Inferring Unknown Race in Central Cancer Registries

Completeness of race information is a criterion for data certification among United States central cancer registries. This paper presents a method for reducing unknown race information by as much as 75%, with 95% accuracy, race-specific sensitivity of 81-99%, and race-specific positive predictive value of 88-97%. The method, Bayesian Improved Surname Geocoding (BISG), has been in wide use in the social sciences and public health for more than 15 years. We use the publicly available North Carolina voter rolls as a proxy for cancer patients, drawing a sample of these persons that mimics the national distribution of cancer incidence by race and ethnicity. BISG has the potential to increase the accuracy of racespecific cancer incidence rates by as much as 3% in registries with the highest levels of missingness; in other registries, the effects will be negligible. The method has been incorporated into freely available computer code.