Why Pairing Can Improve Precision

Why Pairing Can Change Precision

Pairing links two observations that share something relevant: they may come from the same person at two times, or from two participants matched on baseline characteristics. The purpose is not simply to organize observations into rows. When the two measurements in a pair tend to move together, their difference can vary less than a difference between unrelated observations. Less variability can mean a more precise estimate of the average difference.

The key quantity is the within-pair covariance. It describes whether the two measurements in the same pair tend to be high or low together. This tutorial derives how that covariance enters the variance of a difference and explains when its effect improves precision. A positive covariance helps only in comparison with an appropriate independent-sample baseline; pairing alone does not guarantee a benefit.

Definition: Let $X_i$ and $Y_i$ be the two measurements in pair $i$, and define the within-pair difference as $D_i=X_i-Y_i$. The within-pair covariance is $\operatorname{Cov}(X_i,Y_i)=E[(X_i-\mu_X)(Y_i-\mu_Y)]$. A positive covariance means that, within pairs, values of $X_i$ above their mean tend to occur with values of $Y_i$ above its mean, and values below their mean tend to occur together.

Deriving the Variance of a Difference

For any two random variables with finite variances, the variance of their difference is the sum of their variances minus twice their covariance. The result does not require $X_i$ and $Y_i$ to be independent. Indeed, the covariance term is exactly what distinguishes the paired calculation from an independent one.

Theorem: For $D_i=X_i-Y_i$, $$ \operatorname{Var}(D_i) = \operatorname{Var}(X_i)+\operatorname{Var}(Y_i) -2\operatorname{Cov}(X_i,Y_i). $$

To derive the formula, first subtract the mean of $D_i$. Since $E(D_i)=\mu_X-\mu_Y$, we have $D_i-E(D_i)=(X_i-\mu_X)-(Y_i-\mu_Y)$. Squaring this centered quantity creates a cross-product term. Taking its expectation gives the covariance.

\begin{aligned} \operatorname{Var}(D_i) &=E\left[\left\{(X_i-\mu_X)-(Y_i-\mu_Y)\right\}^2\right]\\ &=E[(X_i-\mu_X)^2]+E[(Y_i-\mu_Y)^2]\\ &\qquad-2E[(X_i-\mu_X)(Y_i-\mu_Y)]\\ &=\operatorname{Var}(X_i)+\operatorname{Var}(Y_i) -2\operatorname{Cov}(X_i,Y_i). \end{aligned}

The minus sign has a direct interpretation: when paired measurements move together, subtracting one from the other cancels some of their shared variation. If the two measurements tend to move in opposite directions, subtraction can instead amplify variation.

From One Pair to the Mean Difference

Suppose there are $n$ pairs, and the pairs are independent of one another. Assume the pairs are drawn from the same population, so each pair has the same marginal variances and within-pair covariance. The average of the pair differences is $\bar{D}=(D_1+\cdots+D_n)/n$. Independence across pairs—not independence of the two observations inside a pair—allows the variance of this average to be calculated by dividing the variance of one difference by $n$.

\operatorname{Var}(\bar{D}) = \frac{\operatorname{Var}(D_i)}{n} = \frac{\sigma_X^2+\sigma_Y^2-2\operatorname{Cov}(X_i,Y_i)}{n}.

This formula describes the sampling variability of the mean within-pair difference. Its units are squared units of the measurement; the standard deviation of $\bar{D}$ is the square root of the variance and has the original measurement units. A smaller variance means that repeated samples, under the stated model, would produce estimates more concentrated around the population mean difference.

Worked Example: Positive Association and Lower Variance

Suppose a hypothetical study pairs participants and records a measurement under conditions $X$ and $Y$. Assume the marginal standard deviations are $\sigma_X=10$ units and $\sigma_Y=8$ units, and that the correlation within a pair is $0.75$. For comparison, suppose an independent-sample design would have the same marginal variances and the same number of observations per condition.

First find the covariance using $\operatorname{Cov}(X_i,Y_i)=\rho\sigma_X\sigma_Y$:

\operatorname{Cov}(X_i,Y_i) =(0.75)(10)(8)=60\text{ units}^2.

The variance of one within-pair difference is therefore

\operatorname{Var}(D_i)=10^2+8^2-2(60) =100+64-120=44\text{ units}^2.

With independent observations, the covariance would be zero, giving a difference variance of $100+64=164\text{ units}^2$. For $n=25$ pairs, the standard deviation of the paired mean difference is $\sqrt{44/25}=1.327$ units. The independent-sample comparison has standard deviation $\sqrt{164/25}=2.561$ units. Under these assumptions, the paired estimate varies less because measurements within a pair share positive variation that cancels in their difference.

Covariance, Correlation, and the Size of the Gain

Covariance depends on the measurement units, so correlation is often easier to use when describing the direction and strength of the within-pair association. If both standard deviations are positive, the relationship is

\operatorname{Cov}(X_i,Y_i)=\rho\sigma_X\sigma_Y, \qquad \operatorname{Var}(D_i)=\sigma_X^2+\sigma_Y^2-2\rho\sigma_X\sigma_Y,

where $\rho$ is the within-pair correlation. Holding the two marginal variances fixed, a positive $\rho$ makes the variance of the difference smaller than it would be at zero correlation. A stronger positive correlation generally means more cancellation. A negative $\rho$ makes the variance larger than the zero-correlation value. The marginal standard deviations also matter: the effect of a given correlation depends on their sizes.

