Understanding adverse impact: what the four-fifths rule does and does not tell you

What adverse impact means

A neutral selection procedure — a test, an interview stage, a screening question, an education requirement — has adverse impact when it excludes members of a protected group at a substantially higher rate than others. Intent is not part of the definition, and that is the whole point of the concept. A procedure nobody designed to exclude anyone can still exclude, and it does so quietly, because everyone involved is applying the same rule to everyone.

The unit of analysis is the step, not the funnel. Each stage has its own selection rates, and a process that looks balanced at the offer stage can contain a screening step with a large gap that a later stage happens to offset. Aggregating hides exactly what you are looking for. Under the Uniform Guidelines on Employee Selection Procedures, the components of a selection process are examined where the overall process shows impact — and examining each step is good practice regardless of what the total says.

The obligation attaches to use, not to authorship. EEOC guidance is explicit that an employer remains responsible for a selection procedure it did not build, and that a vendor's documentation does not transfer that responsibility. "The tool said so" is not a defence — structurally the same point as AI decision-support, where someone has to own the decision the machine helped prepare.

The four-fifths rule, precisely

The arithmetic is small enough to do on paper. Compute the selection rate for each group at the step: the number selected divided by the number who were considered. Take the highest rate. If any group's rate falls below four-fifths — 80% — of that highest rate, US federal enforcement agencies will generally regard it as evidence of adverse impact. That is all of it, and its simplicity is precisely why it travels further than it should.

Two features of it are easy to lose. The denominator is everyone considered at that step, not everyone who applied to the role; mixing the two describes your funnel rather than your step, and the two can point in opposite directions. And the comparison is always against the highest-performing group at that step, which is not necessarily the group you expected and can change from one cohort to the next — one reason a ratio should never be read without the underlying counts beside it. Worked through on a single screening step:

• Group A: 100 considered, 60 selected — a 60% selection rate, the highest.
• Group B: 50 considered, 21 selected — a 42% selection rate.
• 42 ÷ 60 = 0.70. Below the 0.80 threshold, so this step flags for review.

Two things the rule does not say

First, it does not say a flagged step is unlawful. The Uniform Guidelines describe the ratio as establishing a numerical basis for an initial inference, and state that it speaks only to the question of adverse impact and is not intended to resolve the ultimate question of unlawful discrimination. It opens an inquiry. It does not close one.

Second, it does not say an unflagged step is clean. The guidance notes that where large numbers of selections are made, relatively small differences in selection rates may nevertheless constitute adverse impact if they are both statistically and practically significant. A high-volume screening stage sitting at 0.85 can be excluding more people than a low-volume stage at 0.70.

So the ratio is a screening device with a threshold chosen for administrative convenience, applied to an outcome that has no natural threshold. Treating 0.79 as a violation and 0.81 as an all-clear reads a precision into it that its own authors declined to claim. It is a good place to start looking and a poor place to stop.

Why a flag is a prompt to investigate, not a verdict

What a flag establishes is that a difference exists in the data you have. What it does not establish is why. The gap might reflect the step measuring something the job does not require. It might reflect who applied, which is a sourcing question rather than an assessment one. It might reflect a small sample producing an unstable number. Or it might reflect a genuine, job-related requirement that the role cannot be performed without.

Separating those is the investigation, and the order matters because each stage saves work on the next. Check stability first, since an unstable number makes the rest of the analysis meaningless. Then look at what the step actually measures and whether that is required for the job as it is performed — the job-relatedness question the Uniform Guidelines are built around, and the one a documented chain from role analysis to task to rubric is designed to answer. Then ask whether an alternative that measures the same thing with less impact is available; EEOC guidance names adopting a less discriminatory alternative among employer responsibilities where one exists.

The two failure modes are symmetrical and both are common. One team treats a flag as an accusation, panics, and removes a step that was doing real work — often replacing it with an unstructured interview that has more impact and less evidence behind it. Another treats the flag as a formality, records that it was reviewed, and changes nothing. Neither has investigated anything.

