Full Hierarchical Bayes integrates over uncertainty in parameters by drawing samples. Practitioners frequently summarize respondent level draws with a simple mean point estimate. This presentation shows how we can use the distribution of a respondent’s posterior draws as a prior, coupled with Empirical Bayes to find point estimates that are most predictive as measured by leave-one-out cross-validation holdout tasks. We also show how we can apply constraints during this Empirical Bayes modeling rather than during HB estimation.
