Generalized Estimating Equations (GEE) for Missing Longitudinal Data with High-Dimensional Covariates

• Activity Number: 234
• Type: Topic Contributed
• Date/Time: Monday, August 10, 2015 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Ming Wang*
• Companies: Penn State
• Keywords: Longitudinal data ; Generalized Estimating Equation ; Multiple imputation ; Missing data ; Quasi-likelihood

Abstract

Missing longitudinal data with high-dimensional covariates has gained research attentions in clinical and biomedical studies. Generalized Estimating Equation (GEE), a marginal statistical method, is commonly used for longitudinal data analysis, and multiple imputation (MI) is popularly way to handle missingness. Selection on working correlation structure and covariates plays a vital role in improving the efficiency of the parameter estimates and the model goodness-of-fit; however, limit work exists on development of model selection strategies for GEE with MI. In this work, we propose a MI-based weighted Quasi-likelihood approach to account for sampling and imputation uncertainty. Also, we extend this proposal to the cases with high-dimensional covariates using penalized techniques for further evaluation. In addition, several existing alternatives including the quasi-likelihood under the independence model criterion (QIC) and the missing longitudinal information criterion (MLIC) are compared and evaluated to show our proposal's outperformance. Finally, the proposed method is illustrated by a real data on colorectal cancer.