Pooled Proportion and Why It Is Used

Why a Two-Proportion Test Pools the Data

In Defining Both Parameters Clearly, you identified \(p_1\) and \(p_2\) as the true proportions for two groups and kept their order clear. In Hypotheses for Comparing Two Population Proportions, you saw that the null hypothesis for no difference can be written \(H_0:p_1=p_2\). That equality has an important consequence for the test calculation: if the null is true, both groups share one common population proportion.

The sample proportions, \(\hat{p}_1\) and \(\hat{p}_2\), will not usually be exactly equal. They come from different samples and can vary by chance. But when calculating a test under \(H_0\), we proceed as if both groups have the same true proportion. To estimate that shared proportion from the data, we combine the successes and observations from both samples. This estimate is called the pooled proportion.

Let \(x_1\) and \(x_2\) be the numbers of successes in the two samples, and let \(n_1\) and \(n_2\) be their sample sizes. Since \(\hat{p}_1=x_1/n_1\) and \(\hat{p}_2=x_2/n_2\), the pooled proportion is:

The numerator counts every success across both groups. The denominator counts every observation across both groups. The pooled proportion is therefore a proportion for the combined data, used as an estimate of the common proportion under the null. It does not say the observed sample proportions are equal, and it does not erase the distinction between the two groups.

Why the Sample Sizes Matter

A pooled proportion is a weighted combination of the two sample proportions. A larger sample contributes more observations, so it has more influence on the combined estimate. If sample sizes are unequal, simply taking the average of \(\hat{p}_1\) and \(\hat{p}_2\) gives the two groups equal weight even though one group contributes more data. That is generally not the pooled proportion.

This formula makes the weighting visible: multiply each sample proportion by its sample size to recover its success count, add those counts, and divide by the total sample size. If the two sample sizes happen to be equal, the pooled proportion is the ordinary average of the sample proportions. Otherwise, do not assume it is.

Pooling Belongs to the Null Model

The pooled proportion is used in a two-proportion hypothesis test because the test calculations are made under the assumption that \(H_0\) is true. If \(p_1=p_2\), both groups have the same underlying success probability, so their observed successes can be combined to estimate it. This shared estimate is then used in the test’s null model.

The pooled proportion is not the estimate of the difference \(p_1-p_2\). For the observed samples, the point estimate of that difference remains \(\hat{p}_1-\hat{p}_2\), in the group order defined for the problem. The pooled proportion answers a different question: what common success proportion is estimated when we impose the equality stated by \(H_0\)?

Pooling is specific to the null hypothesis of equal proportions. A two-proportion confidence interval does not assume \(p_1=p_2\); it estimates the difference that may exist. As explained in Common Errors With Two-Proportion Intervals, the interval uses separate sample proportions in its standard error rather than pooling. Do not carry the pooled calculation over to an interval just because both methods compare two proportions.

What Pooling Does—and Does Not—Mean

Combining counts is a calculation for a particular model, not a claim that the two groups are interchangeable in every respect. The groups remain separately identified, their sample sizes remain \(n_1\) and \(n_2\), and the difference in their sample proportions remains available to compare with the null model. Pooling uses the equality in \(H_0\) to estimate one common success probability.

A pooled proportion can also be used to understand the Large Counts condition for a two-proportion \(z\)-test. Under the equal-proportions null model, the expected numbers of successes in the two groups are \(n_1\hat{p}_{\text{pool}}\) and \(n_2\hat{p}_{\text{pool}}\); the expected numbers of failures are \(n_1(1-\hat{p}_{\text{pool}})\) and \(n_2(1-\hat{p}_{\text{pool}})\). The test’s Large Counts check considers whether these expected counts are at least 10 in each group. This is distinct from checking observed successes and failures for a two-proportion confidence interval.

For instance, in the transit example, the pooled estimate is about \(0.4643\). Under the null model, the expected success and failure counts are about \(80(0.4643)=37.14\) and \(80(1-0.4643)=42.86\) for Group 1, and \(60(0.4643)=27.86\) and \(60(1-0.4643)=32.14\) for Group 2. These are expected counts under the null, not the actual observed counts. Each exceeds 10. The randomization or random-sampling and independence conditions still need to be considered separately, as covered in earlier tutorials.

Common Mistakes and AP Exam Tips

• Averaging the two proportions without weighting: Unless the sample sizes are equal, \((\hat{p}_1+\hat{p}_2)/2\) generally is not pooled. Add success counts and divide by the combined sample size.
• Adding sample sizes but not successes: The numerator must include successes from both groups; the denominator must include all observations from both groups.
• Using the pooled value as the observed difference: The pooled proportion estimates a common success probability under \(H_0\). The observed difference is \(\hat{p}_1-\hat{p}_2\).
• Assuming the observed proportions must match: \(H_0:p_1=p_2\) concerns the population proportions. Sample proportions can differ because of sampling variability.
• Pooling for a confidence interval: The interval does not impose the equality assumption in the null hypothesis. Use the separate sample proportions for a two-proportion interval.
• Rounding too early: Keep the fraction or several decimal places during calculations, then round the reported pooled estimate. This helps avoid small discrepancies later.

Key Takeaway

The equality in \(H_0:p_1=p_2\) gives a two-proportion test one common success proportion to estimate. Pool the two groups’ success counts and divide by their combined sample size; this automatically weights each group according to how many observations it contributes.