Parametric empirical Bayes estimates of disease prevalence using stratified samples from community populations

Studies of chronic diseases in a community setting often employ stratified sample designs to enable the study to attain multiple research goals at a reasonable cost. One important goal is estimation of disease prevalence in the whole community and in important subgroups. Some adjustment for the sample design is necessary; if the design has many strata with very disparate sampling fractions, simply upweighting observed stratum prevalences may lead to unstable estimators. We propose a parametric empirical Bayes estimator in the spirit of the work of Efron and Morris, and we compare it to the direct upweighted estimator and a regression-smoothed estimator. Simulation studies in realistic settings suggest that the new estimator performs best, giving estimates with low bias and good precision under a variety of models.