Abstract: The Baysian analysis of loglinear models requires the evaluation of high-dimensional integrals. Such an evaluation is frequently computationally prohibitive even with modern computers. We provide a parameterization of the loglinear model which renders these integrations amenable to the numerical methods of adaptive important sampling. This approach is applied in the analysis of two-way contingency tables using Goodman's RC model. We base the analysis on the full posterior distribution for the loglinear model and obtain the posterior distribution of a goodness-of-fit measure for Goodman's RC model.
Key words and phrases: Adaptive importance sampling, categorical data, Goodman's RC model, loglinear models, singular value decomposition.