Measure the Barrier, Not the Group: Disability Has No Denominator

Why can't the four-fifths rule answer this?

Because the rule needs two selection rates and disability gives you neither reliably. The four-fifths rule is one paragraph in the 1978 Uniform Guidelines, and it covers race, sex and ethnic group under Title VII ¹. The ADA has no equivalent arithmetic. Even if you borrowed it, disability is not one class, so a barrier that stops a blind applicant is invisible to a deaf one and the pooled ratio averages both into nothing.

The denominator is the second problem. Self-identification is voluntary, it mostly reaches you through the Section 503 federal-contractor form, and the people least likely to tick the box are the people the analysis exists to find. A count built from that describes the applicants comfortable telling you, and it says nothing about the ones who were not. What a low response rate can and cannot support is worked through in whether a self-ID response rate is usable.

None of that makes the exposure theoretical. Screening out is statutory language: 42 U.S.C. 12112(b)(6) makes it discrimination to use a test or other selection criterion that screens out an individual with a disability, or a class of them, unless the criterion as used is job related for the position in question and consistent with business necessity ². Read the scope. It reaches one person, which is why a comfortable group ratio settles nothing.

Name the capability each stage demands

Walk the process end to end and write beside each stage the capability it silently demands. A timed exercise demands speed. A one-way video demands speech a machine can hear and a face pointed at a lens. A gamified test demands vision and fine motor control. Then ask, for each demand, whether the job actually requires it, because that is the question the statute asks too.

The EEOC's 2022 technical assistance on the ADA and algorithmic tools gives the cleanest illustration, a hypothetical built on a programmed cut score: a business requiring 90 percent on a gamified memory assessment rejects a blind applicant who cannot play the game, even though that applicant may have an excellent memory ³. The same document lists three common routes to a violation when an algorithm sits in the stack, and two of them turn on accommodation: failing to provide one that lets a person be rated fairly and accurately, and relying on a tool that rejects somebody who could do the job with one ³. That document came off eeoc.gov in January 2025 and is read here from an archived capture, so it carries an agency's reasoning and no force of law. The mechanism it describes did not go anywhere, and the statute it reads did not change.

The same document is blunt about what a vendor's fairness work covers. Steps taken against Title VII adverse impact on race, sex or national origin are typically distinct from what disability requires, because each disability is unique and an individual can still be screened out however well other disabled people fare on the same assessment ³. A passed audit is therefore not an answer to this question, which is the ground what a bias audit covers and what it misses works over.

Where a model does the screening, the barrier can sit inside the tool. Asked to compare a control CV against the same CV plus a disability-related leadership award, scholarship and panel appearance, GPT-4 put the objectively stronger version first in only 15 of 70 trials, and in none of the ten autism runs ⁴. That is one researcher's CV, one job description and ten trials per condition, so read it as a demonstration rather than a rate. Getting at the same question inside a purchased tool is its own problem, taken up in auditing a screening tool when the vendor holds the data.

What do you actually write down?

Two columns and a date. In the first, every stage of the current process. In the second, the capability that stage requires beyond the job itself, written as a sentence somebody could disagree with. Then a third mark against each demand: whether it appears in the job description as an essential function. A requirement living only in the assessment and nowhere in the role description is the one to argue about first.

Alongside it, one measurement that needs nobody's diagnosis: stage-level abandonment. Pass rate mixes ability with everything else. Started-and-did-not-finish, per stage, is clean, and the comparison that matters is between the stages carrying a capability demand and the stages carrying none.

• Invitation to start. Who opens the link and never begins. Drop-off here is usually about what the invitation said the assessment involves.
• Start to submit. Who begins and abandons. This is where a countdown, a camera requirement or a lockdown shows up.
• Elapsed time to start. A stage where that gap runs long is a stage that takes arranging: equipment, a quiet room, help.
• Reschedule and withdrawal reasons, in the candidate's own words, unedited, and kept out of the hiring file.

None of those four asks anyone to disclose anything, and they stay that way only while they are read as stage totals. One candidate's clock is not evidence about that candidate, and building a private guess about somebody's health out of it walks toward the pre-offer inquiry rules for no gain. They point at a fixable object rather than at a population. Reading a funnel this way, stage by stage, is the general method in measuring bias stage by stage.

Say what a small pipeline cannot support

Write the limit into the same document as the finding. Forty candidates cannot support a significance claim in either direction, and a clean-looking ratio from forty candidates is not evidence that nothing is wrong. On that second point the agencies who wrote the four-fifths rule are explicit: in 1979 they called it a rule of thumb, not intended as a legal definition, and said it does not resolve the ultimate question of unlawful discrimination ⁵.

So when the instruction to produce a disability number arrives, the answer that survives scrutiny is the honest one. The group rate is unavailable for this class, here is the barrier audit that ran instead, here is what changed, here is the date. That document is stronger than a ratio nobody can defend, because it names an object somebody can fix. The small-sample version of the same wall is worked through in too few hires to measure bias.

The deciding question is whether the job needs the capability at all, which is the same question the statute puts to an employer, so the answer belongs in writing and in front of counsel before it becomes policy. Where the job does not need it, removing the demand is right. Where it does, deleting the demand was never the move, and the question becomes whether somebody could do the work another way, which is the ground offering an accommodation without forcing a disclosure covers. The person currently doing the job knows whether the ninety-minute clock is real. Ask them before deleting the timer, write down the answer, and keep it, because the next person to ask will be a candidate who deserves a straight reply.

See a sample report >