Predictive Bayes Factors with an Application Model-Assisted Design

Presented by Shouhao Zhou, Department of Public Health Services, Division of Biostatistics and Bioinformatics, Penn State Cancer Institute, Pennsylvania State University

Wednesday, March 22, 2023
4:00 PM ET
AUST 163
Webex Meeting Link

Bayes factors approach has been widely used as an applied model selection tool by comparing posterior model probabilities. However, it is often criticized for counterintuitive results in statistical hypothesis testing. We propose a new approach, predictive Bayes factors, for predictive model selection of fitted Bayesian models via posterior model probability. In theory, the approach asymptotically estimates the log predictive density ratio. In practice, it dramatically reduces sensitivity to variations in the prior and totally avoids the Lindley’s paradox in testing point null hypothesis, with decent small sample performance. We also develop a novel Bayesian model-assist trial design using this new approach, to determine the escalation and de-escalation boundaries for dose-finding trials. It overcomes the limitations of the previous model-assisted designs and serves as the first model-assisted design to guarantee global optimality and the asymptotic convergence to true MTD. Simulation results also demonstrate the superior operating characteristics based on a large number of scenarios.

Speaker Bio:

Shouhao Zhou is an Assistant Professor of Biostatistics at Pennsylvania State University. He got his PhD in Statistics from Columbia University and previously was a junior faculty at MD Anderson Cancer Center, where he learned the state-of-the-art Bayesian adaptive designs. He is also interested in Bayesian hypothesis testing and model selection, meta-analysis, and statistical modeling of preclinical data.