A random sample of 40 adults aged 18 to 39 from the National Health and Nutrition Examination Survey (NHANES) of 2017 to 2018 contains 10 with hypertension, a sample proportion of $\hat{p} = 10/40 = 0.250$. The interval on the course's formula sheet, the one in nearly every introductory text, is
$$\hat{p} \pm z_{1-\alpha/2}\sqrt{\frac{\hat{p}(1 - \hat{p})}{n}} = 0.250 \pm 1.96 \times 0.0685 = (0.116,\ 0.384).$$
It looks like the interval for a mean in Confidence Intervals and Coverage, and it is built the same way: an estimate, plus or minus a normal quantile times an estimated standard error. For a mean, that recipe gives intervals whose coverage is close to the stated level once the sample is moderately large. Does it do the same for a proportion?
It does not. The coverage probability of an interval method is the share of samples whose interval contains the true value. Over all samples of 40 from this population, this interval contains the true proportion 0.2294 only 91.09% of the time, not 95%. At other sample sizes and proportions it does worse, sometimes far worse.
First we derive the textbook interval and compute its coverage exactly, which the binomial distribution makes possible without any simulation. Then we find the two reasons it fails. Next we derive the three intervals that replace it:
- Wilson's score interval (1927);
- the Agresti–Coull interval, built on Wilson's center;
- the Clopper–Pearson "exact" interval (1934).
Finally we compare all four, and ask what to report when a sample has no events at all. The examples use the hypertension outcome and population defined in The Normal Approximation to the Binomial: 1,696 adults aged 18 to 39, of whom a proportion $p = 0.2294$ have hypertension.
The Wald interval
The formula sheet's interval comes from the normal approximation, as the confidence intervals for a mean did. Since $(\hat{p} - p)/\sqrt{p(1-p)/n}$ is approximately standard normal,
$$P\!\left(-z_{1-\alpha/2} \le \frac{\hat{p} - p}{\sqrt{p(1-p)/n}} \le z_{1-\alpha/2}\right) \approx 1 - \alpha.$$
The standard error contains the unknown $p$. The textbook interval replaces it by $\hat{p}$, giving the estimated standard error $\operatorname{SE} = \sqrt{\hat{p}(1-\hat{p})/n}$, and then solves for $p$ as one would for a mean:
$$\hat{p} - z_{1-\alpha/2}\operatorname{SE} \le p \le \hat{p} + z_{1-\alpha/2}\operatorname{SE}.$$
The same construction, an estimate divided by its estimated standard error, is the large-sample test that Abraham Wald proposed for any parameter estimated by maximum likelihood. So this is called the Wald interval (Brown, Cai and DasGupta, 2001, §2). The formula sheet writes the critical value as $z_{\alpha/2}$, meaning the point with $\alpha/2$ above it. This article writes $z_{1-\alpha/2}$, the point with $1 - \alpha/2$ below it, as the earlier units did. Both are 1.960 for 95% confidence.
For the sample of 40: $\operatorname{SE} = \sqrt{0.250 \times 0.750/40} = 0.06847$, the margin is $1.96 \times 0.06847 = 0.1342$, and the interval is $(0.1158,\ 0.3842)$. It contains the true 0.2294. But one interval that covers says little about the method. How often does it cover?
Coverage computed exactly
For an interval for a mean, coverage had to be estimated by simulation: draw many samples, count the intervals that contain the truth. For a proportion it can be computed exactly. Here is the idea. The interval depends on the sample only through the count $X$, which takes the values $0, 1, \ldots, n$. Each value gives one interval, and that interval either contains $p$ or not. So the coverage probability is the total binomial probability of the counts whose intervals contain $p$:
$$C(p, n) = \sum_{x = 0}^{n} \binom{n}{x} p^x (1-p)^{n-x}\ \mathbf{1}\big[L(x) \le p \le U(x)\big].$$
Here $L(x)$ and $U(x)$ are the interval's limits when $X = x$, and $\mathbf{1}[\cdot]$ is 1 if the condition holds and 0 otherwise. The calculation is exact, with no simulation error, for any method and any $n$ and $p$.
