A simulation study on robust alternatives of least squares regression

We applied four methods of linear regression; the least squares, Huber M, least absolute deviations and nonparametric to several distributional assumptions. The same sets of simulated data were used and MSE, MAD and biases of these methods were compared. The least absolute deviations, Huber M and nonparametric regression shown to be more appropriate alternatives to the least squares in heavy tailed distributions while the nonparametric and LAD regression were better choices for skewed data. However, no best method could be suggested in all situations and using more than one method of exploratory data analysis is recommended in practice. © 2007 Asian Network for Scientific Information.