What is Banding?
Also called Score banding, Confidence banding
Banding reports a score as a range rather than a point, on the reasoning that differences smaller than the measurement error are not real differences. Candidates whose scores fall inside the same band are treated as equivalent, and any ordering among them is decided on other evidence rather than on the score.
The argument for it
Every score carries error. If the standard error of measurement on a scale is four points, then a candidate on 71 and a candidate on 68 are not distinguishable by that instrument, and a shortlist built by sorting the column has manufactured a distinction the data cannot support.
Banding makes that explicit. It also changes behaviour: a reader handed a band asks what else they know about the two people, whereas a reader handed 71 and 68 simply picks 71.
The honest objection
Banding does not remove the arbitrary line, it moves it. Two candidates either side of a band boundary may differ by one point and be treated as different, which is the same problem the bands were introduced to solve, now occurring less often.
The defensible position is that fewer arbitrary lines are better than many, and that a band boundary is at least placed where the measurement error says it can be. What is not defensible is publishing bands and then quietly ranking within them.
The rule you use inside a band is regulated, in the United States at least
Banding hands the decision inside a band to some other rule, and which rule that is turns out to be a legal question as well as a measurement one. 42 U.S.C. §2000e-2(l), added by section 106 of the Civil Rights Act of 1991 and read at source on 16 September 2026, makes it an unlawful employment practice “to adjust the scores of, use different cutoff scores for, or otherwise alter the results of, employment related tests on the basis of race, color, religion, sex, or national origin”.
A banding scheme whose within-band selection rule refers to one of those characteristics is therefore within the words of that provision, whatever the intent behind it. A scheme that decides inside a band on independent job-related evidence — a work sample, a structured interview, a reference on a specific behaviour — is not doing the thing the provision names. The distinction is the selection rule, not the band.
This is United States federal law, it is not legal advice, and other jurisdictions regulate the question differently. The general point survives the jurisdiction: a band with no stated rule for deciding inside it is not a fairness measure, it is a place for an unexamined rule to live.
How many people end up in one band, and the thing banding does not do
The number nobody puts next to the word banding is how many candidates it declares equal. Campion and six co-authors — Outtz, Zedeck, Schmidt, Kehoe, Murphy and Guion — put it in print in Personnel Psychology in 2001, in a paper structured as ten questions with the authors disagreeing on the record. Read at source on 28 September 2026: with a test reliability of 0.80 and the two-SED band width Cascio and colleagues recommended in 1991, taking the top score as the one at the 99.9th percentile, the resulting 95% band contains 38% of the score range and about 25% of the job applicants.
A quarter of everyone who applied, treated as having equal true scores. That is the correct consequence of the arithmetic above and it is also the reason the choice of within-band rule carries so much weight: the rule is not a tiebreaker applied to two people, it is the actual selection procedure for a quarter of the field.
The second finding in the same paper is the one vendors leave out. Banding is frequently sold as a diversity measure. Frank Schmidt's answer in that paper is that banding without minority preferences — traditional or SED — typically does little to reduce adverse impact, and preferences inside the band are the part courts have struck down. Outtz and Zedeck, arguing the other side in the same document, agree the effect depends on being able to select within bands on factors that do not show group differences. Both sides therefore agree on the mechanism: the band itself changes nothing, and everything turns on the rule you put inside it.
If a vendor offers banding as an adverse-impact remedy, that is the sentence to quote back. Ask what the within-band rule is, what evidence supports it as job-related, and what the band width is at their published reliability.
How wide is a band, actually — and the reliability below which it swallows the whole scale
- The width of a confidence band is not a matter of taste. It is fixed by the test's own error: Band width = C × SD × √(1 − rxx) × √2, where SD is the standard deviation of observed scores, rxx is the reliability, √2 converts a one-person interval into a two-person one, and C is the normal deviate for the confidence level — 1.96 for 95%.
- Take a scale with SD = 10, and read the band width straight off the reliability:
- rxx = 0.90 → band = 8.8 points (0.88 SD). rxx = 0.80 → band = 12.4 points (1.24 SD). rxx = 0.70 → band = 15.2 points (1.52 SD). rxx = 0.60 → band = 17.5 points (1.75 SD).
- Set the band width equal to one SD and solve for the reliability: 1.96 × √(2 × (1 − rxx)) = 1 gives rxx ≈ 0.87.
So any instrument with a reliability below roughly 0.87 produces a 95% band wider than the entire standard deviation of its own score distribution. That is not a quirk of this example — it is the published objection. In Campion et al. (2001), John Kehoe writes that the 1.96 SDdiff bandwidth is frequently equal to or larger than the standard deviation of observed scores. The uncomfortable direction is the one to sit with: banding makes the band WIDER as the test gets WORSE, so the least trustworthy instrument declares the largest group of people equal.
Not the same as rounding
Rounding is cosmetic and keeps the ordering intact. Banding deliberately destroys the ordering inside a band, because the instrument cannot support it.
How AssessAll handles it
AssessAll's behavioural reports carry a confidence band rather than a bare score, and the validity checks on the behavioural engine feed that band: divergent consistency pairs, infrequency endorsements and missed instructed-response items widen the band the report should be read with. They never auto-reject anyone. Whether a flag lowers confidence or triggers a rejection is the most useful single question to ask any assessment vendor.
Sources
- Campion, M. A., Outtz, J. L., Zedeck, S., Schmidt, F. L., Kehoe, J. F., Murphy, K. R., & Guion, R. M. (2001). The controversy over score banding in personnel selection: Answers to 10 key questions. Personnel Psychology, 54, 149–185. The publisher's copy is paywalled; an openly hosted full text was read at source on 28 September 2026 and is the copy linked here. Source of the 25%-of-applicants figure, the Kehoe bandwidth-versus-SD observation and the adverse-impact positions
- Cascio, W. F., Outtz, J., Zedeck, S., & Goldstein, I. L. (1991). Statistical implications of six methods of test score use in personnel selection. Human Performance, 4(4) — the origin of the SED banding procedure and of the two-SED band width. Not opened: only the publisher's abstract page is reachable, so everything stated here about it is taken from Campion et al. (2001), which is
- 42 U.S.C. §2000e-2(l) — the score-adjustment prohibition added by section 106 of the Civil Rights Act of 1991, read at source 16 September 2026
Read next
Related terms
More terms beginning with B
Check a selection process against the four-fifths rule
Free, no signup, computed in your browser — with the remedy, not just the verdict.
Last reviewed