As a check, the script reproduced every coverage value Brown, Cai and DasGupta (2001, §2.1) report for the Wald interval:
| $p$ | $n$ | Coverage, script | Coverage, Brown, Cai and DasGupta |
|---|---|---|---|
| 0.2 | 30 | 0.9463 | 0.946 |
| 0.2 | 98 | 0.9281 | 0.928 |
| 0.5 | 17 | 0.9510 | 0.951 |
| 0.5 | 18 | 0.9037 | 0.904 |
| 0.5 | 40 | 0.9193 | 0.919 |
| 0.106 | 100 | 0.9522 | 0.952 |
| 0.107 | 100 | 0.9109 | 0.911 |
| 0.005 | 591 | 0.9449 | 0.945 |
| 0.005 | 592 | 0.7922 | 0.792 |
The rows are already alarming:
- At $p = 0.5$, the most favorable value for the normal approximation, the coverage is 0.951 with 17 observations and 0.904 with 18.
- A sample of 98 does worse than a sample of 30.
- For a rare outcome with $p = 0.005$, adding one observation to a sample of 591 drops the coverage from 0.945 to 0.792.
These are scattered cases. What does the whole picture look like?
Coverage as the sample grows
The figure computes the coverage for the young-adult population, $p = 0.2294$, at every sample size from 10 to 200.
| Interval | Lowest coverage, $10 \le n \le 200$ | Mean coverage | Share of $n$ below 0.93 |
|---|---|---|---|
| Wald | 0.8202 (at $n = 13$) | 0.9374 | 18.3% |
| Wilson | 0.9107 (at $n = 13$) | 0.9505 | 1.0% |
| Agresti–Coull | 0.9316 (at $n = 16$) | 0.9532 | 0% |
| Clopper–Pearson | 0.9521 (at $n = 160$) | 0.9635 | 0% |
The Wald interval's coverage does creep toward 0.95 on average, but not steadily. Between neighboring sample sizes it can jump, from 0.9474 at $n = 12$ to 0.8202 at $n = 13$. Brown, Cai and DasGupta call such sample sizes "lucky" and "unlucky." Their point is that a user cannot know in advance which kind a study has.
Coverage across p
That was one value of $p$. Fixing $n = 40$ instead and computing the coverage at 2,000 values of $p$ spread across $(0, 1)$ gives the second figure.
| Interval ($n = 40$) | Mean coverage over $p$ | Lowest coverage | Share of $p$ below 0.93 | Lowest for $0.1 \le p \le 0.9$ | Mean expected length |
|---|---|---|---|---|---|
| Wald | 0.8907 | 0.0100 | 52.3% | 0.8599 | 0.2341 |
| Wilson | 0.9520 | 0.8434 | 2.3% | 0.9278 | 0.2367 |
| Agresti–Coull | 0.9589 | 0.9330 | 0% | 0.9330 | 0.2426 |
| Clopper–Pearson | 0.9710 | 0.9503 | 0% | 0.9503 | 0.2584 |
The pattern has two parts. Across the middle of the range, the coverage of every method oscillates. The set of counts whose intervals contain $p$ changes abruptly as $p$ crosses an interval's endpoint. That oscillation is unavoidable for any interval computed from a whole-number count. What differs between methods is the level it oscillates around: something below 0.95 for the Wald interval, and 0.95 for Wilson's.
Near the ends the Wald interval fails completely. Before reading on, try to predict how badly.
Check your understanding. Suppose $p = 0.01$ and $n = 40$. The chance of seeing no events at all is $0.99^{40} = 0.669$. What is the Wald interval when $x = 0$, and what does that imply for its coverage?
When $x = 0$ the Wald interval is $0 \pm 0$, the single point zero, which cannot contain a positive $p$. So the coverage is at most $1 - 0.669 = 0.331$, and it is exactly that. At $p = 0.005$ it is 0.18. More generally, the coverage of the Wald interval tends to zero as $p$ approaches 0 or 1, for any fixed $n$ (Brown, Cai and DasGupta, 2001, §2.3).
Why the Wald interval fails
The failure has two sources, both visible in the normal approximation article.
Skewness. For $p$ below one half, $\hat{p}$ has a long right tail and a short left one, but the Wald interval is symmetric around $\hat{p}$. So its two error rates are unequal. The interval misses on one side far more often than $\alpha/2$ and on the other side less often, and the two do not cancel.
The estimated standard error moves with the estimate. Suppose $\hat{p}$ happens to fall closer to 0 than $p$ is. Then it also produces a smaller $\hat{p}(1-\hat{p})$, so the interval is narrow exactly when it most needs to be wide. A sample with an unusually low count gives an interval too short to reach back up to the truth. This is the defect the next interval removes.
