Concluding remarks
We have proposed a new model of finite mixture of bivariate Poisson regressions. The idea is that the data consist of subpopulations of different regression structures. A potential use for such a model is for examining the clustering of observations, taking into consideration the effect of certain covariates while also taking into account the dependence between the response variables. The model corrects for the zero inflation and overdispersion present in the real automobile insurance data set used in the application. The model can also be used to model negative correlation.
The AIC reported here indicates that the 2-finite mixture of bivariate Poisson regression with covariates in the mixing proportion is the best model for describing the data set. This model has a number of interesting features:
Firstly, being a finite mixture it has a nice cluster interpretation. Secondly it is semi-parametric, in the sense that we avoid a parametric assumption on the mixing distribution. Thirdly, it is flexible enough allowing for overdispersion, zero inflation but also for negative correlation. Finally, our EM algorithm makes its fit straightforward while for other models their fitting procedure is not easy for large samples and many covariates as in our case. Such an example is the bivariate Poisson lognormal model which needs the evaluation of a bivariate integral for each observation, and of course each iteration during maximization which leads to a lot of computational problems. Concluding the proposed model provides a very interesting structure while being computationally feasible.
The problem of overdispersion arises because of the presence of unobserved heterogeneity in many real data sets. In insurance data sets, an insurance company cannot keep track of the many differences between policyholders. However, the model proposed in this paper accounts for unobserved heterogeneity by choosing a finite number of subpopulations. We assume the existence of two types of policyholder described separately according to each component in the mixture.
The phenomenon of excess of zeros may also be seen as a consequence of this unobserved heterogeneity. The model proposed here, as a finite mixture of bivariate Poisson regression model, embraces the zero-inflated bivariate Poisson regression model as a special case. The main difference with zero-inflated models is that the two-component mixture model reported here allows mixing with respect to both zeros and positives. This interpretation is more flexible and holds better in our application. The group separation is characterized by low mean (policyholders considered as a ‘‘good’’ drivers) and high mean (policyholders considered as a ‘‘bad’’ drivers).
Moreover, as it seems that the data set may have been generated from two distinct subpopulations, the model allows for a net interpretation of each cluster separately. Note that different regression coefficients can be used to account for the ‘‘observed’’ heterogeneity within each population.
Finally, we would like to mention various ways in which this paper might be extended. Although in the present paper we limit our analysis to the bivariate case, it could be extended to include larger dimensions. Following the general model presented by Karlis and Meligkotsidou (2007), covariates might be added and this finite mixture of multivariate Poisson regressions could be used to cluster high-dimensional data. A particularly interesting case occurs if we consider there to be no dependence within a cluster, whereby within-cluster independent Poisson regressions are considered.