Expected Counts Below 5: Chi-Square Rules and What to Do
The chi-square p-value is an approximation that works when expected counts are large enough. Two rules of thumb say what "large enough" means, and they disagree.
Checked on November 21, 2026. Results computed by our statistics engine, version 1.0.0.
In short
- Both common rules, stated precisely
- Rules apply to expected counts, not observed
- Worked examples of each fix
On this page
Why small expected counts matter
The p-value from a chi-square test comes from the chi-square distribution, which is only an approximation to the true distribution of the statistic. The approximation improves as expected counts grow. With very small expected counts, the true distribution is lumpy and the p-value from the smooth curve can be noticeably too small or too large.
The rules are about expected counts, the E in (O − E)² / E, not the observed counts you entered. A table can have a zero observed count and still be fine, or have no small observed counts and still fail. Many web pages get this wrong.
The two rules
- The strict textbook rule: every expected count should be at least 5. OpenStax states it this way (OpenStax 11.2), and R's
chisq.testprints a warning whenever any expected count is below 5 (R chisq.test source). - Cochran's rule: Cochran's 1954 paper on strengthening chi-square tests is usually summarized as allowing a minimum expected count of 1 when relatively few expected counts are below 5, for example one cell in five or two in ten (Cochran 1954). It is often paraphrased as "no more than 20% of expected counts below 5, and none below 1".
- The calculators on this site check both and tell you which your table passes: green when every expected count is at least 5, amber when some are below 5 (the message says whether Cochran's rule still holds), and red when any expected count is below 1.
Example: one small cell
R's goodness-of-fit example tests counts 89, 37, 30, 28, 2 against proportions 0.4, 0.2, 0.2, 0.19, 0.01. The expected counts are 74.40, 37.20, 37.20, 35.34, 1.86. One of five is below 5 but none is below 1, so the table passes Cochran's rule and fails the strict one. R gives χ² = 5.7947, p = 0.2150, with a warning that the approximation may be incorrect. Report the result with a note, or use an exact multinomial test if the conclusion is borderline.
Example: a tiny 2x2 table
| Column 1 | Column 2 | |
|---|---|---|
| Row 1 | 3 | 0 |
| Row 2 | 0 | 3 |
Every expected count here is 1.5. The uncorrected χ² is 6.00 (p = 0.0143), which looks significant, but the approximation cannot be trusted with expected counts this small. Fisher's exact test gives p = 0.1000: with only 6 observations, even a perfect split is not good evidence. The Yates-corrected χ² of 2.6667 (p = 0.1025) moves the same way. This is a constructed example to show the effect.
What to do instead
- 2x2 table with an expected count below 1: use Fisher's exact test.
- 2x2 table with expected counts of at least 1 but some below 5: Campbell's 2007 simulation study recommends the N−1 chi-square, χ² × (N − 1) / N, over both Fisher's test and the Yates correction (Campbell 2007). The 2x2 calculator shows it.
- Larger tables: an exact test (Fisher-Freeman-Halton) or a Monte Carlo p-value, which R offers with
simulate.p.value = TRUE. - Merge categories only when the merged category still makes sense for your question, and decide before looking at the result. Merging to rescue significance is not legitimate.
- Collect more data if you can. Expected counts grow in proportion to N.
The N−1 chi-square in practice
| Column 1 | Column 2 | |
|---|---|---|
| Row 1 | 1 | 9 |
| Row 2 | 9 | 1 |
Here N = 20 and every expected count is 5, right at the strict cutoff. Campbell recommends the N−1 version for any 2x2 table whose expected counts are all at least 1. Pearson's χ² is 12.80; multiplying by (20 − 1) / 20 gives 12.16, p = 4.883 × 10−4. Fisher's exact test gives p = 0.0011. Both show strong evidence of an association; the N−1 version is less conservative. This is also a constructed example.
Common questions
Does a zero observed count break the test?
No. Only small expected counts matter. A zero row or column total does break it, because every expected count in that row or column is then zero.
Is the rule 5 or 10?
Most sources use 5. A few older texts suggest 10 for 2x2 tables; Campbell's results suggest the N−1 chi-square works well with expected counts down to 1.
Sources
- OpenStax, Introductory Statistics 2e, section 11.2: goodness-of-fit test checked November 21, 2026
- R source code, chisq.test (stats package) checked November 21, 2026
- Cochran W.G. (1954). Some methods for strengthening the common chi-square tests. Biometrics 10:417-451 checked November 21, 2026
- Campbell I. (2007). Chi-squared and Fisher-Irwin tests of two-by-two tables with small sample recommendations. Statistics in Medicine 26:3661-3675 checked November 21, 2026