Selection Models with Augmented Gibbs Samplers for Continuous Repeated Measures with Nonignorable Dropout
Premature withdrawal is a notable problem in biomedical research with longitudinal design. It would complicate statistical analysis with biased or invalid inferences if missing values due to dropout are simply ignored. As one solution to potentially nonignorable droput, the selection model assumes a mechanism of outcome-dependent dropout and jointly models the process of repeated measures and the mechanism of dropout. Previous applications of selec- tion models mainly resort to likelihood-based inferences using optimization methods such as simplex or EM-algorithm. This paper implements the modeling strategy using mixed-effects repeated-measures models with Bayesian inferences. Specifically, the selection model with random-effects and the one with autoregressive covariance structure are introduced. Markov Chain Monte Carlo (MCMC) algorithms based on augmented Gibbs samplers are developed in fitting the models. For demonstration, both simulated and practical data sets are analyzed.