An Application of Bayesian Method in Control-Based Pattern Mixture Models for Longitudinal Studies with Missing Data

• Activity Number: 222
• Type: Topic Contributed
• Date/Time: Monday, August 4, 2014 : 2:00 PM to 3:50 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Frank Liu*+
• Companies: Merck
• Keywords: missing data ; , sensitivity analysis ; multiple imputation ; Bayesian method

Abstract

Missing data can be inevitable and complicates the statistical analysis in longitudinal clinical trials. General guidelines issued by regulatory agencies recommended conducting sensitivity analysis to assess the robustness of the primary analysis results that are often based on maximum likelihood methods under the assumption of missing data are missing at random. Pattern mixture models (PMM) contain a class of methods for sensitivity analysis on handling missing data. A special type of PMM is control-based imputation in which the missing data are imputed based on the models from control group. It assumes that patients who discontinued from the test treatment may have response profile similar to that of the control group. This will yield a conservative estimate and test for treatment difference. The method is often implemented by multiple imputation approach. In this talk, we will apply a Bayesian approach to fit the control-based imputation model, and describe how to use MCMC procedure for statistical inference. We will discuss the pros and cons of this approach and compared with the conventional method, and illustrate the approach in analysis of a longitudinal clinical trial.