Bayesian analysis of hierarchical pattern-mixture models for clinical trials data with attrition and comparisons to commonly used ad-hoc and model-based approaches.

This article addresses the problem of making scientifically sound inferences from clinical trials data with attrition, when reasons of attrition seem related to the outcomes of interest. The problem is particularly difficult when the effect of a covariate is to be estimated and if the dropout mechanism appears to operate differently at different levels of the covariate. In this context, multiple imputation under a multilevel pattern-mixture model that allows random variation across dropout groups is presented. Using simulated data generated around an alcoholic hepatitis trial, we compare the performance of this model and commonly used ad-hoc and model-based approaches.