A generic approach to nonparametric function estimation with mixed data

Most nonparametric function estimators can only handle continuous data. We show that making discrete variables continuous by adding noise is justified under suitable conditions on the noise distribution. This principle is widely applicable, including density and regression function estimation.