This paper presents model specification checking procedures for count data regression models which are consistent in the direction of nonparametric alternatives. The discussion is motivated in the context of a model of demand for health care in Spain. The parameters of the regression model are estimated by maximum likelihood based on Poisson and Negative Binomial specifications as well as by ordinary least squares and semiparametric generalized least squares. However, our interest is not only centered on the estimation ofthe regression parameters, but also the conditional probabilities of counts. Therefore, the specification of the conditional distribution function of counts is the main focus of attention. A useful preliminary diagnosis tool consists of comparing the conditional probabilities estimates by nonparametric regression and by maximum likelihood methods based on alternative models. We present formal specification procedures based on new developed testing methods for regression model checking. The test statistics are based on marked empirical processes which are not distribution free, but their critical values are well approximated by bootstrap. Such tests are valid for testing the functional form of the conditional mean and conditional probabilities resulting from alternative distributional specifications. In our health care demand model, the linear exponential regression model with a Negative Binomial seems to be appropiate for this data set.
展开▼
