Count Data Modeling in Longitudinal Studies: Performance of Conditional and Marginal Approaches

• Activity Number: 468
• Type: Contributed
• Date/Time: Wednesday, August 6, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Leila Amorim*+ and Daniele B. Trindade and Raydonal Ospina
• Companies: Universidade Federal da Bahia and Universidade Federal de Pernambuco and Universidade Federal de Pernambuco
• Keywords: Poisson ; Negative Binomial ; Mixed Effects Model ; Generalized Estimating Equations ; Longitudinal Studies

Abstract

Poisson and Negative Binomial models are often used for analysis of count data when observations are independent. If measures are taken repeatedly over time it is of interest to assess rates of change and factors that may affect this change. Important features of longitudinal data refer to the correlation between repeated measurements of the same individual and the chronological ordering of them. Many methodological advances for analysis of longitudinal data have been observed lately and introduction of new computational tools have become many methods accessible to researchers. Two approaches commonly used to analyze longitudinal data are mixed and marginal models. The mixed model incorporates random effects to deal with intra-individual dependence, whereas in marginal models the response variable is modeled separately from the correlation between the measurements of each sample unit. In this work the performance of estimators from mixed models and generalized estimating equations for analysis of count longitudinal data are evaluated through simulation studies. An application related to modeling CD4 cell counts in patients with AIDS in Brazil is discussed.