Choose a Paired T Procedure

Why the Pairing Changes the Procedure

Suppose a study records a quantitative outcome under two conditions. The choice of procedure depends not only on having two sets of measurements, but also on how those measurements were collected. If each measurement in one condition has a meaningful partner in the other, the analysis can focus on the difference within each pair. If observations in one condition have no such partners, the two groups are independent samples and a two-sample procedure is generally considered instead.

This distinction matters because measurements from the same person, or from deliberately matched individuals, may be related. A paired analysis preserves that relationship by turning each pair into one difference. It then asks whether the population mean of those differences is zero, or how large that mean difference may be. The number of pairs, not the number of individual measurements across both conditions, is the sample size for this analysis.

Definition: Paired quantitative data consist of two quantitative measurements linked by a meaningful pairing for each unit or matched set. Define a within-pair difference \(D_i\) by subtracting one measurement from the other in a stated order. The paired t procedure makes inference about the population mean of these within-pair differences, denoted \(\mu_D\).

The direction of subtraction is essential. If \(D_i\) is defined as “after minus before,” then a negative mean difference describes a decrease from before to after. Reversing the subtraction changes the sign of every difference and of \(\bar{x}_d\), but not the strength of the evidence or the width of a confidence interval. Always define the order before interpreting a result.

When to Choose a Paired t Procedure

Choose paired t inference when the response is quantitative, the observations can be organized into meaningful pairs, and the scientific question concerns the population mean within-pair difference. Common structures include measuring the same person twice, comparing two methods on the same specimen, or comparing individuals who were matched before their outcomes were measured. The details of the design determine whether the pairing is meaningful; having two columns in a data table does not by itself make the data paired.

The procedure is a one-sample t analysis of the differences. First calculate one difference for each pair. Then use the sample mean and sample standard deviation of those differences. Do not enter the two columns as if they were independent groups: doing so discards the pair-by-pair relationship and uses a different estimate of variability.

Paired t model and conditions: The analysis uses the \(n\) within-pair differences \(D_1,\ldots,D_n\) as the observations. They should be independent across pairs and come from a suitable random sample or randomized design for the intended inference. For t-based inference, the population distribution of differences should be approximately normal, especially when \(n\) is small; with larger samples, the procedure is generally more robust, provided the differences do not show severe skewness or influential outliers. The paired differences must be quantitative.

The independence condition concerns the pairs, not the two measurements within a pair. The two values in a pair are expected to be related; that is why they are paired. The normality condition also concerns the distribution of the differences, not the separate distributions of the two original measurements. With a small number of pairs, inspecting the differences for strong skewness or unusual values is particularly important.

The Paired t Statistic and Interval

Let \(d_i\) be the observed difference for pair \(i\). The sample mean difference is \(\bar{x}_d\), and the sample standard deviation of the differences is \(s_d\). If the null hypothesis specifies a population mean difference \(\mu_{D,0}\), the paired t statistic is:

t=\frac{\bar{x}_d-\mu_{D,0}}{s_d/\sqrt{n}}, \qquad df=n-1

For a test of no average difference, \(H_0:\mu_D=0\). The alternative hypothesis expresses the research question: it might be \(H_a:\mu_D\ne 0\), \(H_a:\mu_D>0\), or \(H_a:\mu_D<0\), according to the defined subtraction order and the question being asked. The standard error \(s_d/\sqrt{n}\) estimates the variability of the sample mean difference. Degrees of freedom are based on the number of pair differences, so \(df=n-1\).

A confidence interval for the population mean difference has the form:

\bar{x}_d \pm t^* \frac{s_d}{\sqrt{n}}

Here \(t^*\) is the critical value for the chosen confidence level and \(df=n-1\). The interval estimates \(\mu_D\), in the units of the outcome, and its interpretation must follow the stated order of subtraction.

