(170) Racial and Ethnic Categories in Propensity Score Analyses: An Application in US Medicare Beneficiaries Initiating DPP4 Inhibitors vs Sulfonylureas

Publication Number: 1215

Background: In propensity score (PS)-based analyses of medication safety and effectiveness, investigators might estimate stratum-specific PSs for sex or age to examine heterogeneous treatment effects and confounding structures. Stratifying by racial and ethnic categories may be useful for similar reasons; specifically, stratification may better reflect the structural drivers of racial and ethnic inequities compared to individual-level adjustment for racial and ethnic categories as main effect terms in a PS model.

Objectives: To compare covariate balance across 3 strategies of including race and ethnicity in propensity scores in an active-comparator new-user cohort study of Medicare beneficiaries.

Methods: We used a 20% random sample of US Medicare beneficiaries age ≥66 years with fee-for-service and Part D coverage who were classified as non-Hispanic (NH) white, NH Black, or Hispanic using CMS’ “RTI race code,” and initiated DPP4 inhibitors (DPP4i) or sulfonylureas (SU) while on metformin during 2008-2015 using a 1-year lookback. Using 56 covariates, we fit logistic regression models to estimate PSs under 3 strategies: (1) “full-cohort” PS model with main effect term for racial/ethnic category; (2) ”full cohort” PS with interaction terms between racial/ethnic category and each covariate; and (3) separate stratum-specific PS models for each racial/ethnic category. After calculating inverse probability of treatment weights for each strategy, we assessed covariate balance between DPP4i/SU initiators separately for each racial/ethnic category, summarized using average and maximum standardized absolute mean differences (ASAMD and MSAMD).

Results: Among 23,807 individuals, 77.9% were classified as NH white, 8.4% NH Black, and 13.7% Hispanic; 28.0%, 31.8%, and 31.6% initiated a DPP4i versus SU, respectively. In crude data, ASAMD and MSAMD were 0.05, 0.31 (NH white); 0.07, 0.47 (NH Black); and 0.10, 0.33 (Hispanic). In strategy #1, ASAMD and MSAMD were 0.01, 0.06 (NH white); 0.05, 0.22 (NH Black); and 0.04, 0.22 (Hispanic); importantly, the covariate with maximum imbalance differed by racial/ethnic category: anticholinergic use for NH white, year for NH Black, and number of metformin pills for Hispanic. Strategies #2 and #3 were functionally identical and yielded the best ASAMD and MSAMD: 0.00, 0.04 (NH white); 0.02, 0.10 (NH Black); and 0.01, 0.07 (Hispanic).

Conclusions: Stratifying by racial/ethnic category may improve confounding control while supporting assessment of structural inequities and heterogeneity of treatment effects. These results underscore the importance of modeling assumptions corresponding to structural factors. This study incorporated an active comparator, but covariate imbalance is likely worse in non-active-comparator settings.