On Testing Sample Selection Bias under the Multicollinearity Problem

Abstract

This paper examines and compares the finite sample performance of the existing tests for sample selection bias, especially under the multi-collinearity problem pointed out by Nawata (1993). The results show that under such multicollinearity problem, (i) the t-test for sample selection bias based on the Heckman and Greene variance estimator can be unreliable; (ii) the standard t-test (Heckman 1979) and the asymptotically efficient Lagrange multiplier test (Melino 1982) have correct size but very little power; (iii) however, the likelihood ratio test following the maximum likelihood estimation remains powerful.