On the Effects of Misspecification Errors in Models with Generated Regressors.

This paper considers the consistency and efficiency of two-step estimators of a structural equation when the auxiliary equation is misspecified. Underspecification generally leads to the two-step estimator and the estimator of the error variance being inconsistent. When the errors of the structural and auxiliary equations are uncorrelated, a two-step estimator based on an overspecified auxiliary model will generally lead to a loss of efficiency compared with the two-step estimator based on a correctly specified auxiliary model. The impact of underspecifying and overspecifying the auxiliary equation used to generate initial parameter estimates for Heckman's two-step estimator of models with sample selection is analyzed. Copyright 1994 by Blackwell Publishing Ltd