Individual differences in behavioral judgements of social valence

On the behavioral level (i.e., responses and RTs; S2 Fig), results did not show a consistent pattern of group differences across tasks (S2 Fig). For responses, we found a significant positive association between SA traits and the ratio of negative choices only in the negative ambiguous social scenes (Exp. 1b, = 0.238, = 0.057, p = .044), but not for any other experiments (ps > 0.05). For RTs, in the clear monadic biological motion experiment (Exp. 5), a repeated-measures analysis of variance (ANOVA) revealed a significant interaction between group and stimulus valence, F(1,68) = 4.651, p = .035, = 0.064. Post hoc comparisons showed that both the high social anxiety (HSA) group and the low social anxiety (LSA) group recognized happy emotions more slowly than angry emotions, with this effect being more pronounced in the HSA group. No other significant group × stimulus valence interactions were observed in the remaining experiments.

HDDM parameters reveal initial biases in SA

Among the HDDM parameters, results showed robust associations between SA traits and initial biases (across experiments (Fig 2C) except for clear social scenes (Exp. 1a) and control scenes (Exp. 2b). Specifically, SA traits were positively correlated with in negative ( = 0.233, p = .005) and positive ambiguous social scenes (Exp. 1b, = 0.205, p = .013), and negatively correlated with in neutral scenes (Exp. 2a, = -0.258, p = .043). For facial expressions, two-sample t-tests revealed that HSA exhibited a significantly more positive in clear facial expressions than LSA (Exp. 3, t(61) = 3.916, p < .001, d = 1.012), but significantly more negative in ambiguous facial expressions (Exp. 4, t(67) = 2.813, p = .006, d = 0.676). In biological motion tasks, HSA consistently demonstrated more positive than LSA (Exp. 5,t(68) = 2.600, p = .040, d = 0.500; Exp. 6, t(62) = -5.610, p < .001, d = -1.370; Exp. 7, t(68) = 40.826, p < .001, d = 9.760; Exp. 8, t(68) = -5.900, p < .001, d = -1.410). These results suggest that initial biases may be key contributors to atypical social cognition in individuals with high SA traits. However, instead of homogeneous contributions, initial biases contributed differently to the biased social cognition depending on the combination of social cues. Besides , we did not find robust group differences for other HDDM parameters (, , ) across experiments (S5 Material in S1 Text, S3, S4 Figs).

The Social Prism Model revealed dichotomous prior biases in SA

We quantified prior bias (∆z) as each participant’s deviation in z from the mean of the LSA group within the same experiment (Fig 2B; Eq. 2). We then applied a linear regression approach, incrementally adding social cue predictors to assess whether dimensions of the Social Prism explained individual differences in prior bias. The full model, which incorporated all six social cues and is hereafter referred to as the Social Prism Model (Fig 3C, Eq. 3), provided the best fit to the data (highest − ∆AIC; R² = 0.139; Fig 3A).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. The Social Prism Model and predictive evaluation of social cognition in SA.

(A) Linear regression model comparison. –ΔAIC for incremental addition of social cues in the Social Prism as predictors, including ANI, SI, EMO, FA, BO, and AMB. The full model, referred to as the Social Prism Model, yielded the largest –ΔAIC, indicating the best overall model fit. (B) Estimated effects of each social cue on Δz in the Social Prism Model. Dots indicate mean effect estimates; error bars represent 95% confidence intervals; ***p < .001. (C) The Social Prism systematically deconstructs the complex social scenarios into hierarchical components of social cues and ambiguity. Social cues were further decomposed into perceptual cues (face, body) and inferential cues (social interaction, emotion, animacy). (D) Interaction plots illustrating the estimated Δz values across levels of SA (LSA vs. HSA) and each social cue. FA = face, AMB = ambiguity, ANI = animacy, SI = social interaction, EMO = emotion, BO = body.

https://doi.org/10.1371/journal.pcbi.1014509.g003

Regression and simple slope analyses revealed that all social cues of the Social Prism were significantly associated with prior biases in SA (Fig 3B). However, the directions of associations were not homogeneous (Fig 3D). Specifically, high SA trait was associated with over-negative prior biases in response to FA (β = 1.27 × 10⁻³, p < .001) and AMB (β = 7.08 × 10⁻³, p < .001), but over-positive prior biases in response to ANI (β = –1.04 × 10⁻³, p < .001), SI (β = –1.25 × 10⁻³, p < .001), EMO (β = –1.04 × 10⁻³, p < .001), and BO (β = 2.02 × 10⁻³, p < .001). Together, these findings demonstrate dichotomous prior biases that depend on social cues among individuals with high SA traits, highlighting the cue-specific nature of prior biases in SA.

