Multivariate Longitudinal Analysis with Missing Data

• Activity Number: 242
• Type: Contributed
• Date/Time: Monday, August 5, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Priya Kohli*+ and Tanya Garcia and Mohsen Pourahmadi
• Companies: Assistant Professor and Texas A&M University and Texas A&M University
• Keywords: Cholesky decomposition ; Expectation-Maximization ; Missing data

Abstract

Multivariate longitudinal studies are common in biological, medical and social sciences where two or more attributes are measured on each subject repeatedly over time. In this work, we are concerned with modeling the mean and covariance structure of incomplete multivariate longitudinal data. This involves dealing with the challenge of modeling the correlations as well as handling missing data. Kim and Zimmerman (2012) proposed an unconstrained parametrization of the covariance matrix for multivariate longitudinal data which is based on the modified Cholesky block decomposition of the covariance matrix. However, this approach is not directly applicable when there are missing values in the multivariate longitudinal data, because a coherent Cholesky factorization will usually not exist for all subjects. We extend this approach to incomplete multivariate longitudinal data by employing a generalized EM algorithm for estimating the parameters of the mean-covariance models and derive the score functions and Fisher information matrices corresponding to these parameters. The performance of the proposed estimators is illustrated through an application to the real data and simulation studies.