Disparate impact testing

The central screening measure in disparate-impact analysis is the adverse impact ratio (AIR): the approval rate for a protected-class group divided by the approval rate for the reference group. The widely cited four-fifths rule treats an AIR below 0.80 as a screening signal. Important provenance: the four-fifths rule’s codified source is the EEOC Uniform Guidelines on Employee Selection Procedures at 29 CFR §1607.4(D)^{[10]Jump to source 10 in the sources list}, which govern employment discrimination under Title VII — not credit. No federal lending regulator (CFPB, OCC, FRB, FDIC, NCUA, HUD) has ever codified the four-fifths rule in a fair-lending rule or examination procedure. Lenders use the 80% threshold as a quantitative anchor borrowed from adjacent civil-rights regulation; courts and examiners have referenced it analogically, but it is not a legal threshold under ECOA or the FHA.

For models, disparate-impact testing is run on the model’s outputs across the populations protected under ECOA — race, color, religion, national origin, sex, marital status, age, receipt of public assistance, and the good-faith exercise of any right under the Consumer Credit Protection Act^{[1]Jump to source 1 in the sources list}. Tests are typically performed at the policy cutoffs — the score thresholds where approval decisions are made — because that is where adverse impact materializes for applicants. A model can show no group differences on its raw score distribution and still produce disparate outcomes at the chosen cutoff, which is why best practice combines the AIR screen with multivariate analysis controlling for legitimate underwriting factors^{[12]Jump to source 12 in the sources list}.

Example: adverse impact ratio

AIR is a screening heuristic, not a regulatory threshold under ECOA or the FHA.

Regulatory status (2025–2026)

The federal supervisory regime around disparate-impact testing changed dramatically between mid-2025 and April 2026. Pursuant to Executive Order 14281 (April 23, 2025)^{[4]Jump to source 4 in the sources list}, the OCC (Bulletin 2025-16, July 2025)^{[5]Jump to source 5 in the sources list}, FDIC (FIL-41-2025, August 2025)^{[6]Jump to source 6 in the sources list}, and NCUA (Letter 25-CU-04, September 2025)^{[7]Jump to source 7 in the sources list} ceased examining for disparate-impact liability and removed references from their supervisory manuals. The CFPB went furthest: its April 22, 2026 final rule amending Regulation B (effective July 21, 2026) struck the “effects test” commentary and stated affirmatively that ECOA does not authorize disparate-impact liability^{[3]Jump to source 3 in the sources list}.

Where disparate-impact testing still matters

Despite the federal supervisory pullback under ECOA, disparate-impact testing remains material for several distinct exposure surfaces:

• Fair Housing Act activity. HUD’s rule at 24 CFR §100.500^{[8]Jump to source 8 in the sources list} and the Supreme Court’s decision in Texas Department of Housing v. Inclusive Communities Project, 576 U.S. 519 (2015)^{[9]Jump to source 9 in the sources list} keep disparate-impact alive for any mortgage or housing-credit activity.
• State fair-lending statutes. New Jersey, California, New York, and Massachusetts retain disparate-impact frameworks under state UDAAP / civil-rights laws applicable to multistate lenders.
• Private litigation and DOJ referrals. Private plaintiffs and state attorneys general retain standing; the Federal Reserve has reserved the authority to refer cases to DOJ.
• GSE / FHA / VA program eligibility. Lenders that sell to the GSEs or originate FHA / VA loans remain subject to those programs’ fair-lending expectations, including disparate-impact analysis where applicable.

Burden-shifting and less discriminatory alternatives

Where disparate impact still applies, courts and HUD use a three-step burden-shifting framework: the plaintiff makes a prima facie showing of disparate effect; the defendant articulates a substantial, legitimate business interest; the plaintiff may rebut with a less discriminatory alternative (LDA) that achieves the same legitimate purpose with reduced impact^{[8]Jump to source 8 in the sources list}.Inclusive Communities added a critical guardrail: a claim based solely on statistical disparity fails absent a robust causal link between the defendant’s policy and the disparity^{[9]Jump to source 9 in the sources list}. For ML practitioners, LDA search — systematically exploring alternative model specifications, feature sets, or hyperparameters that reduce AIR while maintaining predictive accuracy — remains best practice for FHA-covered models, state-law-exposed activity, and documentation of fair-lending diligence even where federal supervisory demand has receded.