Model comparisons of the predictability of the changes in judgments.

(A) Parameter estimates in the “linear model” indicating that participants assign a significant weight to the differences in judgments (D), their confidence level (C), and the interaction between them (D × C), but not to the interaction with group size (D × G) and scenario types (D × S) (Eq 7). Error bars represent the standard error of the mean (SEM). (B) The “surprise model” hypothesized that participants who have greater confidence in their judgments are more likely to be surprised by unexpected differences in judgments with a social group and to adapt their judgments to the given judgments of groups. However, the parameter estimates in this model suggest that the different level of confidence could not explain the different level of judgment adaptation (Eq 8). (C) The “Bayesian model” (Eq 6) provides 2 individual, specific functions that allow us to track the changes in credibility. First, individual jurors were more likely to assign credibility to their individual information (J1) when they had a higher level of confidence (left panel). Second, participants assigned more credibility on social information (JS) when it was provided by a larger (n = 20) compared to a smaller (n = 5) group (**, P < 0.001; right panel). (D) The likelihoods of each model prediction were estimated by a leave-one-out-cross-validation (LOOCV) procedure, and the Bayesian information criteria (BIC) were compared between models. The BIC showed that predicted changes in judgments of the “Bayesian model” fit better to the actual changes in judgments than the other models (**, P < 0.001). The data underlying this figure can be found in S1 Data.