Dynamic Estimation of Health Expenditure: A new approach for simulating individual expenditure

This study compares estimates of outpatient expenditure computed with different models. Our aim is to predict annual health expenditures. We use a French panel dataset over a six year period (2000-2006) for 7112 individuals. Our article is based on the estimations of five different models. The first model is a simple two part model estimated in cross section. The other models (models 2 to 5) are estimated with selection models (or generalized tobit models). Model 2 is a basic sample selection model in cross section. Model 3 is similar to model 2, but takes into account the panel dimension. It includes constant unobserved heterogeneity to deal with state dependency. Model 4 is a dynamic sample selection model (with lagged adjustement), while in model 5, we take into account the possible heteroskedasticity of residuals in the dynamic model. We find that all the models have the same properties in the cross section dimension (distribution, probability of health care use by gender and age, health expenditure by gender and age) but model 5 gives better results reflecting the temporal correlation with health expenditure. Indeed, the retransformation of predicted log transformed expenditures in homoscedastic models (models 1 to 4) generates very poor temporal correlation for " heavy consumers ", although the data show the contrary. Incorporation of heteroskedasticity gives better results in terms of temporal correlation.