BackgroundEcologic health studies often rely on outcomes from health service utilization data that are limited by relatively coarse spatial resolutions and missing geographic information, particularly neighborhood level identifiers. When fine-scale geographic data are missing, the ramifications and strategies for addressing them are not well researched or developed. This study illustrates a novel spatio-temporal framework that combines a geographic identifier assignment (i.e., geographic imputation) algorithm with predictive Bayesian variable selection to identify neighborhood factors associated with disparities in emergency department (ED) visits for asthma.MethodsED visit records with missing fine-scale spatial identifiers (~ 20%) were geocoded using information from known, coarser, misaligned spatial units using an innovative geographic identifier assignment algorithm. We then employed systematic variable selection in a spatio-temporal Bayesian hierarchical model (BHM) predictive framework within the NIMBLE package in R. Our novel methodology is illustrated in an ecologic case study aimed at identifying neighborhood-level predictors of asthma ED visits in South Carolina, United States, from 1999 to 2015. The health outcome was annual ED visit counts in small areas (i.e., census tracts) with primary diagnoses of asthma (ICD9 codes 493.XX) among children ages 5 to 19 years.ResultsWe maintained 96% of ED visit records for this analysis. When the algorithm used areal proportions as probabilities for assignment, which addressed differential missingness of census tract identifiers in rural areas, variable selection consistently identified significant neighborhood-level predictors of asthma ED visit risk including pharmacy proximity, average household size, and carbon monoxide interactions. Contrasted with common solutions of removing geographically incomplete records or scaling up analyses, our methodology identified critical differences in parameters estimated, predictors selected, and inferences. We posit that the differences were attributable to improved data resolution, resulting in greater power and less bias. Importantly, without this methodology, we would have inaccurately identified predictors of risk for asthma ED visits, particularly in rural areas.ConclusionsOur approach innovatively addressed several issues in ecologic health studies, including missing small-area geographic information, multiple correlated neighborhood covariates, and multiscale unmeasured confounding factors. Our methodology could be widely applied to other small-area studies, useful to a range of researchers throughout the world.