Disparate Impact Ratio

The foundational threshold for adverse impact, originating from the Uniform Guidelines on Employee Selection Procedures (1978). A selection rate for a protected group that is less than 80% of the rate for the group with the highest selection rate is generally regarded as evidence of adverse impact.

• Formula: DIR = (Selection Rate of Protected Group) / (Selection Rate of Reference Group)
• Threshold: DIR < 0.80 triggers further investigation
• Example: If 60% of Group A is approved but only 30% of Group B, the DIR is 0.50, signaling a potential violation.

Disparate impact focuses on facially neutral policies that produce discriminatory effects, regardless of intent. This differs fundamentally from disparate treatment, which requires proof of intentional discrimination.

• Disparate Impact: Unintentional discrimination caused by a neutral policy (e.g., a height requirement that disproportionately excludes women).
• Disparate Treatment: Intentional discrimination (e.g., explicitly refusing to hire based on race).
• Legal Context: The DIR is a tool to prove the former, where motive is absent but statistical disparity exists.

In AI auditing, the DIR is a core metric for evaluating binary classifiers in high-stakes domains like lending, hiring, and criminal justice. It is often used alongside Equalized Odds and Demographic Parity.

• Pre-deployment Audit: Calculate DIR on test data before model release.
• Post-deployment Monitoring: Track DIR in production to detect concept drift that introduces bias.
• Limitation: DIR does not account for the ground truth base rate differences between groups, which is addressed by metrics like Equalized Odds.

When a model exhibits a DIR below the 0.80 threshold, several technical and procedural mitigations can be applied to restore fairness without necessarily sacrificing overall accuracy.

• Pre-processing: Reweighing or resampling training data to balance representation.
• In-processing: Adding fairness constraints directly into the model's loss function during training.
• Post-processing: Adjusting decision thresholds independently for different groups to equalize acceptance rates.
• Business Justification: If a low DIR is unavoidable, the model owner must demonstrate the policy is a business necessity with no less-discriminatory alternative.

For high-risk AI systems under the EU AI Act, conducting a disparate impact analysis is a critical component of the mandatory Algorithmic Impact Assessment and Fundamental Rights Impact Assessment.

• Conformity Assessment: Providers must demonstrate that bias has been detected and mitigated.
• Transparency Obligations: The methodology for calculating and addressing the DIR must be documented in the model card and technical documentation.
• Post-Market Monitoring: Continuous tracking of the DIR is required to ensure the system remains compliant after deployment.

A fairness criterion requiring that a model's positive prediction rate is equal across all protected demographic groups, regardless of the ground truth. This is the statistical ideal that the Disparate Impact Ratio directly measures.

• Formula: P(ŷ=1|A=a) = P(ŷ=1|A=b)
• Key Distinction: Focuses solely on the model's decision, not its accuracy.
• Trade-off: Can force a model to make deliberate errors to balance outcomes, potentially violating equalized odds.

A fairness metric requiring that a model's true positive rate and false positive rate are equal across different protected groups. Unlike demographic parity, this metric is conditioned on the ground truth.

• Formula: P(ŷ=1|Y=1, A=a) = P(ŷ=1|Y=1, A=b) AND P(ŷ=1|Y=0, A=a) = P(ŷ=1|Y=0, A=b)
• Advantage: Penalizes a model that is only accurate for a privileged group.
• Relationship: A model can satisfy demographic parity but fail equalized odds, and vice versa.

A guideline established by the U.S. Equal Employment Opportunity Commission (EEOC) to operationalize the Disparate Impact Ratio. A selection rate for a protected group that is less than 80% of the rate for the group with the highest selection rate is generally regarded as evidence of adverse impact.

• Calculation: (Selection Rate of Protected Group) / (Selection Rate of Reference Group) < 0.80
• Context: Originates from Title VII of the Civil Rights Act of 1964.
• Limitation: It is a rule of thumb, not a definitive legal test, and is sensitive to small sample sizes.

A causal definition of fairness stating a prediction is fair if it remains the same in a counterfactual world where an individual's protected attribute was changed. This moves beyond correlation to model a causal relationship.

• Structural Equation Modeling: Requires a causal model of the world to generate the counterfactual.
• Example: Would a loan application be denied if the applicant's race were different, holding all other causally dependent variables constant?
• Strength: Directly addresses the intuition behind individual fairness and anti-discrimination.

A non-protected feature that inadvertently encodes a protected attribute, leading to masked discrimination. A model can achieve a perfect Disparate Impact Ratio while still being discriminatory if it relies on proxies.

• Classic Example: Redlining via zip codes, which correlate strongly with race.
• Detection: Requires rigorous feature importance analysis and correlation testing with protected attributes.
• Mitigation: Simply removing the protected attribute is insufficient; proxy variables must be identified and addressed through techniques like adversarial debiasing.

A structured process for evaluating the potential societal, ethical, and legal consequences of an automated decision system before it is deployed. Calculating the Disparate Impact Ratio is a core quantitative component of this broader qualitative assessment.

• Regulatory Driver: Mandated by proposed legislation like the Algorithmic Accountability Act.
• Scope: Includes fairness analysis, stakeholder consultation, and documentation of mitigation strategies.
• Output: A formal report that often includes a Model Card and a Datasheet for Datasets.