Longitudinal Data Analysis: Tackling Dropouts and Those Pesky Outliers

• Activity Number: 344
• Type: Contributed
• Date/Time: Tuesday, August 3, 2010 : 10:30 AM to 12:20 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Devan V. Mehrotra*+ and Xiaoming Li and Jiajun Liu
• Companies: Merck Research Laboratories and Merck Research Laboratories and Merck Research Laboratories
• Keywords: Missing Data ; M-estimation ; Multiple Imputation ; Non-normality ; Outliers ; Robust regression

Abstract

In a comparative clinical trial, endpoints of interest (e.g., bone density) are measured at baseline and fixed post-baseline time points. The resulting longitudinal data, often incomplete due to dropouts, are commonly analyzed using likelihood-based methods (e.g., REML) that assume multivariate normality of the response vector, conditional on covariates. If the normality assumption is untenable, semi-parametric methods like GEE or weighted GEE are sometimes recommended. To retain good performance under normality, but significantly increase efficiency under non-normality, we propose an easy-to-implement alternate approach that uses multiple imputation to tackle missing data and a subsequent robust parametric analysis to tackle potential non-normality and/or outliers. The efficiency gains using the proposed versus standard methods are illustrated using simulations and a real example.