On B-Robust Instrumental Variable Estimation of the Linear Model

The aim of this paper is to demonstrate how to obtain robust (with respect to outlying observations) consistent estimates of the linear model when the fundamental orthogonality condition is not fulfilled. With this end in view, we develop two estimation procedures: Two Stage Generalized M (2SGM) and Robust Generalized Method of Moments (RGMM). Both estimators are consistent, asymptotically normally distributed, and B-robust, i.e. their associated influence function is bounded. Our simulation results indicate that the relatively efficient RGMM estimator (in regressions with heteroskedastic and/or autocorrelated errors) provides accurate parameter esrtimates of a panel data model whose explanatory factors are subject to measurement errors, even if a substantial portion of the data is contaminated with aberrant observations. The traditional estimation techniques such as 2SLS and GMM break down when outliers corrupt the data.