Best practices for analyzing real-world data: Comparing methodologies for imputing ethnicity in a high volume urban clinic

Objective: To compare methodologies for imputing ethnicity from administrative data in a high volume urban clinic. Methods: Using data from 19,165 patients charts in a retrospective study, we compared the accuracy of three methodologies for imputing ethnicity of patients: 1) surname analysis based on tabulation from the 2000 U.S. Census 2) geocoding analysis based on block coding from the 2010 U.S. Census 3) a previously published Bayesian approach involving a combination of surname and geocoding. These results were compared with the ‘gold-standard' of patient self-reported ethnicity. Results: Overall agreement of imputed and self-reported ethnicity was fair for surname analysis (kappa=0.23), moderate for geocoding (kappa=0.58), and strong for the combined model (kappa=0.76). Surname analysis determined Asian ethnicity (sensitivity (SE) 80%; positive predictive value (PPV) 77%) and Latino ethnicity (SE 78%; PPV 68%) with reasonable accuracy but had poor reliability for Caucasians (SE 12%; PPV 92%) and African-Americans (SE 96%; PPV 47%). Geocoding determined African-American ethnicity (SE 74%; PPV 89%) and Caucasian ethnicity (SE 91%; PPV 70%) with reasonable accuracy, but had poor reliability for Asians (SE 10%; PPV 26%) and Latinos (SE 35%; PPV 41%). The Bayesian approach determined African-American (SE 84%; PPV 94%), Caucasian (SE 92%; PPV 82%), Asian (SE 83%; PPV 79%) and Latino (SE 77%; PPV 71%) ethnicity with the highest accuracy of the three methods. Conclusion: A methodology combining surname analysis and geocoding data to determine ethnicity is a valid and accurate means of imputing African-American, Caucasian, Asian and Latino ethnicity when self-reported ethnicity is not available.