A Quasi-experimental Comparison of Econometric Models for Health Care Expenditures

Individual health care expenditures have complex non-normal distributions with severe positive skewness and leptokurtosis. These features present severe challenges to reliable modeling of expenditures for prediction purposes. We compare a variety of methods using quasi-experimental techniques. Our quasi-experiments combine the distributional realism of actual data on health care expenditures with the reliability of Monte Carlo experimental results. We find that models based on Gamma densities predict substantially better than models based on linear regression with and without transformation of the dependent variable. Models based on finite mixtures of Gamma densities show further improvement in predictive properties.