Written and maintained by CASRAI Editorial Board
Published
On this page
KR-20 (Kuder-Richardson Formula 20) is an internal-consistency reliability coefficient for tests and questionnaires whose items are scored dichotomously, for example right or wrong, or yes or no. It compares the sum of the item variances with the variance of the total scores and gives a value that normally falls between 0.0 and 1.0, where higher values mean a more internally consistent instrument.
KR-20 (Kuder-Richardson Formula 20): definition
Kuder and Richardson published the formula in 1937 (Psychometrika 2(3): 151-160); the name records that it is the 20th numbered formula in that paper. The SAGE Encyclopedia of Research Design describes KR-20 as an index of internal consistency for instruments with dichotomously scored items. It is a single-administration method: one group takes the test once, so no retest or parallel form is needed, which is what separates it from test-retest reliability. It is also a close relative of split-half reliability, since both ask whether the items hang together, but KR-20 uses every item rather than one arbitrary split.
KR-20 is a special case of Cronbach’s alpha. When every item is scored 0 or 1, alpha and KR-20 are the same number, as the worked example below shows. Use KR-20 for right/wrong data and alpha for items with more than two score points.
KR-20 formula with every symbol explained
KR-20 = [k / (k − 1)] × [1 − (Σ pjqj) / σ²]
- k: the number of items on the test.
- pj: the proportion of examinees who answered item j correctly (scored 1).
- qj: 1 − pj, the proportion who did not.
- pjqj: the variance of item j when scores are 0 and 1.
- Σ pjqj: the sum of those item variances across all k items.
- σ²: the variance of the examinees’ total scores.
Population or sample variance for σ²
The item term pjqj is a population variance (it divides by n). The total-score variance must use the same convention. If you compute σ² with n − 1, as Excel’s VAR.S does, multiply each pjqj by n/(n − 1) as well. Mixing the two conventions changes the answer, as the example shows.
KR-20 worked example with the arithmetic shown
The data below are an illustrative composite invented for this guide, not a real study: ten examinees answered six right/wrong items (1 = correct). Each row is one examinee.
| Examinee | Item 1 | Item 2 | Item 3 | Item 4 | Item 5 | Item 6 | Total score |
|---|---|---|---|---|---|---|---|
| 1 | 1 | 1 | 1 | 1 | 1 | 1 | 6 |
| 2 | 1 | 1 | 1 | 1 | 1 | 0 | 5 |
| 3 | 1 | 1 | 1 | 1 | 0 | 0 | 4 |
| 4 | 1 | 1 | 1 | 0 | 1 | 0 | 4 |
| 5 | 1 | 1 | 1 | 0 | 0 | 0 | 3 |
| 6 | 1 | 1 | 0 | 1 | 0 | 0 | 3 |
| 7 | 1 | 1 | 0 | 0 | 0 | 0 | 2 |
| 8 | 1 | 0 | 1 | 0 | 0 | 0 | 2 |
| 9 | 1 | 0 | 0 | 0 | 0 | 0 | 1 |
| 10 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
Step 1: item proportions and Σpq
Counting the 1s in each column gives 9, 7, 6, 4, 3 and 1 correct answers out of 10, so p = 0.9, 0.7, 0.6, 0.4, 0.3 and 0.1. Each q is 1 − p, and each product pq is listed in the data table below this guide. The sum is Σpq = 0.09 + 0.21 + 0.24 + 0.24 + 0.21 + 0.09 = 1.08.
Step 2: variance of total scores
The totals are 6, 5, 4, 4, 3, 3, 2, 2, 1, 0. Their sum is 30, so the mean is 3.0. The squared deviations from the mean are 9, 4, 1, 1, 0, 0, 1, 1, 4, 9, which sum to 30. The population variance is 30 / 10 = 3.000, and the sample variance is 30 / 9 = 3.333.
Step 3: the coefficient
With the population variance: KR-20 = (6 / 5) × (1 − 1.08 / 3.000) = 1.2 × (1 − 0.36) = 1.2 × 0.64 = 0.768.
