Finite sample behavior of two step estimators in selection models.

Abstract

The problem of speci cation errors in sample selection models has received considerable attention both theoretically and empirically. However, very few is known about the nite sample behavior of two step estimators. In this paper we investigate by simulations both bias and nite sample distribution of these estimators when ignoring heteroskedasticity in the sample selection mechanism. It turns out that under conditions traditionally faced by practitioners, the misspeci ed parametric two step estimator (Heckman, 1979) performs better, in nite sample sizes, than the robust semiparametric one (Ahn and Powell, 1993). Moreover, under very general conditions, we show that the asymptotic bias of the parametric two step estimator is linear in the covariance between the sample selection and the participation equation.

Investigación empírica de métodos de estimación paramétricos y semiparamétricos de modelos de selección muestral

“…Hurd (1979) shows the consequences of omitted heteroskedasticity and Goldberger (1983) describes the effects of non-normality. Newey (1999) analyzes the impact of misspecification of distribution in two-step estimators, Fernández, Rodríguez-Póo and Villanua (2002) show the impact of ignoring heteroskedasticity and Nawata and Nagase (1996) compare the performance of the two estimators by a simulation study.…”

“…A deeper analysis could be performed, but our interest here is to note that these significant changes can be due to misspecification errors in the parametric equation. Of course, as already remarked in a simulation study in Nawata and Nagase (1996) and Fernández, Rodríguez-Póo and Villanua (2002), these errors may seriously affect the properties of the estimators of the wage equation.…”