Equalized Odds is the fairness criterion requiring that an AI classifier have equal true positive rates and equal false positive rates across all protected groups — stronger than demographic parity because it requires not just equal outcomes but equal accuracy across groups, ensuring the model makes comparably correct and incorrect decisions regardless of group membership.

What Is Equalized Odds?

• Definition: A model satisfies equalized odds when both the True Positive Rate (TPR) and False Positive Rate (FPR) are equal across protected groups — neither group is systematically favored in correct predictions or systematically burdened with incorrect positive predictions.
• Publication: Introduced by Hardt, Price, and Srebro (NeurIPS 2016) as a mathematically precise fairness criterion addressing limitations of demographic parity.
• Two Conditions: Equal TPR (sensitivity): P(Ŷ=1 | Y=1, A=0) = P(Ŷ=1 | Y=1, A=1) AND Equal FPR (1-specificity): P(Ŷ=1 | Y=0, A=0) = P(Ŷ=1 | Y=0, A=1).
• Relaxation — Equal Opportunity: If only TPR equality is required (ignoring FPR), the criterion is called "equal opportunity" — appropriate when false positives are less consequential than false negatives.

Why Equalized Odds Matters

• Recidivism Prediction: The COMPAS controversy (ProPublica, 2016) showed that a criminal risk assessment tool had higher FPR for Black defendants (falsely flagged as high-risk at nearly 2x the rate) — a direct equalized odds violation with devastating civil liberties implications.
• Medical Screening: A cancer screening AI with lower TPR for minority patients means those patients are less likely to be flagged for follow-up when actually at risk — an equal opportunity violation with life-or-death consequences.
• Loan Approval: Equalized odds requires that both qualified applicants from all groups have equal approval rates AND unqualified applicants from all groups have equal rejection rates.
• Superior to Demographic Parity: Demographic parity can be achieved by making a model less accurate for one group to match another. Equalized odds requires genuine accuracy parity — a higher standard.

Mathematical Formulation

For classifier Ŷ, true label Y, and sensitive attribute A ∈ {0,1}:

Equal TPR (Equal Opportunity): P(Ŷ=1 | Y=1, A=0) = P(Ŷ=1 | Y=1, A=1)

Equal FPR: P(Ŷ=1 | Y=0, A=0) = P(Ŷ=1 | Y=0, A=1)

Equalized Odds = Equal TPR AND Equal FPR simultaneously.

The Impossibility Result

Chouldechova (2017) proved that when base rates differ across groups, it is mathematically impossible to simultaneously satisfy: 1. Equalized odds (equal TPR and FPR) 2. Calibration (score = probability of positive outcome) 3. Demographic parity (equal positive rates)

This means every fairness metric involves a genuine trade-off — there is no algorithm that is simultaneously "fair" by all definitions when group base rates differ.

Post-Processing for Equalized Odds

Hardt et al. proposed a practical post-processing solution:

• After training a base classifier, derive separate classification thresholds for each group.
• Solve a linear program to find threshold combinations that equalize TPR and FPR across groups.
• Result: A randomized classifier that satisfies equalized odds exactly.
• Trade-off: Post-processing always decreases overall accuracy relative to the unconstrained optimal classifier.

Equalized Odds vs. Related Metrics

Implementation Tools

• IBM AI Fairness 360: Provides equalized odds post-processing as a built-in mitigation algorithm.
• Fairlearn (Microsoft): Implements equalized odds constraints via exponentiated gradient reduction.
• Google What-If Tool: Visualizes TPR/FPR across groups interactively on any classifier.
• Themis-ML: Academic library for fairness-aware machine learning with equalized odds support.

Equalized odds is the gold standard fairness metric for high-stakes classification — by requiring accuracy parity rather than mere outcome parity, it ensures AI systems do not systematically punish one group with higher false positive rates or deny another group with lower true positive rates, addressing the most concrete mechanisms through which algorithmic discrimination causes real harm.