Bayesian Superpopulation Models for Population-Based Diabetes Surveillance Using Multi-Site Electronic Health Record Data
Abstract
Electronic health records (EHRs) are increasingly used for disease surveillance. However, data collected in EHRs may not be a representative sample of the target population for inference. Methods for EHR-based surveillance remain in their nascence. In addition to simple weighting approaches, researchers have used multilevel regression and poststratification (MRP), a popular technique for finite population quantity estimation that remains understudied in EHRs. We are motivated by a US-based collaborative to modernize diabetes surveillance using EHR data from different health system centers across the country. To estimate type 1 diabetes (T1D) prevalence in children and adolescents, we propose and evaluate three Bayesian superpopulation models that are distinguished by how they account for two important features of EHR data: unobserved heterogeneity and data sparsity. The three models include an analysis of variance-inspired approach; a model with geographic effects varying by demographic subgroup; and MRP, which we consider as the reference model. We apply the models to estimate prevalence overall and in demographic- and geographic-based subgroups. Using plasmode simulations, we evaluate model estimation performance while preserving the complexity of the EHR data. Our results show that the superpopulation-based prevalence estimates were higher than those observed in the EHR sample alone, with the magnitude of the differences varying by health system center. The bias and coverage probabilities of MRP markedly worsened when the EHR sample size was smaller or the selection mechanism was more complex. Overall, the superpopulation model based on analysis of variance demonstrated the best estimation performance.