Written and maintained by CASRAI Editorial Board
Last updated
The standard formula taught for a confidence interval around a sample proportion — p̂ ± z√[p̂(1 − p̂)/n] — is called the Wald interval, and it is unreliable in exactly the situations researchers most often use it: small samples, and proportions near 0% or 100%. It can produce a lower bound below zero, an upper bound above one, or a zero-width interval that implies false certainty. Two alternatives — the Wilson score interval and the Clopper-Pearson exact interval — were built specifically to fix this, and a third, the Agresti-Coull interval, offers a simple hand-computable compromise between them. This guide shows exactly where the Wald interval breaks, with worked numbers, and gives concrete guidance on which method to use for a given proportion, sample size, and reporting context.
The Wald interval: the formula everyone learns first
The Wald interval is a direct application of the general confidence-interval formula (point estimate ± critical value × standard error) to a sample proportion:
Wald CI = p̂ ± z × √[p̂(1 − p̂) / n]
where p̂ = x/n is the observed proportion (x successes out of n trials) and z is the standard normal critical value (1.96 for 95% confidence). It is derived by treating the sampling distribution of p̂ as approximately normal — a consequence of the Central Limit Theorem applied to a binomial proportion — and estimating the standard error using the sample proportion itself in place of the unknown true p. It is taught first because it is the most direct extension of the mean-CI formula covered in Confidence Interval Explained, and because for a moderate n with p near 0.5, it works fine. The trouble is that a large share of real proportion data — adverse-event rates, defect rates, response rates on rare outcomes, small pilot samples — does not sit in that comfortable region.
Why the Wald interval fails
The normal approximation the Wald interval relies on breaks down for two related reasons: the binomial distribution is skewed whenever p is not close to 0.5, and its variance p(1 − p)/n shrinks to zero as p approaches 0 or 1 — exactly where the normal approximation is worst. Two concrete failure modes follow directly from this.
The zero-count problem
If x = 0 (no events observed in n trials), p̂ = 0 and the standard error formula √[p̂(1 − p̂)/n] also evaluates to zero, collapsing the whole interval to a single point:
| Scenario | p̂ | Wald 95% CI |
|---|---|---|
| 0 events in 10 trials (x=0, n=10) | 0.000 | [0.0000, 0.0000] |
A zero-width interval says the true rate is known to be exactly zero with 95% confidence — which cannot be correct after only 10 trials. Observing zero events in a finite sample is consistent with a true rate that is small but nonzero; the interval should be wide, not a single point. This is the most common way the Wald interval is mis-cited in practice: rare-event and adverse-event reporting is exactly the setting where x = 0 shows up routinely, and a naively-applied Wald formula silently produces a nonsensical result rather than an error.
Impossible bounds outside [0, 1]
Because the Wald interval is symmetric around p̂ by construction, it does not respect the [0, 1] boundary the true proportion must fall within. With a small n and p̂ away from 0.5, the lower or upper bound routinely lands outside that range:
| Scenario | p̂ | Wald 95% CI |
|---|---|---|
| 1 event in 20 trials (x=1, n=20) | 0.050 | [−0.0455, 0.1455] |
| 98 events in 100 trials (x=98, n=100) | 0.980 | [0.9526, 1.0074] |
A negative lower bound or an upper bound above 1.0 is not a rounding quirk — it is the formula asserting a probability outside [0, 1], a value that cannot exist. Software that reports these bounds without truncation is exposing the method’s failure directly; software that silently truncates at 0 or 1 hides the failure but still leaves the interval’s true coverage below its stated confidence level.
Large n does not fix this near the boundary
The traditional classroom rule of thumb — use the normal approximation once np̂ and n(1 − p̂) are both at least 5, or at least 10 by the more conservative version — is necessary but not sufficient. The n=100, p̂=0.98 example above has np̂=98 and n(1−p̂)=2, which already fails even the loose np̂ ≥ 5 check, but even scenarios that technically clear the rule of thumb can have poor Wald coverage. The most cited analysis of this problem, Brown, Cai, and DasGupta’s 2001 Statistical Science paper “Interval Estimation for a Binomial Proportion,” showed that the Wald interval’s actual coverage probability oscillates erratically as n increases rather than converging smoothly to the nominal 95% — for many combinations of n and p, true coverage can sit well below 95%, sometimes below 90%, even at sample sizes in the hundreds. Passing the np ≥ 5 rule of thumb is not a guarantee of good coverage; it only rules out the most extreme failures like the zero-width interval above.
