Towards an optimum test for non-additivity in Tukey's model (Q2640270)

The existence of an optimal test for nonadditivity under Tukey's model in a two-way layout with one observation per cell is studied. Tukey's F-test for the significance of the interaction is really meaningful only when the interaction is of product form. The same form is assumed in the paper. The object is using the tool of invariance, to try to state some optimum properties of the test. A group leaving the testing problem invariant is described. Then, based on the expansion of the probability ratio of the nonnull to null distributions of a maximal invariant, the local conditional power (given a complete sufficient statistic under the null hypothesis) of any invariant similar test is computed. Although invariant, Tukey's test turns out not to be the optimal invariant similar one, even locally. Under certain conditions two exact F-tests which are locally better than Tukey's test, exist and have been proposed. Roughly speaking, these conditions concern the magnitude of the standardized main effects. The tests proposed are preferable when these magnitudes are small. Otherwise a risk of loss in power exists if these tests are used instead of Tukey's test. Local unbiasedness of Tukey's test as well as of the two new tests has been shown. Although the problem of derivation of an optimal test remains open, this paper is a deep investigation of its intricacies.