On the Partitioning of Goodness-of-Fit Statistics for Multivariate Categorical Response Models