The Paired T Procedure as a One-Sample Analysis

Why a Paired T Procedure Uses One Sample

When observations come in pairs, the question is usually about how the two measurements compare within each pair. The two measurements in a pair may be related: they might come from the same person at two times, or from two participants deliberately matched on relevant characteristics. Treating all measurements as unrelated would ignore that structure.

The paired t procedure handles the structure by reducing each pair to one numerical difference. The resulting differences form a single sample, so the calculation is a one-sample t analysis—not a two-sample analysis of the original measurements. The number of observations in that analysis is the number of complete pairs.

Definition: Choose and state a direction for subtraction, then let \(D_i\) denote the difference for pair \(i\), for \(i=1,\ldots,n\). The population mean difference is \(\mu_d\), the mean of the population distribution of these differences. The observed differences are \(d_1,\ldots,d_n\); their sample mean is \(\bar{d}\), and their sample standard deviation is \(s_d\). Here \(n\) is the number of complete pairs.

The subtraction direction determines the sign of every difference. For example, defining \(d_i=\text{after}_i-\text{before}_i\) makes a decrease negative, while reversing the subtraction makes the same decrease positive. Either convention is valid, but the definition must remain consistent in the data, calculation, and interpretation. Changing the direction changes the signs of \(\bar{d}\) and the t statistic, but not \(s_d\) or the standard error.

The One-Sample T Calculation on Differences

The paired t procedure estimates the population mean difference with \(\bar{d}\). The sample standard deviation \(s_d\) measures how much the observed differences vary across pairs. Because the mean is computed from \(n\) differences, its estimated standard error is \(s_d/\sqrt{n}\). A t statistic compares the observed mean difference with a specified reference value, often zero.

Key result: For a reference value \(\mu_{d,0}\), the paired t statistic is

t=\frac{\bar{d}-\mu_{d,0}}{s_d/\sqrt{n}}, \qquad df=n-1.

The numerator and standard error have the same units, so \(t\) is unitless. The degrees of freedom are based on the number of pair differences, not the number of individual measurements.

The one-sample t framework requires assumptions about the differences. For the exact small-sample model, the pair differences are an independent sample from a normally distributed population with mean \(\mu_d\) and standard deviation \(\sigma_d\). Independence here concerns differences from different pairs; the two measurements within a pair need not be independent. In applications, whether these assumptions are reasonable is assessed using the design and the distribution of the differences. The relevant distribution is the distribution of differences, not either measurement column considered by itself.

Model and distribution: If \(D_1,\ldots,D_n\) are independent and normally distributed with mean \(\mu_d\) and variance \(\sigma_d^2\), then under the null hypothesis \(H_0:\mu_d=\mu_{d,0}\), the statistic \( (\bar{d}-\mu_{d,0})/(s_d/\sqrt{n}) \) has a central t distribution with \(n-1\) degrees of freedom. If \(\mu_d\ne\mu_{d,0}\), the statistic instead has a noncentral t distribution; normality alone does not make it central.

This distinction matters: the reference value in the numerator becomes the population mean only under the null hypothesis. In the examples below, the t statistic is calculated relative to zero to show how the procedure works. Calculating a statistic is not, by itself, a decision about a research question.

Worked Calculations

Worked Example: Mean Change in a Hypothetical Blood Pressure Measurement

In a hypothetical exercise, six participants have a measurement before and after a program. Define \(d_i=\text{after}_i-\text{before}_i\), in mmHg. Suppose the six observed differences are \(-4,-6,-2,-5,-3,-4\). Calculate the paired t statistic relative to a reference mean difference of zero.

There are \(n=6\) pairs. The difference mean is \( \bar{d}=(-4-6-2-5-3-4)/6=-24/6=-4.000 \) mmHg.

The deviations from \(\bar{d}=-4\) are \(0,-2,2,-1,1,0\). Their squared sum is \(0+4+4+1+1+0=10\). Thus \( s_d=\sqrt{10/(6-1)}=\sqrt{2}=1.414 \) mmHg.

The standard error is \( SE=s_d/\sqrt{n}=1.414/\sqrt{6}=0.5774 \) mmHg. Using \(\mu_{d,0}=0\), the statistic is \( t=(-4.000-0)/0.5774=-6.928 \), with \(df=6-1=5\).

The negative sign follows the chosen after-minus-before coding. This calculation summarizes how far the observed mean difference is from zero in estimated standard-error units; it does not establish that the program caused a change.

Worked Example: Comparing Two Measurement Methods Within Pairs

Suppose six hypothetical samples are each measured by two methods. Define \(d_i=\text{Method A}_i-\text{Method B}_i\), in mg/L, and suppose the differences are \(2,-1,0,1,3,-2\). Find the mean, standard error, and t statistic relative to zero.

The mean difference is \( \bar{d}=(2-1+0+1+3-2)/6=3/6=0.5000 \) mg/L. The deviations from \(0.5\) are \(1.5,-1.5,-0.5,0.5,2.5,-2.5\). The squared deviations sum to \(2.25+2.25+0.25+0.25+6.25+6.25=17.50\).