Do the textbook rules for when the Wald interval may be used protect against this? The rules are $n\hat{p} \ge 5$ and $n(1-\hat{p}) \ge 5$, or the same with 10, and they do not. Brown, Cai and DasGupta (2001, §2.3) show that under the conditions $np \ge 5$ and $n(1-p) \ge 5$ the lowest coverage, even in the limit of large $n$, is at most 0.875. The case $p = 0.5$, $n = 40$ satisfies every such rule, and its coverage is 0.919. They conclude that the standard interval "is just too risky" (§3) and that it should not be used (§5).
Wilson's score interval
Edwin B. Wilson, a statistician at the Harvard School of Public Health, made the decisive observation in a four-page note in 1927. The usual statement, he wrote, makes the observed rate the center of a range for the true rate. But it is the observed rate that varies around the true one. So the probability statement should be made about the observation, with the standard deviation of the true rate. "A better way to proceed," in his words, is to ask for which values of $p$ the observed rate would not be an unusually bad observation (Wilson, 1927, p. 209).
Here is the idea in the course's notation. Instead of substituting $\hat{p}$ for $p$ in the standard error, keep $p$ in it. Then collect every $p$ for which the observed $\hat{p}$ lies within $z = z_{1-\alpha/2}$ standard errors:
$$\frac{|\hat{p} - p|}{\sqrt{p(1-p)/n}} \le z.$$
This is the set of null values $p_0$ that the one-sample $z$ test of Tests and Sample Size for a Proportion would not reject, since that test uses exactly this standard error, $\sqrt{p_0(1-p_0)/n}$. The interval is the inversion of the test. What remains is algebra: the inequality is quadratic in $p$, and its roots are the limits.
The formal version. Squaring both sides, which is legitimate because both are nonnegative,
$$(\hat{p} - p)^2 \le \frac{z^2}{n}\,p(1 - p).$$
Write $t = z^2/n$ and expand: $\hat{p}^2 - 2\hat{p}p + p^2 \le tp - tp^2$, or
$$(1 + t)\,p^2 - (2\hat{p} + t)\,p + \hat{p}^2 \le 0.$$
The left side is a quadratic in $p$ with positive leading coefficient, so it is negative between its two roots. The quadratic formula gives them. The discriminant is
$$(2\hat{p} + t)^2 - 4(1 + t)\hat{p}^2 = 4\hat{p}t - 4t\hat{p}^2 + t^2 = 4t\hat{p}(1 - \hat{p}) + t^2,$$
and so the roots are
$$p = \frac{2\hat{p} + t \pm \sqrt{4t\hat{p}(1-\hat{p}) + t^2}}{2(1 + t)} = \frac{\hat{p} + t/2 \pm \sqrt{t\hat{p}(1-\hat{p}) + t^2/4}}{1 + t},$$
which is Wilson's formula, in his notation with $t = \lambda^2/n$. Putting $z$ back,
$$\frac{\hat{p} + \dfrac{z^2}{2n} \pm z\sqrt{\dfrac{\hat{p}(1-\hat{p})}{n} + \dfrac{z^2}{4n^2}}}{1 + \dfrac{z^2}{n}},$$
the form the NIST/SEMATECH e-Handbook of Statistical Methods gives (§7.2.4.1). Multiplying numerator and denominator by $n$ gives an equivalent form that is easier to compute, with $x$ the count:
$$\frac{x + z^2/2}{n + z^2} \pm \frac{z\sqrt{n}}{n + z^2}\sqrt{\hat{p}(1-\hat{p}) + \frac{z^2}{4n}}.$$
What the formula does
Check your understanding. The Wald interval for the sample of 40 is centered on $\hat{p} = 0.250$. Where should the center of Wilson's interval sit for the same sample: at 0.250, below it, or above it?
Above it. The center, $(x + z^2/2)/(n + z^2)$, is the proportion obtained by adding $z^2/2$ successes and $z^2/2$ failures to the data. That pulls the center from $\hat{p}$ toward one half, here to 0.2719. The shift is the skewness correction: when $\hat{p}$ is small, the interval reaches further up than down, matching the long right tail. Wilson noticed that with the usual allowance of two standard deviations, $z = 2$, the center is $(x + 2)/(n + 4)$. He offered it in place of Laplace's rule of succession, $(x + 1)/(n + 2)$ (Wilson, 1927, p. 211).
The interval also cannot leave $[0, 1]$, because every value in it is a $p$ between 0 and 1 that satisfies the inequality. And it never collapses to a point: for $x = 0$ it is $(0,\ z^2/(n + z^2))$.
