Online Program

Bayesian multivariate hierarchical transformation models for ROC analysis

*James O'Malley, Dartmouth College 
Kelly Zou, Pfizer 

Keywords: Receiver operating characteristic curve, Box-Cox transformation, Hierarchical model, Optimal diagnostic test

A Bayesian multivariate hierarchical transformation model is developed for receiver operating characteristic (ROC) curve analysis based on clustered continuous diagnostic outcome data with covariates. The model incorporates non-linear monotone transformations of the outcomes and allows multiple correlated outcomes may be analyzed. The general framework is illustrated by focusing on the estimation of: (1) the diagnostic accuracy of a covariate-dependent univariate test outcome requiring within-cluster Box-Cox transformations to satisfy the model assumptions; (2) an optimal composite diagnostic test using multivariate clustered outcome data. The proposed methodology is illustrated on prostate cancer biopsy data from a multi-center clinical trial.