The four-fifths rule.

Most concepts in employment law resist arithmetic. Adverse impact is the exception: divide one group's selection rate by another's, and if the quotient falls below 0.8, U.S. enforcement agencies generally treat the gap as evidence that the hiring procedure producing it needs defending. Every number the division requires already sits in the applicant tracking system, applicants in and offers out, held by the organization itself. Yet most organizations have never run the division, and the first person to run it is often opposing counsel.

That asymmetry is the subject of this article. The four-fifths rule is where selection science and employment law meet most directly, because what the law asks of a procedure that trips the ratio — evidence of job-relatedness and a search for less impactful alternatives — is precisely what the research literature on selection methods studies. An organization that understands the rule, the research behind the defense, and the design strategies that move both numbers is practicing measurement, on data it already owns.

What follows walks through the rule itself with the division worked in full, the two-part defense the framework asks for, the inconvenient pairing the research calls the diversity–validity dilemma, the strategy menu that loosens it, and the monitoring habit that turns a litigation surprise into a routinely watched number.

A legal exposure you can compute in a spreadsheet

The rule lives in the Uniform Guidelines on Employee Selection Procedures, adopted jointly in 1978 by the Equal Employment Opportunity Commission, the Civil Service Commission, and the Departments of Labor and Justice (EEOC et al., 1978). The Guidelines apply to any procedure used to make an employment decision, tests, interviews, résumé screens, minimum qualifications, and they define the operative quantity plainly: a selection rate is the share of applicants from a group who pass the procedure. The comparison they prescribe is the four-fifths rule. A selection rate for any race, sex, or ethnic group that is less than four-fifths of the rate for the group with the highest rate is generally regarded by the enforcement agencies as evidence of adverse impact.

The operation deserves to be seen once in full, on illustrative numbers computed for this article. Suppose applicants from Group A file 500 applications and 150 advance, a selection rate of 30%. Applicants from Group B file 300 and 60 advance, a rate of 20%. The ratio is 20 divided by 30, roughly 0.67. That is below 0.8, and impact is indicated. Had Group B's rate been 25% instead, the ratio would have been 0.83, above the line, and the rule of thumb would be satisfied. Figure 1 draws both cases. The pools are invented; the operation is the real one, and it is the whole computation.

Three things the rule is not. It is not a statistical test; the Guidelines frame four-fifths as an enforcement rule of thumb, a screen agencies use to decide where to look harder, and psychometricians have formalized the distinction, since significance tests for selection-rate differences exist and can disagree with the ratio in small samples (Morris & Lobsenz, 2000). It is not a quota, because it compares rates of advancement, not composition of the workforce, and it says nothing about whom to select. And an indicated ratio is not a verdict; it is a shifted burden. The procedure is questioned, not condemned, and the organization is invited to produce a specific kind of answer.

The rule's crudeness is also why it has lasted. A plaintiff and an employer disagree about nearly everything by the time selection rates are examined; the division itself is not among the disputable items. Because the quantity is administrable, computable identically by an agency, a litigant, and the employer's own analyst from the same funnel export, the four-fifths comparison works as a shared table of contents for the harder argument that follows. It tells both sides which procedures, which gates, and which cohorts the real dispute will be about.

One boundary on scope. The four-fifths rule is a creature of the U.S. framework (the Uniform Guidelines operate in the enforcement context of Title VII, where litigation speaks of disparate impact in hiring), and other jurisdictions draw their discrimination standards differently, though the discipline of comparing selection rates travels well. A newer statutory layer aimed specifically at automated and AI-driven hiring tools now mandates similar audits in some places; that terrain belongs to a companion article on AI hiring regulations. And a threshold matter of practice: whether a particular procedure survives legal challenge in a particular case is a judgment for employment counsel, not for a research briefing. What the framework asks for, though, is public knowledge, and it is the next section.

The ratio is a trigger, and what it triggers is evidence

When impact is indicated, the Guidelines ask for two things. The first is validation: evidence that the procedure is job-related, meaning it measurably predicts performance on the job in question or captures knowledge, skills, or behaviors the job demonstrably requires (EEOC et al., 1978). This is not a legal invention grafted onto measurement; it is measurement. The evidence that satisfies a validation requirement (validity coefficients, job analyses, documented links between what is tested and what the work is) is the same evidence a competent buyer should demand before using an instrument at all. Our companion review of what predicts job performance examines that evidence base method by method, and the science behind this platform documents the same standards applied to one battery.

