Analyzing longitudinal binary data in clinical studies

Abstract

In clinical studies, it is common to have binary outcomes collected over time as repeated measures. This manuscript reviews and evaluates two popular classes of statistical methods for analyzing binary response data with repeated measures: likelihood-based Generalized Linear Mixed Model (GLMM), and semiparametric Generalized Estimating Equation (GEE). Recommendations for choice of analysis model and points to consider for implementation in clinical studies in the presence of missing data are provided based on a comprehensive literature review, as well as, a simulation study evaluating the performance of both GLMM and GEE under scenarios representative of typical clinical trial settings. Under Missing at Random (MAR) assumption, GLMM is preferred over GEE, and the SAS PROC GLIMMIX marginal model is recommended for implementing GLMM in analyzing clinical trial data. When there is an underlying continuous variable used to define the binary response, and the missing proportion is high and/or unbalanced between treatment groups, a two-step approach combining Multiple Imputation (MI) and GEE (MI-GEE) is recommended.