We develop new robust discrete choice tools to learn about the average willingness to pay for a price subsidy and its effects on demand given exogenous, discrete variation in prices. Our starting point is a nonparametric, nonseparable model of choice. We exploit the insight that our welfare parameters in this model can be expressed as functions of demand for the different alternatives. However, while the variation in the data reveals the value of demand at the observed prices, the parameters generally depend on its values beyond these prices. We show how to sharply characterize what we can learn when demand is specified to be entirely nonparametric or to be parameterized in a flexible manner, both of which imply that the parameters are not necessarily point identified. We use our tools to analyze the welfare effects of price subsidies provided by school vouchers in the DC Opportunity Scholarship Program. We find that the provision of the status quo voucher and a wide range of counterfactual vouchers of different amounts can have positive and potentially large benefits net of costs. The positive effect can be explained by the popularity of low-tuition schools in the program
removing them from the program can result in a negative net benefit. We also find that various standard logit specifications, in comparison, limit attention to demand functions with low demand for the voucher, which do not capture the large magnitudes of benefits credibly consistent with the data.