Decomposing the Common-Noise Component

Why Separate the Noise Components?

When several treatments are compared with one control group, every estimated treatment-control contrast uses the same estimated control mean. The treatment-arm observations differ from contrast to contrast, but the control observations do not. Separating these sources of sampling error makes the dependence between contrasts explicit: the control-mean error enters each contrast with the same sign.

The key result does not require normally distributed outcomes. It follows from the definitions of the sample means and from independence between groups. Normality may be relevant to particular inferential procedures, but it is not needed for the decomposition or covariance calculation below.

Definition: Consider $K$ treatment arms, indexed by $j=1,\ldots,K$, and one control arm, indexed by $0$. Let $\bar{x}_j$ be the sample mean in arm $j$, and let $\mu_j$ be that arm’s population mean. Define the estimated treatment-control contrast $D_j=\bar{x}_j-\bar{x}_0$, its true value $\delta_j=\mu_j-\mu_0$, and each arm’s mean error $e_j=\bar{x}_j-\mu_j$. Then $$ D_j=\delta_j+e_j-e_0. $$ The term $e_j$ is specific to treatment arm $j$; the term $-e_0$ is the same control-mean error in every treatment-control contrast.

Use this decomposition under a sampling model in which observations are independent across arms, the sample size in each arm is fixed, and each sample mean is unbiased for its arm’s population mean. For the variance formulas, assume arm $j$ has finite variance $\sigma_j^2$ and sample size $n_j$, so that $\operatorname{Var}(e_j)=\sigma_j^2/n_j$. The variances may differ across arms. Independence of arms implies that $e_0,e_1,\ldots,e_K$ are mutually uncorrelated; independent observations within an arm give the stated sample-mean variance. These conditions can fail, for example, if the same participant contributes observations to multiple arms or if the data are clustered in a way that is ignored by the model.

How the Shared Term Produces Covariance

For two distinct treatment arms $j$ and $k$, the errors in their estimated contrasts are $e_j-e_0$ and $e_k-e_0$. Expanding their covariance shows which terms remain. The arm-specific errors have zero covariance with one another and with the control error under the stated independence assumption. The control error, however, is paired with itself.

\begin{aligned} \operatorname{Cov}(D_j,D_k) &=\operatorname{Cov}(e_j-e_0,e_k-e_0)\\ &=\operatorname{Cov}(e_j,e_k)-\operatorname{Cov}(e_j,e_0) -\operatorname{Cov}(e_0,e_k)+\operatorname{Var}(e_0)\\ &=\operatorname{Var}(e_0) =\frac{\sigma_0^2}{n_0}, \qquad j\ne k. \end{aligned}

The same decomposition gives the variance of an individual contrast. Its treatment-arm error and control-mean error are independent, so their variances add:

\operatorname{Var}(D_j) =\operatorname{Var}(e_j-e_0) =\frac{\sigma_j^2}{n_j}+\frac{\sigma_0^2}{n_0}.

Thus the covariance between two distinct contrasts is exactly the variance contributed by their common control mean. This result is about covariance, not a claim that a particular observed pair of contrasts must move together in every sample. It also does not, by itself, establish a joint distribution for the contrasts.

Worked Example: Separating Three Sources of Mean Error

Suppose a hypothetical trial has one control arm and two treatment arms. Assume independent observations across arms, a common outcome standard deviation of 10 points, and sample sizes $n_0=100$, $n_1=80$, and $n_2=50$. These assumptions make the mean-error variances $1$, $1.25$, and $2$ points squared, respectively.

Imagine, for illustration, that the three sample-mean errors in one hypothetical realization are $e_0=0.4$, $e_1=-0.2$, and $e_2=0.7$ points. The errors in the two estimated treatment-control contrasts are therefore $$ e_1-e_0=-0.2-0.4=-0.6\text{ points}, \qquad e_2-e_0=0.7-0.4=0.3\text{ points}. $$ The control error is subtracted from both contrasts. The treatment-specific errors differ, so the two resulting contrast errors need not have the same value or sign.

