Developing and evaluating methods to impute race/ethnicity in an incomplete dataset

The availability of race data is essential for identifying and addressing racial/ethnic disparities in the health care system; however, patient self-reported racial/ethnic information is often missing. Indirect methods for estimating race have been developed, but they usually only consider geocoded and surname data as predictors, may perform poorly among racial minorities, they do not adjust for possible errors for specific datasets, and are unable to provide race estimates for subjects missing some of this information. The objective of this study was to address these limitations by developing novel methods for imputing race/ethnicity when this information is partially missing. By viewing the unobserved race as missing data, we explored different multiple imputation methods for imputing race/ethnicity, and we applied these methods to a subset of Rhode Island Medicaid beneficiaries. Current race imputation methods and newly developed ones were compared using area under the ROC curve statistics and racial composition estimates to identify methods and sets of predictors that yield superior race imputations. Family race was identified as an important predictor and should be included in race estimation models when possible. Bayesian regression models (BRM) provide better race estimates than previously proposed methods. Missing race was multiply imputed using joint modeling and fully conditional specification. Post-imputation analyses showed that fully conditional specification with a BRM is superior to joint modeling for race imputation. The proposed fully conditional specification method is a flexible, effective way of estimating race/ethnicity that allows for propagation of imputation error and ease of interpretation in further analyses.