Predicting How People Update Beliefs: Comparison of Individual-, Subgroup-, and Population-Level Dynamic Bayesian Networks

Abstract

Modelling cognitive processes such as belief updating requires a decision about the level at which a model is fitted: a single model estimated from the whole sample obscures systematic differences between individuals, whereas individual-specific models yield noisy parameter estimates that hardly generalise beyond the observed participants. Clustering individuals with similar patterns of belief updating across individuals, including when, how, and under which conditions they revise their beliefs and fitting one model per cluster offers a compromise, yet its predictive value has received little systematic attention. In this paper, we cluster individuals in dynamic Bayesian networks (DBNs) that share a known and identical graph structure, so that clustering reduces to the conditional probability table (CPT) space, and we apply hierarchical agglomerative clustering to participants’ CPTs, which yields a hierarchy of nested groupings and thus does not require the number of groups to be fixed in advance. Using a simulated belief updating study based on a cooperative gaze-following paradigm, we compare predictions for unseen trials obtained from (i) individual-specific DBNs, (ii) cluster-specific DBNs estimated from the identified groups, and (iii) a single DBN estimated from the entire sample. We evaluate them with classical measures of predictive accuracy and with theory-guided characteristics of belief updating. Cluster-based DBNs predict new trials about as accurately as individual-specific DBNs and significantly better than sample-level models. Thus, they recover interpretable subgroup-specific patterns of belief updating while retaining the generalisability needed to make predictions for previously unseen individuals.

Resources for this Paper

Slides Poster Code