We further evaluated the performance of predicting SA traits through machine learning (ML, Fig 4). The ML model took three feature sets: behavioral features (e.g., RTs), latent features derived from HDDM parameters (e.g., and ), and their combined features, as training inputs to predict SA traits. Results showed that latent features consistently outperformed behavioral features in predicting SA traits across experiments, except Experiments 2, 4, and 5. For instance, in Exp. 1b, latent features achieved a = 0.627 ( = 0.199) compared to = 0.316 ( = 0.288) for behavioral features (t(38) = 3.967, p < .001). Moreover, across all experiments except Exp. 6, models using latent features performed comparably to the combined feature models (Exp. 6, t(38) = 2.327, p = .025). For full results, see S7 Material in S1 Text and S1 Table.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Machine learning results.

Model performance comparison across feature sets using explained variance scores (EVS). Models were trained using behavioral features (green), latent features from HDDM (blue), and combined features (purple). The error bar indicates the standard error of EVS. Statistical significance: *p < .05, **p < .01, ***p < .001.

https://doi.org/10.1371/journal.pcbi.1014509.g004

Bayesian simulation captures biased social cognition

To simulate biased social cognition arising from biased prior expectations, we developed a Bayesian generative model to simulate and reproduce valence judgements in ambiguous social scenes (Exp. 1b). This approach aimed to replicate the behavioral findings that HSA individuals were associated with increased ratios of negative judgements. Specifically, we estimated likelihood parameters by projecting CNN-extracted valence features into a low-dimensional space, where a K-nearest neighbors model identified cluster centroids and computed parameters based on distances to ambiguous images (Fig 5A). These likelihoods, together with prior expectations, were incorporated into a one-step Bayesian inference model to simulate participants’ binary decisions (Fig 5B).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Bayesian simulation for atypical social cognition in SA under ambiguous social scene conditions.

(A) Likelihood parameters estimation. Valence features were extracted from social scenes using a fine-tuned convolutional neural network, and projected into a low-dimensional valence space via Uniform Manifold Approximation and Projection (UMAP). A K-nearest neighbors model (KNN) trained on clear images identified the centroids of the positive and negative valence clusters, and the mean and standard deviation of the likelihood were computed based on Minkowski distances from each ambiguous image to the class centroids. (B) A one-step Bayesian inference framework was used to simulate participants’ binary decisions based on prior expectations and stimulus-driven likelihoods. For clear condition, prior expectation and signal accumulation both contributed to the social cognition. Under ambiguous condition, the visual signal is no longer reliable, and it heightens reliance on prior expectations during social cognition. For participant i and stimulus s, the prior was derived from the individual-level HDDM-estimated initial bias (), while the likelihood was determined by the valence features of the stimulus. Simulated decisions were sampled from the resulting posterior distribution, and binarized based on a threshold (see Methods), where 0 indicates a positive judgement and 1 indicates a negative judgement. (C) Accuracy of simulated valence judgements compared to empirical valence judgements across participants. Each dot represents one individual. (D) Simulated negative response rates closely matched actual negative response rates across participants, indicating high generative fidelity of the model. (E) Simulated behavior showed that SA traits were associated with increased negative response rates, particularly to negative ambiguous stimuli, which replicated patterns observed in actual behavioral data.

https://doi.org/10.1371/journal.pcbi.1014509.g005

The Bayesian simulations successfully reproduced key behavioral patterns observed in Experiment 1b. At the individual level, the simulated trial-wise responses demonstrated high accuracy when comparing model-generated responses against empirical responses as a reference (Fig 5C, ), significantly above chance level (t(72) = 22.092, p < .001). Simulated and actual negative response rates were highly correlated across participants (Fig 5D, r = 0.840, p < .001), demonstrating the model’s ability to capture participants’ overall response pattern. SA traits were significantly associated with simulated negative response rates for negative stimuli ( = 0.250, p = .033), but not for positive stimuli ( = 0.174, p = .140) (Fig 5E). Together, the results provide converging evidence that biased social cognition in SA can be accounted for by a Bayesian inference framework.

