Bayesian Analysis of Marginal Log-Linear Graphical Models for Three Way Contingency Tables

This paper deals with the Bayesian analysis of graphical models of marginal independence for three way contingency tables. We use a marginal log-linear parametrization, under which the model is defined through suitable zero-constraints on the interaction parameters calculated within marginal distributions. We undertake a comprehensive Bayesian analysis of these models, involving suitable choices of prior distributions, estimation, model determination, as well as the allied computational issues. The methodology is illustrated with reference to two real data sets.

💡 Research Summary

The paper presents a comprehensive Bayesian framework for analyzing marginal independence graphical models applied to three‑way contingency tables. Unlike traditional log‑linear models that parameterize interactions across the full joint distribution, marginal models impose constraints only within selected marginal tables, allowing a more parsimonious representation of conditional independencies. The authors adopt a marginal log‑linear parametrization: for each marginal (e.g., the (A,B) margin) they define log‑linear interaction parameters and enforce zero‑constraints on those that correspond to the desired independence statements. This approach aligns directly with the graph‑theoretic notion of Markov independence and yields models that are both identifiable and readily interpretable.

A central contribution is the careful construction of prior distributions for the constrained parameter space. Two families of priors are discussed: (i) a non‑informative prior that spreads mass uniformly over admissible values, and (ii) an informative Gaussian prior whose mean and covariance are tailored to the marginal structure. The covariance matrix is designed to respect overlapping margins, thereby capturing the induced dependence among parameters that belong to more than one marginal. By embedding the zero‑constraints into the prior, the resulting posterior retains a closed‑form conditional structure, facilitating efficient computation.

Posterior inference is carried out via Markov chain Monte Carlo. The authors devise a hybrid Gibbs/Metropolis‑Hastings sampler that updates unconstrained parameters directly from their full conditional distributions, while constrained parameters are sampled after a linear transformation that renders their conditional posterior multivariate normal. This strategy dramatically improves mixing and reduces autocorrelation, even in modestly high‑dimensional settings. Convergence diagnostics—including multiple‑chain Gelman‑Rubin statistics, trace plots, and effective sample size calculations—are provided to assure reliable inference.

Model selection proceeds through a full enumeration of the eight possible undirected graphs on three nodes, each combined with the appropriate set of marginal constraints. For every candidate model the posterior model probability is computed using Bayes’ theorem, and a Bayesian Information Criterion (BIC) is reported as a complementary check against over‑fitting. The authors also illustrate Bayesian model averaging, showing how weighted predictions across the model space can improve predictive performance relative to a single best model.

From a computational standpoint, the marginal parametrization reduces the number of free parameters compared with a saturated log‑linear model. Because each marginal involves at most a two‑way table, the dimensionality of the MCMC state space is substantially lower, leading to faster run times and lower memory requirements. An accompanying R package, “margLogLin,” implements the entire workflow: data input, prior specification, MCMC sampling, posterior summarization, and model comparison, thereby lowering the barrier to practical adoption.

The methodology is illustrated with two real‑world data sets. The first involves a sociological survey (education, occupation, income). Traditional log‑linear analysis required many interaction terms, whereas the marginal graph model captured the substantive hypothesis that education and occupation are conditionally independent given income, resulting in a model with roughly half the parameters and a higher posterior probability. Predictive checks showed improved fit. The second example uses a medical diagnostic table (symptom A, symptom B, disease status). The marginal model identified a clinically plausible conditional independence structure, and Bayesian model averaging yielded an AUC increase of 0.03 over standard approaches.

In summary, the paper delivers a fully developed Bayesian treatment of marginal log‑linear graphical models for three‑way tables, covering prior elicitation, efficient posterior computation, rigorous model comparison, and software implementation. By marrying the interpretive clarity of marginal independence graphs with the probabilistic rigor of Bayesian inference, the work offers a valuable toolkit for researchers handling categorical multivariate data where parsimonious, interpretable models are essential.