Simultaneous Variable Selection for Joint Models of Longitudinal and Survival Outcomes

• Activity Number: 628
• Type: Invited
• Date/Time: Thursday, August 7, 2014 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Zangdong He*+ and Wanzhu Tu and Sijian Wang and Haoda Fu and Zhangsheng Yu
• Companies: Indiana University Fairbanks School of Public Health and Indiana University School of Medicine and University of Wisconsin and Eli Lilly and Company and Indiana University School of Medicine
• Keywords: Cholesky decomposition ; EM algorithm ; Gaussian quadrature ; Mixed e ect selection ; Penalized likelihood

Abstract

Joint models with longitudinal and survival outcomes are used with increasing frequency in clinical investigations. Variable selection in joint model settings, however, has not been developed. Herein, we proposed a doubly penalized likelihood (DPL) method with adaptive LASSO and adaptive selection operator (ASO) penalty functions to simultaneously select fixed and random effects. Reparameterization of covariance matrix by Cholesky decomposition was used to ensure its positive-definitness. Likelihood was penalized by row vector L-2 norm of the decomposed matrix. ASO was then incorporated into the penalized likelihood (PL) to enhance selection accuracy. To correct the estimation bias due to the penalty, we proposed a two-stage procedure to reduce bias. For computation, we approximated the PL by Gaussian quadrature and optimized it by expectation-maximization (EM) algorithm. Simulation showed that the procedure has excellent selection result and that the two-stage estimation substantially reduced the bias. To illustrate, we analyzed a real data set with brain natriuretic peptide as the longitudinal outcomes and death as the survival outcomes from an electronic medical record database.