Therefore \( s_d=\sqrt{17.50/(6-1)}=\sqrt{3.5}=1.871 \) mg/L, and \( SE=1.871/\sqrt{6}=0.7638 \) mg/L.

For \(\mu_{d,0}=0\), \(t=(0.5000-0)/0.7638=0.6546\), with \(df=5\). The paired calculation concerns the six within-sample differences. It does not treat the 12 method readings as 12 independent observations.

Worked Example: Using the Difference Sum of Squares

In a hypothetical matched-pair exercise, eight pairs have differences \(1,2,0,3,1,2,4,3\) points, with the first condition minus the second. Find the sample standard deviation, standard error, and t statistic relative to zero.

The sum of the differences is \(16\), so \(\bar{d}=16/8=2.000\) points. The sum of their squares is \(1+4+0+9+1+4+16+9=44\). A computational form of the sample variance is \( s_d^2=[\sum d_i^2-(\sum d_i)^2/n]/(n-1) \).

Substitution gives \(s_d^2=[44-(16^2/8)]/7=(44-32)/7=12/7=1.714\) points squared, and \(s_d=\sqrt{12/7}=1.309\) points. Hence \(SE=1.309/\sqrt{8}=0.4629\) points.

The statistic is \(t=(2.000-0)/0.4629\approx4.32\), with \(df=8-1=7\). Both the standard error and the t statistic use eight differences because there are eight pairs.

Why the T Reference Distribution Applies

The t distribution enters because the population standard deviation \(\sigma_d\) is not known and is estimated by \(s_d\). Under the independent normal-differences model, the sample mean and sample variance have known sampling properties. These properties justify the exact reference distribution when the null hypothesis is true.

Standardize the sample mean using the population standard deviation.
Under the normal model, \(Z=(\bar{D}-\mu_d)/(\sigma_d/\sqrt{n})\) has a standard normal distribution.

Account for estimating the standard deviation.
For independent normal differences, \(Q=(n-1)S_d^2/\sigma_d^2\) has a chi-square distribution with \(n-1\) degrees of freedom, and \(Z\) and \(Q\) are independent.

Form the t ratio under the null.
When \(H_0:\mu_d=\mu_{d,0}\) is true, \(Z/\sqrt{Q/(n-1)}=(\bar{D}-\mu_{d,0})/(S_d/\sqrt{n})\), which has a central t distribution with \(n-1\) degrees of freedom.

If the true mean differs from the reference, the numerator is not centered at zero in standard-error units. The resulting t statistic has a noncentral t distribution, not the central reference distribution just derived. This is why the statement about a central t distribution must be tied to the null hypothesis, as well as to the independent normal-differences model.

Putting the Calculation in Context

A paired t analysis is a one-sample analysis of the differences, but it still answers a question about the paired design. The interpretation of a positive or negative estimate depends on the chosen subtraction order and on the measured outcome's units. For instance, a negative mean for after-minus-before represents a lower average after measurement in that coding; it does not on its own explain why the measurements differ.

The calculation can be organized consistently:

State exactly how each pair difference is defined.
Count the complete pairs to determine \(n\).
Calculate \(\bar{d}\) and \(s_d\) from the differences.
Calculate \(SE=s_d/\sqrt{n}\), then compare \(\bar{d}\) with the reference value using the t formula.
Use \(df=n-1\), and keep the result connected to the defined outcome and units.

The one-sample framing is useful because it isolates the quantity of interest: the mean within-pair difference. It also prevents a common error—using a two-sample formula that treats the two measurements in each pair as though they came from unrelated units.

Common Mistakes

Using the number of measurements instead of the number of pairs. If there are 12 complete pairs, \(n=12\), not 24; the degrees of freedom are 11.
Using separate group standard deviations in the paired formula. First calculate each within-pair difference, then find the sample standard deviation of those differences.
Changing subtraction order partway through. A consistent reversal changes the signs of the mean and t statistic, so define the order before calculating.
Claiming a central t distribution without the null. Independent normal differences provide the model assumptions, but the statistic has the central t distribution only under \(H_0:\mu_d=\mu_{d,0}\). Under alternatives it is noncentral.
Treating a computed statistic as a causal conclusion. The arithmetic describes paired differences; conclusions about causes depend on the study design and the research question.

Key takeaway: In a paired t procedure, each complete pair contributes one difference. The difference mean, difference standard deviation, and number of pairs determine the standard error and t statistic; the exact central t reference applies under the null and the independent normal-differences model.

Check Your Understanding

Use the paired t framework to answer these questions.

A study has 15 complete matched pairs. What are \(n\) and the degrees of freedom for the paired t statistic?
If differences are coded as treatment minus comparison, what does a negative sample mean difference indicate about the observed averages?
For \(\bar{d}=3.2\), \(s_d=4.0\), and \(n=16\), calculate the standard error and the t statistic relative to zero.
Why is the standard deviation used in a paired t calculation the standard deviation of the differences rather than either group's standard deviation?
Under what conditions does the statistic have a central t distribution, and what distribution does it have under an alternative mean?