A Performance Comparison of Nonparametric vs. Generalized Linear Models in Longitudinal Studies

• Activity Number: 216
• Type: Contributed
• Date/Time: Tuesday, August 5, 2003 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Keith Williams*+ and Mark Austen
• Companies: University of Arkansas for Medical Sciences and University of Arkansas
• Keywords: longitudinal ; nonparametric ; GLM ; simulation ; power

Abstract

Recently there has been a unification of theory for nonparametric models that can be applied to longitudinal studies. These methods have a few trivial assumptions to be met for their application. Generalized linear models (GLM) are also in common use to model longitudinal data. Associated with these GLM models is a set of assumptions about the data. These assumptions include properties about the form, distribution, and covariance structure of the observations. One question that naturally arises in consideration of nonparametric methods is: In what situations are nonparametric methods more or less powerful than GLM methods, especially when the assumptions of the GLM are violated but GLM models applied regardless? This research focuses on comparing performance characteristics such as power for new nonparametric methods versus appropriately and inappropriately applied GLM models and offers performance information guidelines in the use of nonparametric methods. In simulation studies, it was found that for even moderate sample sizes, the performance of these nonparametric methods is comparable to GLM methods and in some instances outperforms them.