Ethics statement

All experiments have been approved by the Ethics Committee of Peking University (#2022-02-10). All participants provided written informed consent before participating in the experiments. Experiments were preregistered on the Open Science Framework (OSF) before data collection. The full pre-registration is available at https://osf.io/nypaz/overview?view_only=9ce4a968b6f145b5be0b7ca10f5e010f (Experiments 1–2) and https://osf.io/kyw7a/overview?view_only=aa6ea58c555d40f2884eea012ed28646 (Experiments 3–8).

Social anxiety questionnaire

The Chinese Social Anxiety Questionnaire for Adults (CSAQ-A) [21,22] was used to measure the individual levels of SA trait. Total scores range from 30 to 150, with higher scores indicating greater SA trait levels. The normative mean for Chinese young adults was 94.93, and the cut-off scores for the high and low SA groups were 110 and 81, respectively [22]. Group definitions are provided in S1 Material in S1 Text.

Sample size determination

Sample size determinations were based on a combination of a priori power analyses and post hoc sensitivity analyses using G*Power (Version 3.1) [63]. For Experiments 1–2, no prospective power analysis was conducted. Instead, a post hoc sensitivity analysis was performed to estimate the minimum detectable effect size for the linear regression models. To ensure a conservative estimate, we based the analysis on Experiment 2, which had the smallest sample size (N = 62; Exp. 1: N = 73). Assuming two predictors (i.e., SA trait scores and stimulus valence), α = .05 and 80% power, the analysis indicated sensitivity to effects of Cohen’s f² ≥ 0.164, corresponding to a moderate effect size. This suggests that the study had sufficient power to detect moderate effects. For Experiments 3–8, sample size planning was informed by a priori power analysis for a 2 (group) × 2 (stimulus valence) repeated-measures ANOVA. Assuming an interaction effect of f = 0.19, α = .05, 80% power, a correlation among repeated measures of 0.5, and nonsphericity correction ε = 1, the analysis indicated that a minimum sample of N = 58 would be sufficient.

Valence judgement tasks and experiment design

Behavioral analysis

Responses and RTs. Statistical analyses were conducted using R version 4.2.1 [71]. To examine whether consistent behavioral patterns emerged across experiments in relation to SA traits, we first conducted linear regression models on the ratio of negative choices and RTs in social scene tasks (Experiments 1–2), using SA traits and stimulus valence category (negative vs. positive, or neutral vs. control) as predictors. For facial expression and biological motion tasks (Experiments 3–8), we performed repeated-measures ANOVAs on accuracy and RTs, with stimulus valence (angry vs. happy) as a within-subject factor and SA group (HSA vs. LSA) as a between-subject factor.

Behavioral modeling using HDDM

Behavioral modeling was conducted using the HDDM package (version 0.8.0) [12]. We employed the stimulus-coding variant of the HDDM, in which the upper and lower decision boundaries correspond to the two response categories (Fig 2A). Participants’ responses and RTs were modeled using the Wiener First Passage Time (WFPT) distribution, the analytical solution of the HDDM with default diffusion noise.

The model included four free parameters (Fig 2A): initial bias (), drift rate (), threshold boundary (), and non-decision time (). indicates the relative starting point between two decision boundaries, ranging from 0 to 1, where 0 corresponds to the lower decision boundary (positive responses) and 1 to the upper decision boundary (negative responses). A value of indicates no initial bias, values exceeding 0.5 indicate initial biases toward the upper decision boundary, requiring less evidence to reach the boundary and make negative judgements. The drift rate represents the speed and direction of signal accumulation, with a larger absolute value indicating faster and more efficient processing. The sign of denotes the direction of accumulation, where negative values reflect evidence favoring the lower decision boundary, and positive values favor the upper decision boundary. The threshold boundary represents the total amount of evidence required to make a decision, and greater values indicate more conservative strategies. The non-decision time captures sensory encoding and motor execution latencies.

