Advanced Bayesian Methods for Generalized Choice Response Time Models of Decision-Making

Indiana University, Bloomington IN

Investigators

Abstract

This research project will develop new tools to study human judgment and decision making. A significant challenge in studying decision making is testing how well a hypothesis or theory is supported by data. In many studies, we only see the outcome of a decision of interest, which makes it difficult to understand the process that led to that decision and to directly test theories about that process. Computational models have proven valuable as tools to encode theories and test them against observations. However, computational models present technical barriers that limit their usability and constrain the scope of questions that can be addressed. This project will provide a suite of efficient yet generally usable computational approaches and tools that facilitate model construction and analysis. The new methods to be developed will broaden the scope of investigations that are possible and the researchers who can carry them out. To ensure their broadest possible usability, the tools will be disseminated in freely available software packages. The investigators will use these tools to study the properties of multi-alternative, multi-attribute choice and assess how context influences people's choices in naturalistic settings. Students supported by this project will be trained in state-of-the-art computational methods which are becoming increasingly more common in science and industry. This project will develop advanced Bayesian methodologies for performing parameter estimation for choice-response time (RT) models. The time it takes for people to make decisions (RTs) provides valuable information about the dynamic process responsible for those decisions. For this reason, models that predict both choices and RTs are used to study decision processes. However, it is challenging to fit these types of models to data, which is a necessary step in assessing the quality of the theories they encode. As a result, model-based approaches are most often applied using decades old models in conjunction with simple experimental designs, both to maintain tractability. To address these issues, this project will develop a set of accessible, high-quality probabilistic methods that are documented to be effective at performing Bayesian parameter estimation for a wide variety of choice-RT models. Researchers will be able to construct more complex choice-RT models and utilize more complex experimental designs, which in combination can facilitate new scientific investigations. As an example of this, the investigators will study the role of context in decisions involving naturalistic information encoded in, for example, semantic or image-based stimuli. Machine-learning models of language or image representations will be integrated with choice-RT models encoding different assumptions about contextual dependencies. These models are fit to complex data sets derived from large-scale experimental designs involving large numbers of participants making naturalistic decisions. The results of this study will help resolve the debate about whether naturalistic choices show context-dependency, which is observed with more artificial stimuli. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.