Estimating the Standard Error of a Proportion

Why Estimate a Standard Error?

A sample proportion is not usually identical to the population proportion it is intended to describe. If we repeatedly took samples of the same size under the same conditions, the sample proportions would vary. The standard error describes the standard deviation of that sampling distribution: it measures the typical amount of variation in the sample proportion from sample to sample.

The population proportion is generally unknown, so its standard error cannot usually be calculated directly. Instead, we use the observed sample proportion to estimate it. This estimate gives a scale for sampling variability; it is not the difference between the sample result and the true population value, and it is not itself an interval for the population proportion.

From the True Standard Error to Its Estimate

Suppose each observation records whether a particular outcome occurred. Code an occurrence as 1 and a non-occurrence as 0. Under a model in which a fixed sample of \(n\) independent observations has the same probability \(p\) of an occurrence, the sample proportion is the number of occurrences divided by \(n\). The standard deviation of this sample proportion is called its true standard error.

Definition: For \(n\) independent binary observations with a common population proportion \(p\), the sample proportion is \(\hat{p}\), and its true standard error is \(\sqrt{p(1-p)/n}\). Because \(p\) is usually unknown, the estimated standard error is \(\sqrt{\hat{p}(1-\hat{p})/n}\).

The assumptions matter: the sample size is fixed, each observation has two possible outcomes, and the observations are independent with the same probability \(p\) of the outcome being counted. A random sample from a sufficiently large population can often be treated this way. If observations are dependent, or sampling without replacement from a small finite population, this formula may not describe the sampling variability exactly.

The formula for the estimate follows by replacing the unknown \(p\) in the true standard error with the observed \(\hat{p}\). Since \(\hat{p}=x/n\), where \(x\) is the number of observed occurrences, the calculation can also be written directly in terms of the count:

\widehat{SE}(\hat{p}) = \sqrt{\frac{\hat{p}(1-\hat{p})}{n}} = \sqrt{\frac{(x/n)(1-x/n)}{n}}

The result is in proportion units. For example, an estimated standard error of 0.04 corresponds to 4 percentage points, not 0.04 percentage points. It describes the estimated standard deviation of sample proportions across repeated samples of the same size and under the same model.

Worked Examples

Worked Example: Estimating a Proportion’s Standard Error

In a hypothetical sample of 120 adults, 36 report meeting a specified daily exercise target. Estimate the standard error of the sample proportion reporting that they meet the target.

The observations are binary: each adult either reports meeting the target or does not. Assume the 120 observations are independent and have a common probability \(p\) of reporting that outcome. Under this model, the plug-in standard error formula applies.

First calculate the sample proportion:

\hat{p}=\frac{x}{n}=\frac{36}{120}=0.30

Then substitute \(\hat{p}=0.30\) and \(n=120\) into the estimated standard error formula:

\widehat{SE}(\hat{p}) = \sqrt{\frac{0.30(1-0.30)}{120}} = \sqrt{\frac{0.21}{120}} = \sqrt{0.00175} \approx 0.0418

The estimated standard error is about 0.0418, or 4.18 percentage points. In repeated independent samples of 120 adults under the same model, the sample proportions would have an estimated standard deviation of about 4.18 percentage points. This does not say that this particular sample proportion is 4.18 percentage points away from the population proportion.

Worked Example: Interpreting the Estimate in Context

In a hypothetical quality check, 18 of 40 independently selected devices pass a specified test. Estimate the standard error of the sample proportion that pass.

Each device has one of two outcomes, pass or not pass. Assume the selected devices can be treated as independent observations with the same probability \(p\) of passing. The fixed sample size is \(n=40\), so use the estimated standard error formula.

The observed proportion passing is:

\hat{p}=\frac{18}{40}=0.45

The estimated standard error is:

\widehat{SE}(\hat{p}) = \sqrt{\frac{0.45(1-0.45)}{40}} = \sqrt{\frac{0.2475}{40}} = \sqrt{0.0061875} \approx 0.0787

The estimate is about 0.0787, or 7.87 percentage points. This indicates the estimated spread of sample pass proportions over repeated samples of 40 devices from the same process. It does not mean that 7.87% of these 40 devices were measured incorrectly, nor does it give the proportion of devices that pass.

