Choosing Pooled, Welch, and Paired t Procedures

Start With the Design, Not the Formula

Comparing two means requires more than subtracting one sample average from another. The calculation depends on how the observations were collected and on what assumptions are reasonable about the population variances. Two independent groups call for an independent-samples procedure; measurements taken in linked pairs call for an analysis of within-pair differences. Among independent-samples procedures, the pooled method assumes equal population variances, whereas Welch’s method does not.

The central selection question is therefore: what is the relationship between the observations, and what variance assumption is defensible? Choosing a procedure because the sample variances happen to look similar, or because a test statistic is more convenient, can lead to a misleading standard error or degrees of freedom.

Definition: For two populations with means $\mu_1$ and $\mu_2$, an independent-samples design observes separate, unrelated units in the two samples. A paired design observes linked units, such as the same unit under two conditions, and analyzes the differences within pairs. The pooled two-sample t procedure assumes independent samples and equal population variances. Welch’s t procedure assumes independent samples but does not require equal population variances.

Choosing the Procedure

First decide whether observations are independent across groups or paired by design. Pairing is not established merely because two samples have the same size, or because observations can be placed in an arbitrary order. There must be a meaningful link: for example, two measurements on the same person or matched units selected as pairs.

For independent samples, the pooled procedure is appropriate when the populations are reasonably modeled as normal for small samples and their variances can be treated as equal. Welch’s procedure is appropriate when population variances may differ. Both procedures require independence between the samples and, within each sample, independent observations. With large samples, the t procedures can often be useful approximations without exact normal populations, but severe skewness or influential outliers can make small-sample results unreliable.

Procedure	Design and variance assumption	Standard error and degrees of freedom
Pooled two-sample t	Independent samples; equal population variances	Uses a common pooled variance; $n_1+n_2-2$ degrees of freedom
Welch t	Independent samples; variances may differ	Uses separate sample variances; approximate degrees of freedom
Paired t	Linked observations; analyze within-pair differences	One-sample t calculation on the differences; $n-1$ degrees of freedom

Independent Samples: Pooled and Welch Procedures

Let the two independent samples have sizes $n_1,n_2$, sample means $\bar{x}_1,\bar{x}_2$, and sample variances $s_1^2,s_2^2$. To test $H_0:\mu_1-\mu_2=\Delta_0$, the pooled procedure estimates the common population variance by weighting each sample variance by its degrees of freedom:

s_p^2=\frac{(n_1-1)s_1^2+(n_2-1)s_2^2}{n_1+n_2-2}, \qquad t=\frac{(\bar{x}_1-\bar{x}_2)-\Delta_0} {s_p\sqrt{\frac{1}{n_1}+\frac{1}{n_2}}}, \qquad \text{df}=n_1+n_2-2.

Here $s_p=\sqrt{s_p^2}$. The degrees of freedom are the total degrees of freedom in the two within-sample variance estimates. The pooled formula is not appropriate merely because equal variances are convenient; it relies on a common population variance.

Welch’s procedure estimates the variance of the difference in sample means using each sample’s own variance. Its degrees of freedom are approximated by the Welch–Satterthwaite formula:

t_W= \frac{(\bar{x}_1-\bar{x}_2)-\Delta_0} {\sqrt{\frac{s_1^2}{n_1}+\frac{s_2^2}{n_2}}}, \qquad \text{df}\ \approx\ \frac{\left(\frac{s_1^2}{n_1}+\frac{s_2^2}{n_2}\right)^2} {\frac{(s_1^2/n_1)^2}{n_1-1}+ \frac{(s_2^2/n_2)^2}{n_2-1}}.

The resulting degrees of freedom need not be an integer. Use them as an approximation when obtaining a critical value or p-value. Welch’s standard error is not a pooled estimate: each variance contributes according to its own sample size.

Worked Example: Independent Samples With a Common Variance

A hypothetical analyst compares a measurement from two independent groups. The observations are $10,11,12,13,14$ in Group 1 and $7.5,8.5,9.5,10.5,11.5$ in Group 2. Assume independent observations from normal populations with a common variance. At the 5% significance level, test $H_0:\mu_1-\mu_2=0$ against a two-sided alternative.

Method and assumptions. The groups contain different, independent units, and the stated model supplies the equal-variance assumption. Use the pooled two-sample t procedure.

Sample summaries. The means are $\bar{x}_1=60/5=12$ and $\bar{x}_2=47.5/5=9.5$. In each group, deviations from its mean are $-2,-1,0,1,2$, so each sum of squared deviations is $4+1+0+1+4=10$. Thus $s_1^2=s_2^2=10/(5-1)=2.5$.

