Comparing Variability for Proportions and Means

Two Standard Errors, Two Kinds of Measurement

A sample proportion and a sample mean are both statistics that vary from sample to sample. Their standard errors describe the spread of those statistics across repeated samples, but the formulas use different population quantities. A proportion depends on the population proportion \(p\); a mean depends on the population standard deviation \(\sigma\). Their units differ as well: a proportion is a fraction (or can be reported in percentage points), while a mean has the units of the variable being averaged.

Keeping those distinctions clear matters whenever two estimates are compared. A standard error of 0.02 for a proportion and a standard error of 2 for a mean are not directly comparable just because 2 is numerically larger. They measure spread on different scales.

The Formulas and Their Population Quantities

For a binary outcome, code each observation as 1 if the outcome is present and 0 otherwise. The sample proportion \(\hat{p}\) is the fraction of observations coded 1. If the population proportion is \(p\), then the variance of one such indicator is \(p(1-p)\). For a sample mean, the relevant population variance is \(\sigma^2\), the variance of the quantitative variable.

The first formula applies when each observation records whether a specified outcome occurs, with the same population probability \(p\). The second applies to a quantitative variable with population mean \(\mu\) and finite population standard deviation \(\sigma\). In both cases, the formulas as written assume independent observations, as with independent sampling or sampling with replacement. When sampling without replacement from a finite population, the finite population correction may be needed, as covered in the earlier material on sampling with and without replacement and the finite population correction.

In practice, the population quantities are often unknown. For a proportion, use the observed proportion \(\hat{p}\) in place of \(p\) to estimate the standard error. For a mean, use the sample standard deviation \(s\) in place of \(\sigma\). These give estimated standard errors, not the exact theoretical standard errors:

Why the Formulas Differ

The formulas share the factor \(1/\sqrt{n}\), but their numerators describe different kinds of population spread. For a binary indicator, the only possible values are 0 and 1, so its population variance is determined entirely by \(p\): it equals \(p(1-p)\). For a general quantitative variable, values can have many possible magnitudes, and their spread is summarized by \(\sigma\).

The comparison can be justified directly from the variance of an average. For \(n\) independent, identically distributed observations \(X_1,\ldots,X_n\), each with mean \(\mu\) and variance \(\sigma^2\), the sample mean is \(\bar{x}=(X_1+\cdots+X_n)/n\). Independence means the variance of the sum is the sum of the variances:

Taking the square root gives \(SE(\bar{x})=\sigma/\sqrt{n}\). For a binary indicator \(I_i\), its average is exactly the sample proportion: \(\hat{p}=(I_1+\cdots+I_n)/n\). Since \(P(I_i=1)=p\), its mean is \(p\), and its variance is

Applying the same variance-of-an-average calculation gives \(\operatorname{Var}(\hat{p})=p(1-p)/n\), and therefore \(SE(\hat{p})=\sqrt{p(1-p)/n}\). This also shows that a sample proportion is a special case of a sample mean: it is the mean of 0–1 indicators. In that special case, the mean's population standard deviation is \(\sqrt{p(1-p)}\), so the two formulas agree.

Worked Comparisons

What Changes With Sample Size?

For a fixed population, both standard errors are proportional to \(1/\sqrt{n}\). If sample size is multiplied by four, the standard error is divided by two. For example, the first worked example's proportion standard error at \(n=400\) is about 0.02291. At \(n=1600\), with the same \(p=0.30\), it is

which is half as large. The same sample-size scaling holds for the mean formula when \(\sigma\) is unchanged. But the formulas do not imply equal standard errors: the spread of the underlying measurements, or the value of \(p(1-p)\), still matters.

For proportions, \(p(1-p)\) is largest when \(p=0.50\) and becomes smaller as \(p\) approaches 0 or 1. Consequently, at a fixed sample size, proportions near 0.50 have greater standard errors than proportions closer to either boundary. For means, the population standard deviation \(\sigma\) sets the scale: a variable measured in larger units or with more population spread can have a larger standard error.

Common Mistakes and a Practical Checklist

A frequent mistake is to compare standard errors as if their numerical values were dimensionless scores. For instance, saying that a mean's standard error of 1.5 is larger than a proportion's standard error of 0.049 does not establish that its sampling distribution is more variable in a meaningful, scale-independent sense. The correction is to state the units and interpret each standard error in context. A proportion's standard error can be converted to percentage points; a mean's standard error remains in the original measurement units.

Another mistake is to use the same population quantity in both formulas. A standard error for \(\hat{p}\) is based on \(p(1-p)\), not on a standard deviation for an unrelated quantitative outcome. A standard error for \(\bar{x}\) uses \(\sigma\), or its estimate \(s\), rather than \(p\). Likewise, distinguish a theoretical calculation using population quantities from an estimate based on sample data.

Before calculating or interpreting either standard error, check the following:

• Is the statistic a sample proportion for a binary outcome, or a sample mean for a quantitative variable?
• Does the formula use the appropriate population quantity: \(p\) for a proportion or \(\sigma\) for a mean?
• Are the observations independent under the sampling plan? If sampling without replacement from a finite population, is a finite population correction needed?
• Have the result's units been reported and interpreted on the correct scale?

These checks make the comparison precise: both standard errors describe sampling-distribution spread and both shrink at the \(1/\sqrt{n}\) rate under the stated conditions, while their population quantities and units remain distinct.