Mixed Effects Regression Trees

• Activity Number: 18
• Type: Contributed
• Date/Time: Sunday, August 3, 2003 : 2:00 PM to 3:50 PM
• Sponsor: General Methodology
• Author(s): Sudeshna Adak*+
• Companies: GE Global Research
• Keywords: repeated measures ; longitudinal ; growth curves ; regression trees ; mixed effects ; score test

Abstract

Mixed effects models and growth curve models have been popularly used in biomedical, sociological, and epidemiological applications to study longitudinal and clustered data. There has been recent interest in nonparametric and semiparametric methods for analyzing such data using splines or kernels. We introduce mixed effects regression trees (MERT)--a general tree growing methodology for longitudinal/clustered data that uses the Laird and Ware (1982) formulation of mixed effects models. It is shown that MERTs can be very easily tuned to growth curve models and used to determine which factors influence growth rate parameters. The advantages of using trees instead of the more continuous spline or kernel based approach is that it allows us to determine significant covariates and identify important strata in the covariates that result in similar outcomes. While the use of trees with independent data is well established, it has not been extended to longitudinal/clustered data. MERT allow us to extend the success of CART to longitudinal/clustered data. An application of this is demonstrated to determine significant covariates that affect the rate of decline in Alzheimer's patients.