Conditional Estimation for Joint Models for a Primary Endpoint and Multivariate Longitudinal Data

• Activity Number: 378
• Type: Contributed
• Date/Time: Wednesday, August 9, 2006 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Erning Li*+ and Naisyin Wang and Nae-Yuh Wang
• Companies: Texas A&M University and Texas A&M University and Johns Hopkins University School of Medicine
• Keywords: multivariate random effects model ; generalized linear model ; asymptotic bias ; conditional score ; variance components ; measurement error

Abstract

We study the association between a primary endpoint and features of multiple longitudinal processes using joint models. In the joint model, subject-specific random effects from a multivariate linear random effects model for the multiple longitudinal processes are used as predictors in a generalized linear model for the primary endpoint. An asymptotic bias analysis indicates that the estimators obtained by Li et al. (2004, Biometrics 60, 1-7), which make no distributional assumption on random effects but assume independent within-subject errors in the longitudinal covariate process, can yield biased inference when these within-subject errors are in fact correlated. To overcome this drawback, we generalize their results to joint models with more flexible multivariate longitudinal covariate processes and develop inferential methods for the regression parameters which are easy to implement.