Random Mean Models for Mixed Continuous and Discrete Clustered Outcomes

DOI: 10.5176/2251-1911_CMCGS40

Authors: Dandan Luo, Peng Zhang and Jianghu Dong

Abstract: Observations of mixed responses over time occur often in health and medicine sciences. The difficulty in joint modeling of continuous and discrete response variables is the lack of a natural multivariate distribution. This paper develops a flexible class of generalized linear latent variable models for multivariate mixed responses, that has underlying Gaussian latent processes. This class of models accommodates any mixture of outcomes from the exponential family. Monte Carlo EM (MCEM) algorithm is proposed for estimating the regression parameters and the variance components of the latent processes. We demonstrate the methodology with the kidney study data, which have observations of the continuous response, estimated glomerular filtration rate (eGFR), and the binary response, retransplant status, over time for each patient, through jointly modeling the clustered observations on binary and continuous responses.

Keywords: Correlated mixed data, Importance sampling, Laplace approximation, Monte Carlo EM algorithm.