Data Analysis with Transformations

Abstract

Abstract : The lambda transformation of Box and Cox was applied to 56 sets of data from five areas of biological research to determine the optimum transformation for a given type of data. Data from one area of research (mask) were improved by a transformation of lambda 1 = 0. This corresponds to a log transform and has been applied routinely to such data. Data from other areas of research were less affected by the transform. For these non-mask data, significance of main effects was not changed, interactions were generally unaffected, and variance homogeneity was achieved in only two of the six possible cases where the lambda transform was compared with no transform. The study corroborates the analysis that has been performed regularly on data from one area of research but indicates that an analysis on untransformed data would generally be as meaningful for the other four areas of research examined. Where it was helpful, the main influence of the transform was in stabilizing variance. For the small (eight) number of cases examined no real improvement in additivity was noted.