A new look at experimental design robustness

Abstract

SUMMARY A common practical experimental philosophy when the underlying response model is unknown is initially to consider a first order model. A total robustness approach requires that the experimenter account for model underspecification due to possible curvature or interactions. It is reasonable that the experimental design possess protection against the selection of an oversimplified model. However, detection of possible underspecification is also important. This paper introduces a new criterion of experimental design robustness and provides the framework for an alternative class of experimental designs that utilize this criterion. This alternative class of designs is called ‘rotated designs’.