Accuracy and repeatability of commercial geocoding

The authors estimated accuracy and repeatability of commercial geocoding to guide vendor selection in the Life Course Socioeconomic Status, Social Context and Cardiovascular Disease study (2001-2002). They submitted 1,032 participant addresses (97% in Maryland, Minnesota, Mississippi, or North Carolina) to vendor A twice over 9 months and measured repeatability as agreement between levels of address matching, discordance (%) between statistical tabulation areas, and median distance (d, in meters) and bearing (θ, in degrees) between coordinates assigned on each occasion (H_o:Σ _{i = 1 → n}[θ_i/n] = 180°). They also submitted 75 addresses of nearby air pollution monitors (77% urban/suburban; 69% residential/commercial) to vendors A and B and then measured accuracy by comparing vendor- and US Environmental Protection Agency (EPA)-assigned geocodes using the above measures. Repeatability of geocodes assigned by vendor A was high (kappa = 0.90; census block group discordance = 5%; d < 1 m; θ = 177°). The match rate for EPA monitor addresses was higher for vendor B versus A (88% vs. 76%), but discordance at census block group, tract, and county levels also was, respectively, 1.4-, 1.9-, and 5.0-fold higher for vendor B. Moreover, coordinates assigned by vendor B were further from those assigned by the EPA (d = 212 m vs. 149 m; θ = 131° vs. 171°). These findings suggest that match rates, repeatability, and accuracy should be used to guide vendor selection.