658 – Missing Data Methods
Simulation Study on Selection of Latent Class Models with Missing Data
Mark Reiser
Arizona State University
Jun Zhang
Arizona State University
Longitudinal biomedical studies often encounter substantial missing data. An increasing number of articles introducing methods for handling missing data have discussed and used latent class models as a flexible way of modeling correlated multivariate categorical data. However, one key assumption of latent class modeling, the validity of the number of latent classes for missing data, has not been examined. The aim of this paper is to investigate the "correct" number of latent classes through simulation studies with missing values. We apply Monte Carlo simulation to generate a longitudinal study with 6 time points and two different missing mechanisms: missing completely at random and missing not at random. A linear mixed model with random intercept and slope is assumed for each latent class. We choose the most efficient approach to evaluate model performances with different latent classes: information criteria. Furthermore, we have investigated how the following factors influence the selection of latent classes for missing data: covariates effects, missing probabilities and the degree of associations among repeated measures. Due to the difficulties to identify the missing mechanism(s) in practice, missing patterns are also investigated in fitting latent class models.