Date of Award

12-24-2025

Date Published

January 2026

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Psychology

Advisor(s)

Michael Kalish

Keywords

Bayesian modeling;belief revision;conceptual change;education;learning;relational knowledge

Subject Categories

Cognitive Psychology | Psychology | Social and Behavioral Sciences

Abstract

Are people rational Bayesian learners that produce a final posterior belief that is a compromise of their prior and the amount of new data? If they are then can one create a model that can predict their final belief given an estimate of their prior belief and the new data? This study aims to explore this. The participants were presented with two relational structures, the initial and the alternative which contradicted each other to a certain degree. In condition 1 the alternative structure was similar to the initial and in condition 2 the structure had the same number of contradictions as condition 1 but the items that contradicted were random. In condition 3, all of the relation pairs contradicted the initial structure. The participants were given a cover story that aimed to reduce pre-experiment bias. The relational structures were presented as directional relational item pairs. During training the participants were given feedback such as (yes, the items interact) or (no, the items do not interact). Given their initial trained belief about the relational item pairs and the conflicting evidence do they update their beliefs in a predictable way? The model generated a prediction and an estimate to compare. The model’s estimated the participant’s belief using their test responses of the item pairs and a vague prior. The model could also generate a prediction given their prior, an estimate from their test responses of the initial phase, and the feedback they would see in the alternative phase to generate a posterior, prediction. The model prediction and the estimate for each person are compared to see if they are credibly the same, if the HDI of the prediction fits with the HDI of the estimate. A number of possible pre-experiment biases were explored and it seems most likely that people use a beta (2,2). Before the experiment began people might assume that there is a 50% chance of the items interacting or not. There was a lot of overlapping between the participants test responses to the initial and alternative phases, but they were affected by if they saw conflicting or consistent feedback. The difference in the participants’ responses between the initial and the alternative phases were also affected by the condition they were in. The condition affected the model’s ability to make predictions. The model had the most trouble with condition 3, where all of the item pairs conflicted. This makes sense since the model also had more trouble with making predictions for the conflicting item relations than the consistent.

Access

Open Access

Download

Share

COinS