Planning the Number of Pairs for Precision

Why Plan the Number of Pairs?

A paired study is often designed to estimate an average change or an average within-pair difference. Planning the number of pairs before collecting data helps answer a practical question: how many pairs should be included to estimate that mean with a desired level of precision? The answer depends not just on the confidence level and target margin of error, but on how variable the differences between paired observations are expected to be.

For example, a study might pair each participant’s measurement before and after an intervention, or match participants and compare their outcomes. In either case, the analysis works with one difference per pair. The sample size for planning precision is therefore the number of pairs, not the number of individual measurements.

Definition: Let $d_i$ be the difference recorded for pair $i$, with its sign defined to match the study question. The mean paired difference is the average of these pair-level differences, and the population standard deviation of the differences is $\sigma_d$. For planning, $n$ denotes the number of independent pairs that contribute a difference.

A target margin of error $E$ is a chosen maximum half-width for a confidence interval, expressed in the outcome’s units. If the target is $E=2$ minutes, for instance, the aim is an interval whose half-width is about 2 minutes or less. Planning estimates whether that precision is plausible for a proposed number of pairs; it cannot guarantee that the observed interval will meet the target.

The Planning Formula

For paired inference, the confidence interval is based on the distribution of the pair-level differences. Its estimated standard error is $s_d/\sqrt{n}$, where $s_d$ is the sample standard deviation of those differences. A two-sided $100(1-\alpha)\%$ t interval has margin of error $t^*_{n-1}s_d/\sqrt{n}$. When planning, $s_d$ is not yet available, so it is replaced by an anticipated value for $\sigma_d$, often informed by earlier data or a pilot study.

E_{\mathrm{planned}} \approx t^*_{n-1}\frac{\sigma_d}{\sqrt{n}}

Because the t critical value depends on the degrees of freedom $n-1$, the exact t-based planning calculation is usually checked across integer values of $n$. The planned count is the smallest integer for which the anticipated margin is no larger than the target $E$.

Planning rule: For a chosen confidence level and anticipated $\sigma_d$, choose the smallest integer $n$ satisfying $$ t^*_{n-1}\frac{\sigma_d}{\sqrt{n}} \le E. $$ The calculation plans for $n$ pairs with usable pair-level differences.

A convenient first calculation uses a standard normal critical value $z^*$, rather than a t critical value. Rearranging $z^*\sigma_d/\sqrt{n}\le E$ gives an approximate count. Round up, since a fraction of a pair cannot be collected. Then check that count with the t critical value; if it misses the target, check larger integers until the t-based requirement is met.

n_{\mathrm{normal}} \approx \left(\frac{z^*\sigma_d}{E}\right)^2

This normal-based formula is a useful starting point, not a substitute for the t-based check when a t interval is planned. The calculation uses the standard deviation of the differences, not the standard deviations of the two measurements treated separately.

Why the Pair-Level Standard Deviation Matters

Pairing changes which variation controls the precision. For each pair, subtract one observation from the other according to the chosen coding; the resulting difference is the unit analyzed. The standard deviation $\sigma_d$ describes how much those differences vary across pairs. If differences are relatively consistent, their mean can often be estimated precisely with fewer pairs. If they vary widely, more pairs are needed to achieve the same target margin of error.

When both measurements are quantitative, the variance of their difference depends on the variability of each measurement and their within-pair association:

\sigma_d^2=\sigma_1^2+\sigma_2^2-2\rho\sigma_1\sigma_2

Here $\sigma_1$ and $\sigma_2$ are the population standard deviations of the two measurements, and $\rho$ is their within-pair correlation. This identity follows from $\operatorname{Var}(X_1-X_2)=\operatorname{Var}(X_1)+\operatorname{Var}(X_2)-2\operatorname{Cov}(X_1,X_2)$, together with $\operatorname{Cov}(X_1,X_2)=\rho\sigma_1\sigma_2$. Thus, the amount of variation in each measurement alone does not determine the precision of the mean difference: the within-pair relationship matters too.

Worked Example: Exact t-Based Planning for a Target Margin

A hypothetical study will estimate the mean change in a continuous outcome, measured in points. Based on relevant prior information, the researchers anticipate $\sigma_d=6$ points. They want a 95% confidence interval with a margin of error no greater than 2 points. Find the planned number of pairs.

For a 95% two-sided interval, $z^*\approx1.96$. The normal-approximation count is

n_{\mathrm{normal}} \approx \left(\frac{1.96(6)}{2}\right)^2 = (5.88)^2 = 34.5744.

Rounding up gives 35 as the initial candidate. Because a t interval is planned, check the t-based margin. With $n=37$, $df=36$, and $t^*_{36}\approx2.0281$,

E_{\mathrm{planned}} \approx 2.0281\frac{6}{\sqrt{37}} \approx 2.0005\text{ points}.

This is just above the 2-point target, so 37 pairs do not meet the requirement. With $n=38$, $df=37$, and $t^*_{37}\approx2.0262$,

E_{\mathrm{planned}} \approx 2.0262\frac{6}{\sqrt{38}} \approx 1.973\text{ points}.

Therefore, 38 pairs is the smallest count that meets the target under these planning assumptions: 37 narrowly misses, while 38 meets it.

Worked Examples With Different Planning Choices

Worked Example: A 90% Interval and a Larger Normal Approximation

A hypothetical paired study measures a laboratory outcome in units. Its anticipated standard deviation of differences is 4.5 units, and the desired margin of error is 1.5 units for a 90% confidence interval. First obtain a normal-based starting count, then check the t requirement.