Let denote the decision time. The WFPT density for responses terminating at the boundary is [72]:

(1)

The specific HDDM model structure (i.e., parameterization of , , and ) was adapted to each experiment’s stimulus type, while retaining the same core HDDM framework to ensure comparability across experiments (for details see S2 Material in S1 Text). All models demonstrated satisfactory convergence ( ≤ 1.05) [73], and the estimated HDDM parameters were analyzed using the same statistical procedures as those applied to behavioral data.

Modeling prior bias with the social prism

To investigate how distinct social cues in the Social Prism shape prior biases in SA, we conducted a series of nested linear regression analyses using the “statsmodels” package (version 0.14.1) [74]. The dependent variable was the prior bias, denoted as , computed as the difference between a participant’s initial bias and the LSA group mean in the same experiment (Fig 2B), as follows:

(2)

where indexes the individual participant, denotes the experiment, and refers to the set of LSA participants in experiment .

We estimated six nested linear models of increasing complexity (Models 1– 6; see S3 Material in S1 Text). Model 6, referred to as the Social Prism Model, incorporated all hypothesized social cues and their interactions with SA traits, as formulated in Eq 3:

(3)

where:

In all models, social cues, including ANI, SI, EMO, FA, BO, and AMB, served as moderators of SA’s effect on prior biases. All binary-coded cues (ANI, SI, EMO, FA, BO) were categorical. AMB was treated as a continuous predictor and calculated using principles from signal detection theory (S4 Material in S1 Text, S1 Fig), reflecting the degree to which positive and negative interpretations of a social stimulus could be distinguished, with higher values indicating greater ambiguity. AMB was modeled both as a main effect and as a moderator interacting with other social cues, with the corresponding results reported in S6 Material in S1 Text. Categorical predictors were dummy-coded, and continuous predictors were used in their raw form.

Model performance was assessed by comparing the goodness-of-fit across models using the Akaike Information Criterion (AIC). Improvements in model fit were evaluated based on changes in AIC (−ΔAIC) relative to a baseline intercept-only model. Nested models were compared sequentially, and a greater decrease in AIC indicated better explanatory power. This approach enabled a principled evaluation of the additional value contributed by incorporating individual social cues and their higher-order interactions with SA.

Predicting social anxiety scores with machine learning

We implemented eXtreme Gradient Boosting (XGBoost) using the “scikit-learn” package (version 1.7.0) [75] in Python to predict individual SA traits from behavioral and cognitive-level features. XGBoost was selected for its effectiveness in handling multicollinearity and delivering robust performance in structured behavioral datasets [76]. The output variable was the SA traits. Three batches of input predictors were selected to examine the contributions to SA trait prediction: behavioral features only, computational features estimated via HDDM only, and a combined set comprising both. Demographic variables were included in all three clusters.

For each stimulus type, separate models were trained and evaluated using repeated hold-out (Monte Carlo cross-validation). Data were randomly split into training (80%) and testing (20%) sets across 20 independent iterations to provide a robust estimate of model performance. Within each iteration, all preprocessing and model selection were restricted to the training set. Feature and target variables were standardized on the training data and applied to the test data. XGBoost hyperparameters were optimized using Bayesian optimization, with 5-fold cross-validation on the training set as the optimization objective. Optimization proceeded for up to 100 iterations or until convergence of the mean squared error [77]. The best hyperparameter configuration was then refit on the full training set and evaluated once on the held-out test set, which was not accessed during preprocessing, tuning, or model selection. Model performance was evaluated using the average Explained Variance Score (EVS) across all cross-validation splits. EVS measured the proportion of variance in the target variable explained by the model, with values ranging from negative infinity to 1.0. Higher EVS values indicated better predictive performance. EVS scores were compared between feature sets using an independent-samples t-test within each experiment.

Social cognition stimulation with a Bayesian framework

Bayesian simulations were implemented in Python to model social cognitive processes in ambiguous social scenes (Exp. 1b), following the Bayesian inference framework. This experiment was selected because it incorporated the full set of social cues defined in the Social Prism and exhibited distinct behavioral patterns. Valence judgements were formalized as a one-step Bayesian inference process [41], in which trial-wise posterior distributions were computed by combining prior and likelihood distributions, both modeled as Gaussian (Fig 5B). A one-step framework was adopted because each ambiguous social scene was treated as an independent event, with judgments shaped by relatively stable prior expectations rather than sequential belief updating across trials.

