Lack of group-to-individual generalizability in pseudocontingencies

TL;DR

People often form beliefs based on pseudocontingencies, but individual differences mean group-level models may overestimate this tendency.

Contribution

The paper introduces a method to quantify individual reliance on pseudocontingencies, revealing significant differences from group-level estimates.

Findings

Individuals rely on pseudocontingencies, but less so than group-level models suggest.

Median reliance on pseudocontingencies is 22-28% lower at the individual level.

Group-level models may overestimate pseudocontingency reliance due to lack of generalizability.

Abstract

Decades of research have shown that people use a basic learning process called pseudocontingency inference to form beliefs about relationships between variables. Rather than relying on co-occurrences, people infer relationships based on separate occurrences of each variable. However, a fundamental question remains unanswered: how do individuals differ in their reliance on pseudocontingencies when forming beliefs? Existing computational models on pseudocontingencies have focused on group-level patterns, obscuring how individual differences affect belief formation. To this end, we formalize the degree to which people rely on actual contingencies or on pseudocontingencies. We focus on the belief that unhealthy food tastes better, a pseudocontingency effect observed even when actual contingencies suggest no or a negative relationship. Using data from previous experiments, we estimate the reliance on pseudocontingencies by calibrating a bias strength parameter at both individual and group levels. Our results reveal that people generally rely on pseudocontingencies instead of actual contingencies, but they do so to varying degrees. Bootstrapped estimates suggest that the median reliance on pseudocontingencies was 22-28% lower in individual-level compared to group-level models. The findings have implications for normative models that assume that people form beliefs about relationships based on actual contingencies. The significant lack of group-to-individual generalizability warrants concerns about the validity of group-level models as these may overestimate the reliance on pseudocontingencies.

Figures