What is Cut score?

Also called Pass mark, Threshold, Cutoff

A cut score is the point on a score scale at which a decision changes — pass or fail, shortlist or reject, certified or not. It is a policy choice made by the organisation using the test rather than a property of the test itself, and moving it changes who is selected without changing anyone's score.

It is a decision, and it belongs to you

Vendors supply scores; the organisation supplies the threshold. That division matters legally as well as practically, because the cut score is what determines selection rates, and therefore what determines whether a procedure shows adverse impact.

A cut score should be recorded with the reasoning behind it: what standard it represents, what method set it, who was in the room, and when it was last reviewed. 'It was 70% when I joined' is the most common answer to that question and the least defensible.

Where cut scores go wrong

The first failure is treating a round number as a standard. Nothing about a test makes 70% meaningful; the figure is inherited from school marking conventions and it survives because nobody re-examines it.

The second is ignoring measurement error at the boundary. If the standard error of measurement is four points, a candidate who scored 68 against a cut of 70 is not reliably below the standard, and a re-sit could easily place them above it. That is an argument for a review band around the cut, not for pretending the boundary is sharp.

The third is setting the cut once and never re-cutting it against what happened downstream. If everybody who passed also passed probation, the bar is not doing work and may simply be shrinking the pool.

Moving the cut moves the selection ratio, and that is what changes the value of the test

A cut score is usually discussed as a standard: does this candidate meet the bar. It is also, and simultaneously, the control that sets your selection ratio — the share of applicants you end up hiring — and the selection ratio is the number that decides how much your instrument is worth. A test can only sort people you are in a position to reject, so the more of the pool the cut lets through, the less any amount of validity can do.

The arithmetic is unforgiving in one direction in particular. At a base rate of 50%, an instrument with a validity of 0.31 lifts the success rate among hires to 71.5% when the cut admits one applicant in ten, to 60.0% at one in two, and to 56.2% at seven in ten. Dropping a cut to fill headcount does not cost you a little of the instrument's value; between those first and last figures it costs roughly three quarters of it, silently, while the test carries on producing the same scores.

Two consequences worth writing into a cut-score record. First, note the selection ratio the cut produced, not only the pass rate: they are the same number seen from opposite ends and only one of them is usually minuted. Second, be sceptical of vendor examples drawn at very low selection ratios — Schmidt and Hunter observed in 1998 that actual selection ratios are typically in the 0.30 to 0.70 range, which is the unflattering half of that scale.

Move the cut score by less than the test's own error, and see how many people change side

A scale with a standard deviation of 10 and a reliability of 0.85, so the [standard error of measurement](/guides/glossary/s/standard-error-of-measurement) is 3.87 points. Scores normally distributed.
Cut at the mean: 50 candidates in 100 pass. Move the cut UP by one standard error and 34.9 pass. Move it DOWN by one standard error and 65.1 pass.
That is a span of 30 candidates in every 100 — for a movement of the line smaller than the error already attached to every score on it.
Now a selective cut, one standard deviation above the mean, which is the shape most screens actually use. At the line, 15.9 in 100 pass. One standard error up: 8.3. One standard error down: 27.0.
Same instrument, same candidates, and the pass rate is 3.3 times larger at one end of that span than the other.
Repeat it on a ten-item form, where the reliability is about 0.69 and the standard error is 5.58. At the same selective cut: 6.0 in 100 at the top of the span against 32.9 at the bottom — a factor of 5.5.

The point is not that cut scores are arbitrary. It is that the sensitivity of the outcome to the line is computable, and it is larger than almost anybody assumes — so a cut score published without the scale's standard deviation and a reliability coefficient beside it cannot be audited by the employer who is liable for it. Two figures turn a threshold into something a reader can check. A shorter test widens the span rather than narrowing it, which is the [ten-item arithmetic](/guides/how-reliable-is-a-10-question-aptitude-test) run at the boundary instead of at the score.

Not the same as standard setting

A cut score is the number. Standard setting is the documented method that produced it — Angoff, bookmark, contrasting groups, or a criterion study. The distinction is the whole audit trail: two defensible methods run by competent panels on the same 20-item test and the same 190 examinees produced cuts of 17.67 and 18.82, and 106 people passed against 88. A vendor who supplies a cut score without naming the method has given you a number with no provenance, and the eighteen people between those two figures are the reason that matters.

Standard setting

How AssessAll handles it

AssessAll Certified decides its levels from per-pillar cuts with a documented algorithm rather than a single overall percentage: a weight-normalised composite across pillars, pillar floors, a borderline margin and a near-miss retake band, with a nominated pillar carrying a floor for the highest level (lib/certified/wc-levels.ts). A section flagged for review is excluded from the composite rather than scored as zero, and no certificate is issued while any section is unscored.

Sources

More terms beginning with C

Construct validity Criterion validity Cronbach's alpha

Check a selection process against the four-fifths rule

Free, no signup, computed in your browser — with the remedy, not just the verdict.

Open the calculator

Last reviewed 2026-09-30