Pooled vs unpooled standard error

By Jude Wallis · Published

A pooled standard error combines both samples into one shared estimate before measuring spread. An unpooled standard error keeps each sample's own estimate. In AP Statistics you pool in exactly one place, the two-proportion z-test, and use the unpooled form for every interval and for two means.

This page spans Units 3 and 4 of the Fall 2026 AP Statistics course. The two-proportion interval in topic 3.10 uses the unpooled standard error and the two-proportion test in topic 3.13 uses the pooled one. The two-sample t procedures in topics 4.7 and 4.10 use the unpooled standard error, the only one the formula sheet lists for a difference of means.

Pooled vs unpooled standard error: the short answer

Any procedure that compares two groups needs a standard error for the difference between two statistics. There are two ways to build it.

The unpooled standard error estimates each group's variability from that group alone and adds the two pieces. It assumes nothing about how the groups relate.

The pooled standard error first merges the two samples into a single estimate of something the groups are assumed to share, then uses that one estimate for both groups. It is correct only when the sharing assumption is justified.

In the AP course the assumption is justified in one procedure. The null hypothesis of a two-proportion z-test says p1=p2p_1 = p_2, and a test is computed as if the null were true, so the test pools. Confidence intervals make no such assumption and stay unpooled. Two-sample procedures for means stay unpooled as well, in both the interval and the test.

The pooled standard error

For two proportions, pooling means computing one combined sample proportion,

p^c=x1+x2n1+n2\hat{p}_c = \frac{x_1 + x_2}{n_1 + n_2}

where x1x_1 and x2x_2 are the success counts, and then using it in place of both p^1\hat{p}_1 and p^2\hat{p}_2 ("p-hat"):

SEpooled=p^c(1−p^c)(1n1+1n2)SE_{\text{pooled}} = \sqrt{\hat{p}_c(1 - \hat{p}_c)\left(\frac{1}{n_1} + \frac{1}{n_2}\right)}

This is the denominator of the two-proportion z statistic. The reasoning comes from the null hypothesis. If p1p_1 and p2p_2 are equal, both samples are estimating the same number, and the estimate based on all n1+n2n_1 + n_2 observations is better than either separate one. The large counts condition for the test uses p^c\hat{p}_c for the same reason.

For two means there is also a pooled version, which merges the two sample standard deviations into one pooled standard deviation sps_p. It assumes the two populations have equal standard deviations. The null hypothesis μ1=μ2\mu_1 = \mu_2 says nothing about the standard deviations, so that assumption has no support from the test itself, and the course does not use it.

The unpooled standard error

The unpooled form gives each group its own term. For two proportions,

SEunpooled=p^1(1−p^1)n1+p^2(1−p^2)n2SE_{\text{unpooled}} = \sqrt{\frac{\hat{p}_1(1 - \hat{p}_1)}{n_1} + \frac{\hat{p}_2(1 - \hat{p}_2)}{n_2}}

and for two means, with sample standard deviations s1s_1 and s2s_2,

SEunpooled=s12n1+s22n2SE_{\text{unpooled}} = \sqrt{\frac{s_1^2}{n_1} + \frac{s_2^2}{n_2}}

The proportion version belongs to the confidence interval for p1−p2p_1 - p_2. An interval estimates how far apart the two proportions are, so it cannot begin by assuming they are equal. Its large counts check uses each group's own successes and failures.

The means version is used for both the two-sample t interval and the two-sample t-test. The AP formula sheet lists only this standard error for a difference of means, which is why a calculator's Pooled prompt is answered No, as pooled or not on a two-sample t-test explains.

The differences side by side

FeaturePooled standard errorUnpooled standard error
How it is builtOne combined estimate used for both groupsA separate estimate from each group
AssumesThe groups share the pooled quantityNothing about how the groups relate
Two proportions, testYesNo
Two proportions, intervalNoYes
Two means, test and intervalNot used in APYes
Proportion formulap^c(1−p^c)(1n1+1n2)\sqrt{\hat{p}_c(1-\hat{p}_c)\left(\frac{1}{n_1}+\frac{1}{n_2}\right)}p^1(1−p^1)n1+p^2(1−p^2)n2\sqrt{\frac{\hat{p}_1(1-\hat{p}_1)}{n_1}+\frac{\hat{p}_2(1-\hat{p}_2)}{n_2}}
Large counts check usesp^c\hat{p}_cp^1\hat{p}_1 and p^2\hat{p}_2

The two versions usually give close numbers, and neither is always the larger one. With 45 successes out of 150 and 30 out of 150, the pooled standard error is 0.0500 and the unpooled is 0.0497. With 60 out of 200 and 45 out of 250, the pooled is 0.0401 and the unpooled is 0.0405. Close is not the same as interchangeable: a borderline test statistic can cross a cutoff on the difference.

