An Exact Significance Test for Three-Way Interaction Effects

Abstract

A modification of Fisher's exact test for the 2 x 2 x 2 contingency table is proposed as a test of the null hypothesis of no three-way statistical in teraction among variables, controlling for the two-way or first-order correlations. The test uses a truncated hypergeometric distribution, limited by the bivariate marginal totals of the variables. Possible generalizations to L x M x N tables are discussed. The test is also applicable to the null hypothesis of no difference in the magnitude of correlation in a comparison of two bivariate distributions. Illustrations of each application are provided. One obvious use in cross-cultural or survey research is as a test of the replication of a correlation in different subsamples of a population.