Imputation of race/ethnicity to enable measurement of HEDIS performance by race/ethnicity

Abstract

Objective To improve an existing method, Medicare Bayesian Improved Surname Geocoding (MBISG) 1.0 that augments the Centers for Medicare & Medicaid Services’ (CMS) administrative measure of race/ethnicity with surname and geographic data to estimate race/ethnicity. Data Sources/Study Setting Data from 284 627 respondents to the 2014 Medicare CAHPS survey. Study Design We compared performance (cross‐validated Pearson correlation of estimates and self‐reported race/ethnicity) for several alternative models predicting self‐reported race/ethnicity in cross‐sectional observational data to assess accuracy of estimates, resulting in MBISG 2.0. MBISG 2.0 adds to MBISG 1.0 first name, demographic, and coverage predictors of race/ethnicity and uses a more flexible data aggregation framework. Data Collection/Extraction Methods We linked survey‐reported race/ethnicity to CMS administrative and US census data. Principal Findings MBISG 2.0 removed 25‐39 percent of the remaining MBISG 1.0 error for Hispanics, Whites, and Asian/Pacific Islanders (API), and 9 percent for Blacks, resulting in correlations of 0.88 to 0.95 with self‐reported race/ethnicity for these groups. Conclusions MBISG 2.0 represents a substantial improvement over MBISG 1.0 and the use of CMS administrative data on race/ethnicity alone. MBISG 2.0 is used in CMS’ public reporting of Medicare Advantage contract HEDIS measures stratified by race/ethnicity for Hispanics, Whites, API, and Blacks.

Racial and ethnic disparities in care for health system‐affiliated physician organizations and non‐affiliated physician organizations

“…We report estimates for all groups but AI/AN and multiracial due to their low precision. 34 Standard errors were clustered on PO. We used "recycled predictions" to estimate group means on the scale of percent (binary and binomial outcomes) or rates (count outcomes).…”

“…Beneficiary race and ethnicity were estimated using the Medicare Bayesian Improved Surname Geocoding (MBISG 2.0) methodology, which combines Medicare administrative data and US census data to derive probabilities of membership in each of six mutually exclusive racial and ethnic groups (white, black, Hispanic, API, American Indian/Alaska Native (AIAN), and multiracial) 34 . For the racial and ethnic groups included in the current analysis, correlations between the MBISG algorithm's estimated race/ethnicity and self‐reported race/ethnicity range from 0.88 to 0.95.…”

Assessing algorithmic fairness with unobserved protected class using data combination

“…Note that the comparative predictiveness of different proxies for protected class labels, such as race, has been thoroughly studied (Consumer Financial Protection Bureau 2014, Dembosky et al 2019, Elliott et al 2008, Imai and Khanna 2016 as it was understood this impacts the quality of corresponding proxy assessments of disparity. The above result highlights that the proxies' predictiveness of outcomes is also crucial, or even sufficient.…”

“…BISG estimates conditional race membership probabilities given surname and geolocation (e.g., census tract or ZIP code) using data from the US census, and then imputes the race labels based on the estimated probabilities. Since its invention (Elliott et al 2008(Elliott et al , 2009), the BISG method has been widely used in assessing racial disparities in health care (Brown et al 2016, Fremont et al 2016, 2005, Haas et al 2019, Weissman and Hasnain-Wynia 2011. In 2009, the Institute of Medicine also suggested it as an interim strategy until routine collection of relevant data is feasible (Nerenz et al 2009).…”

“…As surname and geolocation proxies complement each other, hybrid approaches like BISG were proposed to combine both(Elliott et al 2008(Elliott et al , 2009) and extended to further include first name(Voicu 2018). In terms of the accuracy of race imputation, BISG has been shown to outperform surname-only and geolocation-only analysis in many datasets, including medicare administration data(Dembosky et al 2019), mortgage data (Consumer Financial Protection Bureau 2014), and voter registration records(Imai and Khanna 2016).…”