For a two-sided 90% interval, $z^*\approx1.6449$. Thus,

n_{\mathrm{normal}} \approx \left(\frac{1.6449(4.5)}{1.5}\right)^2 = (4.9347)^2 \approx 24.35.

Rounding up gives 25 as a starting candidate. At $n=25$, $df=24$, and $t^*_{24}\approx1.7109$, the anticipated t-based margin is

1.7109\frac{4.5}{\sqrt{25}} =1.7109(0.9) \approx1.540\text{ units}.

This exceeds the target. At $n=26$, $df=25$, and $t^*_{25}\approx1.7081$, the margin is approximately $1.507$ units, still above 1.5. At $n=27$, $df=26$, and $t^*_{26}\approx1.7056$,

1.7056\frac{4.5}{\sqrt{27}} \approx1.477\text{ units}.

The planned minimum is 27 pairs. The normal approximation provided a useful starting value, but the t-based checks showed that a few more pairs were needed.

Worked Example: How Pair-Level Variation Changes the Plan

Suppose two hypothetical outcomes each have a standard deviation of 10 units. Consider two possible within-pair correlations: $\rho=0.80$ and $\rho=0.20$. Compare the anticipated standard deviation of the difference, then use a 95% normal approximation to plan for a margin of error of 3 units.

For $\rho=0.80$, the difference variance is

\sigma_d^2=10^2+10^2-2(0.80)(10)(10)=40, \qquad \sigma_d=\sqrt{40}\approx6.325\text{ units}.

The approximate count is

n_{\mathrm{normal}} \approx \left(\frac{1.96(6.325)}{3}\right)^2 \approx17.07,

so the normal-based starting count is 18 pairs. For $\rho=0.20$,

\sigma_d^2=10^2+10^2-2(0.20)(10)(10)=160, \qquad \sigma_d=\sqrt{160}\approx12.65\text{ units}.

This gives

n_{\mathrm{normal}} \approx \left(\frac{1.96(12.65)}{3}\right)^2 \approx68.30,

or 69 pairs as the normal-based starting count. These are approximate counts; a study using a t interval should check them against the t-based rule. The comparison shows why planning should use the variability of paired differences: although the two individual outcomes have the same standard deviations in both scenarios, their different within-pair correlations produce very different difference variability and, consequently, different sample-size needs.

Assumptions and Practical Interpretation

The calculation describes precision for the mean of a population of pair-level differences. Its standard assumptions should be considered before treating the resulting count as a sensible plan:

The pair is the analysis unit. Each pair contributes one defined difference, and $n$ counts those differences. Two measurements from one pair are not two independent pairs.
Pairs are independent of one another. The difference from one pair should not provide unaccounted-for dependence with the difference from another pair. The two observations within a pair are expected to be related; that relationship is part of the pair-level difference.
The anticipated $\sigma_d$ is relevant. It should represent differences measured and defined in a way appropriate to the planned study. A poor estimate of $\sigma_d$ can lead to a poor precision plan.
The interval procedure is suitable. For a small number of pairs, the distribution of differences should be reasonably compatible with the paired t procedure, without severe skewness or influential outliers. With more pairs, that procedure is generally more robust, though serious problems in the differences still deserve attention.

The value $t^*_{n-1}\sigma_d/\sqrt{n}$ is an anticipated margin based on the planning standard deviation. In the study itself, the interval uses the observed $s_d$, which may be larger or smaller than anticipated. If the observed differences are more variable than expected, the achieved margin will tend to be wider; if they are less variable, it will tend to be narrower. A planned count therefore does not ensure that the realized interval will have exactly the target width.

The planning target is also about precision, not about whether a hypothesis test will be statistically significant. It answers how narrow an interval is expected to be under a specified variability assumption and confidence level. It does not, by itself, guarantee a particular conclusion about the size or direction of the mean difference.

Common Mistakes

Counting observations instead of pairs. If a study plans 30 pairs, the paired analysis has $n=30$ differences, even though it involves 60 measurements. Use 30 in the paired standard error.
Using the standard deviation of one measurement in place of $\sigma_d$. The planning formula needs the standard deviation across pair-level differences. If only the two measurement standard deviations are known, the within-pair covariance also matters.
Rounding down the approximate count. If the formula gives 34.57, planning 34 pairs cannot meet that normal-approximation inequality. Round up to 35, then check the t-based requirement.
Calling the planned margin guaranteed. The calculation substitutes an anticipated standard deviation. The realized sample standard deviation can differ, so the final interval may be wider or narrower than planned.
Treating the t critical value as fixed. For a t interval, degrees of freedom depend on $n$. Use the critical value for $df=n-1$, and verify the final integer count.

Key takeaway: For precision planning in a paired study, choose the number of independent pairs using the anticipated standard deviation of their differences. Start with the normal approximation if useful, then check the smallest integer $n$ satisfying $t^*_{n-1}\sigma_d/\sqrt{n}\le E$.

Check Your Understanding

Try these questions without looking back.

A paired study will collect two measurements from each participant. If the plan calls for 42 pairs, what value of $n$ belongs in the paired standard error, and why?
In the t-based planning rule, what does $\sigma_d$ measure? Why is it not automatically the standard deviation of either individual measurement?
A normal approximation gives $n=28.2$. What integer count should be checked first, and what further check is needed if a t interval is planned?
Two outcomes have the same marginal standard deviations in two study settings, but one setting has a stronger positive within-pair correlation. Which setting tends to have the smaller standard deviation of differences? Explain using the variance formula.
Why does a planned margin of error not guarantee that the final observed confidence interval will have that margin or less?