School Seminars and Colloquia

Model and Working Correlation Structure Selection in GEE Analyses of Longitudinal Data

Statistics Seminar

by James Cui


Institution: Department of Epidemiology and Preventive Medicine, Monash University
Date: Tue 18th March 2008
Time: 1:15 PM
Location: Room 213 Richard Berry Building, The University of Melbourne

Abstract: The GEE method is one of the most commonly used statistical methods in the
analysis of longitudinal data. A working correlation structure for the
repeated measures of the outcome variable of a subject needs to be
specified in this method. However, statistical criteria for selecting the
best correlation structure and the best subset of explanatory variables in
GEE are only available recently. Maximum likelihood based model selection
methods, such as AIC, are not applicable directly to GEE. Based on the QIC
method proposed by Pan (2001), we systematically developed a general
computing program to calculate the QIC value for a range of different
distributions, link functions and correlation structures. The QIC value
can be used to select both the best correlation structure and the best
subset of explanatory variables. The program was written in Stata
software. In this talk, I will introduce the QIC method and program and
demonstrate how to use it to select the most parsimonious model in GEE
analyses through several representative examples. The program has also
been included in the book “Negative Binomial Regression” (2nd edition) and
used in various short courses on longitudinal data analyses.

For More Information: Guoqi Qian g.qian@ms.unimelb.edu.au