Sample proportion vs population proportion

Q: Which symbol is the sample one, p or p-hat?

p-hat (p-hat) is the sample one. The hat marks a value computed from a sample, so it is a statistic you know. Plain p is the population proportion, a parameter that is fixed and almost always unknown.

Q: Can p-hat ever equal p exactly?

Yes, and it happens fairly often with a small n, since p-hat can only take values x/n. In the district example above one of five samples landed exactly on 0.60. The catch is that you cannot tell from inside the sample, so it changes nothing about how you report the result.

Q: Why do hypotheses always use p and never p-hat?

A hypothesis is a claim about the population, and p is the only one of the two symbols that describes the population. p-hat is the number you compute from your sample to test that claim, so it belongs in the test statistic, not in H₀ or H_a.

Q: What is the difference between sigma sub p-hat and the standard error?

They are the same formula with different inputs. σ_p-hat = sqrt(p(1-p)/n) uses the true or hypothesized p, which you have in a significance test. The standard error sqrt(p-hat(1-p-hat)/n) substitutes your sample value, which is what a confidence interval has to do because p is unknown there.

By Jude Wallis · Published August 11, 2026

The population proportion p is a fixed parameter describing the whole group, and you almost never know it. The sample proportion p-hat is a statistic you compute from one sample, so it changes from sample to sample. p-hat estimates p: its values center on p, with a spread set by the sample size.

AP Statistics: Unit 3 (topics 3.1 Estimators, 3.2 Sampling Distributions for Sample Proportions). In the Fall 2026 AP Statistics course, this pair opens Unit 3, Inference for Categorical Data: Proportions. Topic 3.1 is titled Estimators and topic 3.2 is titled Sampling Distributions for Sample Proportions, so the estimator $\hat{p}$ and the distribution of its values around $p$ are the unit's first two topics, in that order.

p vs p-hat: the short answer

Both are proportions, both are written with the letter p, and only one of them is a number you will normally get to compute.

The population proportion is written $p$ , said "p". It is the fraction of the entire population with the characteristic you care about. It is a parameter: one fixed number that exists whether or not anyone measures it, and in practice you do not know its value.

The sample proportion is written $\hat{p}$ , said "p-hat". It is the fraction of your sample with that characteristic. It is a statistic: you can compute it from the data in front of you, and a different sample gives a different value.

So $p$ is what you want to know and $\hat{p}$ is what you have. The hat is the marker, and it always means sample. For the wider vocabulary of parameters and statistics, see parameter vs statistic.

What p is: fixed, and usually unknown

The population proportion is $p = \frac{\text{individuals in the population with the trait}}{\text{size of the population}}$ . Right now, for a fully defined population, that fraction has one exact value.

What makes $p$ hard is access, not existence. To know it you would have to check every member of the population, which is a census, and a census is the exception rather than the rule (see census vs sample survey). Everywhere else, $p$ stays unknown and you reason toward it.

One consequence shows up on every inference question you will write. Hypotheses are claims about $p$ , never about $\hat{p}$ . Writing $H_0: p = 0.25$ asserts something about the population that data can bear on. Writing $H_0: \hat{p} = 0.25$ asserts something about a number you already have in your hand, which is not worth testing.

What p-hat is: computable, and different every time

The sample proportion is $\hat{p} = \frac{x}{n}$ , where $x$ counts the successes in your sample and $n$ is the sample size. Both are counts of individuals, so $\hat{p}$ always lands between 0 and 1, and it can only take the values $0/n$ , $1/n$ , $2/n$ , and so on.

Because it depends on which individuals happened to land in your sample, $\hat{p}$ is a moving target. Draw again and you get a different value. That is sampling variability, not a mistake, and it is the thing the rest of this page is about. The mechanics of getting $\hat{p}$ out of a word problem, a two-way table, or a reported percentage are worked through in how to find p-hat.

The differences side by side

Feature	Population proportion $p$	Sample proportion $\hat{p}$
Kind of quantity	Parameter	Statistic
What it describes	Every member of the population	Only the individuals you sampled
Value	Fixed	Changes from sample to sample
Do you know it	Almost never	Yes, you computed it
How you would get it	A census of the population	$x/n$ from your data
Role in inference	The unknown a hypothesis is about	The evidence you compute

Every row is the same split. One number is the target, the other is the estimate aimed at it.

One population, five samples, five p-hats

A district has 2,000 seniors, and district records show that 1,200 of them hold a driver's license. So $p = \frac{1200}{2000} = 0.60$ , and for once you actually know it. Now take five separate random samples of $n = 50$ seniors.