Why the Paired Formula Is a One-Sample t Formula

Once each pair has been reduced to a difference, the data are a single sample of \(n\) values: \(d_1,\ldots,d_n\). The quantity of interest is the mean of the population from which these differences arise. Under the null hypothesis, the hypothesized mean is \(\mu_{D,0}\). The estimated standard error of \(\bar{x}_d\) is \(s_d/\sqrt{n}\), so the number of estimated standard errors between the observed mean and the null value is:

\frac{\bar{x}_d-\mu_{D,0}}{s_d/\sqrt{n}}

When the differences are independent and normally distributed, this statistic follows a t distribution with \(n-1\) degrees of freedom under the null. The degrees of freedom arise because \(s_d\) is calculated from differences around their estimated mean \(\bar{x}_d\), leaving \(n-1\) independent deviations. This explains both the paired t test and its confidence interval: each is ordinary one-sample t inference applied to the within-pair differences.

Worked Example: Before-and-After Measurements

In a hypothetical study, six participants have a quantitative outcome measured before and after a program. Define each difference as after minus before. The observed differences, in points, are \(-6,-4,-5,-3,-7,-5\). The question is whether the program is associated with a change in the population mean outcome.

Choose the procedure. Each participant supplies both measurements, so the two values are naturally paired. The outcome is quantitative, and the question concerns the mean change. The paired t procedure is appropriate if the six participant-level differences are independent across participants and the distribution of such differences is sufficiently well behaved. With only six pairs, normality and the absence of unusual differences matter considerably; the given values show no obvious extreme difference, though a real analysis should examine the differences and consider the study’s sampling or assignment design.

Calculate the summary statistics. The mean difference is \(\bar{x}_d=(-6-4-5-3-7-5)/6=-30/6=-5.00\) points. The deviations from \(-5\) are \(-1,1,0,2,-2,0\), whose squared values sum to \(10\). Thus \(s_d=\sqrt{10/(6-1)}=\sqrt{2}=1.414\) points, and \(SE=1.414/\sqrt{6}=0.5774\) points.

Test for a mean change. For \(H_0:\mu_D=0\), the statistic is \(t=(-5.00-0)/0.5774=-8.660\), with \(df=6-1=5\). The two-sided p-value is less than 0.001. Under the model assumptions, this is strong evidence against a zero population mean change. Because the differences were defined as after minus before, the sample mean indicates a decrease of 5.00 points.

Estimate the mean change. A 95% confidence interval uses \(t^*=2.571\) for \(df=5\): \(-5.00\pm2.571(0.5774)=-5.00\pm1.484\), giving \((-6.484,-3.516)\) points. This interval estimates the population mean after-minus-before difference; it does not assign a 95% probability to this particular interval containing the fixed parameter.

Worked Example: Matched Individuals

In a hypothetical educational comparison, eight students are matched in pairs using relevant background information. Each member of a pair completes a task under a different study condition, and the quantitative outcome is a score. Define each pair’s difference as score under condition A minus score under condition B. The eight differences, in points, are \(2,1,3,0,2,1,4,3\). The target is the mean difference for the population of comparable matched pairs.

Choose the procedure and check the setup. Each A observation is deliberately linked to a B observation by the matching, so differences within pairs are meaningful. The outcome is quantitative, and the question is about a mean difference. The paired t procedure is the choice, rather than a two-sample t procedure that treats the two sets of scores as unrelated. In applying it, the eight pairs should be independent of one another, and the distribution of pair differences should have no severe nonnormality or influential outlier. The small sample makes that distributional check important.

Calculate and test. The mean difference is \(\bar{x}_d=(2+1+3+0+2+1+4+3)/8=16/8=2.00\) points. Deviations from 2 are \(0,-1,1,-2,0,-1,2,1\), with squared deviations summing to \(12\). Therefore \(s_d=\sqrt{12/7}=1.309\) points and \(SE=1.309/\sqrt{8}=0.4629\) points. For \(H_0:\mu_D=0\), \(t=(2.00-0)/0.4629=4.320\), with \(df=7\). The two-sided p-value is approximately 0.0035.

