Analysis and Comparison of Six Robust Regression Techniques

Abstract

The ordinary least squares regression can be misleading when there are outliers, heteroscedasticity and non-normality. Problems with ordinary least squares are briefly explained and six robust regression techniques, Theil-Sen, least median of squares, least trimmed squares, least absolute value, least trimmed absolute value and M-regression, that are not affected by these common problems are investigated and compared in terms of actual significance level and relative efficiency over ordinary least squares. Results are discussed and some recommendations are given.