Student Seminar Series - November 10, 2005
University of Minnesota
School of Statistics
College of Liberal Arts

Combining Linear Models in Longitudinal Data Analysis


Song Liu


Thursday, November 10, 2005
9:30 AM, 13 Rapson Hall (Architecture)
Minneapolis, East Bank Campus

Refreshments at 9:15 AM
300 Ford Hall


Abstract

Longitudinal data are the samples from a number of subjects, each measured repeatedly over a time period, very common in biomedical studies. Varieties of methods analyzing longitudinal data have been proposed. However, there exist gaps in model selection techniques for longitudinal data analysis, even for linear (mixed) models. Among the several existing selection techniques information criteria, such as AIC and BIC, have been widely applied. To deal with uncertainty in model selection, model combining methods have been developed in recent years. Its advantages over model selection have been demonstrated both empirically and theoretically. In this paper, we illustrate the uncertainty problem with model selection for longitudinal data by examples, and we propose a model combining method for estimating marginal means. Its theoretical properties are examined and numerical examples are presented. The results support model combining methods when model selection uncertainty is non-negligible.