Accuracy of commercial geocoding: assessment and implications

BACKGROUND Published studies of geocoding accuracy often focus on a single geographic area, address source or vendor, do not adjust accuracy measures for address characteristics, and do not examine effects of inaccuracy on exposure measures. We addressed these issues in a Women's Health Initiative ancillary study, the Environmental Epidemiology of Arrhythmogenesis in WHI. RESULTS Addresses in 49 U.S. states (n = 3,615) with established coordinates were geocoded by four vendors (A-D). There were important differences among vendors in address match rate (98%; 82%; 81%; 30%), concordance between established and vendor-assigned census tracts (85%; 88%; 87%; 98%) and distance between established and vendor-assigned coordinates (mean rho [meters]: 1809; 748; 704; 228). Mean rho was lowest among street-matched, complete, zip-coded, unedited and urban addresses, and addresses with North American Datum of 1983 or World Geodetic System of 1984 coordinates. In mixed models restricted to vendors with minimally acceptable match rates (A-C) and adjusted for address characteristics, within-address correlation, and among-vendor heteroscedasticity of rho, differences in mean rho were small for street-type matches (280; 268; 275), i.e. likely to bias results relying on them about equally for most applications. In contrast, differences between centroid-type matches were substantial in some vendor contrasts, but not others (5497; 4303; 4210) p(interaction) < 10(-4), i.e. more likely to bias results differently in many applications. The adjusted odds of an address match was higher for vendor A versus C (odds ratio = 66, 95% confidence interval: 47, 93), but not B versus C (OR = 1.1, 95% CI: 0.9, 1.3). That of census tract concordance was no higher for vendor A versus C (OR = 1.0, 95% CI: 0.9, 1.2) or B versus C (OR = 1.1, 95% CI: 0.9, 1.3). Misclassification of a related exposure measure--distance to the nearest highway--increased with mean rho and in the absence of confounding, non-differential misclassification of this distance biased its hypothetical association with coronary heart disease mortality toward the null. CONCLUSION Geocoding error depends on measures used to evaluate it, address characteristics and vendor. Vendor selection presents a trade-off between potential for missing data and error in estimating spatially defined attributes. Informed selection is needed to control the trade-off and adjust analyses for its effects.