How Sample Size Changes Proportion Variability

Why Sample Size Matters

A sample proportion varies from sample to sample. If a survey is repeated many times under the same conditions, its sample proportions will not all be identical, even though they come from a population with the same underlying proportion $p$. Sample size controls how much these proportions typically vary: larger samples generally produce less variable sample proportions.

The standard error gives a numerical measure of that sampling variability. The key comparison is not merely that “more data are better,” but how the standard error changes when the sample size changes. Under the independent-sampling model, multiplying the sample size by four halves the standard error. This is a statement about typical variability across repeated samples, not a guarantee about any one sample.

The Model and Standard Error

Suppose each observation records whether an individual has a specified characteristic, called a success. Assume the observations are independent, and that each has the same probability $p$ of success. For example, success might mean that a randomly selected household uses a particular service. The number of observations $n$ is fixed, and the sample proportion $\hat{p}$ is the number of successes divided by $n$.

These conditions describe independent Bernoulli observations. They are appropriate for independent sampling, such as sampling with replacement, or when sampling from a population large enough that dependence from sampling without replacement can be ignored. If observations are strongly dependent, or a substantial fraction of a finite population is sampled without replacement, the formula below may not apply as written.

Definition: Under independent observations with common success probability $p$, the sampling distribution of the sample proportion $\hat{p}$ has mean $p$ and standard error $$ SE(\hat{p})=\sqrt{\frac{p(1-p)}{n}}. $$ The standard error measures the standard deviation of $\hat{p}$ across repeated samples under this model. It is expressed in proportion units; for example, $0.04$ is also 4 percentage points.

For comparisons of sample size, hold $p$ fixed: the same population and the same definition of success are being considered. The sample size appears in the denominator inside the square root. Thus the variance is $p(1-p)/n$, while the standard error is its square root. Confusing these two quantities can lead to an incorrect claim about how rapidly variability changes.

How the Standard Error Changes with Sample Size

Let $n_1$ and $n_2$ be two positive sample sizes for the same population proportion $p$. When $0<p<1$, both standard errors are positive, so their ratio is

\frac{SE_{n_2}(\hat{p})}{SE_{n_1}(\hat{p})} = \frac{\sqrt{p(1-p)/n_2}}{\sqrt{p(1-p)/n_1}} = \sqrt{\frac{n_1}{n_2}}.

The common factor $p(1-p)$ cancels. If $n_2>n_1$, then $n_1/n_2<1$, so its square root is also less than 1. This proves that, for an interior population proportion, increasing the sample size strictly decreases the standard error. The amount of decrease depends on the sample-size ratio, not on which interior value of $p$ is being compared.

At the boundary values $p=0$ and $p=1$, the numerator $p(1-p)$ is zero. The standard error is therefore zero for every positive sample size. It does not strictly decrease at these boundaries; it remains zero. In either boundary case every observation has the same outcome under the model, so every sample proportion equals the population proportion.

Theorem: For a fixed population proportion with $0<p<1$, multiplying $n$ by a factor $k>0$ divides the standard error by $\sqrt{k}$. For $p=0$ or $p=1$, the standard error is zero at every positive sample size. In particular, quadrupling $n$ halves the standard error for $0<p<1$, because $\sqrt{4}=2$.

The result also describes diminishing returns from adding observations. For example, increasing a sample size from 100 to 200 does not halve the standard error; it multiplies it by $\sqrt{100/200}$, about 0.707. To halve the standard error, the sample size must be multiplied by four. To reduce it to one-third, the sample size must be multiplied by nine.

Worked Example: Quadrupling the Sample Size

Suppose a hypothetical population has proportion $p=0.30$ with a specified characteristic. Compare the standard errors of sample proportions from independent samples of size 100 and 400.

The model uses the same $p$ for both sample sizes, with independent observations. The value $p=0.30$ is between 0 and 1, so the standard error formula applies with a positive value. First, $p(1-p)=0.30(0.70)=0.21$.

For $n=100$,

SE_{100}(\hat{p}) =\sqrt{\frac{0.30(0.70)}{100}} =\sqrt{0.0021} \approx 0.04583.

For $n=400$,

SE_{400}(\hat{p}) =\sqrt{\frac{0.30(0.70)}{400}} =\sqrt{0.000525} \approx 0.02291.

The second standard error is half the first, since the sample size was multiplied by four. In percentage-point terms, the standard errors are about 4.58 and 2.29 percentage points. Both sampling distributions are centered at the same population proportion, $0.30$; it is their spread, not their center, that changes with sample size.

Comparing Sample Sizes Directly

The ratio formula is often the quickest way to compare two sample sizes. It avoids recalculating $p(1-p)$ and makes the size of the change easy to see. Use it only when comparing the same population proportion under the same sampling model. If either $p$ or the sampling design changes, the cancellation shown in the derivation need not give the full comparison.

Worked Example: Comparing Standard Errors at Two Sample Sizes

Imagine repeated independent samples from a population in which $p=0.60$. Compare sample sizes $n_1=25$ and $n_2=100$. The population proportion is interior, and the observations are assumed independent with a common success probability, so the standard-error formula is appropriate.

