Bias in Criminal Risk Scores Is Mathematically Inevitable, Researchers Say

Abstraction: Mathematical proof that predictive parity and equal error rates are incompatible fairness criteria

Key points:

• ProPublica analysis of COMPAS showed Black defendants twice as likely to be falsely labeled high-risk as white defendants
• Four independent research groups (Stanford, Cornell/Harvard, Carnegie Mellon, U Chicago/Google) all proved the same result: a risk score cannot simultaneously satisfy "predictive parity" and "equal error rates" when base rates differ between groups
• Northpointe defended COMPAS by citing equal predictive accuracy (~60%) across races, but this ignores differential false-positive rates
• Hardt, Price, Srebro paper "Equality of Opportunity in Supervised Learning" proposed equalizing error rates between groups as definition of nondiscrimination
• Kleinberg et al. "Inherent Trade-Offs in the Fair Determination of Risk Scores" provides formal mathematical proof of the incompatibility

Connections: Propublica · Northpointe · Algorithmic Fairness · Bias In AI · Machine Learning

Source: https://www.propublica.org/article/bias-in-criminal-risk-scores-is-mathematically-inevitable-researchers-say