For the sample of 40, with $z^2 = 3.8415$, the center is $(10 + 1.9207)/43.8415 = 0.2719$. The half-width is
$$\frac{1.96\sqrt{40}}{43.8415}\sqrt{0.250 \times 0.750 + \frac{3.8415}{160}} = 0.1300,$$
and the interval is $(0.1419,\ 0.4019)$. At either limit, the $z$ statistic computed with the null standard error is exactly $\pm 1.96$, as the derivation requires.
How well does it work? Wilson's interval has coverage close to the nominal level across nearly all of $(0, 1)$, as the figures show. It costs almost nothing in length: its mean expected length over $p$ at $n = 40$ is 0.2367, against the Wald interval's 0.2341. Its weakness is a few narrow downward spikes in coverage very close to 0 and 1. The lowest, 0.843 at $n = 40$, is at $p = 0.004$. Brown, Cai and DasGupta (2001, §4.1.1) remove these spikes with a modified lower limit for counts of 1, 2 or 3. The NIST handbook (§7.2.4.1) gives the Wilson interval as its method for a proportion. Brown, Cai and DasGupta recommend it for samples of 40 or fewer, or the Jeffreys interval, a Bayesian construction they find very similar.
The Agresti–Coull interval
The Wilson formula is harder to remember than the Wald one. Can the Wald form be kept and repaired? Agresti and Coull (1998, §3) proposed keeping the form and changing what goes into it: for a 95% interval, add two successes and two failures, then compute the ordinary interval. Brown, Cai and DasGupta (2001, §3.1.2) give the version for any confidence level, using Wilson's center. With $\tilde{n} = n + z^2$ and $\tilde{p} = (x + z^2/2)/\tilde{n}$, it is
$$\tilde{p} \pm z\sqrt{\frac{\tilde{p}(1 - \tilde{p})}{\tilde{n}}}.$$
For the sample of 40, $\tilde{n} = 43.8415$, $\tilde{p} = 0.2719$, and the interval is $0.2719 \pm 0.1317 = (0.1402,\ 0.4036)$, slightly wider than Wilson's. It always contains the Wilson interval, so its coverage is a little higher and its length a little longer. At $n = 40$ its coverage never fell below 0.933 at any $p$ the script tried, and its mean expected length was 0.2426. Brown, Cai and DasGupta recommend it for samples larger than 40, where it and the Wilson interval nearly coincide, as the simplest good interval to teach. Unlike Wilson's, it can extend slightly below 0 or above 1. For $x = 0$ and $n = 40$ it is $(-0.017,\ 0.104)$, and such limits are cut back to the possible range.
The Clopper–Pearson interval
Every interval so far rests on the normal approximation. Can we avoid it entirely? Clopper and Pearson (1934) built an interval from the binomial distribution itself. For 95% confidence they chose the limits for each count so that the binomial tail beyond each limit holds probability 0.025. That makes the statement that $p$ lies between them correct in at least 95% of cases in the long run (pp. 406–407). In later terms, the interval inverts two exact one-sided tests instead of approximate ones.
Having observed $x$ events in $n$ trials, the two limits are defined by tail probabilities:
- the lower limit $p_L$ is the value of $p$ for which a count as large as $x$ or larger has probability exactly $\alpha/2$;
- the upper limit $p_U$ is the value for which a count as small as $x$ or smaller has probability $\alpha/2$.
In symbols,
$$P_{p_L}(X \ge x) = \frac{\alpha}{2}, \qquad P_{p_U}(X \le x) = \frac{\alpha}{2},$$
with $p_L = 0$ when $x = 0$ and $p_U = 1$ when $x = n$. The NIST handbook (§7.2.4.1) gives the same two equations as its exact method for small samples or few defects. Each equation has one solution, because $P_p(X \ge x)$ increases steadily with $p$ and $P_p(X \le x)$ decreases.
The limits have a closed form through the beta distribution. Binomial tail probabilities are incomplete beta functions, $P_p(X \ge x) = I_p(x,\ n - x + 1)$, where $I_p(a, b)$ is the distribution function of a $\text{Beta}(a, b)$ variable at $p$. So the lower limit is the $\alpha/2$ quantile of $\text{Beta}(x,\ n - x + 1)$, and the upper limit is the $1 - \alpha/2$ quantile of $\text{Beta}(x + 1,\ n - x)$ (Brown, Cai and DasGupta, 2001, §4.2.1). For the sample of 40 these are 0.1269 and 0.4120. The script checks them by computing the tail probabilities: $P(X \ge 10) = 0.02500$ at $p = 0.1269$ and $P(X \le 10) = 0.02500$ at $p = 0.4120$.
