Longitudinal Data Analysis--Is "Optimal" Modeling of the Covariance Structure Necessary?

• Activity Number: 163
• Type: Contributed
• Date/Time: Monday, August 9, 2004 : 2:00 PM to 3:50 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Devan V. Mehrotra*+ and Geert Molenberghs and Russell D. Wolfinger
• Companies: Merck & Co., Inc. and Limburgs Universitair Centrum and SAS Institute Inc.
• Keywords: compound symmetry ; missing data ; REML ; repeated measures ; sandwhich estimator ; unstructured covariance

Abstract

In a typical comparative clinical trial, subjects are randomized to receive either treatment A or B. For each subject, the response of interest (e.g., blood pressure) is measured at baseline and at fixed time points after initiation of treatment. The presumed primary objective is to test the hypothesis that the true mean response at the end of the treatment period is the same for A and B. To address this objective, the response data are commonly analyzed using a "cell means" linear model approach based on the restricted maximum likelihood (REML) methodology. A key component of the REML-based analysis is modeling of the dependency between intrasubject responses at different time points, often referred to as the "covariance structure." Two strategies are commonly used in practice. In the first, the covariance structure to be used is specified in advance (compound symmetry, AR(1), etc.). In the second, an attempt is made to select the "optimal" covariance structure for the data in hand from a menu of available structures, based on a particular information criterion (e.g., AIC). We discuss two alternate strategies, one based on the so-called "sandwich" covariance.