Generally applicable modes of analysis for incomplete binary longitudinal clinical trial data

Abstract

Many clinical trials result in incomplete longitudinal data. Common analysis methods are complete case (CC) and last observation carried forward (LOCF), resting on strong and unrealistic assmumptions. Many full longitudinal methods, valid under MAR, have been developed. We foucs on non-Gaussian outcomes, a setting more complicated than the Gaussain counterpart, due to the lack of an analogy for the linear mixed model. Model choices include the random-effects based generalized linear mixed models (GLMM) and the marginal generalized estimating equations (GEE). Since the latter is non-likelihood based, it requires modification (weighted GEE) to be valid under MAR. Both methods provide similar results for hypothesis testing, but the estimated parameters have different interpretation. Current statistical computing brings GLMM and WGEE within reach and their implementation in depression trials is presented, showing they are viable alternatives for CC and LOCF, even when a single time point only (e.g., the last) is of interest. Even then, all information from all profiles, complete and incomplete, is used, showing this approach is fully compatible with the intention-to-treat principle.