Small groups produce unstable numbers

The four-fifths figure is a ratio of two proportions, and proportions computed on small denominators move a great deal. With eight candidates in a group, one person selected or not can swing the ratio across the threshold in either direction. The number is not wrong; it is carrying more sampling noise than a threshold comparison can absorb, and a threshold comparison has no way to tell you so.

The Uniform Guidelines' own questions and answers anticipate this. They note that a difference of more than 20% in selection rates may not provide a basis for finding adverse impact where the number of people selected is very small, and that agencies will not assume the existence of adverse impact, absent other evidence, where the difference could have occurred by chance. Read the arithmetic first, then read how much data produced it — and display the two together, so nobody has to remember to ask.

In practice that means two habits. Set a floor below which a single-cohort flag is recorded but not acted on alone; flagging every group of four generates noise that trains people to ignore the whole exercise, which is worse than not running it. And accumulate across cycles: a gap that persists over several cohorts is a different object from the same gap in one cohort of eleven, and only the first is worth restructuring a process around.

What an employer can actually do

Most of the useful work sits upstream of the measurement. A step flags because of what it asks, and what it asks was decided when it was designed — which means the monitoring you run afterwards is mostly telling you about design choices you can still revisit.

What is deliberately absent from the list below is adjusting scores by group to make a ratio look better. Group-based score adjustment in employment testing is legally fraught in the United States and is not a remedy to reach for casually. The framework's own answer to impact is job-relatedness and less-impactful alternatives — changes to what you measure and how — not arithmetic applied to the output after the fact.

• Start from the job. Derive each step from a documented analysis of what the role requires, so job-relatedness is a record you already hold rather than a reconstruction under pressure.
• Measure the work, not proxies for it. Requirements that stand in for competence — a degree, years in a title, a fluent-writing bar for a role that does not write — are where impact tends to enter without measuring anything the job needs.
• Score against anchored criteria. Consistent rubrics with worked examples give several evaluators the same standard and shrink the space in which unexamined preference operates.
• Monitor by step and by group continuously, not once at launch, and keep the group sizes visible next to the ratios.
• Test less-impactful alternatives that measure the same construct before defending the current step.
• Keep the record: what each step measures, why, who decided, and what you found when you looked.

What SkillCort shows, and what it deliberately does not

Where a role blueprint links assessments together, SkillCort computes selection-rate ratios and effect sizes across a candidate grouping the organisation chooses, and reports them per group with the group size shown next to the number. Small groups are displayed but not flagged, because a flag on eleven people is a claim the data cannot support and a flag people learn to dismiss is worse than no flag at all.

What it is not: this is not protected-class adverse impact analysis. That requires demographic data collected lawfully and with consent, which is the organisation's decision and its data — not something we infer from a name, a photograph, or anything else. What the panel shows is a disparity check on a field you selected. A difference there is a prompt to investigate, not a finding of discrimination, and we do not present it as one.

The rest is the same discipline the platform runs on elsewhere. The chain from blueprint to task to rubric is what makes job-relatedness demonstrable rather than asserted; reliability is reported where there is enough data to compute it and marked unavailable where there is not; and every decision carries the name of the person who made it. That is content validity and an audit trail — not a claim that an assessment predicts later job performance, which we do not make.

Which leaves the last boundary, about what any tool can promise. What is required of you depends on where you operate, which laws apply, and the specific facts of your process. Nothing here is legal advice, and no assessment product — ours included — can make a hiring process compliant. Software can compute the ratios, hold the record of what each step measures and who decided what, and make alternatives easier to test. The obligation and the decision stay with the employer, and any vendor claiming otherwise is selling something that does not exist.

Sources & further reading

• Uniform Guidelines on Employee Selection Procedures, 29 CFR Part 1607 — Electronic Code of Federal Regulations
• Employment Tests and Selection Procedures (EEOC-NVTA-2007-2) — U.S. Equal Employment Opportunity Commission
• Questions and Answers to Clarify and Provide a Common Interpretation of the Uniform Guidelines on Employee Selection Procedures — U.S. Equal Employment Opportunity Commission