Standard Error of Measurement: Why Your IQ Score Is a Range, Not a Number

You took a test. It said 127. What you now have is not the number 127. What you have is a band, and 127 happens to be its centre.

This isn't a hedge or a disclaimer. It's the single most important technical fact about IQ scores, it's computable to the decimal place, and it's so well established that the United States Supreme Court has ruled that ignoring it is unconstitutional in capital cases.

And yet: virtually every IQ site on the internet reports a bare number. Some advertise precision they cannot possibly possess. This article gives you the actual arithmetic — enough to check any test's claims yourself, including ours. It's the technical foundation under everything in IQ test accuracy.

This is why we won't tell you your IQ is exactly anything. Our free IQ test (~30 questions, ~20 min, instant, no email) gives you an honest estimate — and this page tells you exactly how much slack to put around any number you're given, from us or anyone.

The idea: true score plus error

Classical test theory makes one deceptively simple assumption. Any score you observe is two things added together:

The true score is the stable thing you'd get if you could somehow average infinite testings with no fatigue, no memory, no luck. The error is everything unrepeatable: whether you slept well, whether the specific items happened to suit you, whether you guessed the coin-flip questions right, whether the room was too warm, whether you were anxious for the first ten minutes.

Error isn't bias. It doesn't push in a consistent direction — that's a different problem, covered in IQ test bias. Error is noise: random, sometimes favourable, sometimes not. And crucially, you cannot tell from a single score which way it went for you. That's what makes it error rather than something correctable.

The standard error of measurement is simply the typical size of that noise, expressed in IQ points. It's a property of the test, and it's why every professional score report presents an interval.

The formula — and what it produces

Here it is, and it's worth actually looking at, because everything downstream follows from it:

On the IQ scale the standard deviation is 15, so the only variable that matters is reliability — how consistently the test measures whatever it measures, on a 0-to-1 scale. Feed in a reliability, get out an error band. A 95% confidence interval is then ±1.96 × SEM.

Every figure in the table below is computed from that formula. None of it is estimated, sourced from a vibe, or rounded to look better:

Look at the bottom rows. On a test with reliability .80 — which is what an unvalidated online quiz is plausibly achieving, if it's decent — an observed 120 is consistent with a true score anywhere from 107 to 133. That's the difference between "above average" and "gifted range." The number didn't lie to you. It just never had the resolution you assumed it had.

For context, the Wechsler Adult Intelligence Scale (WAIS-IV) reports Full-Scale internal-consistency reliability of about .98, with short-interval test–retest around .96. The Stanford–Binet 5 reports comparable Full-Scale reliability near .98. These are the ceiling — decades of development, huge standardisation samples, trained administrators. And they still can't do better than roughly ±4.

Testing the "accurate to ±3 points" claim

Some tests advertise precision figures. The formula lets you check them, so let's check one.

Suppose a test claims a 95% confidence interval of ±3 IQ points. Work backwards. A ±3 interval requires SEM = 3 ÷ 1.96 = 1.53. Then solve 1.53 = 15 × √(1 − r), which gives √(1 − r) = 0.102, so 1 − r = 0.0104, so:

That is higher than the WAIS-IV — an instrument with decades of development, a nationally representative standardisation sample, and one-on-one administration by a trained psychologist over a couple of hours.

This is the practical value of knowing one formula: you can audit anybody's precision claim in about thirty seconds, and you should.

When is a score difference actually real?

This is where SEM stops being academic. People compare scores constantly — to a sibling, to a threshold, to their own result from three years ago. Most of those comparisons are noise.

The rule: if two confidence intervals overlap substantially, don't treat the scores as different. On a well-normed instrument (reliability ≈ .95–.97), rough guidance runs:

• Gaps of ≤5 points: within the error band of a single administration. Not interpretable as a real difference.
• Gaps of 6–10 points: check the confidence intervals before concluding anything.
• Gaps of 11–14 points: likely meaningful.
• Gaps of 15+ points (a full SD): clearly meaningful.

And those thresholds get wider as reliability drops. On a reliability-.80 online test, a 5-point gap may not be distinguishable from zero at all.