Worked Example: Checking the Variance Identity in Paired Data

Consider four hypothetical pairs of measurements, in units:

Pair $i$	$X_i$	$Y_i$	$D_i=X_i-Y_i$
1	8	7	1
2	10	9	1
3	12	10	2
4	14	12	2

The sample means are $\bar{x}=11$ and $\bar{y}=9.5$. Using the sample variance formula, the deviations for $X$ are $-3,-1,1,3$, so $s_X^2=(9+1+1+9)/3=20/3=6.667$ units squared. The deviations for $Y$ are $-2.5,-0.5,0.5,2.5$, so $s_Y^2=(6.25+0.25+0.25+6.25)/3=13/3=4.333$ units squared.

The sample covariance is the sum of paired cross-products of deviations divided by $n-1$:

s_{XY} = \frac{(-3)(-2.5)+(-1)(-0.5)+(1)(0.5)+(3)(2.5)}{3} = \frac{16}{3} = 5.333\text{ units}^2.

The sample differences are $1,1,2,2$, with mean $\bar{d}=1.5$. Their sample variance is $s_D^2=[(-0.5)^2+(-0.5)^2+(0.5)^2+(0.5)^2]/3=1/3=0.3333$ units squared. The variance identity gives the same result:

s_X^2+s_Y^2-2s_{XY} = \frac{20}{3}+\frac{13}{3}-2\left(\frac{16}{3}\right) = \frac{1}{3} = 0.3333\text{ units}^2.

The positive sample covariance reflects the fact that larger $X$ values tend to be paired with larger $Y$ values in these data. The calculation illustrates the identity; with only four pairs, it is not a basis for a broad conclusion about a population association.

When Pairing Does Not Improve Precision

The comparison with an independent-sample design is clear when the marginal distributions are held fixed. In that comparison, the independent difference has variance $\sigma_X^2+\sigma_Y^2$, while the paired difference has variance $\sigma_X^2+\sigma_Y^2-2\operatorname{Cov}(X_i,Y_i)$. Thus, pairing lowers the variance exactly when the within-pair covariance is positive. If covariance is zero, the variance is the same; if it is negative, the paired difference has greater variance than the independent benchmark.

Worked Example: Negative Association Can Increase Variability

In a hypothetical setting, suppose paired measurements have standard deviations $\sigma_X=6$ units and $\sigma_Y=4$ units, with within-pair correlation $\rho=-0.50$. The covariance is

\operatorname{Cov}(X_i,Y_i) =(-0.50)(6)(4)=-12\text{ units}^2.

The variance of one paired difference is

\operatorname{Var}(D_i)=6^2+4^2-2(-12) =36+16+24=76\text{ units}^2.

With zero covariance and the same marginal variances, the difference variance would be $36+16=52\text{ units}^2$. The negative association therefore increases, rather than decreases, the paired difference variance relative to that benchmark. For $n=16$ independent pairs, the corresponding standard deviation of the mean difference is $\sqrt{76/16}=2.179$ units, compared with $\sqrt{52/16}=1.803$ units at zero covariance.

A positive covariance is not by itself a guarantee that a particular matching strategy is worthwhile. The comparison must be meaningful: if pairing changes which individuals are represented, or changes the marginal variability of the measurements, the simple same-marginals comparison no longer captures the whole design. Also, observations from different pairs need to be independent for the stated variance of the mean to apply. These are reasons to think carefully about how pairs are formed, rather than treating the word “paired” as a promise of higher precision.

Common Mistakes and Why Precision Matters

Dropping the covariance term. Using only $\sigma_X^2+\sigma_Y^2$ treats the two measurements within a pair as uncorrelated. The paired variance includes $-2\operatorname{Cov}(X_i,Y_i)$.
Assuming pairing always helps. Pairing reduces variance relative to the independent benchmark when covariance is positive, has no effect at zero covariance, and can increase variance when covariance is negative.
Confusing association with causation. Positive within-pair association describes co-movement; it does not establish that one measurement causes the other.
Mixing variance and standard deviation units. Variances and covariances are in squared units, while standard deviations are in the original units. Take a square root before interpreting a standard deviation.

Precision matters because a less variable estimate tends to be more informative: repeated samples under the model would produce estimates closer together. Pairing can achieve this by removing shared variation from each difference, while still targeting the mean difference of interest. Whether it does so depends on the actual within-pair relationship and the design assumptions.

Key takeaway: The variance of a within-pair difference is the sum of the two marginal variances minus twice their covariance. With the marginal variances held fixed, positive within-pair covariance reduces difference variability; pairing is not automatically more precise.

Check Your Understanding

Work through these questions using the variance identity and its assumptions.

If $\sigma_X^2=25$, $\sigma_Y^2=16$, and $\operatorname{Cov}(X_i,Y_i)=5$, what is $\operatorname{Var}(X_i-Y_i)$?
For fixed marginal variances, how does a negative within-pair covariance compare with zero covariance in its effect on the variance of a difference?
Suppose the within-pair correlation is $0.60$, with $\sigma_X=5$ units and $\sigma_Y=3$ units. Find the covariance and the variance of the difference.
Why does the variance of the mean of $n$ pair differences equal the variance of one difference divided by $n$ under the stated model?
Explain why positive within-pair association does not, by itself, prove that a particular matching strategy will improve precision in every application.

Pair \(i\)	\(X_i\)	\(Y_i\)	\(D_i=X_i-Y_i\)
1	8	7	1
2	10	9	1
3	12	10	2
4	14	12	2