Date of Award
Doctor of Philosophy (PhD)
Michael C. Mozer
Clayton H. Lewis
Whether driving a car, making critical medical decisions in the ER, answering questions in a marketing survey, or selecting shots in a basketball game, people make decisions and actions that are biased by the sequence of recent experience. These sequential effects are ubiquitous in human behavior and have been demonstrated in a wide range of experimental paradigms. This dissertation begins with a synthesis of the vast computational modeling landscape in this domain. Building upon one of the core principles revealed in this synthesis, I explore how sequential effects in simple choice tasks reflect an individual's attempt to optimize behavior in an ever-changing world. A Bayesian model is proposed which asserts that humans are sensitive to multiple environmental regularities and adapt their behavior according to expectations derived from these sensitivities. Through analyses of two experiments that question how far into the past these sensitivities extend, I demonstrate that events far in the past can exert an observable bias on behavior. This finding is surprising given the prevailing perspective in the literature that sequential effects are relatively ephemeral, fading after roughly 4-6 intervening events. To accommodate this new perspective, a hierarchical generalization of the model is presented that allows for long-range sensitivities exhibiting power decay. Given an expanded understanding of the mechanisms underlying sequential effects, the final chapter focuses on how this understanding can be put to practical use. I address sequential biases in judgment tasks and develop techniques for removing the biases from a sequence of responses. By decontaminating the responses using a novel hierarchical Bayesian model that exploits knowledge of sequential effects, a set of new responses is obtained that is more representative of the individual's true opinions. For each question or stimulus that is judged, the goal is to uncover what the individual would have responded in the absence of any sequential context. Given the growing interest in collecting human judgments and using them to predict individual preferences (e.g., Netflix, Amazon), the ability to effectively decontaminate sequences of judgments is of significant value because it produces more reflective estimates of an individual's internal state with fewer total judgments required.
Wilder, Matthew H., "Probabilistic Modeling of Sequential Effects in Human Behavior: Theory and Practical Applications" (2013). Computer Science Graduate Theses & Dissertations. 63.