Disparate Impact Ratio

The most common threshold for adverse impact. A selection rate for a protected group that is less than 80% of the rate for the group with the highest selection rate generally constitutes evidence of adverse impact.

• Calculation: DIR = (Selection Rate of Protected Group) / (Selection Rate of Reference Group)
• Example: If a hiring model approves 60% of male applicants but only 30% of female applicants, the DIR is 0.30 / 0.60 = 0.50. This violates the 80% rule.
• Origin: Codified in the Uniform Guidelines on Employee Selection Procedures (1978).

Beyond the raw ratio, statistical tests determine if an observed disparity is due to chance. A small sample size can produce a low DIR that is not statistically significant.

• Fisher's Exact Test: Used for small sample sizes to test the independence of selection and group membership.
• Two-Standard Deviation Test: A simpler heuristic; if the difference between expected and actual selections exceeds two standard deviations, it's flagged.
• Practical Significance: A DIR may be statistically significant but not practically meaningful if the absolute number of affected individuals is tiny.

A more granular analysis that controls for legitimate, job-related qualifications. It isolates whether the disparity exists within similarly qualified subgroups.

• Mechanism: Stratify the population by a legitimate explanatory variable (e.g., years of experience) and calculate the DIR within each stratum.
• Purpose: Distinguishes between systemic bias and a genuine difference in the distribution of qualifications.
• Example: If the overall DIR is 0.60, but within each experience band (0-5 yrs, 5-10 yrs) the DIR is > 0.90, the overall disparity is explained by a difference in experience distribution, not direct bias.

When a DIR violation is detected, remediation can occur at different stages of the ML lifecycle. The choice of intervention has distinct trade-offs.

• Pre-processing: Reweighing or resampling the training data to remove historical bias before model training.
• In-processing: Adding a fairness constraint directly to the model's loss function to penalize disparate outcomes during training.
• Post-processing: Adjusting the model's decision threshold for different groups after prediction to equalize selection rates.
• Trade-off: Post-processing is easiest to implement but can violate individual fairness by treating similar individuals differently.

The DIR is a narrow, group-level fairness metric. It does not capture all forms of algorithmic harm.

• Fairness Gerrymandering: A system can satisfy the DIR globally but still discriminate against specific subgroups (e.g., older women).
• Individual Fairness: The DIR ignores whether similar individuals are treated similarly; it only looks at aggregate group outcomes.
• Base Rate Neglect: The metric does not account for true differences in the target variable between groups, which can lead to reverse discrimination if applied blindly.
• Intersectionality: A single binary DIR calculation often fails to capture compounded bias against individuals at the intersection of multiple protected characteristics.

The original Uniform Guidelines threshold from 1978. A selection rate for a protected group that is less than 80% of the rate for the group with the highest selection rate constitutes evidence of adverse impact. This is the direct historical precedent for the Disparate Impact Ratio. If the ratio falls below 0.80, a prima facie case of discrimination is established under U.S. employment law.

A stricter fairness criterion than simple demographic parity. It requires that a model's true positive rate and false positive rate are equal across protected and unprotected groups. Unlike the Disparate Impact Ratio, which only looks at overall selection rates, equalized odds ensures the model is equally accurate for all groups. This is a key metric in the Hardt et al. (2016) fairness framework.

A fairness definition requiring that a model's positive prediction rate is identical across all demographic groups. This is the strictest interpretation of the Disparate Impact Ratio (targeting a ratio of 1.0). While it eliminates disparate impact, it can sometimes conflict with predictive parity if base rates differ between groups. It is often mandated in highly regulated lending and hiring contexts.

An extension of demographic parity that allows for legitimate differentiating factors. It requires equal selection rates across protected groups only after controlling for a set of legitimate risk factors (e.g., driving experience for auto insurance). This aligns with the business necessity defense in disparate impact law, where a facially neutral policy is permissible if it is job-related and consistent with business necessity.

Often used synonymously with Disparate Impact Ratio, particularly in U.S. employment law and EEOC investigations. The calculation is identical: (Selection Rate of Protected Group) / (Selection Rate of Reference Group). An AIR below 0.80 triggers a legal burden on the employer to prove the selection tool's job relevance. It is the foundational metric for Title VII compliance audits.