The Bayesian-optimality of decision-making behavior: across tasks, time, and psychopathology.

Abstract

From medical diagnosis to catching a baseball, people make decisions all the time. But on what basis are those decisions made and how optimal are they? This dissertation focuses on one theoretical approach to assess the optimality of decisions: Bayesian Decision Theory (BDT). BDT postulates that posterior estimation relies on two categories of information: knowledge gained over time (i.e., prior information) and current sensory input (i.e., likelihood), with greater weight given to the category associated with less uncertainty. Here, I test the possibility that while decisions are typically Bayesian in a qualitative sense, they often deviate from BDT quantitatively. Three empirical studies explored the Bayesian optimality of decision-making behavior in visual search (Chapter 2), across sensorimotor and visual search tasks (Chapter 3), and in patients with Borderline Personality Disorder (Chapter 4). Chapter 2 created a novel hybrid search-decision task, in which participants made a target present/absent response on a display of items with partial occlusion. I found that while participants considered both the target's prevalence ("prior") and the degree of occlusion ("likelihood"), they gave disproportionate weights to visible information, showing a mixture of Bayesian inference and under-matching. Chapter 3 tested behavior in the visual search task and a sensorimotor "coin-catching" task within the same set of individuals across two time points, assessing the degree to which they relied on prior vs. likelihood. I found consistent individual differences within a task, with some measures of both tasks displaying good test-retest reliability, but not between tasks, arguing against a domain-general Bayesian weight. Chapter 4 showed that while patients with BPD performed like controls in the coin-catching task in a qualitatively Bayesian manner, both fell short of quantitative BDT predictions. Overall, these findings demonstrate that BDT is a powerful framework for understanding a range of decision-making behaviors including sensorimotor and attentional decisions, across patients and typical groups. However, they also show that Bayesian weights may not be domain-general, and factors other than prior and likelihood may influence decisions.