Abstract: We propose methods for estimating parameters in two types of models for discrete longitudinal data in the presence of nonignorable missing responses and covariates. We first present the generalized linear model with random effects, also known as the generalized linear mixed model. We specify a missing data mechanism and a missing covariate distribution and incorporate them into the complete data log-likelihood. Parameters are estimated via maximum likelihood using the Gibbs sampler and a Monte Carlo EM algorithm. The second model is a marginal model for correlated binary responses and discrete covariates with finite range, both of which may be nonignorably missing. We incorporate the missing data mechanism and the missing covariate distribution into the multivariate probit model defined by Chib and Greenberg (1998). We use the EM by method of weights (Ibrahim, 1990) and sample the latent normal variables conditional on a particular response and covariate pattern. The M-steps for each model are like a complete data maximization problem, and standard methods are used. Standard errors for the parameter estimates are computed using the multiple imputation method of Goetghebeur and Ryan (2000). We discuss the advantages and disadvantages of each model and give some guidance as to when one model might be chosen over the other. We illustrate both models using data from an environmental study of dyspnea in Chinese cotton factory workers.
Key words and phrases: Generalized linear mixed model, Gibbs sampling, Monte Carlo EM algorithm, multivariate probit model, nonignorable missing data.