Generalized Semiparametric Binary Prediction

Jeff Racine

This paper proposes a semiparametric approach to the estimation of ‘generalized’ binary choice models. A ‘generalized’ binary choice model is one with separate indices for each conditioning variable which constitutes a generalization of the standard single-index approach typically employed in applied work. The choice probability distribution is therefore a joint distribution across these indices as opposed to the typical univariate distribution on a scalar index commonly found in applied work. Interest lies in estimating choice probabilities and the gradient of choice probabilities with respect to the conditioning information, and these are estimated nonparametrically using the method of kernels. A data-driven cross-validatory method for bandwidth selection and index-parameter estimation is proposed for maximization of the nonparametric likelihood function. The functional form of the indices enters this nonparametric likelihood function thereby permitting data-driven determination of the index functions in addition to the shape of the joint cumulative distribution
function itself. Applications are considered.