Analysis of Longitudinal Data with Non-Random Missingness Using Shared Random Effects Models

• Activity Number: 461
• Type: Contributed
• Date/Time: Wednesday, August 3, 2011 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Xiaoyun Li*+ and Stuart Lipsitz and Dipankar Bandyopadhyay and Geert Molenberghs and Debajyoti Sinha
• Companies: Merck & Co., Inc. and Brigham and Women's Hospital and Medical University of South Carolina and Universiteit Hasselt/Katholieke Universiteit Leuven and Florida State University
• Keywords: Bridge density ; Logistic link ; Longitudinal data ; Non-random missingness ; Shared random effects

Abstract

Incomplete data is common in longitudinal studies, owing to subjects missing one or more follow up visits. When the missing-data mechanism depends on the outcome of interest, inferences that only use the complete data will no longer be valid. Shared random-effects models that uses latent variables to link the binary longitudinal response to the missing-data mechanism have been proposed, within a Gaussian framework. However, such models only describe the conditional model of the binary response given random effects and do not yield a closed-form marginal model for either the binary longitudinal response or the missing-data mechanism. We propose a shared random-effects model that, unlike the existing models, allows to preserve a logistic regression form for, at the same time, the marginal probability of the binary response, the missing-data indicator, and the conditional probability of the