Misspecification and Heterogeneity in Single-Index, Binary Choice Models

We propose a nonparametric approach for estimating single-index, binary-choice models when parametric models such as Probit and Logit are potentially misspecified. The new approach involves two steps: first, we estimate index coefficients using sliced inverse regression without specifying a parametric probability function a priori; second, we estimate the unknown probability function using kernel regression of the binary choice variable on the single index estimated in the first step. The estimated probability functions for different demographic groups indicate that the conventional dummy variable approach cannot fully capture heterogeneous effects across groups. Using both simulated and labor market data, we demonstrate the merits of this new approach in solving model misspecification and heterogeneity problems.