With the sample variance, adjusting Σpq by n/(n − 1) = 10/9 so the conventions match: Σpq becomes 1.20, and KR-20 = 1.2 × (1 − 1.20 / 3.333) = 1.2 × (1 − 0.36) = 0.768. The two consistent versions agree. If you divide the unadjusted 1.08 by the sample variance 3.333 instead, you get 1.2 × (1 − 0.324) = 0.811, which overstates the result by about 0.04 in this example.
KR-20 versus KR-21 and Cronbach’s alpha on the same data
KR-21 assumes all items have the same difficulty and replaces Σpq with k × p̄ × (1 − p̄), where p̄ is the mean of the item proportions. Here p̄ = 0.5, so KR-21 = 1.2 × (1 − 6 × 0.25 / 3.000) = 1.2 × 0.5 = 0.600, lower than KR-20 because these items clearly differ in difficulty. Cronbach’s alpha computed on the same 0/1 matrix, using sample item variances (which sum to 1.20) and the sample total variance 3.333, is 1.2 × (1 − 0.36) = 0.768, identical to KR-20.
How to interpret KR-20
The SAGE Encyclopedia of Research Design states that KR-20 values generally range from 0.0 to 1.0, with higher values representing a more internally consistent instrument. The same entry notes that, in very rare cases and typically with very small samples, values below 0.0 can occur, which indicates an extremely unreliable measurement. Here is what the sources do and do not let you conclude.
| Result | Reading | Source |
|---|---|---|
| Between 0.0 and 1.0 | The usual range; a higher value means the items are more internally consistent. | SAGE Encyclopedia of Research Design, 2nd ed. |
| Below 0.0 | Rare, typically with very small samples; signals an extremely unreliable measurement. | SAGE Encyclopedia of Research Design, 2nd ed. |
| The 0.768 in the example | The reliability of these raw scores for this group; it is specific to this test and these examinees. | Linacre (1997), Rasch Measurement Transactions 11:3 |
No single cut-off, such as a minimum acceptable value, is endorsed here, because the sources read for this guide do not fix one for all purposes. Decide what level your decision requires from the stakes of the test and the conventions of your field, and state the value you obtained rather than only a verdict. Linacre notes that KR-20, as a special case of alpha, reports the reliability of raw scores, which is local and test-dependent, and that it overstates the reliability of test-independent, generalizable measures; Rasch reliability is more conservative.
Assumptions and common mistakes
- Dichotomous scoring only. KR-20 expects items scored 0 or 1. For partial credit or rating scales use alpha.
- Mismatched variance conventions. Using pjqj (population) with a sample total variance inflates the result, as the 0.811 versus 0.768 comparison shows.
- Treating a high value as proof of one construct. Internal consistency assumes the items measure something homogeneous; a high coefficient can occur with a multidimensional scale, especially with many items. Dimensionality needs its own evidence.
- Ignoring what moves the value. The coefficient is affected by item difficulty, the spread of scores in the group and test length. A homogeneous group with a narrow score range lowers it even when the items are sound.
- Using KR-21 by default. KR-21 assumes equal item difficulty and tends to be more conservative; use it only when you lack the item-level data.
- Confusing reliability with validity. A consistent test can still measure the wrong thing; see reliability versus validity.
How to report KR-20 (APA example)
The APA numbers-and-statistics handout for the 7th edition asks for italics on letters used as statistical symbols and for a leading zero only when a statistic can exceed 1. KR-20 cannot exceed 1 in the usual case, so the example below drops the zero. Report the coefficient, the number of items and the sample size.
Internal consistency of the six-item knowledge test was acceptable for this sample, KR-20 = .77 (N = 10 examinees, k = 6 items).
Replace the adjective with a judgment that fits your field and purpose, and keep the value and the sample size visible so readers can weigh it themselves.
How to calculate KR-20 in Excel, R or Python
All three snippets use the population variance of the totals and reproduce 0.768 for the example above.
Excel
Put the 10 × 6 scores in B2:G11 and each examinee’s total in H2:H11 (=SUM(B2:G2)). In B13:G13 enter =AVERAGE(B2:B11) and fill right to get each p. Then:
=(6/5)*(1-SUMPRODUCT(B13:G13,1-B13:G13)/VAR.P(H2:H11))
If you use VAR.S for the totals, multiply the SUMPRODUCT term by 10/9.
