Modelling Household Expenditure on Health Care in Greece

Health expenditure data are known to be afflicted by restricted range, zero values, skewness and kurtosis. Several methods for modelling such data have been suggested in the literature to cope with these problems. This paper compares the performance of several alternative estimators, including two-part models and generalized linear models. The dependent variable is household, not individual, expenditure on health care in Greece, a country where out-of-pocket health expenditure is higher than anywhere else in the European Union, whether as a proportion of GDP, as a share of all health spending, or in per capita terms. To facilitate comparison of model performance, household health expenditure is examined in two different specifications; expenditure on all health care (where zero values are rare) and expenditure on hospital services alone (where zero values are common). Three of the estimators performed almost equally well in terms of mean square error and mean absolute prediction error; a modified two-part model with non-linear least squares in the second part, a constant-variance generalized linear model, and a varianceproportional- to-mean generalized linear model. The findings suggest that no estimator is best under all circumstances, while most alternative estimators produce similar results. The paper concludes by discussing implications for further research.