Issues When Using Multilevel Multiple Imputation to Handle Missing Binary Data in Cluster Randomized Trials

• Activity Number: 184
• Type: Contributed
• Date/Time: Monday, August 4, 2014 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Jinhui Ma*+ and Monica Taljaard and Lisa Dolovich and Janusz Kaczorowski and Larry Chambers and Lehana Thabane
• Companies: Children's Hospital of Eastern Ontario and Ottawa Hospital Research Institute and McMaster University and McMaster University and Élisabeth Bruyère Research Institute and McMaster University
• Keywords: cluster randomized trial ; missing data ; multilevel multiple imputation

Abstract

Missing data are a common occurrence in Cluster randomized trials (CRTs). It is important to handle missing data properly in the analysis to ensure that valid inferences are obtained. Multiple imputation (MI) offers distinct advantages over other imputation methods and has been widely used in health research. It has been widely recognized that multilevel MI, which is implemented in several software packages such as MLwiN, should be used to reflect homogeneity within clusters in the context of CRTs. Multilevel MI technique works well to handle missing continuous outcomes in CRTs. However, limited attention has been paid to its performance for handling missing binary data in CRTs. Based on empirical and simulated data, bias was assessed in the treatment effect estimate and its standard error, and coverage probability when multilevel MI is used for imputing missing binary data in CRTs and analysis proceeds by logistic regression using the marginal or subject-specific approach. Restrictions and recommendations regarding the appropriate use of multilevel MI for missing binary outcomes in CRTs are provided to facilitate the appropriate use of this technique in practice.