R
X = as.matrix(items) # rows = examinees, columns = items scored 0/1
k = ncol(X); p = colMeans(X); tot = rowSums(X)
var_pop = mean((tot - mean(tot))^2)
(k / (k - 1)) * (1 - sum(p * (1 - p)) / var_pop)
Python
from statistics import pvariance
def kr20(rows):
n, k = len(rows), len(rows[0])
p = [sum(col) / n for col in zip(*rows)]
pq = sum(x * (1 - x) for x in p)
var_total = pvariance([sum(r) for r in rows])
return (k / (k - 1)) * (1 - pq / var_total)
Run on the ten rows above, the Python function returns 0.768. Statistical packages usually report KR-20 as Cronbach’s alpha when the items are 0/1 data.
Frequently asked questions
What does KR-20 measure?
It measures internal consistency: the extent to which the items of a test with right/wrong scoring vary together, summarized as a coefficient that normally runs from 0.0 to 1.0.
Is KR-20 the same as Cronbach’s alpha?
KR-20 is the special case of Cronbach’s alpha for dichotomous items. On 0/1 data with matching variance conventions they give the same value, as in the worked example (both 0.768).
What is the difference between KR-20 and KR-21?
KR-21 assumes every item has the same difficulty and uses only the mean proportion correct, so it needs no item-level data. KR-20 uses each item’s own p, and tends to give the higher value when item difficulties differ (0.768 versus 0.600 here).
Can KR-20 be negative?
Yes, but rarely. The SAGE Encyclopedia of Research Design says values below 0.0 typically arise with very small samples and indicate an extremely unreliable measurement.
Does a high KR-20 mean the test measures one thing?
No. A high value can arise with a multidimensional test, particularly with many items, so unidimensionality has to be checked separately.
Sources
- Kuder, G. F., and Richardson, M. W. (1937). The theory of the estimation of test reliability. Psychometrika, 2(3), 151-160 (citation as given in the Wikipedia entry, checked 2026-11-14).
- SAGE Encyclopedia of Research Design, 2nd ed., entry KR-20, DOI 10.4135/9781071812082.n286, https://sk.sagepub.com/ency/edvol/embed/the-sage-encyclopedia-of-research-design-2e/chpt/kr20 (checked 2026-11-14).
- Linacre, J. M. (1997). KR-20 / Cronbach alpha or Rasch reliability: which tells the truth? Rasch Measurement Transactions, 11:3, 580-581, https://rasch.org/rmt/rmt113l.htm (checked 2026-11-14).
- Wikipedia, Kuder-Richardson Formula 20, https://en.wikipedia.org/wiki/Kuder%E2%80%93Richardson_Formula_20 (formula, variance convention, KR-21; checked 2026-11-14).
- Statology, Kuder-Richardson 20, https://www.statology.org/kuder-richardson-20/ (formula and Excel notes; checked 2026-11-14).
- APA Style, numbers and statistics guide (7th ed.), https://apastyle.apa.org/instructional-aids/numbers-statistics-guide.pdf (checked 2026-10-29).
Data: KR-20 worked example: item statistics for six right/wrong items (10 examinees, illustrative data)
Item-level inputs to the KR-20 worked example in this guide: the number of examinees answering each of six items correctly, the proportion correct (p), its complement (q) and the item variance pq. The data are an illustrative composite computed for this guide; the formula follows Kuder and Richardson (1937) as presented in the SAGE Encyclopedia of Research Design and Wikipedia.
| Item | Examinees correct (count of 10) | p (proportion correct) | q (1 - p) | pq (item variance) |
|---|---|---|---|---|
| Item 1 | 9 | 0.9 | 0.1 | 0.09 |
| Item 2 | 7 | 0.7 | 0.3 | 0.21 |
| Item 3 | 6 | 0.6 | 0.4 | 0.24 |
| Item 4 | 4 | 0.4 | 0.6 | 0.24 |
| Item 5 | 3 | 0.3 | 0.7 | 0.21 |
| Item 6 | 1 | 0.1 | 0.9 | 0.09 |
Sources
- SAGE Encyclopedia of Research Design, 2nd ed.: KR-20 (accessed 2026-11-14)
- Wikipedia: Kuder-Richardson Formula 20 (accessed 2026-11-14)
Compiled by CASRAI from the sources above. Free to reuse under CC BY 4.0 with attribution.







