Tutorials › Multi-Arm Multi-Stage (MAMS) Designs › Distinguishing Outcome Correlation From Contrast Correlation

Home  /  Tutorials  /  Shared-Control Correlation

Shared-Control Correlation · Tutorial 103 of 499

Distinguishing Outcome Correlation From Contrast Correlation

Separate outcome-level independence from the correlation of treatment-effect estimates, and calculate how a shared control mean creates dependence between contrasts.

Advanced 10 min read

What You'll Learn

  • Distinguish correlations among patient outcomes from correlations among estimated contrasts.
  • Derive the covariance of two treatment-versus-control estimates under independent-group sampling.
  • Calculate contrast correlations with equal or unequal group variances and sample sizes.
  • Explain how contrast direction affects the sign of the covariance.
  • Identify assumptions and data features that can invalidate the simple shared-control calculation.

Two Levels of Dependence

In a multi-arm trial, different treatment groups can contain distinct patients whose outcomes are independent under the sampling model. Yet estimates comparing each treatment with the same control group can be correlated. There is no contradiction: the first statement concerns individual outcomes, while the second concerns statistics calculated from group means. The common control mean appears in both statistics.

It helps to state explicitly which random quantities are being compared. A patient outcome is one random variable. A group mean is a statistic formed from outcomes in one arm. A treatment-effect estimate is a contrast of group means. Correlation between two statistics does not imply that the underlying patient outcomes in different arms were correlated.

Definition: The covariance of random variables \(X\) and \(Z\) is \(\operatorname{Cov}(X,Z)=E[(X-E[X])(Z-E[Z])]\). If both variables have positive variances, their correlation is \(\operatorname{Corr}(X,Z)=\operatorname{Cov}(X,Z)/\sqrt{\operatorname{Var}(X)\operatorname{Var}(Z)}\). Here, outcome-level correlation refers to dependence between individual outcomes; contrast correlation refers to dependence between estimated comparisons across repeated hypothetical trials.

Model and Shared-Control Contrasts

Index the control arm by \(0\) and treatment arms by \(j=1,\ldots,J\). Let \(Y_{ij}\) denote the outcome for patient \(i\) in arm \(j\), with mean \(\mu_j\) and variance \(\sigma_j^2\). For the calculation below, assume fixed sample sizes \(n_j\), independent patients within each arm, and independent samples across arms. Assume each arm has finite variance, and that the same control patients contribute to every treatment-versus-control comparison. The arm sample mean is \(\bar Y_j\), and the estimated contrast is

$$ \hat\theta_j=\bar Y_j-\bar Y_0. $$

Under these assumptions, outcomes from distinct arms have covariance zero, and the arm means are independent. Within arm \(j\), the variance of the sample mean is \(\sigma_j^2/n_j\). Consequently,

$$ \operatorname{Var}(\hat\theta_j) =\frac{\sigma_j^2}{n_j}+\frac{\sigma_0^2}{n_0}. $$

Now consider treatment arms \(j\) and \(k\), where \(j\ne k\). Their contrast estimates contain different treatment means but the same control mean. Expanding the covariance gives

$$ \begin{aligned} \operatorname{Cov}(\hat\theta_j,\hat\theta_k) &=\operatorname{Cov}(\bar Y_j-\bar Y_0,\bar Y_k-\bar Y_0)\\ &=\operatorname{Cov}(\bar Y_j,\bar Y_k) -\operatorname{Cov}(\bar Y_j,\bar Y_0) -\operatorname{Cov}(\bar Y_0,\bar Y_k) +\operatorname{Var}(\bar Y_0)\\ &=\frac{\sigma_0^2}{n_0}. \end{aligned} $$

The first three terms in the expansion are zero because the arm samples are independent. The remaining term is the variance of the shared control mean. Thus, the covariance is positive for contrasts written in the same direction, treatment minus control. The correlation is that covariance divided by the product of the contrasts’ standard deviations:

$$ \operatorname{Corr}(\hat\theta_j,\hat\theta_k) = \frac{\sigma_0^2/n_0} {\sqrt{\left(\sigma_j^2/n_j+\sigma_0^2/n_0\right) \left(\sigma_k^2/n_k+\sigma_0^2/n_0\right)}}. $$

This is a property of the joint sampling distribution of the estimates, not a claim that their observed numerical values must move together in every particular trial. The mean effects \(\mu_j-\mu_0\) affect the expected contrast values, but not this covariance under the stated model.

Worked Calculations

Worked Example: Independent Outcomes, Correlated Estimates

Suppose a hypothetical trial has one control arm and two treatment arms. Each arm has 40 patients, and the outcome variance is 16 points squared in every arm. The observed group means are 100 points for control, 108 for Treatment A, and 105 for Treatment B. Assume independent patient outcomes within and across arms.

At the outcome level, a patient’s outcome in Treatment A is independent of a different patient’s outcome in Treatment B or control, by assumption. In particular, the cross-arm outcome covariances are zero. The estimates, however, are \(\hat\theta_A=108-100=8\) points and \(\hat\theta_B=105-100=5\) points, and both use the same control mean.

Each contrast variance is

$$ \operatorname{Var}(\hat\theta_A) =\operatorname{Var}(\hat\theta_B) =\frac{16}{40}+\frac{16}{40}=0.8\text{ points}^2. $$

Their covariance is the control-mean variance, \(16/40=0.4\text{ points}^2\). Therefore,

$$ \operatorname{Corr}(\hat\theta_A,\hat\theta_B) =\frac{0.4}{\sqrt{0.8(0.8)}}=0.500. $$

The estimated treatment effects are 8 and 5 points, while their sampling correlation is 0.500. The positive contrast correlation comes from their shared control term; it does not establish correlation between individual outcomes in the treatment arms.

