Cross-study hierarchical modeling of stratified clinical trial data

Hierarchical random-effects models can be used to estimate treatment or other covariate effects in single-study analyses coordinated over multiple clinical units and can also be extended to a wide variety of cross-study applications. After reviewing the single-study case, we use data from five trial protocols to look for units that tend to have treatment effects consistently above or below the study-specific grand mean across several studies. As a first step, we summarize the patient-level data as study-specific and unit-specific estimated treatment effects and standard errors using independent Cox regression models. We then compare the results of a hierarchical model using these data summaries as input to those produced by a more fully Bayesian method that uses the actual patient-level survival data. We also compare various different models using a deviance information criterion, a recent extension of the Akaike information criterion designed for hierarchical models. Our procedure appears to be effective at answering the question whether certain clinical units of the Terry Beirn Community Programs for Clinical Research on AIDS are better than others at identifying treatment effects where they exist.