Sample	Licensed, $x$	$\hat{p} = x/50$	$\hat{p} - p$
1	34	0.68	+0.08
2	28	0.56	-0.04
3	31	0.62	+0.02
4	27	0.54	-0.06
5	30	0.60	0.00

Five samples, five different answers, and $p = 0.60$ the entire time. The population never moved. Nothing went wrong in samples 1 through 4; different seniors landed in different samples, so the fractions differ.

Two things are worth staring at. Sample 5 hit $p$ exactly, and from inside that sample there is no way to tell: all five look identical from the inside, 50 seniors and a fraction. And these five values happen to average to exactly 0.60, which is a nicety of the five I picked, not a demonstration. Being unbiased is a claim about every possible sample, not about a handful of them.

The sampling distribution is the whole payload

Collect $\hat{p}$ from every possible sample of size $n$ and you get the sampling distribution of the sample proportion. When the observations are independent, it has

\mu_{\hat{p}} = p \qquad \text{and} \qquad \sigma_{\hat{p}} = \sqrt{\frac{p(1-p)}{n}}

Read $\mu_{\hat{p}}$ as "mu sub p-hat" and $\sigma_{\hat{p}}$ as "sigma sub p-hat". Those two lines are the entire relationship between the two proportions, and every proportion procedure you will run is built on them.

The first says $\hat{p}$ is unbiased: across all possible samples its values center on $p$ , with no systematic pull above or below. The second says $\hat{p}$ is variable, and says exactly how variable. For the district above, $p(1-p) = (0.60)(0.40) = 0.24$ , then $0.24/50 = 0.0048$ , and $\sqrt{0.0048} = 0.0693$ . All five sample proportions in the table sit within 0.08 of $p$ , which is about 1.15 of those standard deviations.

Unbiased and variable at once is the point. One $\hat{p}$ is usually not $p$ , and on the samples where it is you have no way of knowing. What you get instead is a known amount of scatter around $p$ , and that known amount is what a margin of error and a p-value are both computed from. Raising $n$ shrinks it: at $n = 200$ , $\sqrt{0.24/200} = \sqrt{0.0012} = 0.034641$ , exactly half of 0.069282, so quadrupling the sample halves the standard deviation.

Before leaning on the normal shape, check two conditions. Large Counts wants $np \geq 10$ and $n(1-p) \geq 10$ , here 30 and 20 (see why 10 successes and 10 failures). Independence is usually argued with the 10% condition, sampling no more than a tenth of the population, here 50 out of 2,000.

The mix-ups that cost points

Reporting $\hat{p}$ as if it were $p$ . "68% of district seniors are licensed" is a claim about 2,000 people built from 50 of them. Say "68% of the sampled seniors" and attach the sample size.
Writing $\hat{p}$ inside a hypothesis. Hypotheses are about $p$ , so $H_0: p = 0.6$ , never $H_0: \hat{p} = 0.6$ .
Substituting $\hat{p}$ for $p$ in $\sqrt{p(1-p)/n}$ without saying so. In a test you have a hypothesized $p$ and you use it. In a confidence interval you do not, so you put $\hat{p}$ in its place and the result is called the standard error, not the standard deviation. The distinction is spelled out in standard error vs standard deviation.
Reading a new $\hat{p}$ as evidence that $p$ changed. Samples 2 and 3 above reported 0.56 and 0.62 from a population that never moved, so check a gap against $\sqrt{2p(1-p)/n}$ before reading it as a real shift: at $n = 50$ a 0.06 gap is 0.61 standard deviations and unremarkable, at $n = 1000$ it is 2.73 and hard to explain away.
Blurring the two groups behind the two symbols. If the population and the sample are not clear in your head, start at population vs sample.

Five samples, one parameter, and the spread to expect

A district has 2,000 seniors, 1,200 of whom hold a driver's license. Five separate random samples of $n = 50$ seniors contain 34, 28, 31, 27, and 30 licensed seniors. Find the population proportion and each sample proportion, then find the mean and standard deviation of the sampling distribution of $\hat{p}$ for $n = 50$ and check the conditions.

