Interpreting Standard Error for Proportions

Why a Sample Proportion Varies

Imagine repeatedly drawing samples of the same size from a population in which a fixed proportion of individuals has a particular characteristic. Each sample produces its own sample proportion, $\hat{p}$. Independent samples can produce different values of $\hat{p}$, but whether they usually differ depends on the population proportion and the sample size. For example, if every individual has the characteristic, every sample proportion is 1; if no individual has it, every sample proportion is 0. The standard error describes the amount of variation in $\hat{p}$ across repetitions of the sampling process.

It is important to separate that repeated-sample variation from the result of any one sample. A particular sample gives one value of $\hat{p}$, which may be above or below the population proportion $p$. The standard error is not the difference between those two values. It describes the spread expected across the sampling distribution: the distribution of sample proportions from repeated samples taken under the same model.

Definition: The standard error of a sample proportion is the standard deviation of the sampling distribution of $\hat{p}$. Under independent sampling with a fixed population proportion $p$ and sample size $n$, it is $$ SE(\hat{p})=\sqrt{\frac{p(1-p)}{n}}. $$ It measures the repeated-sample variability of $\hat{p}$, in proportion units.

The Model Behind the Formula

Use the formula when the sample can be represented by $n$ independent observations, each with the same probability $p$ of having the characteristic. The sample size $n$ is fixed, and $\hat{p}$ is the number of sampled individuals with the characteristic divided by $n$. These assumptions describe independent Bernoulli trials. In particular, the probability of the characteristic is the same for each observation, and the outcomes of different observations are independent. The formula is exact for this model.

In a setting where individuals are sampled without replacement from a finite population, the observations are not exactly independent. The formula above may then need an adjustment. Here, the standard error calculations use the stated independent-sampling model.

The formula uses the population proportion $p$, not the observed proportion from one sample. If $p$ is specified, the formula gives the model's standard error directly. When $p$ is unknown, one can estimate the standard error by substituting $\hat{p}$ for $p$; that is an estimated standard error, rather than the theoretical value calculated from a specified $p$.

Why This Is a Standard Deviation

A short derivation connects the formula to its interpretation. For each sampled individual, define an indicator variable $I_i$ that equals 1 if the individual has the characteristic and 0 otherwise. Under the model, $P(I_i=1)=p$, so $E(I_i)=p$ and $\operatorname{Var}(I_i)=p(1-p)$. The sample proportion is the average of these indicators:

\hat{p}=\frac{I_1+\cdots+I_n}{n}.

Independence means that the variance of the sum is the sum of the variances. Therefore,

\operatorname{Var}(\hat{p}) =\operatorname{Var}\left(\frac{I_1+\cdots+I_n}{n}\right) =\frac{1}{n^2}\sum_{i=1}^{n}\operatorname{Var}(I_i) =\frac{n p(1-p)}{n^2} =\frac{p(1-p)}{n}.

The standard deviation is the square root of the variance, giving $SE(\hat{p})=\sqrt{p(1-p)/n}$. Also, $E(\hat{p})=p$, so the sampling distribution is centered at $p$. This establishes why the formula describes spread around the population proportion, rather than the error in a particular observed sample.

Worked Example: Interpreting a Standard Error

Suppose a hypothetical population has $p=0.30$ of its individuals meeting a stated criterion. A sample of $n=100$ is drawn under the independent-sampling model. Find and interpret the standard error of $\hat{p}$.

Check the model. The calculation assumes 100 independent observations, each with probability 0.30 of meeting the criterion. The sample size and population proportion are specified.

Calculate.

SE(\hat{p})=\sqrt{\frac{p(1-p)}{n}} =\sqrt{\frac{0.30(1-0.30)}{100}} =\sqrt{\frac{0.21}{100}} =\sqrt{0.0021} \approx 0.04583.

Interpret. Across repeated independent samples of 100 from this population, the sample proportions have a standard deviation of about 0.0458 proportion units around $p=0.30$. Since one proportion unit is 100 percentage points, this is about 4.58 percentage points. It describes the scale of the sampling distribution's spread; it does not say that every sample proportion will be within 4.58 percentage points of 0.30.

Comparing Standard Errors

To compare two standard errors, calculate each using the population proportion and sample size specified for that sampling model. A larger standard error means greater spread among repeated sample proportions under that model. A smaller one means less spread. The comparison does not tell us which value a particular sample will produce.

For a fixed sample size, the expression $p(1-p)$ is largest at $p=0.50$. One way to see this is to write $p(1-p)=0.25-(p-0.50)^2$. Because a square cannot be negative, $p(1-p)\leq 0.25$, with equality at $p=0.50$. Thus, for a given $n$, the model's standard error is largest when the population proportion is one-half. Proportions equally far below and above one-half give the same standard error: $p(1-p)=(1-p)p$.

