiment 3 (β = 0.714; p_<0 = 0.002; 95%CrI = [0.229, 1.198]; BF₁₀ = 6.226). This suggested that slower RTs were associated with a higher probability of avoiding the angry avatar in the Predictable condition, but not (or to a lower extent in Experiment 3) in the Unpredictable one. Finally, in Experiment 1, we found a credible main effect of RTs (β = 1.050, p_<0 = 0, 95%CrI = [0.571, 1.544], BF₁₀ = 288.812), but not a credible Condition by RTs interaction (95%CrI = [−0.467, 1.013], BF₀₁ = 4.841). For a discussion about the differences across experiment on the RTs effects see Supplementary Discussion 1.

Drift-diffusion model

Overall, model comparison revealed that the best fitting model included the effect of Condition on drift-rate, boundary separation and non-decision time, with the exception of Experiment 2, where the effect on boundary was not included in the winning model (Fig. 2D). Posterior parameter testing showed that participants exhibited a credibly greater drift-rate toward avoidance decisions in the Predictable compared to Unpredictable condition (Condition effect: Exp1: β = 0.344, p_<0 = 0, 95%CrI = [0.226, 0.460]; Exp2: β = 0.303, p_<0 = 0, 95%CrI = [0.134, 0.472]; Exp3: β = 0.185, p_<0 = 0.003, 95%CrI = [0.056, 0.312]). However, the drift-rate in the Unpredictable condition was still credibly above zero (Intercept: Exp1: v = 0.575, p_<0 = 0, 95%CrI = [0.398, 0.753]; Exp2: v = 0.384, p_<0 = 0, 95%CrI = [0.161, 0.605]; Exp3: v = 0.298, p_<0 = 0.001, 95%CrI = [0.131, 0.468]), suggesting enhanced evidence accumulation for avoidance decisions even when information about a predicted outcome was not available.

Second, the non-decision time was not credibly different between conditions in Experiments 1 and 2 (Condition effect on t0, Exp1: β = 0.005, p_<0 = 0.085, 95%CrI = [−0.002, 0.012]; Exp2: β = 0.007, p_<0 = 0.033, 95%CrI = [−0.0004, 0.015]), but it was credibly shorter in the Predictable compared to Unpredictable condition in Experiment 3 (β = −0.011, p_>0 = 0, 95%CrI = [−0.017, −0.006]), although both effects were small.

Finally, we found that the response boundary was slightly credibly higher in the Predictable condition compared to the Unpredictable condition in Experiments 1 and 3 (Condition effect on a, Exp1: β = 0.048, p_<0 = 0.010, 95%CrI = [0.009, 0.089]; Exp3: β = 0.035, p_<0 = 0.019, 95%CrI = [0.002, 0.068]), indicating a greater response caution under predictability (Fig. 2A–C).

Results specific to Experiment 3

Subjective Evaluation Task

A two-sided paired t-test on the mean reported values on the Subjective Evaluation Task revealed that participants significantly preferred to avoid (mean value ± SD = 65.231 ± 20.478) rather than approach angry avatars (27.704 ± 20.490; t₄₅ = −14.105, p < 0.001, d = −2.08, 95%CI_d = [−2.71, −1.71], BF₁₀ = 1.357e + 15; Supplementary Fig. S13).

Electrocardiography

Cluster-based analysis of event related cardiac responses revealed two significant clusters of beta coefficients for the stimulus-locked iHR series (Supplementary Fig. S15). The beta coefficients for the intercept (i.e., the grand average HR activity) showed a significant negative cluster (t_mass = −214.635, p_corr = 0.014), beginning approximately one second before the doors opened and extending approximately 2.5 s after. These negative coefficients indicated a cardiac deceleration around the time of the response in both Conditions and independent of Outcome and RTs. Furthermore, the coefficients for the Condition effect presented a significant negative cluster (t_mass = −188.635, p_corr = 0.005), beginning approximately 0.5 s after the doors opened and ending approximately 3 s after the end of the response time. Hence, participants presented a greater and longer lasting cardiac deceleration around the response in the Predictable than Unpredictable condition (Fig. 3A). Supporting the threat-related nature of this effect, cardiac responses did not significantly differ between Conditions in neutral trials (Supplementary Fig. S16).

