Analysis of Longitudinal Trials with non-MCAR Dropouts and Potentially Non-Normal Data: Is Weighted GEE the Solution?

• Activity Number: 541
• Type: Contributed
• Date/Time: Thursday, August 10, 2006 : 10:30 AM to 12:20 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Robin Mogg*+ and Devan V. Mehrotra
• Companies: Merck Research Laboratories and Merck Research Laboratories
• Keywords: missing data ; weighted estimating equations ; multiple imputation

Abstract

In a typical longitudinal comparative clinical trial, some subjects drop out because of "treatment failure", some are lost to follow-up, etc. The incomplete longitudinal data are often analyzed using REML, the SAS PROC MIXED default, which assumes multivariate normality of the responses. If the normality assumption is untenable and the missing data are "missing at random", the weighted generalized estimating equations (WGEE) method of Robins et al. (1995) can be used as an alternative to REML. In this talk, we will use simulations to compare WGEE with both REML and a new method that we have developed, multiple imputation of the missing values followed by application of the Wei-Lachin (1984) approach with Wilcoxon scores (MI -> WL). The simulations reveal that, with respect to type I error rate control and power, the MI -> WL method is the best, while WGEE is generally the worst.