Deriving Covariance From the Shared Control

Why a Shared Control Creates Covariance

When two treatments are each compared with the same control group, both estimated differences contain the control mean. A change in that one estimated mean affects both comparisons: if the control mean is unusually high in a particular sample, both treatment-minus-control differences tend to be lower. This shared sampling variation is the source of their covariance under independent sampling across arms.

The result is about the sampling distribution of estimates across hypothetical repetitions of the same design. It is not a calculation from the observed treatment means alone, and it does not say that the biological treatment effects are related. The covariance arises because the comparisons reuse the same random control mean.

Definition: Label the control arm $0$ and two treatment arms $1$ and $2$. Let $\bar{x}_a$ be the sample mean in arm $a$, and define the two treatment–control mean differences by $D_1=\bar{x}_1-\bar{x}_0$ and $D_2=\bar{x}_2-\bar{x}_0$. Their covariance is the repeated-sampling quantity $\operatorname{Cov}(D_1,D_2)=E[(D_1-E[D_1])(D_2-E[D_2])]$.

Sampling Assumptions and the Result

Suppose observations in arm $a$ are independent and identically distributed with population mean $\mu_a$, finite population variance $\sigma_a^2$, and sample size $n_a$. Assume also that the samples from different arms are independent. In particular, $\bar{x}_1$, $\bar{x}_2$, and $\bar{x}_0$ are mutually independent. The arm variances and sample sizes need not be equal. No normality assumption is needed for this covariance result.

Under these assumptions, a sample mean in arm $a$ has variance $\sigma_a^2/n_a$. The result of interest follows from covariance algebra; it does not depend on the values of the treatment-arm means or variances.

Theorem: If the three arm sample means are mutually independent and the control observations are independent with finite variance $\sigma_0^2$, then $$ \operatorname{Cov}(D_1,D_2)=\operatorname{Var}(\bar{x}_0)=\frac{\sigma_0^2}{n_0}. $$ The covariance is nonnegative and is strictly positive when $\sigma_0^2>0$.

Derivation From the Shared Mean

The covariance of two differences can be expanded using linearity in each argument and the identity $\operatorname{Cov}(A,A)=\operatorname{Var}(A)$. Showing all terms makes clear which assumptions are doing the work.

\begin{aligned} \operatorname{Cov}(D_1,D_2) &=\operatorname{Cov}(\bar{x}_1-\bar{x}_0,\bar{x}_2-\bar{x}_0)\\ &=\operatorname{Cov}(\bar{x}_1,\bar{x}_2) -\operatorname{Cov}(\bar{x}_1,\bar{x}_0) -\operatorname{Cov}(\bar{x}_0,\bar{x}_2) +\operatorname{Var}(\bar{x}_0). \end{aligned}

Mutual independence of the arm samples implies independence of their sample means. Thus each covariance between means from distinct arms is zero. Only the final term remains. To evaluate it, write the control sample mean as the average of $n_0$ independent observations, each with variance $\sigma_0^2$:

\operatorname{Var}(\bar{x}_0) =\operatorname{Var}\left(\frac{1}{n_0}\sum_{j=1}^{n_0}X_{0j}\right) =\frac{1}{n_0^2}\sum_{j=1}^{n_0}\operatorname{Var}(X_{0j}) =\frac{n_0\sigma_0^2}{n_0^2} =\frac{\sigma_0^2}{n_0}.

The sum-of-variances step uses independence within the control arm. If control observations are dependent, that step does not follow as written. The cross-arm independence assumption is also essential: without it, one or more of the first three covariance terms may not be zero.

Write both contrasts.
Each difference contains its own treatment mean and the same control mean.

Expand the covariance.
The expansion has three cross-arm covariance terms and one control-mean variance term.

Apply independence.
Independence across arms makes the three cross-arm covariance terms zero.

Evaluate the remaining variance.
The control sample mean has variance $\sigma_0^2/n_0$ for independent, identically distributed control observations.

Worked Calculations

Worked Example: Computing Covariance From the Control Variance

Consider a hypothetical trial measuring systolic blood pressure in millimetres of mercury (mmHg). The control arm has $n_0=40$ participants and population variance $\sigma_0^2=100\text{ mmHg}^2$. Assume independent sampling within and across arms, with finite variances. The two treatment arms each have their own independent sample and both comparisons use this same control sample. Find the covariance between the two treatment–control mean differences.

The assumptions permit the theorem to be used: the treatment sample means are independent of each other and of the control sample mean, and the control observations are independent with the stated variance. Therefore,

\operatorname{Cov}(D_1,D_2) =\frac{\sigma_0^2}{n_0} =\frac{100\text{ mmHg}^2}{40} =2.50\text{ mmHg}^2.

