Abstract:
|
It is nearly impossible to directly estimate diabetes incidence rates for most US counties due to the small number of individuals surveyed and the very low fraction of new cases observed in each county. Small area estimation (SAE) methods that use Bayesian hierarchical models help address this problem by borrowing strength from other counties and states. Still, the accuracy of the incidence estimates is limited by the lack of data; some counties have no survey data whatsoever. Using SAE models, county incidence is estimated by drawing from the posterior distribution, which is comprised of prior and likelihood factors. In counties with few or no observations, the posterior draws are heavily influenced by the selected prior distribution. To improve our estimates, we implemented a Bayesian hierarchical SAE model with an informative (power) prior that adds relevant historical information. We used survey and simulated data to examine our proposed model. Compared with the diffuse prior, in terms of relative root mean squared error (RRMSE), the power prior can improve estimation accuracy by 21% when 40% of counties have zero observations, and by 27% when 60% of counties have no data.
|