The Wilson Score Interval: A Better Confidence Interval for Proportions

The standard formula for a confidence interval for a proportion — sample proportion plus or minus a margin of error — is called the Wald interval, and it's what's typically taught first. It also has a real, sometimes dramatic weakness: with a small sample or a proportion near 0 or 1, it can produce a nonsensical interval that extends below 0% or above 100%. The Wilson score interval fixes this, and is the better default choice outside of introductory coursework — many statistical software packages already use it, or something close to it, as their built-in default for exactly this reason.

Where the Wald Interval Breaks Down

The familiar Wald formula for a proportion's confidence interval is:

This formula relies on a normal approximation to the underlying binomial distribution — an approximation that works reasonably well when p̂ is close to 0.5 and n is reasonably large, but degrades badly when p̂ sits near 0 or 1, or when n is small. In the worst cases, the margin of error can exceed the distance from p̂ to 0 (or to 1), pushing the calculated interval below 0% or above 100% — a mathematically impossible range for a genuine proportion, and a clear sign the underlying normal approximation has broken down.

Worked Example: Where Wald Fails

Problem: Out of 10 randomly sampled customers, 1 reports a product defect (p̂ = 0.10). Construct a 95% confidence interval using the Wald formula.

Wald margin of error = 1.96 × √[0.10×0.90/10] = 1.96 × √0.009 = 1.96 × 0.0949 ≈ 0.186

Wald 95% CI = 0.10 ± 0.186 = (−0.086, 0.286)

The lower bound is negative — a nonsensical result, since a proportion can never actually be below 0%. This isn't a rare edge case; it's a routine occurrence whenever a small sample produces a proportion near either extreme, exactly the situation described above.

The Wilson Score Formula

The Wilson score interval derives from a more careful mathematical treatment of the same underlying problem, and its formula looks more involved but handles exactly this situation correctly:

The interval is Center ± Margin. Unlike the Wald interval, which is centered exactly on p̂, the Wilson interval's center is pulled slightly toward 0.5 — this built-in shrinkage is exactly what keeps the interval from overshooting past 0 or 1.

Worked Example: Wilson on the Same Data

Using the same data (p̂=0.10, n=10, z=1.96):

Denominator: 1 + z²/n = 1 + 1.96²/10 = 1 + 0.384 = 1.384

Center = [0.10 + 1.96²/(2×10)] / 1.384 = [0.10 + 0.192] / 1.384 ≈ 0.211

Margin = 1.96×√[0.10×0.90/10 + 1.96²/(4×10²)] / 1.384 = 1.96×√[0.009+0.0096] / 1.384 ≈ 0.193

Wilson 95% CI = 0.211 ± 0.193 = (0.018, 0.404)

The Wilson interval stays entirely within the valid 0-to-1 range, as it always does by construction. It's also visibly asymmetric around the original p̂=0.10 (extending further above than below) — a genuine, correct reflection of the underlying binomial distribution's own asymmetry near a boundary, something the symmetric Wald formula can't represent at all.

Why the Interval Comes Out Asymmetric

Near p=0, a proportion has much more room to be higher than the sample suggests than to be lower (since it can't go below 0 at all) — the true binomial sampling distribution reflects this asymmetry directly, and the Wilson interval, being derived more rigorously from that actual distribution rather than a symmetric normal approximation, correctly reproduces it. This is precisely the mathematical honesty the simpler symmetric Wald formula sacrifices for computational convenience.

When the Difference Actually Matters

For large samples with p̂ reasonably close to 0.5, the Wald and Wilson intervals converge closely, and the choice matters little in practice. The difference becomes practically important specifically when n is small (rough rule of thumb: under about 30-40) or p̂ is within roughly 15-20 percentage points of 0 or 1 — exactly the combination in the worked example above, and a genuinely common situation in real applications like rare-event rates, small pilot studies, or early-stage A/B tests with few conversions so far.

Real-World Uses

Early-stage A/B testing: a new landing page with only a handful of conversions so far is exactly the small-sample, near-zero-proportion scenario where Wald can fail — using Wilson from the start avoids ever reporting a nonsensical negative conversion-rate bound to stakeholders.

Rare disease prevalence and safety monitoring: medical and public health research routinely deals with proportions close to 0 (a rare condition's prevalence, an adverse event rate) — the Wilson interval, or the related Clopper-Pearson interval, is standard practice in these fields specifically because of this well-known Wald limitation.

Online reviews and rating systems: platforms ranking items by an underlying "true" positive-rating proportion, especially for items with few reviews so far, commonly use Wilson-interval-based scoring (sometimes called a "lower bound" score) specifically to avoid overstating confidence in a rating built from a very small number of reviews.

Common Mistakes

Defaulting to the Wald interval out of habit, without checking whether it's appropriate. Many statistical software packages default to Wilson (or an even more refined alternative) specifically because of this well-known Wald weakness — using Wald reflexively, especially with small samples or extreme proportions, risks reporting an interval that's visibly broken.

Reporting a Wald interval that extends below 0 or above 1 without flagging the problem. When this happens, it's a clear signal the normal approximation has failed for this specific sample, not a result to report as-is — switching to Wilson (or truncating the interval at the valid boundary, a cruder but sometimes-used fix) are both better options than reporting an impossible range unchanged.