Why the Estimate Changes

The estimated standard error depends on both the sample size \(n\) and the observed proportion \(\hat{p}\). Holding the proportion fixed, increasing \(n\) makes the estimated standard error smaller because the denominator inside the square root increases. More precisely, the estimate is proportional to \(1/\sqrt{n}\) when \(\hat{p}\) stays the same. For example, quadrupling the sample size halves the estimated standard error.

The proportion also matters. The product \(\hat{p}(1-\hat{p})\) is largest when \(\hat{p}=0.5\) and smaller when \(\hat{p}\) is closer to 0 or 1. To see why, complete the square:

\hat{p}(1-\hat{p}) = \frac{1}{4}-(\hat{p}-0.5)^2 \leq \frac{1}{4}

The squared term is nonnegative, so the product cannot exceed \(1/4\), with equality at \(\hat{p}=0.5\). Thus, for a fixed \(n\), the estimated standard error is largest at a sample proportion of 0.5. This describes the formula’s behavior; it does not imply that every sample proportion near 0 or 1 is known with certainty.

Worked Example: Comparing Sample Sizes

Two hypothetical independent samples from the same process each have an observed proportion of 0.40. One sample has \(n=100\), and the other has \(n=400\). Compare their estimated standard errors. Assume binary outcomes and the same probability of the outcome across observations in each sample.

For the sample of 100:

\widehat{SE}_{100} = \sqrt{\frac{0.40(1-0.40)}{100}} = \sqrt{0.0024} \approx 0.0490

For the sample of 400:

\widehat{SE}_{400} = \sqrt{\frac{0.40(1-0.40)}{400}} = \sqrt{0.0006} \approx 0.0245

The second sample is four times as large, and its estimated standard error is half as large: \(0.0245 \approx 0.0490/2\). In percentage-point terms, the estimates are about 4.90 and 2.45 percentage points. This comparison concerns the sampling variability expected for the two sample proportions, not how close either observed proportion must be to the unknown population proportion.

A Boundary Case to Handle Carefully

If all observations have the same outcome, then \(\hat{p}=0\) or \(\hat{p}=1\). Substitution into the plug-in formula gives an estimated standard error of zero. For example, if 25 out of 25 observations are successes, then:

\hat{p}=1,\qquad \widehat{SE}(\hat{p}) = \sqrt{\frac{1(1-1)}{25}} =0

That zero is the result of the plug-in calculation, not proof that the unknown population proportion is known exactly or that future samples cannot vary. The formula uses the observed proportion as an estimate of \(p\); in a boundary sample, this substitution can understate uncertainty. Recognizing this limitation is important when interpreting the calculation.

Common Mistakes

Using \(p\) as if it were known. The true standard error uses the population proportion, but in ordinary estimation \(p\) is unknown. Use \(\hat{p}\) to calculate the estimated standard error.
Forgetting the sample size inside the square root. The formula is \(\sqrt{\hat{p}(1-\hat{p})/n}\), not \(\sqrt{\hat{p}(1-\hat{p})}/n\).
Calling the standard error the sample’s actual error. The estimated standard error describes a typical spread across repeated samples under the model. It does not reveal the exact difference between this sample proportion and \(p\).
Confusing proportions and percentage points. An estimate of 0.05 is 5 percentage points when expressed as a percentage. State which scale you are using.
Treating zero as certainty. An estimate of zero from an all-success or all-failure sample is a boundary result of the plug-in formula, not evidence that the population proportion has no uncertainty.

Key takeaway: For binary observations under an independent, common-probability model, estimate the standard error of a sample proportion with \(\sqrt{\hat{p}(1-\hat{p})/n}\). Interpret it as an estimated standard deviation across repeated sample proportions, in proportion units or percentage points.

Check Your Understanding

Answer each question using the definition and formula above.

A hypothetical sample has 28 successes among 80 independent binary observations. Calculate \(\hat{p}\) and its estimated standard error.
In context, what does an estimated standard error of 0.03 mean if the sample proportion is a rate reported as a percentage?
If the observed proportion is held fixed and the sample size is multiplied by four, what happens to the estimated standard error? Explain using the formula.
Why is the estimated standard error largest, for a fixed sample size, when \(\hat{p}=0.5\)?
A sample contains only successes, so the plug-in estimated standard error is zero. What does that calculation show, and what does it not establish?