When to use which

Decide in two steps.

  1. Means or proportions? For two means, use the unpooled standard error every time.
  2. For proportions, test or interval? A significance test of H0:p1=p2H_0: p_1 = p_2 uses the pooled standard error. A confidence interval for p1−p2p_1 - p_2 uses the unpooled one.

The rule behind step 2 is general. A test statistic is computed under the null hypothesis, so wherever the null supplies a value, the standard error uses it. In a one-proportion test the null supplies p0p_0. In a two-proportion test the null supplies only the statement that the two proportions match, and p^c\hat{p}_c is the estimate of that common value. An interval has no null hypothesis, so it uses what the samples show.

Because the test and the interval for two proportions use different standard errors, they can disagree in a borderline case: a 95% interval can just contain 0 while the two-sided test at α=0.05\alpha = 0.05 just rejects, or the reverse. That situation is worked through in two-proportion confidence interval contains zero.

The classic mix-up and how to avoid it

The standard error is pooled in the wrong place in three common ways.

  • Pooling in the interval. Using p^c\hat{p}_c inside a confidence interval for p1−p2p_1 - p_2 builds in the assumption that the difference is 0 while estimating that difference.
  • Not pooling in the test. Using the unpooled formula in a two-proportion z-test gives a different test statistic from the one the procedure defines. In the first worked example the correct zz is 2.991 and the unpooled version gives 2.963.
  • Pooling means because proportions are pooled. The habit does not transfer. For means, pooling requires equal population standard deviations, and when the smaller sample has the larger spread the pooled standard error is too small, which overstates the evidence. The second worked example shows the conclusion changing.

A quick check before computing: write the null hypothesis. If you are running a test and the null says two proportions are equal, pool. In every other two-sample situation in the course, keep the groups separate.

Two proportions: pooled for the test, unpooled for the interval

Independent random samples are taken from two large schools. At School 1, 60 of 200 students walk to school. At School 2, 45 of 250 do. Test H0:p1=p2H_0: p_1 = p_2 against Ha:p1≠p2H_a: p_1 \neq p_2 at α=0.05\alpha = 0.05, then build a 95% confidence interval for p1−p2p_1 - p_2. Use the correct standard error for each.

  1. Sample proportions: p^1=60200=0.30\hat{p}_1 = \frac{60}{200} = 0.30 and p^2=45250=0.18\hat{p}_2 = \frac{45}{250} = 0.18, so p^1−p^2=0.12\hat{p}_1 - \hat{p}_2 = 0.12.

  2. For the test, pool: p^c=60+45200+250=105450=0.2333\hat{p}_c = \frac{60 + 45}{200 + 250} = \frac{105}{450} = 0.2333.

  3. Check large counts with p^c\hat{p}_c: 200(0.2333)=46.7200(0.2333) = 46.7, 200(0.7667)=153.3200(0.7667) = 153.3, 250(0.2333)=58.3250(0.2333) = 58.3, 250(0.7667)=191.7250(0.7667) = 191.7, all at least 10.

  4. Pooled standard error: 0.2333×0.7667×(1200+1250)=0.17889×0.009=0.0016100=0.04012\sqrt{0.2333 \times 0.7667 \times \left(\frac{1}{200} + \frac{1}{250}\right)} = \sqrt{0.17889 \times 0.009} = \sqrt{0.0016100} = 0.04012.

  5. Test statistic: z=0.120.04012=2.991z = \frac{0.12}{0.04012} = 2.991. The two-sided p-value is 0.0028, so reject H0H_0.

  6. For the interval, do not pool. Check large counts with each sample: 60 and 140 at School 1, 45 and 205 at School 2, all at least 10.

  7. Unpooled standard error: 0.30×0.70200+0.18×0.82250=0.0010500+0.0005904=0.0016404=0.04050\sqrt{\frac{0.30 \times 0.70}{200} + \frac{0.18 \times 0.82}{250}} = \sqrt{0.0010500 + 0.0005904} = \sqrt{0.0016404} = 0.04050.

