Assessing Bayesian Semi‐Parametric Log‐Linear Models: An Application to Disclosure Risk Estimation

We propose a method for identifying models with good predictive performance in the family of Bayesian log-linear mixed Dirichlet process random effects count data. Their wide applicability makes assessment model crucial many fields, including disclosure risk estimation, which is focus present work. Rather than assessing on whole contingency table, we target specific objective analysis and two-stage selection procedure aimed at limiting form bias arising selection. Our proposal combines two different criteria: first stage, path search space identified through strongly penalized log-likelihood; second, small number semi-parametric evaluated context-dependent score-based information criterion. Tested variety tables, our proves to be able identify few steps, even presence large tables sampling structural zeros. carefully discuss proposed context literature contextualize illustrative application recent debate statistical limitation. Finally, provide examples further applications research areas.