Heteroscedastic Shared-Control Correlation

Why Unequal Arm Variances Matter

When several treatments are compared with one control, each treatment–control contrast includes the same estimated control mean. That shared quantity creates dependence between the contrasts. If outcome variances differ across arms, however, the amount of common noise is determined specifically by the control arm’s variance and sample size, while each contrast’s total variance also includes its own treatment-arm contribution. The resulting correlation is therefore not determined by allocation ratios alone.

The earlier material on shared-control comparisons establishes why the common control mean induces covariance. Here the derivation allows each arm to have its own outcome variance. The distinction matters when treatment groups are more or less variable than the control group: using a common variance when it is not justified can misstate both the contrast variances and their correlation.

Model and Assumptions

Index the control arm by \(0\) and the treatment arms by \(j=1,\ldots,K\). In arm \(j\), let the \(n_j\) outcomes be independent observations with mean \(\mu_j\) and variance \(\sigma_j^2\). Assume the samples in distinct arms are independent, and that every arm has a finite variance. The sample sizes are fixed. The variances need not be equal: \(\sigma_0^2,\sigma_1^2,\ldots,\sigma_K^2\) may all differ.

Write \(\bar{X}_j\) for the sample mean in arm \(j\), and define the treatment–control contrast \(D_j=\bar{X}_j-\bar{X}_0\). Its expected value is \(\mu_j-\mu_0\). The covariance and correlation formulas below describe sampling variation across repetitions of this design; they do not require the arm means to be equal.

Deriving the Covariance and Correlation

For an arm with independent observations and finite variance, the variance of its sample mean is \(\operatorname{Var}(\bar{X}_j)=\sigma_j^2/n_j\). Sample means from different arms are independent under the stated model, so their covariance is zero. Expand the covariance of two contrasts, keeping the common control mean in both expressions:

The treatment means contribute no covariance because they come from independent arms. The two negative cross-covariance terms are also zero for the same reason. What remains is the variance of the shared control mean. For a single contrast, the two arm means are independent, so \(\operatorname{Var}(D_j)=\operatorname{Var}(\bar{X}_j)+\operatorname{Var}(\bar{X}_0)=\sigma_j^2/n_j+\sigma_0^2/n_0\). Dividing the covariance by the product of the contrast standard deviations gives the correlation formula.

This argument uses only independence across arms and finite variances. Normal outcome distributions are not needed to derive these covariance and correlation expressions. Normality may matter for particular distributional or inferential procedures, but it is not an assumption of this calculation.

Reading the Formula

Let \(c=\sigma_0^2/n_0\) be the variance contributed by the shared control mean, and let \(a=\sigma_j^2/n_j\) and \(b=\sigma_k^2/n_k\) be the treatment-mean variance contributions. Then the correlation has the compact form \(\rho_{jk}=c/\sqrt{(c+a)(c+b)}\), provided both contrasts have positive variance. The correlation depends on all three contributions, not just on the fact that the control is shared.

If \(\sigma_0^2>0\), the shared contribution is positive, and the correlation is positive whenever both contrast variances are positive. If \(\sigma_0^2=0\) and both contrast variances are positive, the covariance and correlation are zero. If a contrast itself has zero variance, its correlation with another variable is undefined; it is not correct to assign a correlation by dividing by a zero standard deviation.

For positive \(c\), the correlation cannot exceed one: \((c+a)(c+b)=c^2+c(a+b)+ab\ge c^2\), so \(\rho_{jk}\le 1\). If \(a\) and \(b\) are positive, increasing the common contribution \(c\) increases the correlation toward one. Indeed, \(\rho_{jk}^2=c^2/[(c+a)(c+b)]\), and its logarithmic derivative with respect to \(c\) is \(2/c-1/(c+a)-1/(c+b)>0\). As \(c\) tends to zero the correlation tends to zero; as \(c\) grows without bound it tends to one. These are limiting statements with \(a,b\) held fixed, not a claim that every change to sample size has only one effect: changing \(n_0\) changes \(c\), while changing a treatment sample size changes \(a\) or \(b\).

Checks, Alternatives, and Limitations

The equal-variance model is a simplifying alternative, not a requirement for the shared-control covariance derivation. If all arms have a common variance \(\sigma^2\), substitution gives \(\rho_{jk}=(\sigma^2/n_0)/\sqrt{(\sigma^2/n_j+\sigma^2/n_0)(\sigma^2/n_k+\sigma^2/n_0)}\). The common variance cancels, leaving a formula in the sample sizes. Under heteroscedasticity, that cancellation does not apply: the arm-specific variances must be retained. An equal-variance calculation may then overstate or understate the correlation, depending on the control and treatment variance contributions.

For a practical analysis, check that the model’s sampling structure is credible: observations should be independent across arms, and the variance used for each arm should correspond to that arm’s outcomes. Unequal arm spreads in plots or substantially different arm-specific sample variances can warn against imposing a common-variance model, though sample variation means such differences alone do not prove heteroscedasticity. Dependence between participants across arms, clustering, or other departures from independent sampling can also invalidate this simple covariance calculation.

The formulas are population quantities under the specified model. In applications, the true variances are usually unknown; substituting estimated variances creates estimated covariance and correlation values, not exact known parameters. The expressions also do not by themselves account for interim selection, missing outcomes, or other design features that change the sampling structure. Their scope is the covariance of ordinary treatment–control sample-mean contrasts under the stated independent-arm model.

A Common Mistake

A tempting but incorrect argument is: “Both contrasts use the same control group, so their correlation must be positive.” The missing condition is that the control mean must have nonzero variance. If \(\sigma_0^2=0\) while both contrast variances are positive, then \(\operatorname{Cov}(D_j,D_k)=0\) and \(\operatorname{Corr}(D_j,D_k)=0\). The corrected statement is that the correlation is positive when \(\sigma_0^2>0\) and both contrast variances are positive. Always calculate the shared contribution \(\sigma_0^2/n_0\) before drawing a conclusion about the sign or size of the correlation.