Nonparametric vs Parametric Binary Choice Models: To Drink or not to Drink (Tap Water)?

Recent developments in the nonparametric estimation of conditional probability distribution functions (PDFs) offers practitioners a flexible framework for estimation and inference. The modelling of conditional PDFs can be extremely useful for a range of tasks including direct quantile estimation and prediction of consumer choice, by way of example. In this paper assess the potential of this nonparametric estimator for improved modelling of consumer choice. We model a dataset in which the outcome is a binary consumer choice while the covariates consist of both continuous and categorical (discrete) variables. We assess the relative performance of the nonparametric estimator and the parametric Probit specification that dominates in applied settings. We compare these estimators using a variety of measures and also assess their performance on independent data drawn from the same underlying distribution and test for significant differences. Finally, we demonstrate that the nonparametric estimator reveals certain features present in the data that lie undetected by the parametric Probit model.