Abstract
We adapt the estimation method proposed by Gallant and Nychka to binary-choice models. We present Monte Carlo and asymptotic comparisons with the probit estimator and discuss optimization algorithms, choice of starting values, and strategies for choosing the number of parameters used in approximating the density. Seminonparametric estimation is almost as efficient as probit estimation in normal samples and performs better in nonnormal samples. The estimation of a participation model with 3,658 observations and 21 explanatory variables demonstrates the practicability of this approach on personal computers.