Prior Specification. For each trial, the prior mean was defined by the participant’s HDDM-estimated initial bias conditioned on the stimulus valence. Specifically, each participant had two valence-specific initial bias parameters, and , corresponding to the negative and positive conditions, respectively. These parameters were interpreted as category-specific prior expectations reflecting stable tendencies developed across repeated encounters with negative and positive social stimuli. Let denote the valence of the stimulus. The valence-dependent initial bias for was defined as:

(4)

The trial-wise prior mean was then set to this value:

(5)

The prior distribution was assumed to follow a normal distribution with this mean and a fixed standard deviation of to reflect moderate uncertainty:

(6)

Likelihood Estimation. To derive likelihood estimates for ambiguous social scenes, we implemented a VGG16–UMAP–KNN pipeline. This approach was intended as a computational approximation of sensory evidence rather than a biologically realistic model of human perceptual processing. Likelihoods were derived from the valence features of ambiguous social scenes (Fig 5A), estimated via a K-Nearest Neighbors (KNN) classifier trained on clear social scenes with known valence labels. Valence features were extracted using a fine-tuned VGG16 convolutional neural network [78], which achieved an accuracy of 81.82% in valence judgements for clear social scenes. Feature representations were taken from the final fully connected layer (1 × 4096), and subsequently reduced to a lower-dimensional space (1 × 5) using Uniform Manifold Approximation and Projection (UMAP) [79]. Dimensionality reduction was performed separately for clear and ambiguous images to account for distributional differences. The KNN classifier, trained on clear-scene embeddings and corresponding valence labels, was applied to ambiguous-scene embeddings to produce probabilistic valence estimates.

For each ambiguous scene stimulus , the likelihood mean was defined based on its normalized Minkowski distances to the positive and negative cluster centroids in the 5-dimensional UMAP embedding space:

(7)

where and denote the Minkowski distances from ambiguous stimulus to the centroids of the positive and negative clusters, respectively. This measurement reflected the relative position of the stimulus along the positive-negative valence dimension, and was used as the mean for both types of likelihood distributions.

To quantify classification uncertainty, valence-specific standard deviations were defined as normalized deviations from each cluster center:

(8)

(9)

where and denoted the mean intra-cluster distances of the positive and negative training samples, respectively. To prevent the likelihood from dominating the prior due to differences in scale, we normalized all likelihood standard deviations by a scaling factor , set equal to the prior standard deviation ( = 0.04), and then divided by the average unnormalized standard deviation across all stimuli. The average unnormalized standard deviation was computed as follows:

(10)

where denoted the number of ambiguous images. The normalized standard deviations for positive samples and negative samples were calculated as:

(11)

(12)

The final likelihood distribution used for trial-level simulation was defined as a Gaussian with mean , and a standard deviation determined by the valence of the presented stimulus (, as defined above). Specifically:

(13)

All prior and likelihood distributions were truncated and normalized to the interval [0, 1], as both and reside in a bounded decision space, ensuring compatibility in probabilistic integration.

Posterior Computation and Decision Rule. For each trial, the posterior distribution was computed by pointwise multiplication of the normalized prior and likelihood probability density functions (Fig 5B), followed by normalization:

(14)

A single value was then drawn from the resulting posterior distribution:

(15)

Based on this sampled value, the simulated binary decision as generated using a fixed threshold at 0.5, where 0 indicates a positive judgement and 1 indicates a negative judgement:

(16)

Simulation Evaluation. Given that the simulation was parameterized using HDDM-derived priors estimated from the observed behavioral data, it was intended as a test of generative adequacy rather than an independent out-of-sample prediction. Simulation performance was evaluated in three aspects. First, for each participant, we calculated the accuracy of the model-simulated binary responses compared to the participants’ observed responses, and conducted a one-sample t-test against 0.5 to evaluate whether the model generated random responses. Next, we assessed the correlation between the simulated and empirical negative response rates to examine the model’s ability to capture overall response patterns. Third, to determine whether the simulations reflected key behavioral signatures, we computed the simulated negative response rate for each condition and participant. The simulated negative response rates were then entered into a linear regression model with SA traits and stimulus valence (positive vs. negative) as predictors, testing whether the simulated data replicated the SA-by-valence interaction observed in the empirical data.