Distribution-free and link-free estimation for the sample selection model (Q1110950)

The prevalent estimation methods for the sample selection model rely heavily on parametric assumptions and are sensitive to departures from the underlying parametric assumptions. We propose an alternative estimation method, the corrected maximum likelihood estimate, which is consistent for the slope vector in the outcome equation up to a multiplicative scalar, even though the parametric model on which the estimate is based might be misspecified. As an important corollary, it follows from our result that \textit{R. J. Olsen}'s [A least squares correction for selectivity bias. Econometrica 48, 1815-1820 (1980)] corrected ordinary least squares estimate is consistent if the outcome equation is linear, without requiring Olsen's assumptions on the joint error distribution.