The Wilson score interval
The Wilson score interval (Wilson, 1927) fixes the Wald interval’s core problem by changing what gets estimated. Instead of building a symmetric interval around p̂ using an SE that assumes p̂ is the true p, it inverts the score test for p directly — solving for the values of p at which the observed p̂ would just barely be judged consistent with p at the chosen confidence level. The result is:
Wilson CI = [p̂ + z²/2n ± z√(p̂(1−p̂)/n + z²/4n²)] / (1 + z²/n)
Two structural differences from the Wald formula explain why it performs better: the center of the interval is not p̂ itself but a weighted average that pulls slightly toward 0.5 (more so at small n), and the whole expression is divided by (1 + z²/n), which keeps both bounds inside [0, 1] by construction — the interval can never go negative or exceed 1. Applied to the two failure cases above:
| Scenario | p̂ | Wald 95% CI | Wilson 95% CI |
|---|---|---|---|
| 0 events in 10 trials | 0.000 | [0.0000, 0.0000] | [0.0000, 0.2775] |
| 1 event in 20 trials | 0.050 | [−0.0455, 0.1455] | [0.0089, 0.2361] |
Both Wilson intervals stay within [0, 1] and, critically, both are wide enough to honestly reflect how little the data rules out — zero events in 10 trials genuinely is consistent with a true rate as high as roughly 28%. Brown, Cai, and DasGupta’s coverage analysis found Wilson’s actual coverage tracks its nominal 95% target closely across almost the entire range of n and p, which is the main reason it is now the default or recommended method in most modern statistical software rather than a specialist alternative.
The Clopper-Pearson exact interval
The Clopper-Pearson interval (Clopper & Pearson, 1934) takes a different approach entirely: rather than approximating the binomial with a normal distribution, it inverts the exact binomial distribution itself, using its known relationship to the Beta distribution. The bounds are the Beta quantiles:
Lower bound = Beta−1(α/2; x, n−x+1)
Upper bound = Beta−1(1−α/2; x+1, n−x)
(with the lower bound defined as 0 when x = 0, and the upper bound defined as 1 when x = n). Because it is built from the exact binomial distribution rather than a normal approximation, Clopper-Pearson is guaranteed to have coverage at or above the nominal confidence level for every n and every p — it never under-covers, which is why it is called “exact.” That guarantee has a real cost: because the binomial is discrete and the normal approximation is not, exact inversion is systematically conservative, producing intervals that are wider than they strictly need to be, especially at small n or p near the boundary:
| Scenario | p̂ | Wilson 95% CI (width) | Clopper-Pearson 95% CI (width) |
|---|---|---|---|
| 1 event in 20 trials | 0.050 | [0.0089, 0.2361] (0.227) | [0.0013, 0.2487] (0.248) |
| 15 events in 20 trials | 0.750 | [0.5313, 0.8881] (0.357) | [0.5090, 0.9134] (0.405) |
In both cases Clopper-Pearson is noticeably wider than Wilson while covering the same true confidence level, and in most practical settings the extra width buys accuracy the data doesn’t actually require. That conservatism is a feature, not a defect, in one specific context: regulatory and safety-critical reporting where the standard explicitly calls for a method that never overstates precision — certain GxP, medical-device, and clinical-safety reporting conventions specify Clopper-Pearson (or cite it as the reference method) for exactly this reason. Outside those contexts, its extra conservatism is usually just lost precision.
