Correlation With Equal Allocation and Equal Variance

Why a Shared Control Creates Correlation

In a multi-arm trial, several treatments may each be compared with the same control group. Although the treatment groups are distinct, each estimated treatment effect contains the same estimated control mean. If that control mean happens to be high in a particular sample, all treatment-minus-control estimates tend to be lower; if it happens to be low, they tend to be higher. This shared sampling variation makes the comparisons dependent.

This tutorial derives the special-case correlation when every group has the same sample size and all groups have a common outcome variance. The central result is simple, but its assumptions matter: under this model, any two distinct treatment-control contrasts have correlation \(1/2\). This is a correlation between estimates across repeated samples, not a statement that the true treatment effects themselves are correlated.

Model and Assumptions

For the derivation, assume that each of the \(K+1\) groups has the same sample size \(n\). Let \(X_{g\ell}\) denote the outcome for participant \(\ell\) in group \(g\). Assume that, for each group, outcomes have a finite mean \(\mu_g\) and the same variance \(\sigma^2>0\), and that observations are independent across participants and groups. In particular, the groups are disjoint and there is no additional dependence, such as clustering, that links their outcomes.

The treatment means \(\mu_i\) need not be equal to the control mean \(\mu_0\), and the treatment effects are treated as fixed quantities. The correlation in question comes from the repeated-sampling variation of the estimates. Normal outcomes are not required for this covariance and correlation calculation; independence and the stated finite, common variances suffice. Normality would matter for some inferential procedures, but it is not needed to obtain this result.

The earlier material on deriving covariance from the shared control establishes the key covariance structure: distinct independent treatment means contribute no covariance, while the control mean is shared. Here we combine that structure with the marginal variance of each contrast and simplify the resulting correlation.

Deriving the Equal-Allocation Correlation

For two different treatments \(i\) and \(j\), expand the covariance of their contrasts. The two treatment means are independent of one another and of the control mean, so their cross-covariances are zero. Only the variance of the shared control mean remains.

Each contrast also contains one treatment mean and the control mean, with opposite signs. Since those two means are independent, their variances add when the difference is formed:

The same expression holds for \(D_j\). Substituting these quantities into the definition of correlation gives the result:

The cancellation explains the invariance. Increasing \(n\) or reducing \(\sigma^2\) reduces both the shared-control covariance and each contrast variance in the same proportion, so the correlation remains \(1/2\). Changing treatment means shifts the expected contrasts but does not change their sampling covariance. The correlation is therefore a feature of the sampling structure and allocation assumptions, not of whether the treatments are effective.

What the Correlation Does—and Does Not—Say

A positive correlation means that sampling deviations in the two contrasts tend to move in the same direction. For example, if the control mean is unusually high relative to its population mean, subtracting it tends to make both contrasts unusually low. The independent treatment-group sampling errors also contribute to each contrast’s variability, which is why the correlation is positive but less than one.

The result does not say that the contrasts are identical, nor does it say that the actual treatment effects \(\mu_i-\mu_0\) and \(\mu_j-\mu_0\) vary together. Under the model, these effects are fixed parameters. It also does not give the observed correlation between a list of treatment estimates in one particular data set; one realization of the trial is not a repeated-sampling distribution.

The correlation is not determined by the shared control alone in every possible design. Equal sample sizes and a common variance are essential to the value \(1/2\). If group sizes differ, or if variances differ across groups, the covariance and marginal variances need not have the equal-allocation forms used above. Those cases require their own calculation rather than borrowing this numerical result.

Assumption Checks and Alternatives

Before using the equal-allocation result, check the design and outcome model rather than relying on the label “multi-arm.” The following questions are practical diagnostics for whether this derivation describes the comparisons at hand.

• Are the group sizes equal? Verify that control and every treatment group have the same analysis sample size. Unequal numbers invalidate the equal-allocation calculation.
• Is a common variance plausible? The derivation assumes the same outcome variance in control and treatment groups. Strongly different group variances call for a variance-specific calculation.
• Are group samples independent? Shared participants, repeated measurements treated as independent, or clustering that induces within-group dependence can change the variance structure.
• Are all contrasts oriented the same way? The result concerns \(\bar{X}_i-\bar{X}_0\) for each treatment. Reversing the sign of just one contrast reverses the sign of its correlation with the others.

A useful comparison is a design in which each treatment is compared with a distinct, independent control group of the same size. Under independent sampling across all groups, two treatment-control contrasts then share no random group mean, so their covariance is zero. The shared-control design instead introduces the common control component and, under the assumptions here, correlation \(1/2\). Treating shared-control contrasts as independent would therefore omit a genuine source of dependence.

The equal-allocation result is also limited to this particular contrast structure. It does not automatically apply to comparisons between two treatment arms, to designs with unequal allocation, or to outcomes with different variances by group. Nor does correlation alone determine every joint probability or operating characteristic: those calculations may require additional distributional assumptions and information. The result supplies one important piece of the sampling dependence, not a complete analysis of every multi-arm question.

Common Mistakes

• Assuming “same control” always means correlation \(1/2\). The shared control creates a covariance, but the numerical correlation also depends on the variances of all group means. The value \(1/2\) follows under equal group sizes and common variance.
• Calling the contrasts independent because their treatment groups are separate. The contrasts share \(\bar{X}_0\), so independence of the treatment groups does not make the contrasts independent.
• Confusing a fixed effect with a random estimate. The true effects are fixed parameters in this model. The correlation is between their estimates’ sampling errors across repeated samples.
• Reading correlation as treatment-effect similarity. A correlation of \(0.5\) describes joint sampling variation, not the size or similarity of the true effects.

A compact check is to write both contrasts explicitly and identify every random group mean they contain. For two contrasts oriented as treatment minus control, the shared term is the control mean with the same negative sign. Then verify equal group sizes, common positive variance, and independent group sampling before using the one-half result.