Equal Opportunity

The core mathematical requirement: the probability of a positive prediction given a positive ground-truth label must be identical across all groups.

• Formula: P(Ŷ=1 | Y=1, A=a) = P(Ŷ=1 | Y=1, A=b)
• Focus: Only constrains outcomes for the "qualified" subset of the population
• Contrast with Equalized Odds: Does not constrain False Positive Rates, allowing different error profiles for unqualified individuals
• Real-world example: A loan model where creditworthy applicants have the same approval rate regardless of race

Equal Opportunity belongs to the separation family of fairness criteria, meaning it conditions on the true outcome Y.

• Conditional independence: Prediction Ŷ must be independent of protected attribute A, conditional on Y=1
• Key insight: Acknowledges that base rates may differ between groups, but insists that qualified individuals be treated identically
• Causal interpretation: Blocks the direct path from protected attribute to decision for those who truly merit a positive outcome
• Implementation: Typically enforced as a constraint during model training or as a post-processing adjustment to decision thresholds

The most common implementation strategy involves learning group-specific decision thresholds that equalize TPR across groups.

• Post-processing approach: Train a single classifier, then calibrate thresholds per group to satisfy TPR equality
• In-processing approach: Add a Lagrangian penalty term to the loss function that penalizes TPR disparity
• Trade-off management: Adjusting thresholds may reduce overall accuracy but guarantees equal opportunity for qualified candidates
• Practical tooling: Libraries like Fairlearn and AIF360 provide built-in Equal Opportunity constraint optimizers

Equal Opportunity is philosophically grounded in the principle that merit should be recognized uniformly across demographic lines.

• Moral foundation: Aligns with the intuition that the most qualified candidates should succeed, regardless of identity
• Limitation: Does not address disparities in who becomes "qualified" due to historical or structural barriers
• Complementary metric: Often paired with Demographic Parity or Equalized Odds for a more complete fairness audit
• Use case fit: Well-suited for selective processes like hiring, college admissions, and premium service eligibility where identifying true positives is the primary concern

Equal Opportunity is a relaxed version of the stricter Equalized Odds criterion, which demands equality of both TPR and FPR.

• Equalized Odds: Requires P(Ŷ=1 | Y=y, A=a) = P(Ŷ=1 | Y=y, A=b) for both y=0 and y=1
• Equal Opportunity: Only requires equality for y=1 (the positive class)
• Practical advantage: Easier to satisfy than Equalized Odds, often with less accuracy degradation
• When to prefer: Choose Equal Opportunity when false positives for unqualified individuals carry different costs across groups, or when the primary concern is ensuring access for deserving candidates

Rigorous auditing requires computing group-conditional TPRs and testing for statistically significant differences.

• Metric calculation: TPR_g = True Positives_g / (True Positives_g + False Negatives_g) for each group g
• Disparity measure: Common metrics include the ratio (min TPR / max TPR) or the difference (max TPR - min TPR)
• Statistical testing: Bootstrap confidence intervals or permutation tests to determine if observed disparities are significant
• Documentation: Results should be reported in a Model Card alongside other fairness metrics and dataset composition details