Separating selection and moral hazard effects in Swiss health insurance - a nonparametric bounds analysis
- Presenter:
Chair: James Burgess; Discussant: TBA Tue June 6, 2006 8:00-9:30 Room 121
There is an ongoing debate in health economics whether cost sharing in health insurance changes the behaviour of patients. The well documented negative correlation between the degree of cost sharing and health care costs can be attributed to changes in behaviour (reduction of moral hazard) or selection (more healthy people select higher cost sharing). Previous empirical evidence for Switzerland, which was obtained based on strong identification assumptions using non-experimental data, is mixed. In this paper we employ nonparametric bounding techniques to minimise statistical assumptions. This methodology yields the minimum and maximum value a treatment effect can take given the assumptions made. If nothing can be assumed about the data generating process we obtain the so-called no assumption (or worst case) bounds. These bounds can be tightened if we are willing to impose additional assumptions.
Treatment is ordered in our application. We consider three possible treatment states: low deductible, medium deductible and high deductible. These deductible levels are chosen by the individuals at the start of their health insurance contract. The outcome variable is the probability of going to the doctor. Our empirical analysis shows that we need some assumptions to tighten the bounds; the no-assumption bounds are too wide to be useful. We consider two further assumptions: treatment response is mean independent of an instrument, and treatment response is monotone (i.e. we assume the sign of the treatment effect to be known). The mean independence assumption is also weakened by assuming that mean response varies weakly monotonically with the instrument. Under the first two assumptions we estimate bounds for the treatment effect of the high deductible compared to the low deductible that are below zero. Hence given these two assumptions we conclude that there is a negative treatment effect (i.e. there is a change in behaviour) of at least -0.08. Given that the observed difference is -0.24 at least one third of the difference in the probabilities of going to a doctor can be attributed to a reduced moral hazard effect. Weakening the mean independence assumption reduces the treatment effect to -.04, but it remains different from zero. This finding is in contrast to previous empirical analyses based on the same data.
