Bayesian Methods for Modeling Nonignorably Missing Data in Cluster Randomized Trials with Binary Outcomes

• Activity Number: 447
• Type: Topic Contributed
• Date/Time: Wednesday, August 6, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Catherine Crespi*+
• Companies: University of California, Los Angeles
• Keywords: cluster randomized trials ; nonignorably missing data ; missing data mechanism ; Bayesian modeling ; Bayesian methods ; binary outcomes

Abstract

Missing responses are a common problem in cluster randomized trials (CRTs), in which groups of individuals rather than individuals per se are randomized to different conditions. The multilevel data structure of CRTs makes many missing data strategies, including most multiple imputation methods, inappropriate, because they assume independent observations. The problem is especially acute for binary outcomes; whereas some multiple imputation strategies for multilevel normal data have been developed, such methods are lacking for multilevel binary data. We discuss Bayesian methods for modeling missing binary responses in CRTs. Full Bayesian modeling provides a comprehensive framework for handling the missing responses and exploring the sensitivity of inferences to different assumptions about the missing data mechanism. The methods are illustrated on a CRT of an intervention to promote colorectal cancer screening among individuals at high risk for the disease.