Nonparametric regression for binary dependent variables

revised version of Discussion paper 2001-12#### Finite-sample properties of nonparametric regression for binary dependent variables are analyzed. Nonparametric regression is generally considered as highly variable in small samples when the number of regressors is large. In binary choice models, however, it may be more reliable since its variance is bounded. The precision in estimating conditional means as well as marginal effects is investigated in settings with many explanatory variables (14 regressors) and small sample sizes (250 or 500 observations). The Klein Spady estimator, Nadaraya-Watson regression and local linear regression often perform poorly. Local logit regression, on the other hand, is 10 to 70% more precise than parametric regression. In an application to female labour supply, local logit finds heterogeneity in the effects of children on employment that is not detected by parametric nor semiparametric estimation. Download Discussion Paper: (pdf, 921 kb) former title: Applied higher-dimensional nonparametric regression