Impact of Various Modeling Options for Using GEE to Analyze Over-Dispersed Longitudinal Count Data

• Activity Number: 118
• Type: Topic Contributed
• Date/Time: Monday, July 30, 2012 : 8:30 AM to 10:20 AM
• Sponsor: Biopharmaceutical Section
• Author(s): Scott Miller*+
• Companies: FDA
• Keywords: GEE ; count data ; overdispersion

Abstract

Clinical trials often analyze count data from the same subjects monitored over time. These data are often analyzed via a generalized estimating equation (GEE) approach; however, several aspects of this trial design complicate the analysis. A working correlation structure must be specified to model the anticipated correlation. The empirical\sandwich standard error estimate is said to be robust to the choice of the working correlation, but some have expressed a preference for the model-based standard error estimate. And due to both within-subject and between-subject heterogeneity, the data are often over-dispersed relative to the commonly used Poisson distribution. As a result, the analyst must select between using the Poisson distribution without adjusting the variance, using the Poisson distribution and adjusting the variance via the scale parameter, or using the negative binomial distribution. Thus, there are three major issues (working correlation structure, standard error estimate, and statistical distribution) which may impact the results. This talk will compare several different combinations of these choices for analyzing over-dispersed longitudinal count data.