Confidence Intervals for Mean Paired Differences

Why Use an Interval for Paired Differences?

When two measurements come from the same individual, the difference between them can describe the change or contrast of interest. A confidence interval for the population mean paired difference estimates the average of those individual differences in a target population. It gives both a plausible range of values and a sense of the estimate’s precision.

The order of subtraction matters. If a study records a measurement before and after an intervention, defining each difference as after minus before makes positive values indicate increases. Defining it as before minus after makes positive values indicate decreases. The interval must use the same order as the differences, and its interpretation must preserve that order.

Definition: For \(n\) paired observations, define \(d_i\) as the difference for pair \(i\), using one stated subtraction order. The population mean paired difference, \(\mu_d\), is the mean of these differences in the population of interest. The sample mean difference is \(\bar{d}\), and the sample standard deviation of the differences is \(s_d\).

Paired data are analyzed through the differences: once the \(d_i\) values have been formed, the interval uses the one-sample t method on those values. The sample size \(n\) is the number of complete pairs, not the number of individual measurements. In particular, the standard error describes variation in the sample mean of the differences.

The t Confidence Interval

A t confidence interval for \(\mu_d\) is appropriate when the paired differences are a random or otherwise representative sample of the population differences, and the pairs are independent of one another. For small samples, the distribution of the differences should be approximately normal, without strong skewness or extreme outliers. With a sufficiently large sample, the t procedure is generally more robust to nonnormality, although extreme outliers can still be problematic. The relevant distributional check is on the differences, not separately on the two columns of measurements.

If these conditions are reasonable, a two-sided \(100(1-\alpha)\%\) confidence interval uses the t critical value with \(n-1\) degrees of freedom:

\bar{d}\ \pm\ t^*_{n-1}\frac{s_d}{\sqrt{n}}

Here \(t^*_{n-1}\) is the positive t critical value that leaves area \(\alpha/2\) in each tail of a t distribution with \(n-1\) degrees of freedom. The quantity \(s_d/\sqrt{n}\) is the estimated standard error of \(\bar{d}\). The result is centered at the observed mean difference, and the margin of error is the critical value multiplied by that standard error.

Why the Formula Works

The justification begins with a model for the differences. If the \(n\) differences are independent observations from a normal population with mean \(\mu_d\) and standard deviation \(\sigma_d\), then the standardized sample mean using the population standard deviation has a standard normal distribution. Since \(\sigma_d\) is unknown, it is estimated by \(s_d\). The resulting statistic follows a t distribution with \(n-1\) degrees of freedom:

T=\frac{\bar{d}-\mu_d}{s_d/\sqrt{n}}\sim t_{n-1}

By the definition of \(t^*_{n-1}\), the central area between the two critical values is \(1-\alpha\):

P\left(-t^*_{n-1}\leq \frac{\bar{d}-\mu_d}{s_d/\sqrt{n}} \leq t^*_{n-1}\right)=1-\alpha

The standard error is positive, so multiplying through by it preserves the inequality. Subtracting \(\bar{d}\) and then multiplying by \(-1\) reverses the inequality signs. Rearranging gives:

P\left( \bar{d}-t^*_{n-1}\frac{s_d}{\sqrt{n}} \leq \mu_d \leq \bar{d}+t^*_{n-1}\frac{s_d}{\sqrt{n}} \right)=1-\alpha

Thus, the interval formula is the set of parameter values that satisfy the central t-statistic inequality. Under the model, the procedure produces intervals that cover \(\mu_d\) in \(100(1-\alpha)\%\) of repeated samples. The confidence level describes the long-run performance of the method; after an interval has been calculated, it is not a probability statement that \(\mu_d\) lies within that particular interval.

Worked Examples

Worked Example: Estimating an Average Increase

In a hypothetical study, eight students complete a short task before and after a training session. The recorded completion times, in seconds, are paired by student. Define \(d=\text{before time}-\text{after time}\), so positive differences indicate that a student completed the task faster after training. The observed differences are \(1, 2, 0, 3, 2, 1, 4, 1\) seconds. Construct a 95% confidence interval for the population mean difference.

Conditions. Suppose these students were randomly selected from the population of interest and their outcomes are independent across students. The eight differences are not extremely skewed and contain no obvious extreme outlier, so a t interval is a reasonable method for this hypothetical example. Because there are only eight pairs, the shape of the differences matters.

Calculate the sample summaries. The sum of the differences is \(14\), so \(\bar{d}=14/8=1.75\) seconds. The sum of squared deviations from the mean is \(11.5\) seconds squared. Therefore,

s_d=\sqrt{\frac{11.5}{8-1}} =\sqrt{1.6429} \approx 1.282\text{ seconds}

For a 95% interval, \(df=n-1=7\), and \(t^*_{7}=2.365\). The standard error is \(1.282/\sqrt{8}\approx0.4532\) seconds, giving:

1.75\pm 2.365\left(\frac{1.282}{\sqrt{8}}\right) =1.75\pm1.071 =(0.679,\ 2.821)\text{ seconds}

We are 95% confident that the population mean of before-minus-after completion times is between about \(0.679\) and \(2.821\) seconds. The interval is entirely positive, so its values correspond to an average reduction in completion time after training under this subtraction order. It does not imply that every student became faster.