Under the stated sampling model, the variances of the contrasts are $$ \operatorname{Var}(D_1)=1.25+1=2.25\text{ points}^2, \qquad \operatorname{Var}(D_2)=2+1=3\text{ points}^2. $$ Their covariance is the control mean-error variance, $1\text{ point}^2$. The implied correlation is $$ \frac{1}{\sqrt{(2.25)(3)}}=0.3849. $$ The covariance calculation uses the model’s variances, not the particular hypothetical errors in the realization.

Standardized Errors Still Share the Control Component

Standardization rescales each contrast by its own standard error; it does not remove the shared control noise. Let $v_j=\sigma_j^2/n_j$ for each arm, so that the variance of $D_j$ is $v_j+v_0$. If the variances are known, the standard error is $SE_j=\sqrt{v_j+v_0}$. Define $U_j=e_j/\sqrt{v_j}$ and $U_0=e_0/\sqrt{v_0}$ when these variances are positive. Each $U$ has mean zero and variance one. The standardized contrast error can then be written as follows:

\frac{D_j-\delta_j}{SE_j} = \sqrt{\frac{v_j}{v_j+v_0}}\,U_j - \sqrt{\frac{v_0}{v_j+v_0}}\,U_0.

The control term has a negative coefficient in every standardized contrast. When the errors are uncorrelated as assumed, the covariance of two standardized contrast errors is $$ \frac{v_0}{\sqrt{(v_j+v_0)(v_k+v_0)}}. $$ This is the same covariance result expressed after dividing each contrast by its standard error. In particular, standardizing does not justify treating the resulting statistics as independent.

Worked Example: Standardizing Under Unequal Variances

Suppose a hypothetical study has independent arms, $n_0=40$, $n_1=64$, and $n_2=36$, with standard deviations $\sigma_0=8$, $\sigma_1=6$, and $\sigma_2=9$ points. The mean-error variances are $$ v_0=\frac{8^2}{40}=1.6,\qquad v_1=\frac{6^2}{64}=0.5625,\qquad v_2=\frac{9^2}{36}=2.25 \quad\text{points}^2. $$ The contrast standard errors are $$ SE_1=\sqrt{0.5625+1.6}=\sqrt{2.1625}=1.471\text{ points}, \qquad SE_2=\sqrt{2.25+1.6}=\sqrt{3.85}=1.962\text{ points}. $$ The covariance of the unstandardized contrasts is $v_0=1.6\text{ points}^2$. Their correlation is therefore $$ \frac{1.6}{\sqrt{(2.1625)(3.85)}}=0.5545. $$ Equivalently, in the standardized decomposition the coefficient on the common $U_0$ term has magnitude $\sqrt{1.6/2.1625}=0.860$ for the first contrast and $\sqrt{1.6/3.85}=0.644$ for the second. Their product, $0.860(0.644)=0.5538$, is approximately the correlation; the small difference from $0.5545$ is due to rounding the coefficients. Using unrounded values gives the same result.

The variances differ across arms, but the covariance still comes solely from the shared control mean error. Equal variance across arms is not required for this conclusion.

A Useful Check: The Shared Error Cancels When Subtracted

The decomposition also offers an algebraic check on signs. Subtract the estimated contrast for treatment arm $k$ from that for treatment arm $j$. The control mean appears in both contrasts and cancels:

D_j-D_k =(\delta_j+e_j-e_0)-(\delta_k+e_k-e_0) =(\delta_j-\delta_k)+(e_j-e_k).

This identity does not make the individual treatment-control contrasts independent. It shows instead that the shared control error is absent from their difference. For the first worked example, the variance of $D_1-D_2$ must equal $v_1+v_2=1.25+2=3.25\text{ points}^2$. Checking it from the contrast variances and covariance gives $2.25+3-2(1)=3.25\text{ points}^2$, as required.

