Treatment Effects in Randomized Longitudinal Trials with Different Types of Non-Ignorable Dropout

Randomized longitudinal designs are commonly used in medical and psychological studies to investigate the treatment effect of an intervention method or an experimental drug. Traditional linear mixed-effects models for randomized longitudinal designs are limited to maximum likelihood (ML) methods which assume data are missing at random (MAR). In practice, because longitudinal data are likely to be missing not at random (MNAR), the traditional ML method might lead to severely biased estimation results. In such cases, an alternative approach is to utilize pattern-mixture models. In this dissertation, Monte Carlo simulation studies are undertaken to compare the traditional ML method and two different approaches of pattern-mixture models (i.e., the D-A method and the A-D method) across (1) different variations of mixed-effects models (i.e., with or without adjustments), (2) different missing mechanisms (i.e., MAR, random-coefficient-dependent MNAR or outcome-dependent MNAR), and (3) different types of group-based missing probabilities. Analytical derivations are also provided to explain the source of bias across different estimation methods. Results suggest that the traditional ML method is well suited for MAR data whereas the proposed PM-AD model has the best overall performance for MNAR data. Omitting the group membership predictor for the intercept or including the baseline score as a covariate can lead to larger power and smaller bias in treatment effect estimates when the ML method is used. Applications of different estimation methods are also illustrated using a real data example.