Implications of missingness in self-reported data for estimating racial and ethnic disparities in Medicaid quality measures

Objective: To assess the feasibility and implications of imputing race and ethnicity for quality and utilization measurement in Medicaid. Data Sources and Study Setting: 2017 Oregon Medicaid claims from the Oregon Health Authority and electronic health records (EHR) from OCHIN, a clinical data research network, were used. Study Design: We cross-sectionally assessed Hispanic-White, Black-White, and Asian-White disparities in 22 quality and utilization measures, comparing selfreported race and ethnicity to imputed values from the Bayesian Improved Surname Geocoding (BISG) algorithm. Data Collection: Race and ethnicity were obtained from self-reported data and imputed using BISG. Principal Findings: 42.5%/4.9% of claims/EHR were missing self-reported data; BISG estimates were available for >99% of each and had good concordance (0.87-0.95) with Asian, Black, Hispanic, and White self-report. All estimated racial and ethnic disparities were statistically similar in self-reported and imputed EHR-based measures. However, within claims, BISG estimates and incomplete self-reported data yielded substantially different disparities in almost half of the measures, with BISG-based Black-White disparities generally larger than self-reported race and ethnicity data. Conclusions: BISG imputation methods are feasible for Medicaid claims data and reduced missingness to KEYWORDS Bayesian imputation, health care disparities, HEDIS, Medicaid, quality of health care, race factors What is known on this topic * Missingness of self-reported race and ethnicity in administrative data is often high and due to non-ignorable mechanisms. * Bayesian Improved Surname Geocoding (BISG) indirect estimates of race and ethnicity have high accuracy in predicting racial and ethnic identity. * Reliance on administrative data that does not correct for missingness may misrepresent measures of health disparities. What this study adds * This study demonstrates the feasibility of using BISG in a Medicaid population, producing imputed estimates for more than 99% of beneficiaries. * Estimated magnitudes of disparities were larger in claims-based measures that incorporated BISG than in measures that did not adjust for missingness. * Electronic health records-based measures of disparities that incorporated BISG were statistically similar to those using self-reported race and ethnicity data, perhaps reflecting the low rate of missingness in electronic health record data.
1 | INTRODUCTION Implementation of the Affordable Care Act (ACA) in 2014 was associated with significant reductions in, but not elimination of, racial and ethnic disparities. (1) Analyses of the [...]