Unreplicated fractional factorials, analysis with the half-normal plot and randomization of the run order

Abstract

There is an ongoing discussion whether it is wise to randomize the run order of a factorial experiment if there is concern about a possible time trend in the experiment. It can be argued that a randomized order is not very effective because the trend inflates the error. Some authors even criticize that a randomized order will normally not be orthogonal to trend, they claim that therefore there will be bias under the randomized order. On the other hand, a systematic order will only be useful if the true trend is behaving as is predicted by the model. The present paper investigates the properties of different run order strategies in a simulation study with unreplicated factorial designs. We check to which extend the presence of a time trend might inflate the probability of false rejection of a true nullhypothesis, and we compare the power of significance tests based on the half-normal plot under the various run order concepts.