Comparison of Alternative Approaches to Using Race-and-Ethnicity Data in Estimating Differences in Health Care and Social Determinants of Health.

Objective: The objective of this study was to compare 2 approaches for representing self-reported race-and-ethnicity, additive modeling (AM), in which every race or ethnicity a person endorses counts toward measurement of that category, and a commonly used mutually exclusive categorization (MEC) approach. The benchmark was a gold-standard, but often impractical approach that analyzes all combinations of race-and-ethnicity as distinct groups.

Methods: Data came from 313,739 respondents to the 2021 Medicare Consumer Assessment of Healthcare Providers and Systems (CAHPS) surveys who self-reported race-and-ethnicity. We used regression to estimate how accurately AM and MEC approaches predicted racial-and-ethnic differences in 5 CAHPS patient experience measures and 4 patient characteristics that we considered proxies for social determinants of health (SDOH): age, educational attainment, and self-reported general and mental health. We calculated average residual error proportions for AM and MEC estimates relative to all-combination estimates.

Results: In predicting CAHPS scores by race-and-ethnicity, on average 0.9% of the variance across groups in the AM and MEC approaches represented a departure from the gold standard. In predicting proxy SDOH variables, on average 4.7% of the AM variance across groups and 7.1% of the MEC variance across groups represented departures from the gold standard.

Conclusion: Researchers may want to consider AM over MEC when modeling outcomes by race-and-ethnicity given that AM outperforms MEC in predicting racial-and-ethnic differences in proxy SDOH characteristics and is comparably accurate in predicting differences in patient experience. Unlike MEC, AM does not assume that every multiracial person has similar outcomes and that Hispanic persons have similar outcomes irrespective of race.