Interpret. If the assumptions are reasonable, these results provide evidence that the population mean score difference is not zero. The positive observed difference means that condition A’s scores are higher on average than condition B’s for these matched comparisons. The inference concerns the mean difference, not whether every individual performs better under A.

Worked Example: Same Units, Two Measurement Methods

In a hypothetical measurement study, five samples are each measured using methods A and B. Define the difference as method A minus method B. The observed differences, in millimeters, are \(1,-1,2,0,3\). The question is whether the methods have a nonzero mean difference for the population of comparable samples.

Select the procedure. Both measurements come from the same sample, so each A value is paired with its B value. The response is quantitative, and the target is the average within-sample difference. Therefore, use a paired t procedure. The five samples should be independent; with only five differences, the distribution of differences needs to be reasonably close to normal without an influential outlier. The listed differences do not contain an extreme value relative to the others, but the small sample means that assumptions cannot be assessed confidently from a few observations alone.

Compute the test statistic. The mean difference is \(\bar{x}_d=(1-1+2+0+3)/5=5/5=1.00\) mm. Deviations from 1 are \(0,-2,1,-1,2\), and the squared deviations sum to \(10\). Thus \(s_d=\sqrt{10/4}=\sqrt{2.5}=1.581\) mm and \(SE=1.581/\sqrt{5}=0.7071\) mm. Testing \(H_0:\mu_D=0\) gives \(t=(1.00-0)/0.7071=1.414\), with \(df=4\), and a two-sided p-value of approximately 0.230.

Interpret. This sample does not provide strong evidence against a zero population mean difference at conventional significance levels. That conclusion is not proof that the methods agree exactly; it says that these data, under the assumptions, do not establish a nonzero mean difference. Also, a mean difference addresses average bias, not whether the methods give sufficiently similar readings for every sample.

A Practical Selection Check

Identify the response.
Confirm that the outcome being compared is quantitative and state its units.

Look for meaningful partners.
Determine whether each observation in one condition is linked to one observation in the other by the same unit or a defensible matching structure.

Define the difference.
Write the subtraction order explicitly, such as A minus B or after minus before. This fixes the direction of \(\mu_D\).

Check the differences and design.
Assess independence across pairs, the basis for generalizing or attributing effects, and whether the difference distribution is suitable for t inference.

Use paired t inference if the question is about the mean difference.
Analyze the \(n\) differences, not the two original columns as separate samples.

Common Mistake: Choosing by the Number of Groups

A frequent error is to reason, “There are two conditions, so I need a two-sample t procedure.” That overlooks the pairing. If the same people are measured under both conditions, or if each observation has a meaningful matched partner, the relevant observations for inference are the pair differences. The paired t procedure uses the variation among those differences.

The opposite error is also possible: labeling two data columns “paired” merely because they appear beside each other or have the same number of rows. Pairing must come from the study design and a meaningful link between observations. If people in one group are not linked to particular people in the other group, the fact that the group sample sizes happen to match does not create pairs.

Key takeaway: Choose a paired t procedure when two quantitative measurements form meaningful pairs and the target is their population mean within-pair difference. State the subtraction order, check conditions on the differences and across pairs, and base the calculation on one difference per pair.

Check Your Understanding

For each situation, identify whether paired t inference is appropriate and explain the target parameter or the reason another choice is needed.

A clinic records a quantitative symptom score for each patient before and after treatment. Which procedure would you consider, and how would you define the difference?
Two unrelated groups of patients receive different treatments, with one quantitative outcome measured once per patient. Does having two groups make this paired data?
Sixteen samples are each measured by two laboratory methods. What is the sample size for a paired t analysis, and what is the population parameter?
Why should a small-sample paired t analysis examine the distribution of the differences rather than only the distributions of the two original measurements?
If differences are defined as condition B minus condition A, what does a negative value of \(\bar{x}_d\) mean in context?

What You'll Learn