A Comparison of Two Fully Bayesian Changepoint Models for Early Detection of Prostate Cancer

• Activity Number: 90
• Type: Contributed
• Date/Time: Monday, August 9, 2004 : 9:00 AM to 10:50 AM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Wonsuk Yoo*+ and Elizabeth H. Slate
• Companies: Medical University of South Carolina and Medical University of South Carolina
• Keywords: Prostate Specific Antigen (PSA) ; Bayesian changepoint model ; reversible jump Markov Chain Monte Carlo (MCMC) ; conditional predictive ordinate (CPO) ; pseudo-Bayes factor ; ROC curve

Abstract

This research compares two models for longitudinal prostate specific antigen (PSA) readings according to their ability to detect prostate cancer (Pca) onset. Model I is a fully Bayesian hierarchical changepoint model, similar to that of Slate and Clark (1999), in which cancer onset is represented as a changepoint in the men's PSA trajectories and all men are presumed to experience onset eventually. Model II postulates a mixture for the PSA series, for which one component contains a change point and the other does not. Both models are fit using Markov chain Monte Carlo methods, with a reversible jump implementation for model II. We apply these models to data from the Nutritional Prevention of Cancer Trials, and investigate the effects of covariates (smoking, alcohol usage, and body mass index) on PSA growth by examining credible regions for covariate parameters and by computing conditional predictive ordinate values and pseudo-Bayes factors. We then use ROC curves to compare the performance of diagnostic rules for Pca onset derived from the posterior distributions of the change points for each model.