Pooled variance and test statistic.

s_p^2=\frac{4(2.5)+4(2.5)}{8} =\frac{10+10}{8}=2.5, \qquad SE=\sqrt{2.5\left(\frac15+\frac15\right)} =\sqrt{1}=1.

The observed difference is $12-9.5=2.5$, so $t=(2.5-0)/1=2.5$, with $5+5-2=8$ degrees of freedom. The two-sided 5% critical value is approximately $2.306$. Since $|2.5|>2.306$, reject $H_0$ at the 5% level. The data provide evidence of a difference in population means under the stated model. Check: the standard error is in the same measurement units as the means, and the test statistic is dimensionless.

Common mistake avoided: pooling is justified here by the stated common-variance model, not simply by noticing that the two sample variances match.

Worked Example: Independent Samples With Unequal Variances

Two independent samples measure hypothetical processing times. Group 1 has $n_1=6$, $\bar{x}_1=30$, and $s_1^2=16$; Group 2 has $n_2=8$, $\bar{x}_2=25$, and $s_2^2=4$. Treat the samples as independent and do not assume equal population variances. Test $H_0:\mu_1-\mu_2=0$ against a two-sided alternative at the 5% level.

Method and assumptions. The samples are independent, but their population variances are not assumed equal. Use Welch’s procedure.

Standard error and statistic.

SE_W=\sqrt{\frac{16}{6}+\frac{4}{8}} =\sqrt{2.6667+0.5} =\sqrt{3.1667}\approx1.7795, \qquad t_W=\frac{30-25}{1.7795}\approx2.810.

Degrees of freedom. The two terms in the numerator of the degrees-of-freedom expression are $16/6=2.6667$ and $4/8=0.5$. Therefore,

\text{df}\approx \frac{(2.6667+0.5)^2}{\frac{(2.6667)^2}{5}+\frac{(0.5)^2}{7}} = \frac{10.0278}{1.4222+0.0357} \approx6.88.

For a two-sided test with approximately $6.88$ degrees of freedom, the 5% critical value is about $2.37$. Since $|2.810|>2.37$, reject $H_0$ at the 5% level. Check: the observed mean difference is positive, and the positive statistic is consistent with Group 1’s larger sample mean.

Common mistake avoided: do not replace Welch’s variance estimate with an unweighted average of the sample variances. That would give $(16+4)/2=10$, which is not the pooled variance formula and does not implement Welch’s method.

Paired Observations: Analyze the Differences

For $n$ linked pairs, define $D_i$ as the first measurement minus the second measurement for pair $i$. The question becomes a one-sample test of the mean difference $\mu_D$. The pairs must be independent of one another; within a pair, the two measurements are generally dependent, which is precisely why their difference is informative.

Paired t procedure: Let $\bar{D}$ and $s_D$ be the sample mean and sample standard deviation of the $n$ pair differences. To test $H_0:\mu_D=\Delta_0$, calculate $$ t=\frac{\bar{D}-\Delta_0}{s_D/\sqrt{n}}, \qquad \text{df}=n-1. $$ For a small sample, the differences should be reasonably modeled as coming from a normal population. The relevant distributional condition concerns the differences, not a requirement that each of the two measurement columns be analyzed separately.

Worked Example: Before-and-After Measurements on the Same Units

A hypothetical study measures a response before and after a procedure on each of six participants. Define $D=\text{before}-\text{after}$, so positive values represent a decrease after the procedure. The observed differences are $2,1,3,0,2,4$. Assuming independent pairs and approximately normal differences, test $H_0:\mu_D=0$ against $H_a:\mu_D>0$ at the 5% level.

Method. The measurements are linked within each participant, so analyze the six differences with a one-sample t procedure.

Mean and standard deviation of the differences. Their sum is $2+1+3+0+2+4=12$, giving $\bar{D}=12/6=2$. Deviations from 2 are $0,-1,1,-2,0,2$; their squared sum is $0+1+1+4+0+4=10$. Hence $s_D^2=10/(6-1)=2$, and $s_D=\sqrt{2}$.

Test statistic and decision.

SE=\frac{\sqrt{2}}{\sqrt{6}}=\sqrt{\frac13}\approx0.5774, \qquad t=\frac{2-0}{0.5774}\approx3.464, \qquad \text{df}=6-1=5.

The one-sided 5% critical value for 5 degrees of freedom is approximately $2.015$. Since $3.464>2.015$, reject $H_0$. The data support a positive mean before-minus-after difference under the stated assumptions. Check: positive $D$ was defined as a decrease, and the observed mean difference is positive.

