A power study of a rank transform for the two‐way classification model when interaction may be present

Abstract The power of the usual parametric analysis of variance for the two‐way model when interaction may be present is compared with the power of a test based on performing the usual parametric analysis on ranks, i.e., a rank transform is used. While the mathematical theory has not yet been developed for such an approach, this study shows the value of developing the theory for such an approach to the analysis of experimental designs, and indicates that this approach is probably a valid procedure. Simulation results are présentés for three underlying populations: normal, contaminated normal, and exponential. These results show the rank transform approach has a null distribution similar to that of the F ‐statistic computed on the raw data in all the cases considered. Though little, if any, loss of power occurs from using the rank transform approach, a significant increase in power is apparent when the underlying population is contaminated normal.