EVALUATION OF BIAS RESULTING FROM GEOGRAPHIC IMPUTATION OF PATIENT RESIDENTIAL ADDRESS: IMPLICATIONS FOR ASSESSMENT OF AMBIENT EXPOSURES IN HEALTH ANALYSES

Linkage of environmental and health data is a critical component of epidemiological analysis. When the available spatial information in health records is insufficient, manual assignment of geographic coordinates at a fine resolution, such as the street address- or Census tract-level, is a resource-intensive process. Geographical imputation is one possible solution, with potential to improve data processing time and minimize ecological and other biases. Few studies have assessed geo-imputation impacts on exposure assessment and exposure-disease associations in an epidemiologic analysis. To do this, we conducted a case-crossover analysis of fine particulate matter (PM_2.5) and respiratory disease after applying a probabilistic geo-imputation procedure. Hospitalization data were partially geo-imputed to Census tracts in 1000 iterations based on population density, age, and racial/ethnic distributions in randomized and non-randomized scenarios. Exposure to ambient PM_2.5 distributed across a 12km grid surface was assigned to individuals in a Geographic Information System based on actual and imputed address coordinates. Epidemiologic metrics included the distribution, sensitivity, and specificity of PM_2.5 exposure, and the ratio measure of association (Hazard Ratio (HR)). Geo-imputed data yielded non-significantly higher mean PM_2.5 exposures and estimates of the exposure-disease association (HRs<1% higher for imputed data), but with overlapping confidence intervals. Results suggest that the specificity of binary exposure indicators following imputation is good (90.5-99.8%), supporting probable non-differential bias in HRs. Sensitivity was lowest for dichotomized concentration values at the high end of the PM distribution (62.4-87.7%). Overall, geo-imputation had minimal impact on the distribution of exposure, the validity of exposure assignment, and ratio effect measures. Given its advantages and nominal influence on epidemiological outputs, geo-imputation is a candidate method to reconcile health data with limited spatial attributes for linkage with ambient PM_2.5. Further evaluations should determine if the results are robust to inclusion of different ambient exposures or those at coarser geographic scales.