Disparate Impact

A practical statistical guideline from the U.S. Equal Employment Opportunity Commission's Uniform Guidelines on Employee Selection Procedures. A selection rate for any protected group that is less than 80% of the rate for the group with the highest selection rate constitutes evidence of adverse impact.

• Formula: (Selection Rate of Protected Group) / (Selection Rate of Highest Group) < 0.80
• Example: If 60% of male applicants are approved but only 30% of female applicants, the ratio is 0.50, triggering a disparate impact flag
• Limitation: This is a rule of thumb, not a definitive legal test; courts may consider statistical significance tests like the Z-test or Fisher's exact test for smaller sample sizes

Disparate impact does not require proof of discriminatory intent. The doctrine targets facially neutral policies—rules or algorithms that appear objective on their face but produce discriminatory outcomes in practice.

• Key Distinction: Unlike disparate treatment, which requires evidence of intentional discrimination, disparate impact focuses solely on statistical outcomes
• Algorithmic Context: A credit scoring model that uses seemingly neutral features like zip code or browsing history can produce disparate impact if those features serve as proxies for protected attributes such as race or national origin
• Legal Precedent: Established in Griggs v. Duke Power Co. (1971) , where a high school diploma requirement for employment was found to disproportionately exclude Black applicants without being a valid predictor of job performance

Once a plaintiff establishes a prima facie case of disparate impact, the burden shifts to the defendant to prove the challenged practice is job-related and consistent with business necessity. This is the primary legal defense against disparate impact claims.

• Validation Requirement: The employer or model developer must demonstrate through statistical validation studies that the selection criterion is predictive of successful performance
• Less Discriminatory Alternative: Even if business necessity is proven, the plaintiff can still prevail by showing a less discriminatory alternative exists that serves the same business purpose
• Algorithmic Parallel: In AI governance, this maps to the requirement that model features be predictively valid and that less biased model architectures be evaluated before deployment

Beyond the Four-Fifths Rule, courts and regulators increasingly require formal statistical significance tests to establish that observed disparities are not due to random chance.

• Standard Threshold: A p-value < 0.05 (two-tailed) is commonly used to establish statistical significance
• Common Tests: The two-sample Z-test for proportions, chi-squared test for categorical outcomes, and Fisher's exact test for small sample sizes
• Practical Significance vs. Statistical Significance: Large datasets can produce statistically significant results with trivially small effect sizes; courts may also consider practical significance—the magnitude of the disparity
• Multiple Comparisons: When testing across many subgroups, corrections like the Bonferroni correction must be applied to avoid false positives

Algorithms can produce disparate impact even when protected attributes are explicitly excluded from the training data. This occurs through proxy variables—features that are statistically correlated with protected class membership.

• Redlining by Proxy: Using geographic features like zip code or census tract that correlate with racial demographics
• Behavioral Proxies: Features like browsing history, purchase patterns, or social network data can encode socioeconomic status, which intersects with protected attributes
• Detection Method: Correlation analysis between model features and protected attributes; a feature with a Pearson correlation coefficient > 0.5 with a protected attribute warrants scrutiny
• Mitigation: Techniques like adversarial debiasing or disparate impact remover pre-processing can reduce proxy discrimination

Modern AI regulations explicitly incorporate disparate impact concepts into compliance requirements. The EU AI Act and NIST AI Risk Management Framework both mandate testing for discriminatory outcomes.

• EU AI Act: High-risk AI systems must undergo conformity assessments that include bias testing across protected groups; disparate impact findings can block market access
• NIST AI RMF: The framework's Map function requires organizations to identify potential disparate impacts before deployment
• NYC Local Law 144: Requires bias audits for automated employment decision tools, with results publicly disclosed
• Documentation: Organizations must maintain model cards and algorithmic impact assessments documenting disparate impact testing results and mitigation steps