Why it never undercovers, and what that costs
The Clopper–Pearson interval guarantees coverage of at least $1 - \alpha$ for every $p$ and $n$. Here is the idea: a miss on either side can happen only for counts in a binomial tail, and each tail was built to hold less than $\alpha/2$.
The formal version. Fix the true $p$. The interval misses $p$ from above, $p_U(X) < p$, exactly when the observed count is so small that $P_p(X \le x) < \alpha/2$. That is because $P_{p'}(X \le x)$ decreases in $p'$ and equals $\alpha/2$ at $p' = p_U(x)$. The counts with this property form a lower tail $\{0, 1, \ldots, x^*\}$. Its total probability is $P_p(X \le x^*)$, which is less than $\alpha/2$ by the definition of $x^*$. The same argument bounds the chance of missing from below by $\alpha/2$. So the interval misses with probability less than $\alpha$.
Check your understanding. If one interval guarantees at least 95% coverage for every $p$, why not always use it?
"Less than" is the catch. Because the binomial is discrete, the tail probabilities jump, and the actual error rate on each side is usually well under $\alpha/2$. At $n = 40$ the coverage averaged 0.971 over $p$, and at the young-adult prevalence it was 0.978. The price is length. The mean expected length is 0.2584 at $n = 40$, against Wilson's 0.2367, so the exact interval is about 9% wider while aiming at the same confidence level. Brown, Cai and DasGupta (2001, §4.2.1) call it "wastefully conservative." They do not consider it a good choice for practical use unless strict adherence to coverage of at least $1 - \alpha$ is demanded. The word "exact" refers to its use of the exact binomial distribution, not to exact coverage.
The four intervals compared
Back to the sample of 40 young adults, $x = 10$ of $n = 40$:
| Interval | Lower | Upper | Width | Coverage at $p = 0.2294$, $n = 40$ | Expected length there |
|---|---|---|---|---|---|
| Wald | 0.1158 | 0.3842 | 0.2684 | 0.9109 | 0.2557 |
| Wilson | 0.1419 | 0.4019 | 0.2601 | 0.9628 | 0.2495 |
| Agresti–Coull | 0.1402 | 0.4036 | 0.2634 | 0.9628 | 0.2540 |
| Clopper–Pearson | 0.1269 | 0.4120 | 0.2850 | 0.9780 | 0.2734 |
All four contain the true 0.2294 in this sample. The last two columns describe the methods rather than the sample. Over all samples of 40 from this population, the Wald interval would cover 91% of the time while being, on average, longer than Wilson's. The Wald interval's apparent economy does not exist. It is centered in the wrong place and pays for it in coverage.
The binomial calculation assumes independent observations, while the samples are drawn without replacement from 1,696 people. Does that change the picture? The script drew 10,000 samples of 40 for each of 30 seeds and counted the intervals that contained 0.2294. Median coverage was:
- 0.9142 for Wald (interquartile range 0.9114 to 0.9161);
- 0.9654 for Wilson and Agresti–Coull;
- 0.9799 for Clopper–Pearson.
Each is within 0.004 of the binomial value in the table. Sampling without replacement makes $\hat{p}$ slightly less variable, so coverage is slightly higher.
When no events are observed
Rare outcomes produce samples with $x = 0$, and the choice of interval matters most there. For $n = 40$ and $x = 0$:
| Interval | 95% interval |
|---|---|
| Wald | $(0,\ 0)$ |
| Wilson | $(0,\ 0.0876)$ |
| Agresti–Coull | $(-0.017,\ 0.104)$, cut to $(0,\ 0.104)$ |
| Clopper–Pearson | $(0,\ 0.0881)$ |
The Wald interval claims certainty that the proportion is zero, which no sample can show. A sensible statement is an upper bound. With $x = 0$, the exact one-sided $1 - \alpha$ upper bound solves $(1 - p)^n = \alpha$, so $p_U = 1 - \alpha^{1/n}$. For 95% and $n = 40$ it is $0.0722$.
Since $\ln(1 - p) \approx -p$ for small $p$, the bound is approximately $-\ln(0.05)/n = 2.996/n$. This is the rule of three: when no events are seen in $n$ trials, $3/n$ is an approximate 95% upper confidence bound for the event rate. Here $3/40 = 0.075$. Wilson made the same point in 1927 with his upper limit $\lambda^2/(n + \lambda^2)$ for a sample with no events ("total failure," in his words). For $\lambda = 1.96$ and $n = 40$ it is the 0.0876 in the table.
