Why a test score is a range: the standard error of measurement explained

Every psychometric score carries some measurement error, and the standard error of measurement tells you how much. Understanding it changes how you read a score: not as a single exact point, but as a range within which the candidate's true level most probably lies.

No test measures perfectly

When a candidate scores 55 on a reasoning test, it is tempting to treat that number as a fixed fact about them. It isn't. If the same person could sit an equivalent version of the test the next day, without any practice effect, they would almost certainly get a slightly different score. Their concentration, the particular items they happened to see, a lucky guess or a misread question all push the observed score a little above or below the level we are actually trying to measure.

Classical test theory describes this simply: an observed score is a true score plus error.^[1,2] The standard error of measurement (SEM) is the typical size of that error for a given test. It is expressed in the same units as the scores, which is what makes it so useful in practice. The joint professional Standards for Educational and Psychological Testing treat information about the precision of scores as something test developers should provide and test users should understand before interpreting results.^[3]

From reliability to the standard error of measurement

The SEM is calculated from two things published in any good technical manual: the spread of scores (the standard deviation) and the test's reliability. The formula is the standard deviation multiplied by the square root of one minus the reliability.^[1,2] A perfectly reliable test would have an SEM of zero; a test with no reliability at all would have an SEM as large as the standard deviation itself.

A worked example makes this concrete. Many tests report results on a T-score scale, with a mean of 50 and a standard deviation of 10. If the test's reliability is .90, the SEM is 10 × √0.10, or about 3.2 points. If the reliability is .80, the SEM rises to about 4.5 points. That difference in reliability, which can look modest in a manual, changes the precision of each individual score considerably. It is one reason why reliability is central to judging whether a test is fit for purpose, a topic we cover in what makes a psychometric test valid.

Confidence bands: reading a score as a range

Because measurement error behaves broadly like a normal distribution, the SEM can be used to build a confidence band around a score. Roughly 68% of the time the true score will lie within one SEM of the observed score, and about 95% of the time within 1.96 SEMs.^[1] Using the example above, with an SEM of 3.2, a T-score of 55 is best read as most probably lying somewhere between about 49 and 61.

Two technical refinements are worth knowing. Strictly, a band for the true score should be centred on an estimate pulled slightly towards the group mean rather than on the observed score itself, although centring on the observed score is a reasonable shortcut when reliability is high and the score is not extreme.^[1] And a single SEM is an average: precision usually varies somewhat across the score range, which is why some manuals report conditional standard errors at different score levels.^[1,3]

Why small differences between candidates shouldn't decide outcomes

The practical consequence for hiring is significant. When you compare two candidates, both scores contain error, so the uncertainty about the difference between them is larger than the uncertainty about either score alone. The standard error of a difference between two people on the same test is about 1.4 times (√2) the SEM.^[1] With an SEM of 3.2, that is about 4.5 points, which means two candidates would need to differ by roughly 9 T-score points before you could be 95% confident that one genuinely outperformed the other.

In other words, a candidate on 54 and a candidate on 57 are, as far as the test can tell, indistinguishable. Ranking them strictly by a three-point gap, or rejecting one and inviting the other, gives the number a precision it does not have. This is also why fixed pass marks are hard to defend: two people separated by a single point either side of a line are treated as different, when the test cannot reliably tell them apart. Overlapping bands are a signal to treat candidates as broadly equivalent on that measure and to let other evidence, such as interviews, work samples or other tests, do the differentiating. Non-overlapping bands suggest a real difference, though even that is a probabilistic judgement rather than a certainty.^[1]

Why scores are reported in bands

Reporting scores in broad descriptive bands, for example Very Low, Low, Average, High and Very High, is a direct response to measurement error. A band communicates the level of performance at roughly the precision the test can support, and discourages over-reading small numerical differences. It also makes the comparison group explicit, since bands are normally defined relative to a norm group, which is why the choice of norm group matters so much.

Banding has been debated in personnel selection, particularly in the United States, where some approaches treat all scores within a statistically defined band as equivalent and then choose between candidates on other grounds. Proponents argue this respects the limits of measurement; critics argue it can discard useful information if bands are drawn too wide.^[4] Both sides agree on the underlying point: scores are estimates, and decisions should reflect their precision.

For employers, the lesson is straightforward. Ask for the reliability and SEM of any test you use, interpret each result as a range rather than a point, and combine test scores with other evidence from a structured process built on job analysis. A single number on a single test should rarely be the deciding factor on its own.