Testing the fit of relational models

Abstract: Relational models generalize log-linear models to arbitrary discrete sample spaces by specifying effects associated with any subsets of their cells. A relational model may include an overall effect, pertaining to every cell after a reparameterization, and in this case, the properties of the maximum likelihood estimates (MLEs) are analogous to those computed under traditional log-linear models, and the goodness-of-fit tests are also the same. If an overall effect is not present in any reparameterization, the properties of the MLEs are considerably different, and the Poisson and multinomial MLEs are not equivalent. In the Poisson case, if the overall effect is not present, the observed total is not always preserved by the MLE, and thus, the likelihood ratio statistic is not identical with twice the Kullback-Leibler divergence. However, as demonstrated, its general form may be obtained from the Bregman divergence. The asymptotic equivalence of the Pearson chi-squared and likelihood ratio statistics holds, but the generality considered here requires extended proofs.