Can tests for treatment group equality be improved?: The bootstrap and trimmed means conjecture

Abstract

Westfall & Young's (1993) results suggest that by combining bootstrap methods with methods based on trimmed means researchers can obtain better Type I error control when testing for treatment group equality. Thus, the methods due to Alexander & Govern (1994), Box (1954), and Welch (1951) were compared under conditions of non‐normality and variance heterogeneity when group sizes were unequal, conditions known to produce liberal rates of error. We found that, typically, in one‐way designs better Type I error control can indeed be obtained when the bootstrap method is used in conjunction with test statistics based on trimmed means.