Hermite Regression Analysis of Multi-Modal Count Data
David E. A. Giles · RePEc: Research Papers in Economics · 2010
We discuss the modeling of count data whose empirical distribution is both multi-modal and over-dispersed, and propose the Hermite distribution with covariates introduced through the conditional mean. The model is readily estimated by maximum likelihood, and nests the Poisson model as a special case. The Hermite regression model is applied to data for the number of banking and currency crises in IMF-member countries, and is found to out-perform the Poisson and negative binomial models. I am grateful to participants in a seminar at Simon Fraser University, at the MODSIM07 Congress at the University of Canterbury, and a referee for their helpful comments on preliminary versions of this work.