Agresti-Coull: the “add two successes, two failures” shortcut
The Agresti-Coull interval (Agresti & Coull, 1998) is a deliberately simple approximation to Wilson’s coverage performance, built to be computable by hand. It adjusts the counts before applying the ordinary Wald formula:
p̃ = (x + z²/2) / (n + z²)
Agresti-Coull CI = p̃ ± z√[p̃(1 − p̃) / (n + z²)]
For a 95% interval, z² ≈ 4, so this is the well-known “plus-four” rule of thumb: add 2 to the success count and 2 to the failure count (4 trials total), then run the plain Wald formula on the adjusted numbers. Because the adjustment pulls p̃ away from 0 or 1 before the symmetric Wald step, it largely avoids the zero-width and out-of-bounds failures shown above, at the cost of occasionally producing a bound very slightly outside [0, 1] itself in extreme small-n cases (a residual Wald-style symmetry artifact) — a boundary check-and-truncate step is standard practice when using it. Its main appeal is pedagogical and computational: it gets most of Wilson’s coverage accuracy from a formula simple enough to compute with a calculator, which is why it shows up often in introductory statistics texts as the recommended replacement for the plain Wald formula.
All four methods, side by side
The table below applies all four methods to the same set of scenarios at 95% confidence, computed directly from each method’s formula (Wilson and Agresti-Coull by the closed-form expressions above; Clopper-Pearson via the Beta-quantile relationship).
| Scenario | p̂ | Wald | Wilson | Clopper-Pearson | Agresti-Coull |
|---|---|---|---|---|---|
| x=0, n=10 | 0.000 | [0.0000, 0.0000] | [0.0000, 0.2775] | [0.0000, 0.3085] | [−0.0434, 0.3209] |
| x=1, n=20 | 0.050 | [−0.0455, 0.1455] | [0.0089, 0.2361] | [0.0013, 0.2487] | [−0.0091, 0.2541] |
| x=15, n=20 | 0.750 | [0.5602, 0.9398] | [0.5313, 0.8881] | [0.5090, 0.9134] | [0.5275, 0.8919] |
| x=48, n=100 | 0.480 | [0.3821, 0.5779] | [0.3846, 0.5768] | [0.3790, 0.5822] | [0.3846, 0.5768] |
| x=98, n=100 | 0.980 | [0.9526, 1.0074] | [0.9300, 0.9945] | [0.9296, 0.9976] | [0.9256, 0.9989] |
| x=250, n=500 | 0.500 | [0.4562, 0.5438] | [0.4563, 0.5437] | [0.4553, 0.5447] | [0.4563, 0.5437] |
The pattern to notice: at x=48, n=100 and x=250, n=500 — p̂ near 0.5, moderate-to-large n — all four methods agree almost to the third decimal place, and Agresti-Coull matches Wilson exactly to four decimals at that scale. The disagreement only opens up as p̂ moves toward 0 or 1, or as n gets small — which is precisely the diagnostic for when the choice of method actually matters. If your data sits in the last row’s regime, the plain Wald formula is a defensible shortcut. If it sits in the first three rows’ regime, it is not.
Which method should you actually use?
- Default to Wilson score for general research reporting. It has near-nominal coverage across almost the full range of n and p, stays within [0, 1] by construction, and is what most modern statistical software now defaults to or recommends (R’s
prop.test()uses a continuity-corrected Wilson interval by default, for example). - Use Clopper-Pearson when the reporting context explicitly requires guaranteed coverage — conservative by design, never under-covering — such as certain regulatory, GxP, or clinical-safety reporting conventions that call for an “exact” method by name. Expect wider intervals as the tradeoff.
- Use Agresti-Coull when you need something computable by hand or need to explain the correction intuitively (“add two successes and two failures”) without invoking the score-test derivation behind Wilson. It tracks Wilson closely in practice.
- Avoid the plain Wald interval whenever x = 0 or x = n, whenever p̂ is not reasonably close to 0.5, or whenever n is small (rule-of-thumb: below roughly 30–40, though as shown above this is not a hard cutoff). Its main remaining justification is pedagogical — it is the simplest formula to teach the general CI = estimate ± critical value × SE structure — not that it is the best method to actually report.
A fifth option worth knowing about, though less commonly taught: the Jeffreys interval, a Bayesian credible interval built from a Beta(0.5, 0.5) reference prior. Brown, Cai, and DasGupta’s analysis found it performs comparably to Wilson in most cases and is available in statistical software alongside the four methods above; it is a reasonable substitute where a Bayesian framing is already in use elsewhere in the analysis. See Frequentist vs. Bayesian Statistics for the broader framework distinction.
