Reducing bias without prejudicing sign

Abstract

Jackknife and bootstrap bias corrections are based on a differencing argument which does not necessarily respect the sign of the true parameter value. Depending on sampling variability they can over-correct, producing a final estimator that is negative when one knows on physical grounds that it should be positive. To overcome this problem we suggest a simple, alternative bootstrap approach, based on biased-bootstrap methods. Our technique has similar properties to the standard uniform-bootstrap method in cases where the latter does not endanger sign, but it respects sign in a canonical way when the standard method disregards it. Show more