Job-relatedness can be shown along more than one route. An employer can present evidence that scores on the procedure predict later performance in the role, or that the procedure representatively samples the knowledge, skills, and behaviors the job itself requires; the Guidelines describe validation in these terms rather than prescribing a single method (EEOC et al., 1978). Two features of the requirement deserve emphasis. Validation attaches to the use, not to the instrument in the abstract: an assessment defensible for one role can be indefensible for another if the construct it measures has no demonstrated bearing on that job's work. And purchasing a published instrument does not transfer the obligation. The employer using the procedure answers for it, which makes the vendor's documentation a beginning rather than a defense.

The second requirement is less widely known and more consequential for design. Where two procedures would serve the employer's legitimate interest comparably well, and one produces less impact, the Guidelines expect the alternative to have been considered (EEOC et al., 1978). This is the alternatives question, and it converts selection design from a pure optimization problem into a documented search. An organization that chose its instrument from a menu of one has a weaker answer than an organization that can show what else it weighed, what the validity trade-offs were, and why it landed where it did.

Put the two requirements together and the shape of a defense becomes clear. It is a file, not an argument improvised later: the job analysis, the validity evidence for each gate in the funnel, the alternatives considered at design time, and the selection-rate record itself. Organizations rarely maintain that file because nothing in day-to-day hiring demands it. The Guidelines, in effect, assume a discipline that ordinary practice has never adopted, and the gap between the two is where the exposure compounds quietly, cohort after cohort.

The diversity–validity dilemma is about instruments, not people

The alternatives question sends organizations to a body of research that studies exactly it: how selection methods differ in validity, how they differ in the average score gaps they produce between groups, and whether a battery can be designed that does well on both. The research literature calls this the diversity–validity dilemma, after a review by Ployhart and Holtz (2008) in Personnel Psychology that catalogued the strategies available for reducing subgroup differences and adverse impact while protecting validity. A companion overview in the same issue situates the dilemma in its legal context, and its point deserves restating here: the tension between minimizing impact and maximizing validity is a legal problem as much as a psychometric one, which is why the two-part defense of the previous section reads like a research agenda (Pyburn, Ployhart, & Kravitz, 2008).

The empirical backdrop can be stated without printing a single group comparison, and in this article it will be. A meta-analysis of ethnic group differences across employment and educational samples found that average subgroup differences are largest on traditional cognitive ability measures and that their size varies substantially with context, sample, and measurement conditions (Roth, BeVier, Bobko, Switzer, & Tyler, 2001). One discipline governs how to read that finding, here and everywhere: observed differences describe scores produced by particular instruments under particular conditions, the formats, content, reading demands, stakes, and applicant pools of specific tests, and they are not measurements of the capability of any group. The rest of this article treats them accordingly, as properties of instruments in context, which is also the only framing under which a hiring team can act on them, since instruments are what a hiring team controls.

The validity side of the ledger is better known. Sackett, Zhang, Berry, and Lievens (2022), in the re-analysis that re-estimated the field's meta-analytic validity evidence (they applied less aggressive range-restriction corrections than earlier syntheses had), report structured interviews at .42, job knowledge tests at .40, biodata at .38, work samples at .33, general cognitive ability at .31, and conscientiousness at .19. If the instruments with the largest typical differences were also the weakest predictors, there would be no dilemma; the inconvenient fact is that the pairing is mixed. Some strong predictors carry larger typical differences, and some instruments with smaller typical differences bring real validity of their own.

The mixed pairing is what makes both easy exits fail. An organization that responds to an indicated ratio by discarding its strongest predictors outright has traded away measured validity, and with it real predicted performance, for an improvement it could likely have achieved at lower cost through design. An organization that responds by declaring its instrument valid and stopping there has answered the framework's first question while ignoring the second, since the alternatives inquiry presumes exactly the design search it skipped. Read whole, the framework rejects both reflexes. It asks the organization to hold two variables at once, which is an engineering posture, and engineering is what the strategy families in the next section amount to.

Figure 2 lays the two dimensions against each other, deliberately without numbers. Read it as a map of the design space rather than a scoreboard: instruments toward the right predict better, instruments toward the bottom typically produce smaller average differences as used in practice, and the plain reading is that no single instrument owns the bottom-right corner outright. What the map rewards is combination. A battery anchored by instruments in the favorable region, with careful use of those outside it, is a different legal and scientific position than a single high-difference instrument standing alone.

Five strategy families move both numbers

Ployhart and Holtz (2008) organized the interventions the literature has tested into families, and the durable lesson of their review is that the dilemma responds to design. No strategy eliminates it, several meaningfully reduce it, and every one carries a cost that should be accepted knowingly rather than discovered later. Figure 3 lists the families with their price tags attached.