Rare events in large samples arise well beyond medicine. The library's notes on Sycophancy to Subterfuge describe a study of rare model behaviors that compared several interval methods for counts such as 5 events in 10,000 trials. Adding error bars to evals applies the standard error of a proportion to language model benchmarks, where every question is scored right or wrong.
Which interval to use
| Situation | Interval |
|---|---|
| General use, any $n$ | Wilson |
| $n$ above about 40, and a formula that is easy to teach | Agresti–Coull (add $z^2/2$ successes and failures) |
| Coverage of at least $1 - \alpha$ required for every $p$ | Clopper–Pearson |
| A quick figure for a large sample with $\hat{p}$ far from 0 and 1 | Wald, knowing it runs a little short |
| No events observed | Wilson or Clopper–Pearson; for a one-sided bound, $1 - \alpha^{1/n}$, about $3/n$ |
One-sided bounds follow the same constructions, with $z_{1-\alpha}$ in place of $z_{1-\alpha/2}$ (NIST, §7.2.4.1), or $\alpha$ in place of $\alpha/2$ in the Clopper–Pearson equations. The side must be chosen before seeing the data, for the reasons in One-Sided Confidence Bounds and Sample Size.
What we have learned
The textbook interval for a proportion, the Wald interval, borrows the recipe for a mean, and the recipe does not transfer. Computed exactly, its coverage for samples of 40 young adults is 0.9109, and over all $p$ at $n = 40$ it averages 0.8907. It fails for two reasons: it is symmetric around a skewed estimate, and its estimated standard error shrinks exactly when the estimate is too low. Wilson's interval keeps the unknown $p$ in the standard error. It covers close to 95% at almost the same length. The Agresti–Coull interval gets nearly the same result from the Wald formula with $z^2/2$ added successes and failures. Clopper–Pearson guarantees at least 95% and pays for the guarantee with about 9% more length.
Wilson's interval came from asking which values of $p$ a test would not reject. That test, the one-sample $z$ test, is the subject of Tests and Sample Size for a Proportion. The next article measures how well it keeps its stated error rate, and asks how many observations a study of a proportion needs.
In Python
from scipy import stats
r = stats.binomtest(10, 40)
print(r.proportion_ci(0.95, method='wilson')) # about (0.1419, 0.4019)
print(r.proportion_ci(0.95, method='exact')) # Clopper-Pearson, about (0.1269, 0.4120)
def coverage(lower, upper, n, p):
"""Exact coverage: total probability of the counts whose interval contains p."""
return sum(stats.binom.pmf(x, n, p) for x in range(n + 1)
if lower(x, n) <= p <= upper(x, n))
Every number in this article is computed by tools/experiments/applied-statistics/confidence-intervals-for-a-proportion.py in the site's repository. Its Wilson and Clopper–Pearson limits agree with scipy's binomtest(...).proportion_ci to within $10^{-12}$, and its coverage calculation reproduces the published values above.
References
- Agresti, A., and Coull, B. A. (1998). Approximate is better than "exact" for interval estimation of binomial proportions. The American Statistician, 52(2), 119–126. record
- Brown, L. D., Cai, T. T., and DasGupta, A. (2001). Interval estimation for a binomial proportion. Statistical Science, 16(2), 101–133. record · 4 pages
- Clopper, C. J., and Pearson, E. S. (1934). The use of confidence or fiducial limits illustrated in the case of the binomial. Biometrika, 26(4), 404–413. record · 2 pages
- National Center for Health Statistics (2020). National Health and Nutrition Examination Survey 2017–2018: DEMO_J, BPX_J and BPQ_J data files. Centers for Disease Control and Prevention. https://wwwn.cdc.gov/nchs/nhanes/continuousnhanes/default.aspx?BeginYear=2017 record · 45 pages
- NIST/SEMATECH. e-Handbook of Statistical Methods, §7.2.4.1 (Confidence intervals). National Institute of Standards and Technology. https://www.itl.nist.gov/div898/handbook/ record · 30 pages
- Pagano, M., Gauvreau, K., and Mattie, H. (2022). Principles of Biostatistics (3rd ed.). CRC Press. The course text. record · 35 pages
- SciPy documentation.
scipy.stats.binomtestandBinomTestResult.proportion_ci. https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.binomtest.html record - Wilson, E. B. (1927). Probable inference, the law of succession, and statistical inference. Journal of the American Statistical Association, 22(158), 209–212. record · 3 pages