A Comparison of Methods for Classifying and Modeling Respondents Who Endorse Multiple Racial/Ethnic Categories: A Health Care Experience Application

Abstract

Background: Race/ethnicity information is vital for measuring disparities across groups, and self-report is the gold standard. Many surveys assign simplified race/ethnicity based on responses to separate questions about Hispanic ethnicity and race and instruct respondents to “check all that apply.” When multiple races are endorsed, standard classification methods either create a single heterogenous multiracial group, or attempt to impute the single choice that would have been selected had only one choice been allowed. Objectives: To compare 3 options for classifying race/ethnicity: (a) hierarchical, classifying Hispanics as such regardless of racial identification, and grouping together all non-Hispanic multiracial individuals; (b) a newly proposed additive model, retaining all original endorsements plus a multiracial indicator; (c) an all-combinations approach, separately categorizing every observed combination of endorsements. Research Design: Cross-sectional comparison of racial/ethnic distributions of patient experience scores; using weighted linear regression, we model patient experience by race/ethnicity using 3 classification systems. Subjects: In total, 259,763 Medicare beneficiaries age 65+ who responded to the 2017 Medicare Consumer Assessments of Healthcare Providers and Systems Survey and reported race/ethnicity. Measures: Self-reported race/ethnicity, 4 patient experience measures. Results: Additive and hierarchical models produce similar classifications for non-Hispanic single-race respondents, but differ for Hispanic and multiracial respondents. Relative to the gold standard of the all-combinations model, the additive model better captures ratings of health care experiences and response tendencies that differ by race/ethnicity than does the hierarchical model. Differences between models are smaller with more specific measures. Conclusions: Additive models of race/ethnicity may afford more useful measures of disparities in health care and other domains. Our results have particular relevance for populations with a higher prevalence of multiracial identification.