The expression “robust regression” denotes a set of estimation techniques that are less sensitive than ordinary least squares (OLS) to the effect of possible influential observations. In its various forms, robust regression has a well established tradition in statistics (see, e.g., Maronna et al., 2006). However, apart from median regression and quantile regression in general (Koenker and Bassett, 1978), robust regression was slow to gain popularity in applied econometrics.
Over the past decade, however, a form of robust regression based on Huber’s (1964)MM-estimator was made available in popular software packages1 and has been increasingly used in leading research publications and in industry.2 Unfortunately, most practitioners who have used this estimator seem to be unaware of the fact that its properties depend on strong assumptions about the distribution of the errors.
We show that the specific MM-estimator that has become popular in applied econometrics, henceforth termed MbwMbw-estimator,3 is inconsistent for the parameters of the conditional mean when the errors are skewed and heteroskedastic, and provide simulation evidence on its performance.