Statistics Glossary → Confidence interval for a proportion (Wilson, Clopper-Pearson, Agresti-Coull, Jeffreys)
Confidence interval for a proportion (Wilson, Clopper-Pearson, Agresti-Coull, Jeffreys)concept
An interval expressing the uncertainty about the population value of a proportion observed in a sample (e.g. 12 out of 40), and the various methods for constructing it.
The best-known method is the Wald interval: p̂ ± z·√(p̂(1 - p̂)/n). Simple as it is, it performs poorly for small samples and for proportions close to 0 or 1; its actual coverage can fall well below 95% and the interval can spill outside 0-1. The Wilson (score) interval is obtained by inverting the score test and gives good coverage even for small n. Agresti-Coull offers a similar improvement by adding roughly two successes and two failures to the Wald formula. The Jeffreys interval is a Bayesian interval based on a Beta(0.5, 0.5) prior and also has good frequentist coverage.
The Clopper-Pearson interval is computed exactly from the binomial distribution and guarantees coverage of at least the nominal level; it is therefore conservative, meaning it can be wider than necessary. Brown, Cai and DasGupta (2001) recommended the Wilson or Jeffreys interval for general use, and these rather than Wald for small n. Even when no successes are observed (0/n), these methods give a meaningful upper limit, whereas Wald produces a meaningless interval of zero width. With survey data the design effect must be taken into account; logit-transformed or Korn-Graubard intervals are used for that purpose.
An everyday example
A newly opened restaurant finds that 19 of its first 20 customers say they are satisfied (95%). The Wald formula pushes the upper limit above 100% and gives a lower limit of about 85%. The Wilson interval runs from about 76% to 99%: with so few customers, it honestly shows that the true satisfaction rate could be below 80%.
In MerQur
In the Binomial Test the 'All CI methods' checkbox gives the Clopper-Pearson, Wilson, Agresti-Coull, Jeffreys and Wald intervals together; MR Frequency Analysis offers 'Proportion CI (None, Wilson, Clopper-Pearson)', Weighted Frequency offers 'Proportion CI method (Logit-transformed, Wald, Korn-Graubard)', and the McNemar test offers the 'Agresti-Min / Tango CI' for the paired difference in proportions. Crosstab, the ROC Curve and Confusion Matrix Metrics give Wilson intervals; in the Permutation and Kruskal-Wallis tests a 99% Clopper-Pearson CI is reported for the Monte Carlo p.
Analyses where this concept appears
Permutation TestKruskal-WallisBinomial TestMcNemar TestCrosstabMR Frequency AnalysisROC CurveConfusion Matrix MetricsLikert Scale AnalysisWeighted Frequency (Survey Freq)