|
Abstract:
|
Count data occur naturally in a number of disciplines ranging from economics and the social sciences to finance as well as medical sciences. Most count data are plagued with over-dispersion and excess zeros making it difficult to model them with vanilla linear models. Different models have been proposed to capture this peculiarity in count data viz.: Classical models such as the generalized Poisson regression model and the negative binomial regression model have been used to model dispersed count data. Hurdle and zero-inflated models are also said to be able to capture over-dispersion and excess zeros in count data. In this paper, we compare the performance of Poisson and Negative Binomial hurdle models, zero-inflated Poisson and Negative Binomial models, classical Poisson and Negative Binomial regression models as well as the zero-inflated compound Poisson generalized linear models to modelling frequency of auto insurance claims in a typical emerging market. The model parameters are estimated using the method of maximum likelihood. The models' performances are compared based on their information criteria (AIC and BIC) and Gini index.
|
