ABSTRACT BOOTSTRAP CONFIDENCE INTERVALS IN LINEAR MODELS

Abstract

A bootstrap method for generating confidence intervals in linear models is suggested. The method is motivated by an abstract nonobservable bootstrap sample of true residuals leading to an observable final result. This means that the only error in the method is the pure bootstrap error obtained by replacing the true residual distribution by the empirical one. It is shown that the method is valid, having the same asymptotic conditional distribution as the ordinary bootstrap method. Simulations indicate clearly that the abstract bootstrap method works better than the ordinary bootstrap method for small samples.