De-Biasing the Bias: Methods for Improving Disparity Assessments with Noisy Group Measurements

Missing demographic data is a widespread problem that can hide—or exaggerate—algorithmic bias: The paper demonstrates that when clinical algorithms lack accurate racial and ethnic information, assessments of whether these systems treat different groups fairly become unreliable. In their osteoporosis treatment case study, using imputed race data rather than verified information led to misleading conclusions about disparities.

Statistical methods can quantify uncertainty in bias measurements: The authors developed techniques that don't just estimate bias in algorithms, but also measure how much error comes from uncertainty in group membership data itself. This allows policymakers to see a range of possible disparities rather than a single potentially misleading number.

Sensitivity analysis helps decision-makers act despite incomplete information: Rather than delaying algorithm deployment until perfect demographic data exists, the paper's approach lets organizations test how robust their fairness conclusions are across different scenarios of data quality. This enables more informed decisions about whether an algorithm is safe to use.

Current fairness metrics may be unreliable when demographic data is poor quality: The paper proves theoretical bounds showing that common fairness measurements can be significantly distorted by errors in racial/ethnic classification, suggesting that many existing algorithmic audits may need re-evaluation if they relied on imperfect demographic information.