Conditional Generalized Estimating Equations for Longitudinal Data with Nonignorable Missing Data

• Activity Number: 340
• Type: Invited
• Date/Time: Wednesday, August 14, 2002 : 2:00 PM to 3:50 PM
• Sponsor: Biometrics Section*
• Author(s): Edward Vonesh*+
• Affiliation(s): Baxter Healthcare Corporation
• Keywords: Gaussian random effects ; Informative dropout ; Poisson regression ; Weibull survival models ; Quadratic exponential family ; Shared parameter models

Abstract

Longitudinal studies which focus on modeling serial measurements over time are frequently analyzed using generalized linear or nonlinear mixed-effects models. Such analyses are complicated by the potential for having nonignorable missing data and the possibility that the underlying model is nonlinear in the random effects. To address the first issue, we consider a class of shared parameter models in which we jointly model serial outcome measurements and missing data indicator variables, assuming the joint distribution is from a quadratic exponential family. Such models fall within a multivariate nonlinear mixed-effects framework, and estimation requires maximizing an integrated likelihood function. As an alternative to numerical integration, which may prove difficult to implement, we propose using first- or second-order Conditional Generalized Estimating Equations (CGEE) for estimating fixed and random-effects parameters. We investigate via limited simulation the performance and sensitivity of CGEE in longitudinal settings with nonignorable monotone missing data associated with dropout. The methods are illustrated using data from the Modification of Diet in Renal Disease Study.