Multivariate Gamma-Poisson Models

Abstract This article extends the multivariate gamma-Poisson model of repeated events, developed by Arbous and Kerrich (1951) and Bates and Neyman (1952), by (a) compounding it with a Dirichlet distribution in the analysis of cross-sectional data and by (b) allowing individual rates to shift at the start of the second period in the analysis of two-period longitudinal data. The first extension allows individuals to specialize in particular types of events. The model is shown to describe multivariate distributions of work incapacities, prison infractions, and criminal victimizations. The second extension provides a framework for estimating changes in individual rates. It is used to describe the shift in victimization rates observed in longitudinal studies of the National Crime Survey.