Adverse impact ratio calculator

How the ratio is calculated

Four steps, and the arithmetic is deliberately simple — the difficulty is never the division, it is deciding what counts as “considered” and what counts as “selected”.

1. For each group, divide the number selected by the number considered. That is the group’s selection rate.
2. Find the group with the highest selection rate. That group is the reference.
3. Divide every other group’s rate by the reference rate. That quotient is the impact ratio.
4. Compare each ratio to 0.80.

Two definitional traps do more damage than any error in the maths. The first is the denominator: “considered” means everyone who actually reached the step being tested, not everyone who ever touched the funnel — if you test a shortlist decision, the denominator is the people who were scored, not everyone who clicked apply. The second is testing the whole funnel at once. A pipeline can pass end-to-end while one stage inside it fails badly; run the calculation stage by stage before you run it overall.

What 0.80 means, and what it does not

The threshold is a screening heuristic, not a legal test. Section 4(D) of the 29 CFR 1607.4 says a rate below four-fifths of the highest rate “will generally be regarded by the Federal enforcement agencies as evidence of adverse impact”. Generally, and evidence — not conclusively, and not a finding.

The same section cuts the other way too, and this is the half that marketing pages tend to leave out: smaller differences may still constitute adverse impact where they are significant in both statistical and practical terms, and larger differences may not, where the numbers are small or the difference is unstable. A ratio of 0.83 is not a clean bill of health.

The psychometric literature is blunter still. Roth, Bobko and Switzer, modelling the rule’s behaviour in the Journal of Applied Psychology in 2006, found it produces high rates of false positives at the sample sizes real employers actually have — the rule flags procedures that are not in fact adverse, simply because small groups produce noisy rates. Hauenstein, Holmes and Tison reached a similar conclusion in Public Personnel Management in 2013 when they compared the rule against statistical significance tests. Use the ratio as a trigger for a closer look, never as the finding itself.

Why this calculator also runs a significance test

Because a ratio alone hides how much evidence sits behind it. A ratio of 0.72 on 40 people and the same 0.72 on 4,000 are very different facts, and the four-fifths rule cannot tell them apart. The calculator therefore reports a two-proportion z-test against the reference group alongside the ratio, so you can see whether the gap is larger than sampling noise.

That convention has a legal ancestry: in Castaneda v. Partida and Hazelwood School District v. United States, both decided in 1977, the US Supreme Court treated a difference of more than two or three standard deviations between the expected and observed outcome as enough to call neutral selection into question. Two standard deviations is roughly the p < 0.05 line the calculator marks.

The z-test has its own limit. When expected cell counts are small the normal approximation stops behaving, which is why the calculator raises a small-sample warning below 30 considered or five expected in a cell and points you to Fisher’s exact test instead. A tool that returns a confident p-value on eleven people is worse than no tool.

Where the four-fifths rule applies in 2026

United States, federal. The Uniform Guidelines were adopted in 1978 and 29 CFR 1607.4(D) is still the text on the books as at the verification date below. However, the EEOC carries a final-stage rulemaking on its Unified Agenda — RIN 3046-AB43, “Rescission of Uniform Guidelines on Employee Selection Procedures” — with final action projected for November 2026. Anyone building a compliance process on the guidelines this year should be watching that docket.

Rescinding the guidelines would not make the arithmetic irrelevant, and it is worth being precise about why. Disparate-impact liability under Title VII is statutory, not regulatory; the guidelines are the agencies’ interpretation of it, not its source. A private plaintiff’s claim does not disappear with an interpretive rule.

City and state law is moving the other way.New York City’s Local Law 144 has, since 5 July 2023, required an independent annual bias audit of an automated employment decision tool that reports selection rates and impact ratios by sex and by race or ethnicity, including intersectional categories, with a summary published. California’s FEHA regulations on automated-decision systems took effect on 1 October 2025; they require four years of record retention for selection criteria and ADS data, and they make evidence of anti-bias testing relevant to an employer’s defence — which means the testing is worth doing and worth documenting whatever happens federally.

Elsewhere.Most jurisdictions have no codified four-fifths threshold. India’s equality law, for example, contains no equivalent numerical rule, so the ratio is used there as internal governance rather than as a compliance test. That is still a good reason to compute it: it is the cheapest early warning that a screen is behaving differently for different groups.

What to do when a ratio comes in below 0.80

In order, and the first step is the one most teams skip.

1. Check the numbers before you change anything. Confirm the denominator is the right population and that the groups are large enough for the ratio to be stable. A flag on nine people is usually a flag about nine people.
2. Decompose the funnel. Re-run the calculation stage by stage — sift, assessment, interview, offer. The gap is normally concentrated in one stage, and it is often not the one you assumed.
3. Decompose the instrument. If an assessment is the stage, run it per section or per competency. A battery can show a gap that traces to a single component — a speeded section, a culturally loaded item set — which is a fixable problem rather than an argument about the whole method.
4. Re-examine the cut score. Ask what evidence supports the bar where it currently sits. A cut set by convenience rather than by a documented method such as Angoff is the most common and most fixable cause of an avoidable gap.
5. Ask whether the procedure is job-related. This is the question the Uniform Guidelines actually turn on, and it is a validation question, not a statistical one.
6. Write down what you found and what you did. Under the California rules, evidence of anti-bias testing and of the response to its results is part of the defence. Untested is bad; tested, flagged and ignored is worse.

How AssessAll handles this inside a hiring drive

The same arithmetic runs inside the hiring module as an opt-in monitor, so a team does not have to export a funnel to check it. What it actually does, precisely:

• It is off by default and switched on per requisition. Nothing is collected unless someone deliberately turns it on.
• It asks for self-declared gender and age band only, always optional, always with a “prefer not to say”, on the apply form and at walk-in drive registration.
• It computes each group’s selection rate — reaching shortlisted, offered or hired — and flags any group below 0.80 of the highest rate.
• Groups smaller than 5 are shown but excluded from flagging, for the small-sample reason above.
• Answers are used for aggregate monitoring only and never appear on an individual candidate view or in a report.

What it is not: it is not a Local Law 144 bias audit, which must be performed by an independent auditor and covers race and ethnicity categories this monitor does not collect. AssessAll makes no claim to certify or clear any employer’s selection process. The monitor is an early-warning instrument, and the flag it raises means “look at this”.

Frequently asked questions