The Pooled Two-Sample T Procedure

Why Pool Information About Variability?

To compare two population means, we need to account for the variability in both samples. The pooled two-sample t procedure does this by estimating one common population variance from the two groups’ within-group variation. That estimate is then used to calculate the standard error of the difference between the sample means.

Pooling can be useful when it is reasonable to model the two populations as having the same variance. It is not a way to make unequal variances equal, nor does it mean that the two sample variances must match exactly. The procedure relies on a substantive assumption about the population variances, alongside the independence and distributional conditions for two-sample t inference.

Let the sample sizes be \(n_1\) and \(n_2\), the sample means be \(\bar{x}_1\) and \(\bar{x}_2\), and the sample standard deviations be \(s_1\) and \(s_2\). Under the model, each sample variance estimates the same population variance, but each estimate is based on a different number of degrees of freedom.

Deriving the Pooled Variance

Within group \(i\), the sample variance \(s_i^2\) is the sum of squared deviations from that group’s sample mean divided by \(n_i-1\). Therefore, the group’s within-sample sum of squares is \((n_i-1)s_i^2\). The pooled estimate combines those two sums of squares and divides by their combined degrees of freedom.

The weights \(n_1-1\) and \(n_2-1\) are the degrees of freedom contributed by the two groups. Thus, the pooled variance is generally not the unweighted average of \(s_1^2\) and \(s_2^2\). When the groups have equal degrees of freedom, however, the two weighted contributions are equal and the pooled variance is exactly their unweighted average.

The pooled standard deviation is \(s_p=\sqrt{s_p^2}\). It estimates the common population standard deviation \(\sigma\). It is a useful intermediate calculation, but inference about the difference in means uses the pooled standard error, which also depends on both sample sizes.

Standard Error, Test Statistic, and Interval

Under the common-variance model, the estimated standard error of \(\bar{x}_1-\bar{x}_2\) is the pooled standard deviation multiplied by the square root of the sum of the two sample-size reciprocals. The denominator \(n_1+n_2-2\) in the pooled variance is also the degrees of freedom for the pooled t procedure.

For testing \(H_0:\mu_1-\mu_2=\Delta_0\), where \(\Delta_0\) is the hypothesized difference, the pooled t statistic is the observed difference minus the hypothesized difference, divided by its estimated standard error. If the null hypothesis is true and the model assumptions hold, this statistic has a t distribution with \(n_1+n_2-2\) degrees of freedom.

For a two-sided test, evidence against \(H_0\) is measured by the probability, assuming \(H_0\), of a t statistic at least as far from zero as the observed statistic. For a confidence interval with confidence level \(C\), use the appropriate t critical value with \(n_1+n_2-2\) degrees of freedom.

Why the Degrees-of-Freedom Weighting Works

The weighting follows directly from the within-group sums of squares. For each group, \((n_i-1)s_i^2\) measures the total squared deviation around that group’s own mean. Adding these quantities combines the observed within-group variation without mixing the two group means into one center. Under a shared variance model, the expected within-group sum of squares in group \(i\) is \((n_i-1)\sigma^2\). Adding across groups gives an expected total of \((n_1+n_2-2)\sigma^2\). Dividing by the combined degrees of freedom therefore gives an estimate centered on the common variance.

This justification also explains why the pooled degrees of freedom add: each group contributes its own residual degrees of freedom after estimating its sample mean. The total is \((n_1-1)+(n_2-1)\), not \(n_1+n_2\).

Common Mistakes and When Pooling Is Appropriate

• Using an unweighted average automatically. The reasoning “there are two sample variances, so average them” ignores their degrees of freedom. Use the weighted formula; it coincides with the unweighted average when the groups have equal degrees of freedom.
• Confusing variance with standard deviation. Pool \(s_1^2\) and \(s_2^2\) using the weighted formula, then take the square root to obtain \(s_p\). Do not directly average \(s_1\) and \(s_2\).
• Using the wrong degrees of freedom. The pooled procedure uses \(n_1+n_2-2\), reflecting the two groups’ within-sample degrees of freedom.
• Treating the common-variance assumption as a sample-data fact. Similar observed sample standard deviations do not prove equal population variances, and unequal observed values do not by themselves prove they differ. The assumption must be judged using the design and subject-matter context as well as the data.
• Pooling when the equal-variance model is implausible. The pooled standard error is built on a common population variance. If that model is not defensible, the pooled procedure may give unreliable inference; the earlier material on Welch’s procedure provides an alternative that does not require equal variances.

For instance, the incorrect reasoning “the two sample variances are 4 and 9, so the pooled variance must be 6.5” overlooks how much information each estimate contributes. With 4 and 6 degrees of freedom, respectively, the correct pooled estimate is \((4(4)+6(9))/10=7\), not 6.5. If both groups instead had equal degrees of freedom, their unweighted average would agree with the pooled estimate.

Pooling matters because it can combine information efficiently when a common variance is a reasonable model. But a calculation cannot validate its own assumptions: a correctly computed interval or p-value can still be misleading if the population variances are substantially different or the observations are not independent. State the assumptions, assess whether they fit the setting, and interpret the result in the units of the outcome.