The first family is combination: add valid predictors that typically produce smaller differences rather than dropping valid predictors that produce larger ones. Broadening what is measured, adding structured behavioral measures or biodata alongside an ability test, tends to improve the battery's overall impact profile while holding or improving validity, at the cost of testing time and instrument spend. The second family removes construct-irrelevant variance, the portion of score differences produced by demands the job never makes: dense reading in a role that involves little reading, tight time limits where speed is not the construct. Cutting those demands requires a job analysis rigorous enough to say what the construct actually is, which is a cost, and also exactly the artifact a validation file needs anyway.

The third family is structure: fixed questions, anchored rating scales, trained assessors, consistent scoring. Structure is the rare intervention that tends to move validity up and typical differences down at the same time, and it is cheap; its price is discipline. A companion article examines how structured hiring reduces bias in depth. The fourth family operates before assessment begins: expand and diversify the recruiting pool, because selection-rate ratios are computed on who applies, and a procedure's impact profile can improve or deteriorate with no change to the procedure itself. The fifth family covers scoring and combination choices, weighting predictors and banding scores, where the technical and legal trade-offs are real and contested, and design decisions deserve documentation at the moment they are made.

The families also differ in where they naturally sit in a deployment sequence. Structure and construct-irrelevant cleanup are the plausible first moves for most organizations: both are inexpensive relative to what they return, both tend to strengthen validity evidence rather than merely defend it, and both produce artifacts, the job analysis and the standardized protocol, that a validation file needs regardless. Combination follows as batteries are rebuilt, since adding instruments is a procurement cycle rather than a policy change. Pool expansion runs in parallel on the recruiting side. Weighting and banding come last and most carefully, because they are the family in which measurement choices and legal exposure intertwine most tightly.

Adverse impact is a property of the cohort, not the procedure

The four-fifths rule is computed on a specific applicant pool over a specific period, and this makes it unlike most properties organizations attribute to their hiring tools. Validity generalizes; a ratio does not. A procedure whose ratios cleared 0.8 comfortably last year can trip the threshold this year with no change to a single item, because a recruiting channel shifted, a job posting was rewritten, a competitor's layoff flooded the pool, or a campus program wound down. The instrument held still and the cohort moved.

The operational consequence is that checking once is nearly the same as never checking. Selection-rate ratios belong in the same category as the funnel metrics talent teams already watch continuously, time-to-fill, stage conversion, offer acceptance, and in the same operational posture organizations already apply to assessment security: a standing process with an owner and a cadence, not an annual scramble. Figure 4 draws the loop. Each cohort's outcomes are collected, per-group rates computed, ratios compared against the threshold, and findings routed into procedure or recruiting adjustments before the next cohort opens. Run this way, a deteriorating ratio is an early warning that arrives while the fix is cheap, which is a different experience of the same number than discovery.

Granularity matters as much as cadence. A single funnel-wide ratio can mask offsetting gates: a screening stage that suppresses one group's rate can be hidden by a later stage that passes nearly everyone who survived it, which is why the computation belongs at every decision point rather than only at the offer. Small cohorts call for the opposite caution. Ratios computed on a handful of applicants swing for reasons that have nothing to do with the procedure, and small samples are exactly where the four-fifths ratio and the formal tests part company most often (Morris & Lobsenz, 2000). Serious monitoring therefore pairs the ratio with sample-size awareness: the sober response to a volatile small-cohort ratio is accumulation and attention, not alarm.

Run your ratios before someone else does

The practical program follows from everything above. It requires treating the selection-rate record as a first-class output of the hiring funnel, not a statistician on staff.

• Compute per-group selection rates at every gate, now. Not only offers: the résumé screen, the assessment cut, the interview shortlist. Each gate is a selection procedure under the Guidelines, including the one most organizations have never validated at all, examined in our review of résumé screening accuracy. A funnel that looks fine at the offer stage can be concentrating its impact two gates earlier.
• File validity evidence with the procedure, not in a drawer. The job analysis, the instrument's validity documentation, and the link between the two should live where the procedure lives, current and retrievable, so that job-relatedness is a record rather than a reconstruction.
• Document the alternatives you considered. The alternatives question is answered best contemporaneously: what else was evaluated, what the validity and impact trade-offs appeared to be, why the chosen design won. A page written at design time outweighs a narrative assembled under deadline years later.
• Give the ratio an owner and a cadence. A quarterly review of selection-rate ratios by cohort, with below-threshold findings routed to a named owner, converts the four-fifths rule from a litigation artifact into a dashboard metric. The organizations that fare best under scrutiny are the ones for whom the scrutiny is redundant.

The deeper shift is in initiative. An organization that runs its own ratios holds the timeline: it finds the drifting cohort in a quarterly review, weighs its strategy menu while every option is still open, and writes the record that a defense would later need. An organization that never runs them has not avoided the question.