The Effects of Discarding Inliars when Binomial Data are Subject to Classification Errors

Abstract

This article deals with the biasing effects of errors of measurement on the estimation of a population proportion or on the degree of association between two dichotomous variables. Consider a dichotomy that is formed by making a cut on an ordered scale, where errors of classification occur with greatest frequency in the neighborhood of the cut. It is demonstrated through numerical investigation that it is advantageous, both in terms of bias and of mean square error, to discard those observations which are most subject to error. Guidelines are provided for choosing an optimum exclusion interval