In this paper we propose a new robust estimator in the context of two-stage estimation methods directed towards the correction of endogeneity problems in linear models. Our estimator is a combination of Huber estimators for each of the two stages, with scale corrections implemented using preliminary median absolute deviation estimators. In this way we obtain a two-stage estimation procedure that is an interesting compromise between concerns of simplicity of calculation, robustness and efficiency. This method compares well with other possible estimators such as two-stage least-squares (2SLS) and two-stage least-absolute-deviations (2SLAD), asymptotically and in finite samples. It is notably interesting to deal with contamination affecting more heavily the distribution tails than a few outliers and not losing as much efficiency as other popular estimators in that case, e.g. under normality. An additional originality resides in the fact that we deal with random regressors and asymmetric errors, which is not often the case in the literature on robust estimators.
Bibliographical noteFunding Information:
This research has been supported under the British Academy Award No. 30,649. Tae-Hwan Kim is grateful to the College of Business and Economics of Yonsei University for financial support under the project code 2005-1-0244 and the second author is grateful for the financial support by Spanish Ministry of Sciences and Technology. Project No. BEC2002-03097 and by the Instituto Valenciano de Investigaciones Economicas. We thank participants in a seminar in Alicante and two referees for their useful comments. Usual disclaimers apply.
All Science Journal Classification (ASJC) codes
- Statistics and Probability
- Statistics, Probability and Uncertainty
- Applied Mathematics