Consistency in Self-Reported Race-and-Ethnicity Over Time: Implications for Improving the Accuracy of Imputations and Making the Best Use of Self-Report.

Background: Medicare Bayesian Improved Surname and Geocoding (MBISG), which augments an imperfect race-and-ethnicity administrative variable to estimate probabilities that people would self-identify as being in each of 6 mutually exclusive racial-and-ethnic groups, performs very well for Asian American and Native Hawaiian/Pacific Islander (AA&NHPI), Black, Hispanic, and White race-and-ethnicity, somewhat less well for American Indian/Alaska Native (AI/AN), and much less well for Multiracial race-and-ethnicity.

Objectives: To assess whether temporal inconsistency of self-reported race-and-ethnicity might limit improvements in approaches like MBISG.

Methods: Using the Medicare Health Outcomes Survey (HOS) baseline (2013-2018) and 2-year follow-up data (2015-2020), we evaluate the consistency of self-reported race-and-ethnicity coded 2 ways: the 6 mutually exclusive MBISG categories and individual endorsements of each racial-and-ethnic group. We compare the consistency of self-reported race-and-ethnicity (HOS) to the accuracy of MBISG (using 2021 Medicare Consumer Assessment of Healthcare Providers and Systems data).

Results: Concordance (C-statistic) of HOS baseline and follow-up self-reported race-and-ethnicity was 0.95-0.97 for AA&NHPI, Black, Hispanic, and White, 0.83 for AI/AN, and 0.72 for Multiracial using mutually exclusive categories (weighted concordance=0.956). Concordance of MBISG with self-report followed a similar pattern and had similar values, with somewhat lower AI/AN and Multiracial values. The concordance of individual endorsements over time was somewhat higher than for classification (weighted concordance=0.975).

Conclusions: The concordance of MBISG with self-reported race-and-ethnicity appears to be limited by the consistency of self-report for some racial-and-ethnic groups when employing the 6-mutually-exclusive category approach. The use of individual endorsements can improve the consistency of self-reported data. Reconfiguring algorithms such as MBISG in this form could improve its overall performance.