Computing each interval in R and Python
- R —
prop.test(x, n)returns a Wilson interval with continuity correction;binom.test(x, n)returns the Clopper-Pearson exact interval; thebinompackage’sbinom.confint()function returns all major methods (Wald, Wilson, Clopper-Pearson/exact, Agresti-Coull, Jeffreys) side by side from a single call, which is the fastest way to compare them on real data. - Python —
statsmodels.stats.proportion.proportion_confint(count, nobs, method=...)supports'normal'(Wald),'wilson','beta'(Clopper-Pearson),'agresti_coull', and'jeffreys'as themethodargument, making it straightforward to compute any of the intervals in this guide directly, or to loop across all five for a comparison table like the one above.
Whichever tool computes the interval, check which method is the default before trusting an unlabeled output — the same function name in different packages does not always default to the same method, and the difference is exactly the gap this guide has been demonstrating.
Reporting a proportion’s confidence interval correctly
Whichever method is used, name it explicitly in the methods section (“a Wilson score 95% confidence interval was calculated for each proportion”) rather than leaving the reader to assume the plain Wald formula was used by default — the choice materially affects the reported bounds in exactly the small-n, boundary-adjacent cases where getting it right matters most. See How to Report Confidence Intervals in a Manuscript for the general formatting conventions (APA bracket notation, AMA/biomedical parenthetical notation), and Confidence Interval Explained for the underlying concept, interpretation, and the general CI formula this guide builds on. For sample-size planning before data collection, see Power Analysis and Sample Size Calculation; for the critical-value/z-table mechanics behind every method above, see Z-Score (Standard Score).
Frequently Asked Questions
Is Wilson the same thing as the “adjusted Wald” interval?
Not exactly, though they are closely related and often confused. “Adjusted Wald” usually refers to the Agresti-Coull interval — add pseudo-counts, then apply the plain Wald formula. The Wilson score interval is derived differently (by inverting the score test directly) and has a different, non-Wald-shaped formula, though its numeric results are very close to Agresti-Coull’s in most cases, which is exactly why Agresti-Coull works as a hand-computable stand-in for it.
Why did R’s prop.test() give a different answer than my hand-calculated Wald interval?
prop.test() computes a Wilson score interval with a continuity correction by default, not the plain Wald interval — that is a deliberate design choice by R’s authors, not a bug or a different rounding convention. If a plain Wald interval is genuinely what’s needed, it has to be computed directly from the formula, or with method='normal' in Python’s statsmodels, rather than assumed from a generic “proportion test” function.
Does the continuity correction matter?
The continuity correction adjusts for approximating a discrete binomial distribution with a continuous normal or score-based curve; it widens the interval slightly and improves coverage further at small n, at the cost of a small amount of extra conservatism. It is applied by default in some software’s Wilson implementation (including R’s prop.test()) and omitted in others, which is one more reason to state explicitly which method and which variant was used.
What sample size makes the Wald interval safe to use?
There is no clean cutoff. The traditional np̂ ≥ 5 (or ≥ 10) rule of thumb catches the most extreme failures but, per Brown, Cai, and DasGupta’s analysis, does not guarantee coverage close to nominal even when satisfied. The practical answer is to use Wilson or Agresti-Coull by default rather than trying to certify the Wald interval safe for a specific n and p in advance.
Which method do clinical trial adverse-event tables typically use?
Exact (Clopper-Pearson) intervals are common in adverse-event and safety-reporting tables specifically because of their guaranteed-coverage property, though the exact convention varies by sponsor, protocol, and applicable regulatory guidance — check the statistical analysis plan for the specific method actually prespecified for a given trial rather than assuming.
Can any of these methods produce an interval wider than [0, 1] itself?
Wilson and Clopper-Pearson cannot — both are constructed to stay within [0, 1]. The plain Wald interval routinely does not, as shown above. Agresti-Coull is usually well-behaved but can very slightly exceed the boundary in extreme small-n cases, since its last step is still the symmetric Wald formula applied to adjusted counts; truncating at 0 or 1 if this occurs is standard practice.
