Statistical Parity

Statistical parity is satisfied when the probability of a positive prediction is identical across all groups defined by a protected attribute (e.g., race, gender). Formally, for a binary classifier Ŷ and protected attribute A, the condition is P(Ŷ=1 | A=a) = P(Ŷ=1 | A=b) for all groups a and b. This metric focuses solely on the output distribution of the model, not its accuracy or the true labels. It is the most intuitive definition of equality of outcome in automated decision-making.

A defining characteristic of statistical parity is its deliberate independence from the true label (Y). The metric does not consider whether the positive predictions are correct. This is a critical distinction from accuracy-based fairness metrics like equalized odds. By ignoring the base rate of outcomes, statistical parity enforces a strict proportional representation in the model's decisions, which can be a legal or policy requirement in contexts where historical labels are themselves tainted by systemic bias.

In practice, perfect equality is rare, so fairness is measured using the Demographic Parity Difference (DPD). This is calculated as the absolute difference between the selection rates of two groups: DPD = |P(Ŷ=1 | A=unprivileged) - P(Ŷ=1 | A=privileged)|. A DPD of 0 indicates perfect statistical parity. A common fairness threshold, particularly in tools like Fairlearn and AIF360, is a DPD ratio of 0.8 or less, aligning with the Four-Fifths Rule used in employment law.

Enforcing statistical parity often introduces a direct accuracy-fairness trade-off. If the true base rates of a positive outcome differ between groups, forcing equal prediction rates requires the model to deliberately make more errors for one group—either increasing false positives for the underprivileged group or increasing false negatives for the privileged group. This tension makes statistical parity controversial in high-stakes domains like medicine, where denying a correct diagnosis to achieve a demographic quota is ethically problematic.

Statistical parity is a purely group-level fairness metric. It provides no guarantees of fairness for any specific individual. A model can satisfy statistical parity perfectly while still making wildly unjust decisions for particular people, a concept known as counterfactual unfairness. For example, a model could accept all qualified men and unqualified women while rejecting all unqualified men and qualified women, achieving perfect gender parity in acceptance rates but violating the principle of treating similar individuals similarly.

Statistical parity is closely aligned with the legal doctrine of disparate impact, which identifies facially neutral policies that disproportionately harm a protected class. In the U.S., the Four-Fifths Rule from the Uniform Guidelines on Employee Selection Procedures is a direct application of the parity ratio. Under the European Union Artificial Intelligence Act, high-risk AI systems must be tested for discriminatory impacts, making statistical parity a primary quantitative tool for demonstrating compliance during a bias audit.

A stricter fairness criterion that requires a classifier to achieve equal true positive rates and equal false positive rates across protected groups. Unlike statistical parity, equalized odds conditions on the true outcome, ensuring that errors are distributed evenly. This metric is preferred when the base rates of the target variable differ legitimately between groups, as it penalizes models that mask inaccuracy in one group with higher accuracy in another.

A legal doctrine and quantitative measure originating from U.S. employment law. It identifies facially neutral policies that disproportionately harm protected groups. The Four-Fifths Rule serves as a practical threshold: a selection rate for any group that is less than 80% of the rate for the highest group constitutes evidence of adverse impact. Disparate impact analysis is a cornerstone of regulatory compliance audits.

A causal definition of fairness where a decision is fair if it would remain the same in a counterfactual world where the individual belonged to a different demographic group. This approach uses structural causal models to distinguish discriminatory path-specific effects from legitimate, non-discriminatory influences. It addresses the limitations of purely observational metrics by modeling the underlying causal mechanisms.

The application of technical interventions at three stages of the ML pipeline:

• Pre-processing: Transform training data to remove discriminatory patterns before model training
• In-processing: Add fairness constraints directly to the model's objective function during training
• Post-processing: Adjust model outputs or decision thresholds after prediction to satisfy fairness criteria Each approach involves distinct trade-offs between accuracy and fairness.

Open-source libraries that operationalize fairness evaluation and mitigation:

• AI Fairness 360 (AIF360): IBM's comprehensive suite with 70+ fairness metrics and 10+ bias mitigation algorithms
• Fairlearn: Microsoft's toolkit providing assessment dashboards and interactive visualizations for disaggregated performance analysis Both integrate with standard ML pipelines and support binary classification and regression tasks.

A structured transparency document that reports the intended use, evaluation results, and ethical considerations of a trained model. Model cards disaggregate performance metrics across demographic groups, explicitly reporting fairness evaluations. Originating from Google research, they are now a standard requirement under proposed AI regulations for high-risk systems, enabling downstream auditors to assess suitability for deployment.