The standard error of measurement (SEM) is the typical distance between the score a candidate actually got and the score they would average across many repeat sittings of an equivalent test. It equals the score standard deviation multiplied by the square root of one minus the reliability coefficient. A score is therefore a band, not a point.
That sentence is uncontroversial in psychometrics and almost universally ignored in hiring software. Scores arrive as rank-ordered integers, and the integer is what gets argued about in the shortlisting meeting. The SEM tells you how much of that ordering is real.
This is not the reliability-versus-validity question. Reliability tells you whether the instrument is consistent; validity, whether it measures what you care about. The SEM answers something narrower and more operational: how far wrong is this person's number likely to be, in the units you are about to decide in.
Where the band comes from
The formula is SEM = SD × √(1 − reliability), and it behaves the way intuition says it should. It comes straight out of classical test theory and underpins the precision chapter of the *Standards for Educational and Psychological Testing*. Perfect reliability gives a SEM of zero. Zero reliability gives a SEM equal to the full spread of scores — the test tells you nothing about any individual.
Take a realistic corporate assessment: a 100-point scale, a score standard deviation of 12 points in your applicant pool, and an internal consistency reliability of .85, which is a respectable figure for a well-built 40-item test.
- SEM: 4.65 points
- 68% band (±1 SEM): roughly ±4.7 points
- 95% band (±1.96 SEM): roughly ±9.1 points
So a candidate who scored 68 has a 95% band running from about 59 to 77. A candidate who scored 72 has one running from about 63 to 81. Those are overlapping bands. Treating the four-point gap as a difference in ability is not a judgement call; it is a misreading of the instrument.
Drop the reliability to .70 — common for short screeners and for single-competency subscales pulled from a longer battery — and the SEM rises to 6.57 points, a 95% band of ±12.9. Conversely, for a 95% band no wider than ±5 points on that scale you need a reliability of about .955. Few commercial assessments reach that at subscale level, and almost none reach it in fifteen minutes.
What the big published tests actually report
The large testing programmes publish these numbers, which makes them a useful calibration point.
GRE General Test — ETS's published reliability and SEM table, averaged across multi-stage tests delivered September 2023 to June 2026:
- Verbal Reasoning: reliability .88, SEM 3.17
- Quantitative Reasoning: reliability .94, SEM 2.56
- Analytical Writing: reliability .77, SEM 0.40
Verbal and Quantitative use a 130–170 scale — 41 points wide. A SEM of 3.17 is 7.7% of the whole scale, on a measure with .88 reliability built by one of the best-resourced testing organisations in the world; the 95% band is roughly ±6 points. Admissions committees routinely distinguish a 158 from a 162.
IELTS — the published 2024–2025 test statistics are even more striking, because the reporting scale is coarse:
- Listening: Cronbach's alpha averaging .90, SEM 0.43 bands
- Academic Reading: alpha .92, SEM 0.36 bands
- General Training Reading: alpha .91, SEM 0.47 bands
- Overall composite: reliability .97, SEM 0.17 bands
Employers and universities set IELTS cut-offs in half-band steps — 6.0, 6.5, 7.0. The Listening SEM of 0.43 is very nearly the size of the step the decision is made in. The composite SEM of 0.17 is much tighter, which is the real lesson: the four-skill aggregate supports a half-band decision far better than any single skill does. Inter-rater reliability is reported at .90 for Speaking and .92 for Writing — respectable, and still not zero error.
The band widens when you compare two people
This is the part most score reports omit. When you compare two scores, you are accumulating the error in both. For two measurements with equal precision, the standard error of the difference is SEM × √2.
ETS publishes this separately, which is good practice:
- Verbal Reasoning: SEM 3.17 for an individual score, 4.48 for a score difference
- Quantitative Reasoning: 2.56 individual, 3.62 for a difference
- Analytical Writing: 0.40 individual, 0.57 for a difference
Apply it to the corporate example above. With a SEM of 4.65, the standard error of the difference between two candidates is 6.57 points, and the 95% interval on that difference is ±12.9 points. Two candidates 10 points apart on a 100-point assessment are, on that evidence alone, not reliably different.
The same mathematics applies to pre- and post-training gain scores, which is why a few points of movement after a workshop is usually noise. It applies to within-person profiles too — "she's stronger on analysis than on communication" — because that difference carries the error of both subscales, and subscales are always less reliable than the total.
Near a cut score, precision becomes a classification question
If your decision is pass/fail rather than ranked, the useful statistic is not the band but the classification rates: how often the process puts a candidate in the right category (accuracy), and how often a second equivalent sitting would put them in the same category (consistency).
Florida's statewide assessment technical report for 2020–21 gives a well-documented example from education testing. At a single cut point, performance is good: Grade 3 English Language Arts showed classification accuracy of .936 and consistency of .911 at one of its cuts, and the report describes consistency as "around .9 or slightly below .9" for its lower cuts. But the overall accuracy index — correct placement across all five performance levels at once — ran from .721 to .805 across subjects and grades.
A single, well-separated pass/fail line can be classified very consistently. Carving the same scale into five labelled bands cannot: every cut line you add is another place where candidates sit inside the SEM of a boundary. That is a general property of measurement precision, not a criticism of that programme — and not a number to import into hiring, because employment assessments differ in length, stakes and population. You need your own figures.
It is also the reason AssessAll reports AI proctoring output as integrity bands rather than a single suspicion score: a classification with a stated width is harder to over-read than a decimal.
What the professional standards expect
The Standards for Educational and Psychological Testing (AERA, APA and NCME, 2014 edition) devote their second chapter to reliability and precision, and expect test developers to report standard errors of measurement and, where precision varies across the scale, conditional SEMs at several score levels — because precision is usually worst at the extremes. A Joint Committee to revise the Standards was named in June 2024; the 2014 edition remains the operative one, with no publication date announced for the next. The SIOP Principles (5th edition, 2018) carry the same expectation into employment selection, and the EEOC's guidance on employment tests treats any test used in a selection decision as a selection procedure that has to be defensible.
In practice, five things follow:
- Get the SEM for your own population. A vendor's manual reports it for their norm sample. Your applicant pool may have a narrower spread, which lowers the SD and the reliability estimate together.
- Report a band, or at least a rounded band label. Reporting "72" invites a comparison that "mid-band, 63–81" does not.
- Never rank-order inside one SEM. If your shortlist rule is "top 20 by score", check the gap between rank 20 and rank 25 against the standard error of the difference.
- Use the composite for the decision and the subscales for development. The IELTS numbers show the pattern: the aggregate is far more precise than any component.
- State the conditional precision at your cut. Precision at the boundary is what the decision actually rests on.
If an assessment's documentation does not state a SEM in the units you will decide in, that is the first question to ask — of a vendor, or of whoever is running your custom assessment design.
The takeaway: reliability is a property of the test; the SEM is a property of the decision, telling you in your own scoring units how large a gap has to be before it means anything. Publish the band, and most arguments about two-point differences stop happening.