Treatment of Non-normal Responses for Designed Experiments

Abstract

Many experimental designs, most especially industrial designs produce non-normal response variables. The Least Squares method of modeling may therefore not produce efficient estimates. Models are built using B-technique and Box-Cos methods of data transformation or by using GLM to overcome the non-normality nature of the data. In this paper, these techniques are compared using histograms of the estimated mean responses and by examining the length of the confidence interval about the mean responses. It is observed that B-technique is the best method of data transformation but GLM provides an excellent alternative if the experimental design points are nor replicated.