A performance-based assessment of robust regression methods

A number of robust regression techniques have been proposed as alternatives to least squares when outliers are present in data. No single robust technique is universally accepted as the best. Techniques exhibit strengths and weaknesses depending on the percentage of outliers present, and the location of the outliers in the regressor space. Also, robust techniques perform quite differently when no outliers are present. The purpose of this paper is to evaluate existing and proposed robust methods relative to their performance on a comprehensive group of datasets with and without outliers. Monte Carlo simulation is used to generate datasets with varying outlier characteristics. The coefficient estimates generated by the robust methods are compared to the true coefficient values. Techniques are then ranked based on their performance. The best performing methods consist of an existing method and a method proposed in this paper.