Background

Real-world data (RWD) is increasingly used to identify targets for reducing racial disparities, but data-retrieval processes (e.g., measurement, selection, restriction) can distort prevalence and absolute risk differences, quantities that determine how much disparity an intervention can remove. We asked if RWD distorts the observed reduction in the Black–White gap in uncontrolled diabetes when evaluating a hypothetical intervention to eliminate insurance gaps.

Methods

Using published nationally representative prevalence estimates for race, insurance coverage, and uncontrolled diabetes (HbA1c >9%) as our benchmark, we calibrated a source population propagating parameter uncertainty via probabilistic simulation. Our causal estimand was the proportion of disparity eliminated (PDE): the percentage reduction in the Black–White absolute risk difference under a hypothetical intervention do(E=0). We simulated three RWD data-source mechanisms: (1) EHR data, (2) hospital-cohort, and (3) claims data restriction, representing differential outcome misclassification, selection, and collider stratification bias. We calculated source-driven instability as the probability of observing a minimal or reversed effect (PDE <5% or PDE <0) conditional on a true baseline PDE ≥20%.

Results

In the benchmark, eliminating the coverage gap reduced the racial disparity by a median of 18% (IQR 3%–39%). Simulated EHR misclassification attenuated the benefit to 4% (−5% to 18%), with a 19% probability of inverting the direction of effect. Hospital selection reversed the effect to −5% (−23% to 17%), with a 45% instability rate. Insured restriction eliminated the signal (0%, −1% to 2%), with a 80% reversal rate.

Conclusions

RWD can decrease or reverse the observable effects of structural interventions on disparity reduction. Health equity research using RWD should report absolute disparity estimands (e.g., PDE) and conduct quantitative bias analyses to account for data-source bias mechanisms.