Worked Example: Checking a Covariance Calculation by Cancellation

Use the first worked example’s hypothetical model: $v_0=1$, $v_1=1.25$, and $v_2=2$, all in points squared. The difference between the two estimated treatment-control contrasts has variance $$ \operatorname{Var}(D_1-D_2) =v_1+v_2=1.25+2=3.25\text{ points}^2, $$ because the common control error cancels and the treatment-arm errors are independent.

Alternatively, the variance identity for a difference gives $$ \operatorname{Var}(D_1-D_2) =\operatorname{Var}(D_1)+\operatorname{Var}(D_2)-2\operatorname{Cov}(D_1,D_2) =2.25+3-2(1)=3.25\text{ points}^2. $$ Agreement between the two calculations checks both the positive covariance sign and the control-error subtraction in the decomposition.

Assumptions, Diagnostics, and Alternatives

Before applying the formulas, check that the sampling structure matches the model. Each contrast must use the same control sample mean for the shared-error term to be identical. The arm sample means must be unbiased for the means being compared, and the arm-specific mean errors must be uncorrelated for the covariance to reduce to the control variance alone. If observations are paired across arms, clustered, or otherwise dependent, additional covariance terms may be present. The decomposition $D_j-\delta_j=e_j-e_0$ still follows from the definitions, but the simplified variance and covariance formulas may not.

A direct calculation from the contrast definitions and a calculation from the error decomposition are two routes to the same covariance. The direct route expands $\operatorname{Cov}(\bar{x}_j-\bar{x}_0,\bar{x}_k-\bar{x}_0)$; the decomposition route identifies the mean errors first. The latter is often easier to audit because it makes the shared term visible. Neither route permits an independence assumption that contradicts the sampling design.

Check the sign. Each contrast contains $-e_0$, so the covariance contribution from the control term is $\operatorname{Var}(e_0)$, which is nonnegative. A negative contribution here usually signals an expansion error.
Check the sample sizes and variances. Under independent observations, $v_0=\sigma_0^2/n_0$, not $\sigma_0^2/n_j$. The control variance uses the control sample size.
Do not confuse zero mean with zero covariance. The control error has expectation zero, but it is the same random error in every contrast. A zero expected error does not eliminate its contribution to covariance.
Do not infer more than covariance establishes. The decomposition gives a covariance structure under its assumptions; covariance alone does not specify the full joint distribution or validate a particular testing procedure.

The method’s strength is also its limitation: it isolates the dependence created by a common estimated control mean. It does not address bias, missing outcomes, variance estimation, or dependence caused by other features of the data. In those settings, the model must represent the additional sources of uncertainty rather than attributing all dependence to the shared control.

Key takeaway: For each treatment arm, the contrast error is $e_j-e_0$. The treatment error is arm-specific, while the same control error is subtracted in every contrast. Under independent arms, the covariance between distinct contrasts is therefore the variance of the control sample mean.

Check Your Understanding

Answer each question, then compare with the solutions below.

In a shared-control design, write the error in $D_j=\bar{x}_j-\bar{x}_0$ in terms of the arm mean errors.
Suppose $v_0=0.8$, $v_1=1.2$, and $v_2=2.0$, in squared units. Find the two contrast variances and their covariance under independent arms.
For the values in question 2, calculate the correlation between the two contrasts to three significant figures.
Explain why a zero expected control-mean error does not make the control contribution to covariance equal to zero.
State one sampling feature that would require checking or modifying the simplified covariance formula.

Solutions. (1) $D_j-\delta_j=e_j-e_0$, where $e_j=\bar{x}_j-\mu_j$ and $e_0=\bar{x}_0-\mu_0$. (2) The variances are $v_1+v_0=1.2+0.8=2.0$ and $v_2+v_0=2.0+0.8=2.8$ squared units; the covariance is $v_0=0.8$ squared units. (3) The correlation is $0.8/\sqrt{(2.0)(2.8)}=0.338$. (4) The same random control error appears in both contrasts, so its covariance with itself is its variance, even though its expectation is zero. (5) Examples include paired observations across arms, clustering, or any dependence that makes arm mean errors correlated.