Worked Example: How the Subtraction Order Changes Interpretation

In a hypothetical study, six people have their pulse measured in the morning and evening. Define \(d=\text{morning pulse}-\text{evening pulse}\), measured in beats per minute. The six paired differences are \(3,-1,2,0,1,1\). Construct a 90% confidence interval for the population mean difference.

Conditions. Assume the six people are a representative sample and the pairs are independent. For this small sample, suppose the differences are reasonably symmetric with no extreme outlier. These assumptions make the t interval reasonable for the example.

The mean difference is \(\bar{d}=6/6=1\) beat per minute. The sum of squared deviations from 1 is \(10\), so \(s_d=\sqrt{10/5}=\sqrt{2}\approx1.414\) beats per minute. Here \(df=5\). For a 90% interval, \(t^*_{5}=2.015\), and:

1\pm2.015\left(\frac{1.414}{\sqrt{6}}\right) =1\pm1.163 =(-0.163,\ 2.163)\text{ beats per minute}

We are 90% confident that the population mean morning-minus-evening pulse difference is between about \(-0.163\) and \(2.163\) beats per minute. The range includes zero, so the interval includes both a very small average difference in either direction. If instead the differences had been defined as evening minus morning, every difference and both interval endpoints would change sign: the interval would be \((-2.163,\ 0.163)\) beats per minute. Reversing the order changes the direction, not the underlying paired comparison.

Worked Example: Building an Interval From Software Summaries

A hypothetical laboratory study records paired measurements of a marker before and after a procedure for 12 randomly selected participants. Define \(d=\text{after}-\text{before}\), measured in units per liter. Software reports \(\bar{d}=-4.2\) units per liter and \(s_d=5.1\) units per liter. Construct and interpret a 95% interval.

Conditions. Assume participants’ pairs are independent and representative of the intended population. Suppose the 12 paired differences show no extreme outliers or strong skewness. Those distributional features, rather than the separate before and after measurements, are the relevant small-sample check.

There are \(n=12\) pairs, so \(df=11\). The 95% critical value is \(t^*_{11}=2.201\). The standard error is \(5.1/\sqrt{12}\approx1.472\) units per liter. Substitution gives:

-4.2\pm2.201\left(\frac{5.1}{\sqrt{12}}\right) =-4.2\pm3.239 =(-7.439,\ -0.961)\text{ units per liter}

We are 95% confident that the population mean after-minus-before difference is between about \(-7.439\) and \(-0.961\) units per liter. Because the whole interval is negative, its values indicate a lower average measurement after the procedure under the defined order. This statement concerns the population mean paired difference, not the change for each participant.

Interpreting and Reporting the Interval

A complete report identifies the outcome, subtraction order, population, confidence level, interval, and units. For example: “The 95% confidence interval for the population mean before-minus-after completion-time difference is \(0.679\) to \(2.821\) seconds.” Naming the order prevents readers from having to guess what a positive or negative endpoint means.

The interval estimates an average difference, not a typical individual difference and not the range of all individual changes. Its width reflects the precision of the estimate: all else equal, a smaller standard error or a lower t critical value produces a narrower interval. The t critical value depends on the confidence level and degrees of freedom, while the standard error depends on the variability and number of paired differences.

A confidence interval is also tied to the population and sampling process. A correctly calculated interval does not remove bias caused by an unrepresentative sample, dependence between pairs, or a poorly measured outcome. Conditions support the statistical procedure; they do not guarantee that the study’s design answers every scientific question.

Common Mistakes

Changing the subtraction order mid-analysis. Calculating after-minus-before differences and then describing positive values as decreases reverses the meaning. State the order first and use it consistently.
Using the number of measurements as the sample size. With 12 people measured twice, there are 12 paired differences, so \(n=12\) and \(df=11\), not \(n=24\).
Checking the wrong distributions. The t interval applies to the paired differences. Inspecting only the before and after measurements separately does not check the relevant small-sample condition.
Misstating confidence. A 95% confidence level does not mean there is a 95% probability that the fixed population mean is in the calculated interval. It describes the long-run coverage of the interval procedure.
Confusing an interval for a mean with individual outcomes. The endpoints describe plausible values for the population mean difference, not a range that contains 95% of individual differences.

Key takeaway: Form one difference per pair, keep its subtraction order explicit, and use \(\bar{d}\pm t^*_{n-1}s_d/\sqrt{n}\) when the paired-difference conditions are reasonable. Interpret both endpoints in context and with the same units and direction.

Check Your Understanding

Answer each question by referring to the paired differences and the stated subtraction order.

A researcher defines \(d=\text{after}-\text{before}\). What does a positive value of \(d\) mean?
A study has 15 complete pairs. What are \(n\) and the degrees of freedom for the t interval?
For a small paired sample, which values should be checked for strong skewness or extreme outliers: the before measurements, the after measurements, or the differences?
Why is it incorrect to describe a 95% confidence interval as giving a 95% probability that the fixed population mean lies in the calculated interval?
If every difference is redefined in the reverse order, how do the interval endpoints and their interpretation change?