Worked Example: Comparing Two Population Proportions

Two hypothetical populations are sampled independently, with $n=100$ observations from each. The population proportion is $p=0.50$ in the first population and $p=0.20$ in the second. Compare the standard errors.

Check the model. For each calculation, assume 100 independent observations with the stated common probability of the characteristic.

For the first population:

SE(\hat{p}_1)=\sqrt{\frac{0.50(1-0.50)}{100}} =\sqrt{0.0025}=0.05000.

For the second population:

SE(\hat{p}_2)=\sqrt{\frac{0.20(1-0.20)}{100}} =\sqrt{0.0016}=0.04000.

The standard error is 0.0500, or 5.00 percentage points, for the population with $p=0.50$; it is 0.0400, or 4.00 percentage points, for the population with $p=0.20$. Under these models, repeated sample proportions are more variable for the first population. Both calculations use the same sample size, so the difference comes from the different values of $p(1-p)$.

Worked Example: Comparing Specified Proportions at the Same Sample Size

A hypothetical quality-control process has a population proportion $p=0.25$ of items that do not meet a specification. Another process has $p=0.60$. For each process, take an independent sample of $n=80$. Which sample proportion has the greater standard error?

Check the model. For each process, assume each of the 80 sampled items independently has the stated probability of not meeting the specification.

For the first process:

SE(\hat{p}_1)=\sqrt{\frac{0.25(1-0.25)}{80}} =\sqrt{\frac{0.1875}{80}} =\sqrt{0.00234375} \approx 0.04841.

For the second process:

SE(\hat{p}_2)=\sqrt{\frac{0.60(1-0.60)}{80}} =\sqrt{\frac{0.24}{80}} =\sqrt{0.003} \approx 0.05477.

The second process has the greater standard error: about 0.0548 rather than 0.0484. In percentage-point terms, the repeated-sample standard deviations are about 5.48 and 4.84 percentage points, respectively. Although the population proportions are on opposite sides of 0.50, $0.60$ is closer to 0.50 than $0.25$, so its value of $p(1-p)$ is larger.

Worked Example: Comparing Two Specified Sample Sizes

For a hypothetical population with $p=0.40$, compare the standard errors for independent samples of $n=100$ and $n=400$.

Check the model. Both samples are modeled as independent observations with the same population probability $p=0.40$; only the specified sample size differs.

For $n=100$:

SE_{100}(\hat{p})=\sqrt{\frac{0.40(0.60)}{100}} =\sqrt{0.0024} \approx 0.04899.

For $n=400$:

SE_{400}(\hat{p})=\sqrt{\frac{0.40(0.60)}{400}} =\sqrt{0.0006} \approx 0.02449.

For these specified models, the standard errors are about 4.90 and 2.45 percentage points. The second is smaller. This calculation compares the two stated cases; the formula can be used in the same way whenever $p$ and $n$ are given.

What a Standard Error Does Not Say

A standard error is a standard deviation, so it summarizes spread rather than giving a guaranteed maximum difference. A sample proportion can be more than one standard error from $p$, just as a value from any distribution can be more than one standard deviation from its mean. Without additional information about the shape of the sampling distribution, the standard error alone does not give the probability that $\hat{p}$ falls within a specified distance of $p$.

It is also possible for two independent samples to produce the same sample proportion. Whether two proportions usually agree or differ depends on the model parameters. For instance, when $n=1$ and $p=0.10$, each sample proportion is either 0 or 1. The probability that two independent sample proportions are equal is $0.10^2+0.90^2=0.82$, so they are equal more often than not. If $p=0$ or $p=1$, all sample proportions are identical. These cases do not contradict the standard-error formula: it describes the amount of variability, which is zero at $p=0$ and $p=1$, and can be small in other settings.

Common pitfall: Saying “the standard error is the amount this sample's proportion is wrong” confuses a sampling-distribution standard deviation with one sample's realized difference $\hat{p}-p$. The correction is to interpret the standard error as the spread of $\hat{p}$ across repeated samples under a stated model. One observed sample has one realized difference; its standard error is not that difference.

When reporting a standard error in context, name the proportion, state the repeated-sampling setting, and include the scale. For example, “The sample proportion of items failing inspection has a repeated-sample standard deviation of about 0.0484, or 4.84 percentage points, for samples of 80 under this model” is more informative than reporting only “SE = 0.0484.”

Check Your Understanding

Use the formula and the interpretation principles in this tutorial.

For $p=0.15$ and $n=200$, calculate $SE(\hat{p})$ and interpret it in proportion units and percentage points.
For a fixed sample size, why is the standard error largest when $p=0.50$?
Two independent samples can have the same sample proportion. What determines whether they usually agree or differ?
A learner says, “The standard error is the difference between the observed sample proportion and $p$.” Identify and correct the error.
For $p=0.30$, calculate and compare the standard errors for $n=50$ and $n=200$. State what the comparison means across repeated samples.