Inherent Trade-Offs in the Fair Determination of Risk Scores

Jon Kleinberg, Sendhil Mullainathan and Manish Raghavan study group calibration and equal average scores within positive and negative outcome classes. Theorem 1.1 requires perfect prediction or equal base rates if all three conditions hold in their model.^[1]

Why distinguish the requirements?

These score-level balance conditions differ from equal false-positive and false-negative rates at a chosen threshold. Compare Chouldechova’s threshold-based analysis and Fairness Tradeoffs. The theorem does not choose the morally appropriate criterion or show every conception of fairness impossible.^[1]

The COMPAS entry connects the formal questions to reporting and institutional use.

Historical context

Reviewed original v1, submitted September 19, 2016. Selected definitions/Theorem 1.1/interpretation inspected, including rendered p.4. Later revisions, full proofs, approximation theorem and code not audited.^[1]