  8. Margin of error: 1.960×0.04050=0.07941.960 \times 0.04050 = 0.0794. Interval: 0.12±0.07940.12 \pm 0.0794, from 0.0406 to 0.1994.

The test uses the pooled standard error 0.04012 and gives z=2.991z = 2.991 with p-value 0.0028, so there is convincing evidence that the proportions who walk differ. The interval uses the unpooled standard error 0.04050 and runs from 0.0406 to 0.1994. One data set produced two standard errors because the test assumes p1=p2p_1 = p_2 and the interval does not.

Two means: why the course keeps them unpooled

Group 1 has n1=10n_1 = 10, xˉ1=57\bar{x}_1 = 57, s1=12s_1 = 12. Group 2 has n2=40n_2 = 40, xˉ2=50\bar{x}_2 = 50, s2=8s_2 = 8. Both populations are stated to be approximately normal, and the samples are independent and random. Compute the unpooled and pooled standard errors for xˉ1−xˉ2\bar{x}_1 - \bar{x}_2 and the two-sided test of H0:μ1=μ2H_0: \mu_1 = \mu_2 at α=0.05\alpha = 0.05 under each.

  1. Unpooled standard error: 12210+8240=14.4+1.6=16=4.000\sqrt{\frac{12^2}{10} + \frac{8^2}{40}} = \sqrt{14.4 + 1.6} = \sqrt{16} = 4.000.

  2. Unpooled test statistic: t=57−504.000=1.750t = \frac{57 - 50}{4.000} = 1.750. Technology gives 11.08 degrees of freedom and a two-sided p-value of 0.1077. Fail to reject H0H_0.

  3. Pooled variance: sp2=(10−1)(122)+(40−1)(82)10+40−2=1296+249648=379248=79s_p^2 = \frac{(10 - 1)(12^2) + (40 - 1)(8^2)}{10 + 40 - 2} = \frac{1296 + 2496}{48} = \frac{3792}{48} = 79, so sp=79=8.888s_p = \sqrt{79} = 8.888.

  4. Pooled standard error: 79×(110+140)=79×0.125=9.875=3.142\sqrt{79 \times \left(\frac{1}{10} + \frac{1}{40}\right)} = \sqrt{79 \times 0.125} = \sqrt{9.875} = 3.142.

  5. Pooled test statistic: t=73.142=2.228t = \frac{7}{3.142} = 2.228 with 10+40−2=4810 + 40 - 2 = 48 degrees of freedom and a two-sided p-value of 0.0306. This version would reject H0H_0.

  6. Compare. The pooled standard deviation of 8.888 is pulled toward the large group's 8, so it understates the variability of the small group, whose sample standard deviation is 12. The pooled standard error comes out smaller, 3.142 against 4.000.

The unpooled standard error is 4.000, giving t=1.750t = 1.750 and a p-value of 0.1077: fail to reject H0H_0. The pooled standard error is 3.142, giving t=2.228t = 2.228 and a p-value of 0.0306, which would reject. The AP procedure is the unpooled one, because nothing in the data or the null hypothesis supports assuming equal population standard deviations.

Frequently asked questions

Why is the standard error pooled in a two-proportion z-test?

The null hypothesis says p1=p2p_1 = p_2, and a test statistic is computed assuming the null is true. If the two proportions are equal, both samples estimate the same value, so the combined proportion p^c=x1+x2n1+n2\hat{p}_c = \frac{x_1 + x_2}{n_1 + n_2} is the best estimate of it and is used for both groups.

Do I pool for a confidence interval for two proportions?

No. A confidence interval estimates the difference p1−p2p_1 - p_2 without assuming it is 0, so each group keeps its own sample proportion: p^1(1−p^1)n1+p^2(1−p^2)n2\sqrt{\frac{\hat{p}_1(1-\hat{p}_1)}{n_1} + \frac{\hat{p}_2(1-\hat{p}_2)}{n_2}}.

Is the pooled standard error always smaller than the unpooled one?

No. For two proportions with equal sample sizes the pooled value is at least as large as the unpooled one, and with unequal sizes it can go either way. For two means with equal sample sizes the two standard errors are identical, and with unequal sizes the pooled one is smaller when the smaller sample has the larger standard deviation.