(BSPOT05) Interpreting Effect Estimates Across Racial and Ethnic Categories in Propensity Score-Based Analyses: An Application in US Medicare Beneficiaries Initiating Antidiabetic Medications

Background: Racial and ethnic variables in healthcare data can reflect structural drivers of inequity (e.g., inequitable healthcare access, interpersonal discrimination and bias, and different distributions of social determinants of health) which influence how patients receive one treatment over another. It is unclear how accounting for these variables impacts estimates and interpretation of treatment effect heterogeneity.

Objectives: To compare common approaches for incorporating racial and ethnic categories in propensity score (PS)-based analyses when examining effect estimates in a study of cardiovascular outcomes in patients with diabetes.

Methods: We used a 20% random sample of US Medicare beneficiaries age ≥66 years with fee-for-service and Part D coverage. We included patients on metformin when they initiated DPP4 inhibitors (DPP4i) or sulfonylureas (SU) during 2008-2015 who were classified as non-Hispanic (NH) White (n=18,548), NH Black (n=1994), or Hispanic (n=3265) based on CMS’ “RTI race code” variable. We fit three types of logistic regression models to estimate PS: (1) “base” model with 56 claims- and enrollment-based covariates; (2) “adjusted” model adding a main effect for racial/ethnic category; and (3) “interaction” model adding interaction terms between racial/ethnic category and each other covariate. We then estimated crude and IP-weighted 2-year cumulative incidence (henceforth risk) and 2-year risk differences for a composite outcome (non-fatal MI, stroke, all-cause mortality) in an as-treated analysis.

Results: For SU initiators, crude 2-year risks (95%CI) varied across racial/ethnic categories: 9.2% (8.4%-9.9%) for NH White, 14.0% (11.1%-17.7%) for NH Black, 8.8% (7.0%-11.1%) for Hispanic. Weighted 2-year RDs (DPP4i vs SU) were also heterogeneous across racial/ethnic categories: -1.9% (-3.2%, -0.6%) for NH White, -7.4% (-12.7%, -2.0%) for NH Black, -2.5% (-5.9%, 1.0%) for Hispanic; however, RD estimates were similar across all PS strategies. All-cause mortality accounted for 78% (NH White), 90% (NH Black), and 77% (Hispanic) of outcomes. Covariate distributions were balanced within racial/ethnic categories, but some varied across racial/ethnic categories (e.g., Part D low-income subsidy: 18.4% for NH White, 50.1% for NH Black).

Conclusions: Risk difference estimates appeared greatest in the NH Black category, but this heterogeneity is likely driven by differential distributions of effect measure modifiers or coding practices leading to differential measurement error. PS-based analyses that lack attention to structural inequities reflected by racial and ethnic categories may smooth over important heterogeneous treatment effects or induce residual confounding, thus limiting their validity and interpretability.