Comparing across different tests adds error on top of this — different tests measure overlapping but non-identical constructs, use different items, and are normed on different samples. A 122 on one instrument and a 115 on another may be telling you nothing except that they're different instruments. And if the norms are from different eras, the Flynn effect adds a systematic drift on top of the random error. This is worth remembering whenever you read a precise-looking figure on an IQ score chart or a percentile table.

Regression to the mean: why your retest was lower

Francis Galton noticed it in the 1880s and it has been generating false conclusions ever since. Take an extreme score, measure again, and the second score tends to be less extreme. People experience this as a loss. It isn't one.

The mechanism falls straight out of "true score + error." A very high observed score is high for two reasons: your true score is high, and the error happened to break your way. Favourable error is not expected to recur — that's what random means. So the second score keeps the true part and loses the lucky part.

Critically: there is no force pulling anyone toward average. Nothing about the person changed. It's an arithmetic consequence of two imperfectly correlated measurements. The expected retest deviation from 100 is simply the original deviation multiplied by the test's reliability:

Run the numbers. On a test with reliability .90, an observed 145 predicts an expected retest of 140.5. An observed 130 predicts 127. Drop reliability to .80 and that 145 predicts 136 — a nine-point "fall" that is pure statistics. And it runs both ways: an observed 70 on a .90-reliability test predicts an expected retest of 73. The low scorer "improves" for exactly the same reason.

This matters enormously for the rare-score pages people love — how rare is a 140 IQ, an IQ of 145. The further from 100 a score sits, the more of it is likely to be favourable error, and the more it should be expected to move on retest. Extreme scores are the least stable scores, not the most impressive ones.

It's also the single biggest reason bad research goes wrong: select people for being extreme, apply an intervention, watch them regress, and credit the intervention. Any study without a control group subject to the same regression is uninterpretable.

Practice effects: scores rise, ability doesn't

Pushing against regression is the opposite artefact. Take the same test twice and you'll usually score higher — not because you got smarter, but because you've seen it.

The best estimate comes from Hausknecht and colleagues (2007), who pooled 50 studies, 107 samples, and 134,436 participants: an adjusted retest effect of d = 0.26, or roughly 4 IQ points, from one administration to the next. Scharfen, Peters and Holling (2018) found around a third of an SD (~5 points) first-to-second, and about half an SD by the third sitting. Moderators are consistent and sensible:

• Identical forms roughly double the effect (d ≈ 0.40–0.42) versus alternate forms (d ≈ 0.22–0.23) — only identical forms permit item memory.
• Coaching increases it.
• Shorter intervals increase it; the gain decays as memory and format familiarity fade.
• Not all subtests move equally. On the Wechsler scales, gains are largest on speeded and performance tasks and smallest on novel fluid-reasoning tasks like Matrix Reasoning and on vocabulary.

Studies re-administering the WAIS-IV at 3- and 6-month intervals found Full-Scale IQ up around 7 points, with Processing Speed up about 9 — and the authors were explicit that these reflect practice, not genuine intellectual change.

Put the two artefacts side by side and you get the interpretive trap: regression pushes extreme scores toward 100 while practice pushes all scores up. A person who scored 140 and retests at 138 may have experienced substantial regression partly masked by a practice gain. Neither number is "the real one." This is also why claims that brain-training raises IQ deserve suspicion — see brain training and IQ; a pre/post gain on the same test is exactly what these two artefacts produce for free.

Hall v. Florida: when the error band became law

Freddie Lee Hall was sentenced to death in Florida. After the Supreme Court barred executing people with intellectual disability, Hall presented an IQ score of 71. Florida law required a score of 70 or below before a defendant could present any further evidence of intellectual disability — adaptive functioning, school records, upbringing, anything. Hall was one point over. The door was closed.

In Hall v. Florida (2014), the Supreme Court reversed. The Court held that Florida's rule disregarded established practice in two linked ways: it treated an IQ score as final and conclusive evidence of intellectual capacity when experts would consider other evidence, and it relied on a purportedly scientific measurement while refusing to recognise that measurement's inherent imprecision. Professionals, the Court noted, have long agreed IQ scores should be read as a range.

