A robust instrumental-variables estimator

摘要

The classical instrumental-variables estimator is extremely sensitivento the presence of outliers in the sample. This is a concern because outliers cannstrongly distort the estimated effect of a given regressor on the dependent vari-nable. Although outlier diagnostics exist, they frequently fail to detect atypicalnobservations because they are themselves based on nonrobust (to outliers) es-ntimators. Furthermore, they do not take into account the combined influencenof outliers in the first and second stages of the instrumental-variables estima-ntor. In this article, we present a robust instrumental-variables estimator, ini-ntially proposed by Cohen Freue, Ortiz-Molina, and Zamar (2011, Working paper:nhttp://www.stat.ubc.ca?nruben/website/cv/cohen-zamar.pdf ), that we have pro-ngrammed in Stata and made available via the robivreg command. We have im-nproved on their estimator in two different ways. First, we use a weighting schementhat makes our estimator more efficient and allows the computations of the usualnidentification and overidentifying restrictions tests. Second, we implement a gen-neralized Hausman test for the presence of outliers.