Very Robust Regression: Frameworks for Comparisons

We use a smoothly parameterized series of examples that shows, in a systematic way, how the behaviour of algorithms for very robust regression depends on the closeness of the outliers to the main data. An algorithm based on the Forward Search outperforms Least Trimmed Squares and its reweighted version. An empirical measure of the overlap of the two samples structures our investigation of the bias and variance of the robust estimators. We also consider the power of tests for outliers associated with the estimation methods.JRC.G.2-Global security and crisis managemen