A 71 with an SEM of ~2.5 has a 95% interval of roughly 66–76. Hall's score was entirely consistent with a true score below 70. Florida had taken a number with a known error band and treated it as exact — and a man's life turned on the rounding.

It's a grim way to make the point, but it's the clearest one available: the standard error of measurement is not a technicality. It is the difference between a measurement and a verdict.

How to read your own score

Five things worth internalising:

• Convert the point to a band. If a test doesn't publish its reliability, you can't compute the band — which is itself the most useful thing that silence tells you.
• Never treat a threshold as a line. A 129 and a 131 are the same score wearing different hats. Any cutoff — Mensa's, a gifted programme's, a diagnostic criterion — sits inside somebody's error band.
• Expect extreme scores to move. The further from 100, the more regression to expect. That's not failure; it's arithmetic.
• Discount your retest. Roughly 4 points of any second-sitting gain is practice, more on an identical form.
• Audit precision claims. One formula, thirty seconds. Anything promising ±3 is claiming reliability of .99.

The deeper point: error is not a defect of bad tests that good tests avoid. It's a property of measuring anything, and the best instruments ever built still carry ±4. A test that reports a bare number isn't more precise than one that reports a range. It's just less honest about the same underlying uncertainty. And as the heritability literature and the g factor both suggest from other directions — the number was never the fixed, load-bearing fact people want it to be.

Frequently asked questions

Related reading

← Browse all articles on the blog

References

1. Galton, F. (1886). Regression towards mediocrity in hereditary stature. Journal of the Anthropological Institute of Great Britain and Ireland, 15, 246–263.
2. Lord, F. M., & Novick, M. R. (1968). Statistical Theories of Mental Test Scores. Addison-Wesley.
3. Kulik, J. A., Kulik, C.-L. C., & Bangert, R. L. (1984). Effects of practice on aptitude and achievement test scores. American Educational Research Journal, 21(2), 435–447.
4. Kaufman, A. S., & Lichtenberger, E. O. (2006). Assessing Adolescent and Adult Intelligence (3rd ed.). Wiley.
5. Roid, G. H. (2003). Stanford–Binet Intelligence Scales, Fifth Edition: Technical Manual. Riverside Publishing.
6. Hausknecht, J. P., Halpert, J. A., Di Paolo, N. T., & Moriarty Gerrard, M. O. (2007). Retesting in selection: A meta-analysis of coaching and practice effects for tests of cognitive ability. Journal of Applied Psychology, 92(2), 373–385.
7. Wechsler, D. (2008). Wechsler Adult Intelligence Scale — Fourth Edition: Technical and Interpretive Manual. Pearson.
8. Whitaker, S. (2010). Error in the estimation of intellectual ability in the low range using the WISC-IV and WAIS-III. Personality and Individual Differences, 48(5), 517–521.
9. Calamia, M., Markon, K., & Tranel, D. (2012). Scoring higher the second time around: Meta-analyses of practice effects in neuropsychological assessment. The Clinical Neuropsychologist, 26(4), 543–570.
10. Duff, K. (2012). Evidence-based indicators of neuropsychological change in the individual patient: Relevant concepts and methods. Archives of Clinical Neuropsychology, 27(3), 248–261.
11. Marsden, E., & Torgerson, C. J. (2012). Single group, pre- and post-test research designs: Some methodological concerns. Oxford Review of Education, 38(5), 583–616.
12. Estevis, E., Basso, M. R., & Combs, D. (2012). Effects of practice on the Wechsler Adult Intelligence Scale-IV across 3- and 6-month intervals. The Clinical Neuropsychologist, 26(2), 239–254.
13. Hall v. Florida, 572 U.S. 701 (2014).
14. American Educational Research Association, American Psychological Association, & National Council on Measurement in Education. (2014). Standards for Educational and Psychological Testing. AERA.
15. Scharfen, J., Peters, J. M., & Holling, H. (2018). Retest effects in cognitive ability tests: A meta-analysis. Intelligence, 67, 44–66.