Worked Example: Unequal Sample Sizes and Variances

Consider another hypothetical trial. The control arm has \(n_0=40\) patients and outcome variance \(\sigma_0^2=16\text{ units}^2\). Treatment A has \(n_1=30\) and variance \(9\text{ units}^2\); Treatment B has \(n_2=50\) and variance \(25\text{ units}^2\). Suppose the respective sample means are 12, 15, and 10 units, and assume independent samples across arms and independent patients within each arm.

The observed contrasts are \(15-12=3\) units for A versus control and \(10-12=-2\) units for B versus control. The variances of the estimated contrasts are

$$ \begin{aligned} \operatorname{Var}(\hat\theta_1)&=\frac{9}{30}+\frac{16}{40}=0.700\text{ units}^2,\\ \operatorname{Var}(\hat\theta_2)&=\frac{25}{50}+\frac{16}{40}=0.900\text{ units}^2. \end{aligned} $$

The covariance is \(16/40=0.400\text{ units}^2\), so the correlation is

$$ \frac{0.400}{\sqrt{0.700(0.900)}}= \frac{0.400}{\sqrt{0.630}}=0.504. $$

Thus the two estimates have a sampling correlation of about 0.504 despite different treatment-arm variances and sample sizes. Those differences change the contrast variances, and therefore the denominator of the correlation, but the covariance still comes from the control mean. The calculation relies on the stated variances and independence assumptions; the observed sample means alone do not establish those assumptions.

Worked Example: Contrast Direction Changes the Sign

Suppose a hypothetical three-arm trial has 40 patients per arm, an outcome variance of \(16\text{ points}^2\) in each arm, and independent outcomes within and across arms. Let the control, A, and B sample means be 15, 18, and 12 points. Compare the two contrasts \(\hat\theta_A=\bar Y_A-\bar Y_0\) and \(\hat\phi=\bar Y_0-\bar Y_B\). The observed estimates are both 3 points.

The first contrast contains \(-\bar Y_0\); the second contains \(+\bar Y_0\). The treatment-arm means are independent of each other and of the control mean, so the only nonzero covariance term is the one involving the control mean, with a negative sign:

$$ \operatorname{Cov}(\hat\theta_A,\hat\phi) =\operatorname{Cov}(\bar Y_A-\bar Y_0,\bar Y_0-\bar Y_B) =-\operatorname{Var}(\bar Y_0) =-\frac{16}{40}=-0.400\text{ points}^2. $$

Each contrast has variance \(16/40+16/40=0.800\text{ points}^2\). Therefore their correlation is \(-0.400/\sqrt{0.800(0.800)}=-0.500\). The group outcomes remain independent under the model. It is the opposite signs assigned to the shared control mean in these two contrasts that produce negative contrast covariance.

Checking the Interpretation and Assumptions

A useful way to keep the levels separate is to ask what random quantities are in the covariance calculation. If they are individual outcomes from distinct, independent arms, their covariance is zero under the model above. If they are treatment-versus-control estimates, each estimate is a combination of group means, and the same control mean can contribute to both. The covariance of those combinations need not be zero.

The simple formula is not automatic merely because a study is randomized. It describes a model with independent arm samples and fixed sample sizes. In some randomization-based settings, dependence can arise from the assignment mechanism or finite-population structure; repeated measurements, clustering, or other shared sources of variation can also create outcome-level dependence. The actual covariance structure must match the design and analysis model.

For a practical check, verify that the same control observations really are used in both estimates, identify each contrast’s sign for the control mean, and confirm the independence assumptions for the group means. If the control groups are separate and disjoint, the shared-control covariance term does not apply. If one control group is reused, treating its mean as though it were independently sampled for every comparison discards the source of covariance.

Common pitfall: “The treatment arms contain different patients, so the treatment-effect estimates are independent.” The first clause may support independence of outcomes across those arms, but it does not make the estimates independent: both estimates also contain the same random control mean. Calculate covariance between the contrasts, not just between the treatment-arm outcomes.

This distinction matters whenever joint uncertainty is relevant: for example, when several treatment comparisons are considered together. A calculation that keeps only the separate contrast variances describes their individual uncertainty but omits how the estimates co-vary. Conversely, a positive contrast correlation should not be reported as evidence that treatment-arm patient outcomes themselves were correlated.

Check Your Understanding

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

  1. Under independent arm sampling, what is the covariance between an individual outcome in Treatment A and an individual outcome in Treatment B? What is the source of covariance between their treatment-versus-control estimates?
  2. Two same-direction contrasts each have variance \(1.2\text{ units}^2\), and their shared control mean has variance \(0.3\text{ units}^2\). Find their correlation.
  3. Explain why changing a contrast from treatment minus control to control minus treatment changes the sign of its covariance with a same-direction treatment-minus-control contrast.
  4. Name two conditions that could make the independent-arm outcome assumption inappropriate.

Solutions. (1) The individual-outcome covariance is zero under the stated independence model. The contrast covariance is the variance of the shared control mean, because that random mean occurs in both estimates. (2) The correlation is \(0.3/\sqrt{1.2(1.2)}=0.250\). (3) Reversing the contrast changes the control coefficient from \(-1\) to \(+1\); its covariance contribution with another treatment-minus-control contrast therefore changes sign. (4) Examples include clustered outcomes, repeated measurements on the same patients, or dependence induced by the particular assignment or sampling structure.

Previous

Next

Back to Tutorials
Clinical Biostats