For the first sample size,

SE_{25}(\hat{p}) =\sqrt{\frac{0.60(0.40)}{25}} =\sqrt{0.0096} \approx 0.09798.

For the second,

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

The direct comparison gives the same answer:

\frac{SE_{100}(\hat{p})}{SE_{25}(\hat{p})} =\sqrt{\frac{25}{100}} =0.5.

Thus the standard error falls from about 9.80 to 4.90 percentage points. The sample size is four times as large, so the standard error is half as large. This does not say that every sample of 100 will produce a proportion closer to $0.60$ than every sample of 25; it describes the relative spread over repeated samples.

Choosing a Sample Size for a Target

Sometimes the goal is to make the standard error no larger than a specified positive value $E$. When $0<p<1$, set the standard-error formula equal to $E$ and solve for $n$. If the requirement is “at most $E$,” the inequality must also be respected. Since sample size must be a positive whole number, round a noninteger result upward rather than to the nearest integer.

E=\sqrt{\frac{p(1-p)}{n}} \quad\Longrightarrow\quad E^2=\frac{p(1-p)}{n} \quad\Longrightarrow\quad n=\frac{p(1-p)}{E^2}, \qquad 0<p<1,\ E>0.

The restriction $0<p<1$ matters. If $p=0$ or $p=1$, then the standard error is zero for every positive $n$, so it cannot equal a positive target $E$. In those boundary cases, every positive sample size meets any positive upper-bound target. The rearranged expression must not be used there: it would give $n=0$, which is not a valid sample size.

Worked Example: Finding a Sample Size for a Target Standard Error

Suppose a hypothetical population has known proportion $p=0.40$, and the goal is a standard error no greater than $0.02$, or 2 percentage points. Assume independent observations with common success probability $0.40$. This is an interior value of $p$, so the sample-size rearrangement is valid.

Substitute $p=0.40$ and $E=0.02$:

n\geq \frac{p(1-p)}{E^2} =\frac{0.40(0.60)}{(0.02)^2} =\frac{0.24}{0.0004} =600.

Because 600 is already a positive whole number, a sample size of 600 meets the target exactly:

SE_{600}(\hat{p}) =\sqrt{\frac{0.24}{600}} =\sqrt{0.0004} =0.02.

Therefore, 600 independent observations give a standard error of 2 percentage points under the stated model. A sample size greater than 600 would give a smaller standard error, while a smaller positive sample size would give a larger one. If the target had been “strictly less than $0.02$,” the minimum whole-number sample size would instead be 601.

What the Result Does—and Does Not—Say

Increasing sample size reduces the standard error for $0<p<1$, but it does not change $p$, the population characteristic being estimated. The expected sample proportion remains $p$. More precisely, the sampling distribution is more tightly spread around the same center under the model. This is why larger samples can provide more stable estimates.

Standard error is not a maximum possible error and does not guarantee that an observed sample proportion lies within one standard error of $p$. A particular sample can still produce a proportion that differs substantially from $p$. The claim concerns the distribution of results over repeated samples, not the certainty of any single result.

The formula also depends on the sampling assumptions. If observations are not independent, or if sampling without replacement from a finite population has a substantial effect, simply applying $\sqrt{p(1-p)/n}$ may give an inaccurate description of variability. Always connect the formula to the sampling model rather than treating it as an automatic rule for every data-collection process.

Common Mistake: “Twice the Sample Gives Half the Standard Error”

A frequent error is to treat standard error as if it were directly proportional to $1/n$. The variance is proportional to $1/n$, but standard error is the square root of variance, so it is proportional to $1/\sqrt{n}$ when $p$ is fixed and interior.

For example, doubling the sample size changes the standard error by a factor of $\sqrt{1/2}$, about $0.707$, not $0.5$. Quadrupling the sample size gives a factor of $\sqrt{1/4}=0.5$. Also avoid saying that standard error strictly decreases for every possible $p$: when $p=0$ or $p=1$, it stays at zero for all positive sample sizes.

Key takeaway: Under independent sampling with fixed $p$, $SE(\hat{p})=\sqrt{p(1-p)/n}$. For $0<p<1$, multiplying sample size by $k$ divides standard error by $\sqrt{k}$; at $p=0$ or $p=1$, standard error is zero for every positive sample size.

Check Your Understanding

Use the independent-sampling model and keep the population proportion fixed unless a question says otherwise.

If the sample size increases from 50 to 200 and $0<p<1$, by what factor does the standard error change?
For $p=0.25$, compare the standard errors at $n=64$ and $n=256$. Show the formula and state the comparison in proportion or percentage-point units.
Why is it incorrect to say that doubling sample size halves the standard error?
For $p=1$, what is the standard error for any positive sample size, and why does it not decrease as $n$ grows?
For $p=0.30$ and a positive target $E=0.01$, write the sample-size expression. Why must the result be rounded upward if it is not an integer?