Common mistake avoided: treating the 12 measurements as independent discards the pairing. The inferential sample consists of six independent differences, not 12 independent observations.

Why the Pooled Statistic Has a t Distribution

The pooled procedure has an exact Student t distribution under its normal, independent, common-variance model. This follows by combining the normal distribution of the difference in sample means with the independent estimate of the common variance.

Standardize the difference in means.
Under $H_0:\mu_1-\mu_2=\Delta_0$, normal sampling and independence give $$ Z=\frac{(\bar{X}_1-\bar{X}_2)-\Delta_0}{\sigma\sqrt{1/n_1+1/n_2}}\sim N(0,1). $$

Estimate the common variance.
For independent normal samples with common variance $\sigma^2$, the two within-sample sums of squares, divided by $\sigma^2$, are independent chi-square variables with $n_1-1$ and $n_2-1$ degrees of freedom. Their sum gives $$ \frac{(n_1+n_2-2)s_p^2}{\sigma^2}\sim\chi^2_{n_1+n_2-2}. $$

Form the ratio.
The pooled variance estimate is independent of the difference in sample means under this normal model. A standard normal variable divided by the square root of an independent chi-square variable divided by its degrees of freedom has a Student t distribution. Substituting $s_p$ for $\sigma$ therefore gives the pooled statistic with $n_1+n_2-2$ degrees of freedom.

Welch’s statistic uses separate variance estimates. Its displayed degrees-of-freedom formula approximates the sampling distribution rather than giving the same exact t result as the pooled normal model. The paired statistic, by contrast, has the one-sample t result when the pair differences are independent and normal.

Selection Traps and Practical Importance

Selecting the procedure correctly matters because the standard error and reference distribution determine how strong the evidence appears. A useful order of work is to identify the design, state the variance assumptions, select the procedure, and only then calculate the statistic.

Do not infer pairing from equal sample sizes. Pairing requires a real unit-to-unit link. If two groups contain unrelated units, an arbitrary pairing changes the analysis without changing the data collection.
Do not choose pooled solely because sample variances look close. The pooled method assumes equal population variances. A failure to observe a large difference in two sample variances does not establish that assumption.
Do not pool independent samples when the variance assumption is unsupported. Welch’s procedure keeps the variance contributions separate and is a standard choice when equality is not justified.
Do not treat paired measurements as independent. Doing so ignores within-pair dependence and can produce an incorrect standard error.
Keep the alternative and sign convention consistent. For a one-sided paired test, define the difference so the direction named in the alternative corresponds to positive or negative values as intended.

These distinctions are useful whenever an analysis compares two conditions, groups, or methods. A defensible test is not selected by the size of the observed difference; it follows from the sampling design and the model assumptions. The procedures and calculations here are aligned with the SOA VEE Mathematical Statistics syllabus.

Check Your Understanding

For each problem, identify the appropriate procedure and complete the requested calculation.

Two independent groups have $n_1=n_2=8$, sample means differing by 3, and sample variances $s_1^2=s_2^2=4$. Under an equal-variance model, find the pooled test statistic for $H_0:\mu_1-\mu_2=0$ and its degrees of freedom. (Answer: pooled t; $t=3$, df $=14$.)
Independent samples have $n_1=5,n_2=10$, $\bar{x}_1-\bar{x}_2=4$, $s_1^2=9$, and $s_2^2=1$. Using Welch’s procedure, find the standard error, test statistic, and approximate degrees of freedom. (Answer: $SE=\sqrt{1.9}\approx1.3784$, $t\approx2.902$, df $\approx4.45$.)
Seven linked pairs have differences $2,0,1,-1,3,1,1$, with $D$ defined as measurement A minus measurement B. For $H_0:\mu_D=0$, find the paired t statistic and degrees of freedom. (Answer: $\bar D=1$, $s_D^2=10/6$, $t\approx2.049$, df $=6$.)
A researcher measures the same 12 devices under two settings, and the within-device differences are reasonably modeled as normal. The researcher instead proposes an independent-samples pooled test because there are 12 readings under each setting. State the appropriate procedure and the correct number of independent analysis units. (Answer: paired t procedure on 12 within-device differences; 12 independent pairs.)

Procedure	Design and variance assumption	Standard error and degrees of freedom
Pooled two-sample t	Independent samples; equal population variances	Uses a common pooled variance; \(n_1+n_2-2\) degrees of freedom
Welch t	Independent samples; variances may differ	Uses separate sample variances; approximate degrees of freedom
Paired t	Linked observations; analyze within-pair differences	One-sample t calculation on the differences; \(n-1\) degrees of freedom

What You'll Learn