Unifying Frameworks for Reward-Aversion Judgment: A Bayesian Analysis
Presented at the 2024 BayesiaLab Conference in Cincinnati on April 11, 2024.
Abstract
Using Bayesian Belief Networks (BBN), variables from three graphs based on Relative Preference Theory (RPT) were integrated. The graphs in RPT have analogies to known graphs in other frameworks for human judgment but are derived from the only framework tested and found to meet physics criteria for lawfulness. The graphs are analogous to those found in Prospect Theory, Markowitz portfolio theory, and the Payne, Bettman, and Johnson adaptive strategy selection. RPT produces nonlinear graphs with R2 fits per person, on average > 0.9. RPT graphs calibrate value based on patterns in prior judgments (value function), limits in risk-reward and risk-aversion relations (limit function), and balance between reward and aversion judgments (tradeoff function). Using the same picture rating task administered to three distinct internet cohorts, each with more than 3,000 subjects, graphs were observed calibrating reward/aversion value (value function), associating risk to reward/aversion (limit function), and balancing reward against aversion (tradeoff function). Fifteen mathematical features of these graphs based on engineering approaches or behavioral finance constructs, such as loss aversion and risk aversion, were computed and then used as input for Bayesian Belief Network analysis and correlational analysis to identify consistent relationships between these fifteen mathematical features. When consistent relationships were observed, we then fit mathematical functions to the combined dataset of 10,000+ subjects. This analysis showed that the calibration of value (prospect theory-like graph) anchors the relationship of risk-reward variables in a distinct manner from how it anchors the relationship of risk to aversion, and four distinct clusterings of these graph features can be observed based on their graph of origin with highly interpretable relationships. The fifteen graph features can be combined and applied to predict the outputs of human judgment using machine learning. Thus, it is possible to create an individual profile or “fingerprint,” which can be used to predict behavior or other psychological conditions such as anxiety, depression, and suicidality. Marketers can use the features as a segmentation variable to identify the best prospects and design the most effective messages.
Presentation Video
Authors
Hans Breiter
and Martin Block, Shamal Lalvani, Sumra Bari, Nicole L. Vike, Leandros Stefanopoulos, Byoung-Woo Kim, Aggelos K. Katsaggelos
Medill Integrated Marketing Communications, Northwestern University, Evanston, IL, USA
Department of Electrical and Computer Engineering, Northwestern University, Evanston, IL, USA
Department of Computer Science, University of Cincinnati, Cincinnati, OH, USA
School of Medicine, Aristotle University of Thessaloniki, Thessaloniki, Greece
Department of Radiology, Northwestern University, Chicago, IL, USA
Department of Computer Science, Northwestern University, Evanston, IL, USA
Laboratory of Neuroimaging and Genetics, Department of Psychiatry, Massachusetts General Hospital and Harvard School of Medicine, Boston, MA, USA
About the Presenter
Vikram Suresh, Ph.D., has been a Postdoctoral Fellow in the Department of Computer Science at the University of Cincinnati since August 2023. He earned his Ph.D. in Business Administration (Quantitative Economics) from the University of Cincinnati in 2023. His research focuses on combining Bayesian hierarchical modeling, AI, and econometric methods to analyze various topics, including treatment outcomes in adolescents with depression, socioeconomic predictors of treatment outcomes in adults with major depressive disorder, and the impact of age on antidepressant response. He has several publications in peer-reviewed journals and is working on manuscripts related to statistical approaches in randomized controlled trials, income inequality estimation, and the decline in high school student performance using AI experiments.