Biased evaluations emerge from inferring hidden causes

How do we evaluate a group of people after having positive experiences with some members and negative experiences with others? In particular, how do rare experiences with members who stand out (e.g., negative experiences when most are positive) influence the overall impression we have of the group? Here, we show that such rare events may be overweighted due to normative inference of the hidden, or latent, causes that are believed to generate the observed events. We propose a Bayesian latent-cause inference model that learns environmental statistics by combining highly similar events together and separating rare or highly variable observations. The model predicts that group evaluations that rely on averaging inferred latent causes will overweight variable events. We empirically tested these model-derived predictions in four decision-making experiments, where subjects observed a sequence of social (Exp 1 to 3) or non-social (Exp 4) behaviors and were subsequently asked to estimate the average of observed values. As predicted by our latent-cause model, average estimation was biased toward rare and highly variable events when observing social behaviors. We then showed that tracking of a single summary value, instead of parsing events into distinct latent causes, eliminates the bias. These results suggest that biases in evaluations of social groups, such as negativity bias, may arise from the causal inference process of the group.