Geocoding and selection bias in public health research using geographic information systems

Objective: Describe selection bias in GIS analyses with unrepresentative data owing to missing geocodes. Design: Spatial analysis of prostate cancer incidence among whites and African Americans in Virginia, 1990-1999, using the Virginia Cancer Registry. Outcome measures: Statistical tests for clustering were performed and mapped. The patterns of missing census tract identifiers for the cases were examined by general linear regression models and stratified by rural status. Results: All cases in the VCR are located in counties, and 26,338 (74%) of these cases were successfully geocoded to census tracts. Statistical testing for global clustering was highly significant for the entire study period and for 1990-94 and 1995-99 separately. For each time period, however, cluster patterns appeared markedly different, depending upon whether one used the cases located in the county or those geocoded to the census tract. Multivariate regression analysis showed that, in the most rural counties (where the missing data were concentrated), the percent of a county's population over age 64 and with less than a high school education were both independently associated with a higher percent of missing geocodes (p = 0.016 and p = 0.003, respectively). Discussion: We found statistically significant pattern differences resulting from spatially non-random differences in geocoding completeness across Virginia. Appropriate interpretation of maps, therefore, requires an understanding of this phenomenon, which we call “cartographic confounding.”