A new algorithm for resistant regression
Abstract
The ordinary least squares regression method is nota reliable tool in regression analysis without first diagnosing possible outliers present in the data set. The least median of squares regression technique (Rousseeuw 1984), which is designed to lessen the impact of outlying observations, is presented and some alternatives are given. The output of a Fortran implementation of this regression technique, called PROGRESS (Leroy and Rousseeuw 1984), is illustrated with an example. The results can be interpreted by means of graphical representation of the standardized residuals. It is showed how PROGRESS can be used as a diagnostic tool in regression analysis.