MerQuR+Statistics Glossary

Statistics Glossary → Continuity correction (Yates)

Continuity correction (Yates)concept

A correction that shifts a discrete (count-based) test statistic half a unit toward its expected value when approximating it with a continuous distribution, to reduce the error caused by discreteness.

Counts and rank sums can take only certain values, whereas the normal and chi-square distributions are continuous. In small samples this mismatch can make the approximate p value too small. The continuity correction subtracts 0.5 from the absolute difference between the observed and expected values; in Yates's version for 2x2 tables, X² = Σ (|O - E| - 0.5)²/E. A similar correction is used in the normal approximations of the Mann-Whitney, Wilcoxon, sign, McNemar and runs tests.

The correction makes the p value larger and the test more conservative. Yates's correction for 2x2 tables has often been found overly conservative and is therefore debated; with small expected frequencies an exact test (Fisher) is a better solution. As the sample grows, the effect of the correction vanishes. Effect sizes (e.g. phi, Cramér's V) are usually computed from the uncorrected statistic.

An everyday example

A shop shows two packaging designs to 40 customers: 26 choose A and 14 choose B. Under equal preference the expected count is 20; uncorrected z = (26 - 20)/√10 ≈ 1.90, while corrected z = (26 - 20 - 0.5)/√10 ≈ 1.74. The exact binomial p value (about 0.08) is closer to the corrected approximation (0.08) than to the uncorrected one (0.06).

In MerQur

Offered as 'Yates correction, 2x2' (default on; also applied to G² when the G-test is on) in the chi-square independence test, 'Continuity correction' in the Mann-Whitney U test (default on) and in the Wilcoxon test (default off), and as the 'Continuity-corrected Z' checkbox in the runs test. The 'Asymptotic (χ², Yates)' method in McNemar, the pairwise McNemar post-hoc choices of Cochran's Q, the 'Large-sample Z' approximation of the sign test, the Z of the Mann-Kendall test, the McNemar test in Confusion Matrix Metrics and the Mantel-Haenszel DIF test in item response theory use a continuity correction; the Page test and the pairwise Wilcoxon approximation of the Friedman test are uncorrected.

Analyses where this concept appears

Mann-Whitney UWilcoxon Signed-RankFriedmanSign TestRuns TestChi-Square IndependenceMcNemar TestCochran's Q TestConfusion Matrix MetricsItem Response Theory (IRT / Rasch)Mann-Kendall + Sen's Slope

Related concepts

Asymptotic (normal) approximationExpected frequency