Importantly, the fact that the cardiac deceleration began before the stimulus onset (Fig. 3A) may be related to the onset of the jittered time interval in anticipation of the doors opening (as participants may already be beginning to prepare the response). Therefore, as a robustness check, we repeated the same analysis described above on time series centered on the beginning of the jittered interval. These analyses yielded the same results.

Effect of Heart Rate and Subjective Value on Avoidance Decisions

To explore the hypothesis that cardiac deceleration enhances subjective value integration into choice probability, we fitted a Bayesian hierarchical logistic model predicting avoidance choices by Condition, RTs, maximum iHR change from baseline value around stimulus onset, and participants’ Subjective Value (SV). A credible Condition by HR by SV by RTs four-way interaction emerged (β = −0.931, p_>0 = 0.007, 95%CrI = [−1.685, −0.191], BF₁₀ = 3.236; see Supplementary Table S22), suggesting that greater HR deceleration around stimulus onset was associated with an increased positive effect of SV on the probability of avoiding the angry avatar, with the effect increasing at slower RTs and only when the outcome of the action was predictable (Fig. 3C).

Electromyography

Cluster-based permutation analysis did not reveal any clusters of significant effects of our regressors on peak-centered EMG activity (all p_corr > 0.05), neither when testing the Condition effect nor when testing the Choice effect (i.e., approach vs. avoid angry avatars).

Comparing GD and SR latent classes

Finite mixture modeling revealed that two latent classes best explained our data (Supplementary Fig. S14). Considering the behavioral pattern presented by these classes, we will refer to them as ‘GD Class’ (n = 61, 17 from Experiment 2 and 44 from Experiment 3) and ‘SR Class’ (n = 29, 13 from Experiment 2 and 16 from Experiment 3). Although it appeared that the classes were more evenly balanced in Experiment 2 compared to Experiment 3, this difference was not statistically significant (χ²=1.838, p = 0.175).

Model-free analyses of avoidance rates

A credible three-way Condition by RTs by Class interaction emerged (β = 1.165; 95%CrI = [0.338, 1.995]; p < 0 = 0.004; BF10 = 18.858; see Fig. 4B and Supplementary Table S23). In general, the GD Class, but not the SR one, presented a credibly greater avoidance rate in the Predictable condition compared to the Unpredictable one. Furthermore, the SR Class presented a credibly greater avoidance rate compared to the GD class [SR: P(Avoid|Pred.)= 0.739; P(Avoid|Unpred.)= 0.724; GD: P(Avoid|Pred.)= 0.575; P(Avoid|Unpred.)= 0.515]. This pattern was sustained by the fact that in the GD Class slower RTs were associated with a higher probability of avoidance only in the Predictable condition, whereas this effect was present in both Conditions in the SR Class. Interestingly, in the SR Class, avoidance rate in the Unpredictable condition was credibly greater than chance as reflected by a credible model intercept (β = 0.883; 95%CrI = [0.718, 1.048]; p < 0 = 0; BF₁₀ = 4.696e + 15), and confirmed by a two-sided Wilcoxon-test against chance (V = 405, p < 0.001, d = 1.690, 95%CI_d = [1.310, 2.370], BF₁₀ = 1.5575e + 07). Conversely, in the GD class the avoidance rate in the Unpredictable condition was not credibly different than chance level, as reflected by a credible negative Class effect in the model (β = −0.820; 95%CrI = [−1.015, −0.625]; p > 0 = 0; BF₁₀ = 6.745e + 38) and confirmed by a two-sided Wilcoxon test against chance (V = 924, p = 0.118, BF₀₁ = 3.615805).

**Fig. 4: Main results from classes analysis.**