Know your priors: Task specific priors reflect subjective expectations in Bayesian models of categorization
Nicolás Marchant, Felipe Diego Toro-Hernández, Sergio E. Chaigneau · 2021
The emergence of Bayesian causal models has had an enormous influence on how researchers understand people’s processing of probabilistic information, providing a description of cognitive phenomena and allowing their formalization and ensuing mathematical predictions. In the topic of categorization, the Generative Model (GM) is one such Bayesian formalization. Over three different experiments we tested if the GM can predict human causal-based categorization when subjects are facing probabilistic information about feature base-rates and about feature’s causal-strength. In our experiments we implemented a condition designed to favor subjects’ use of interfeature causal relations when making judgments (i.e., the Consistency condition). We contrasted this condition with a typical Category Membership condition. Our data suggest that on both conditions participants were doing causal-based processing. However, our data also shows that subjects in the Category Membership condition weighted causal information less and feature base-rates more than in the Consistency condition. Our results can be accounted for by assuming that subjects brought task-specific prior probabilities to the category membership task, biasing them into believing that a category’s characteristic features were high-probability features. Other explanations, such as that subjects in the Category Membership condition were doing similarity-based processing or that they did not understand causal-probabilistic information, were possible to rule out based on our experiments.