On Globally Robust Conndence Intervals for Regression Coeecients

1 Abstract We will develop conndence intervals for linear regression coeecients when the para-metric model is violated by the presence of a fraction of outliers and high leverage data points. Our method will be based on a robustiied bootstrap technique. Unlike the classical bootstrap, our robust bootstrap does not produce unduly heavy tails or extreme re-sampled statistics when the original sample is contaminated. Classical bootstrap has not been widely applied in robust regression due to the heavy computation involved in the calculation of robust estimators. Our method greatly simpliies the calculations making robust inference a real alternative for the practitioner concerned with the validity of the underlying regression model.