The covariance is $2.50\text{ mmHg}^2$. It is determined by the control-arm sampling variability, not by the treatment-arm sample sizes or treatment-arm population variances. Those quantities can matter for other properties of the differences, but they do not enter this covariance under the stated independence assumptions.

Worked Example: Estimating Covariance When the Variance Is Unknown

Suppose a hypothetical trial measures a continuous outcome in units. There are $n_0=36$ independent control observations, and the observed control sample standard deviation is $s_0=9$ units. Assume the treatment-arm samples are independent of one another and of the control sample. Because the population variance is unknown, use the control sample variance as a plug-in estimate.

First square the sample standard deviation: $s_0^2=9^2=81\text{ units}^2$. Substituting this estimate for $\sigma_0^2$ gives

\widehat{\operatorname{Cov}}(D_1,D_2) =\frac{s_0^2}{n_0} =\frac{81\text{ units}^2}{36} =2.25\text{ units}^2.

This is an estimated covariance, not the exact population covariance: the latter is $\sigma_0^2/n_0$, and $\sigma_0^2$ is unknown. The estimate is based on the control sample’s observed variability. Its validity as an estimate depends on the sampling assumptions, including independent control observations.

Worked Example: What Changes If Treatment Means Are Dependent?

Consider a hypothetical design in which the sample means, measured in units, have variances $4\text{ units}^2$ for each treatment mean and for the control mean. Suppose the two treatment means have covariance $1\text{ unit}^2$, while each treatment mean is uncorrelated with the control mean. These stated covariances are compatible with the stated variances: the covariance between the treatment means is smaller in magnitude than the product of their standard deviations, which is $2\times2=4\text{ units}^2$.

The general expansion, rather than the independence result, must be used:

\begin{aligned} \operatorname{Cov}(D_1,D_2) &=\operatorname{Cov}(\bar{x}_1,\bar{x}_2) -\operatorname{Cov}(\bar{x}_1,\bar{x}_0) -\operatorname{Cov}(\bar{x}_0,\bar{x}_2) +\operatorname{Var}(\bar{x}_0)\\ &=1-0-0+4 =5\text{ units}^2. \end{aligned}

Here the covariance is $5\text{ units}^2$, not the control-mean variance of $4\text{ units}^2$. The reason is that the treatment means are not independent, so their covariance contributes an additional term. This example illustrates why the assumptions must be checked before applying the simplified formula.

Interpretation, Diagnostics, and Alternatives

Covariance is measured in the product of the variables’ units; here, both contrasts are measured in outcome units, so their covariance is in squared outcome units. A positive covariance indicates that deviations of the two contrasts from their respective expectations tend to have the same sign. In the shared-control setting, a control mean above its expected value subtracts a larger amount from both contrasts, moving both downward; a control mean below expectation moves both upward.

The expression $\sigma_0^2/n_0$ does not describe the probability that both comparisons are significant, nor does it by itself quantify how strongly they are associated on a standardized scale. It is one covariance quantity. Interpretation in any particular analysis also requires attention to the sampling design and the other features of the estimates.

Before using the formula, check how participants and observations were assigned and collected. The simple result requires that treatment-arm samples are independent of one another and that the control sample is independent of both. If, for example, the design creates dependence between treatment-arm means through shared clusters or repeated observations, the cross-arm terms in the expansion may be nonzero. Likewise, clustering or repeated measurements within the control arm can make its sample-mean variance differ from $\sigma_0^2/n_0$. In either case, use a variance model that reflects the actual sampling structure rather than assuming independent observations.

A different design could use a separate, independently sampled control group for each treatment comparison. Under independence of all those arm means, the two contrasts would then contain different control means, and there would be no shared-control variance term in their covariance. That is a design alternative, not a different calculation for a single shared control group.

Common pitfall: “The two treatment means come from different arms, so the two treatment–control differences are independent.” This reasoning overlooks that both differences subtract the same random control mean. Under the stated independence assumptions, the treatment means are independent, but the differences share the control mean and have covariance $\sigma_0^2/n_0$.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

In the independent-arm setting, which term in the covariance expansion remains, and why?
A hypothetical control arm has $n_0=25$ and population variance $\sigma_0^2=64\text{ units}^2$. What is the covariance between the two treatment–control mean differences?
Does the formula $\sigma_0^2/n_0$ require normally distributed outcomes? State the assumptions it does require.
If the two treatment means have nonzero covariance, can the shared-control result be used unchanged? Explain.

Solutions. (1) Only $\operatorname{Var}(\bar{x}_0)$ remains because independence makes the three covariances between means from distinct arms zero. (2) The covariance is $64/25=2.56\text{ units}^2$. (3) No normality assumption is required. The result requires finite control variance, independent observations within the control arm, and independent samples across the three arms. (4) No. The general expansion includes $\operatorname{Cov}(\bar{x}_1,\bar{x}_2)$, which is nonzero in that case, so the covariance need not equal the control-mean variance alone.