Population proportion: $p = \frac{1200}{2000} = 0.60$ . This is a parameter, and it is the same number for all five samples because the district did not change.
Sample proportions, one per sample: $34/50 = 0.68$ , $28/50 = 0.56$ , $31/50 = 0.62$ , $27/50 = 0.54$ , $30/50 = 0.60$ . Five statistics, five values.
Center of the sampling distribution: $\mu_{\hat{p}} = p = 0.60$ . That is what unbiased means, and it does not depend on the sample size.
Spread, digit by digit. Multiply: $(0.60)(0.40) = 0.24$ . Divide by $n$ : $0.24/50 = 0.0048$ . Square root: $\sqrt{0.0048} = 0.069282$ , so $\sigma_{\hat{p}} = 0.0693$ .
Large Counts: $np = 50(0.60) = 30$ and $n(1-p) = 50(0.40) = 20$ . Both are at least 10.
Independence via the 10% condition: $10n = 10(50) = 500$ , which is below the population of 2,000, so sampling 50 seniors without replacement is fine.
Compare the misses. The largest is sample 1: $0.68 - 0.60 = 0.08$ , and $0.08/0.069282 = 1.1547$ , about 1.15 standard deviations. Under the normal approximation roughly 75% of samples land within 0.08 of $p$ , so five out of five doing so is unremarkable.

$p = 0.60$ for all five samples. The sample proportions are 0.68, 0.56, 0.62, 0.54, and 0.60. The sampling distribution of $\hat{p}$ has $\mu_{\hat{p}} = 0.60$ and $\sigma_{\hat{p}} = 0.0693$ , both conditions hold, and the largest deviation is 1.15 standard deviations.

Two samples, two intervals, one p

Now suppose $p$ is unknown. Sample 1 above found 34 licensed seniors out of 50, and sample 4 found 27 out of 50. Build a 95% confidence interval for $p$ from each, using $z^* = 1.96$ , and check whether each captures the true value 0.60.

Sample 1: $\hat{p} = 34/50 = 0.68$ . Because $p$ is unknown, put $\hat{p}$ where $p$ sits in the standard deviation formula; the result is the standard error.
Multiply: $(0.68)(0.32) = 0.2176$ . Divide by 50: $0.2176/50 = 0.004352$ . Square root: $\sqrt{0.004352} = 0.065970$ .
Margin of error: $1.96 \times 0.065970 = 0.129301$ . Interval: $0.68 \pm 0.129301$ , so $(0.5507, 0.8093)$ .
Sample 4: $\hat{p} = 27/50 = 0.54$ . Multiply: $(0.54)(0.46) = 0.2484$ . Divide by 50: $0.2484/50 = 0.004968$ . Square root: $\sqrt{0.004968} = 0.070484$ .
Margin of error: $1.96 \times 0.070484 = 0.138149$ . Interval: $0.54 \pm 0.138149$ , so $(0.4019, 0.6781)$ .
Compare. The two intervals have different centers and different widths, because $\hat{p}$ moved and the standard error moved with it. Both contain 0.60. The interval is what shifts around; $p$ sits still.

Sample 1 gives $\hat{p} = 0.68$ and the interval $(0.5507, 0.8093)$ . Sample 4 gives $\hat{p} = 0.54$ and the interval $(0.4019, 0.6781)$ . Both capture $p = 0.60$ , from two samples that disagreed by 0.14.

Frequently asked questions

Which symbol is the sample one, p or p-hat?

$\hat{p}$ (p-hat) is the sample one. The hat marks a value computed from a sample, so it is a statistic you know. Plain $p$ is the population proportion, a parameter that is fixed and almost always unknown.

Can p-hat ever equal p exactly?

Yes, and it happens fairly often with a small $n$ , since $\hat{p}$ can only take values $x/n$ . In the district example above one of five samples landed exactly on 0.60. The catch is that you cannot tell from inside the sample, so it changes nothing about how you report the result.

Does a bigger sample make p-hat closer to p?

It shrinks the typical distance, but it does not guarantee any one sample is closer. The standard deviation $\sqrt{p(1-p)/n}$ is inversely proportional to $\sqrt{n}$ , so quadrupling $n$ halves it: 0.069282 at $n = 50$ becomes 0.034641 at $n = 200$ when $p = 0.60$ . Size does not fix a biased sampling method, which is a separate problem covered in does a bigger sample fix bias.

Why do hypotheses always use p and never p-hat?

A hypothesis is a claim about the population, and $p$ is the only one of the two symbols that describes the population. $\hat{p}$ is the number you compute from your sample to test that claim, so it belongs in the test statistic, not in $H_0$ or $H_a$ .

What is the difference between sigma sub p-hat and the standard error?

They are the same formula with different inputs. $\sigma_{\hat{p}} = \sqrt{p(1-p)/n}$ uses the true or hypothesized $p$ , which you have in a significance test. The standard error $\sqrt{\hat{p}(1-\hat{p})/n}$ substitutes your sample value, which is what a confidence interval has to do because $p$ is unknown there.