Data Transformation in Two-Way Analysis of Variance

Abstract

A procedure for choosing a power transformation so that data from a replicated two-way classification better satisfy the usual analysis of variance assumptions is described. The statistic proposed in this article is compared with the Box and Cox [7] likelihood procedure by empirical sampling. Its 95 percent confidence intervals performed at their nominal level, which was not true for the likelihood method. Point estimates for both techniques were the same on the average, although those based upon the likelihood statistic were somewhat less variable.