A Workflow for Shared-Control Dependence Calculations

Why Use a Calculation Workflow?

In a multi-arm trial, several treatment-versus-control comparisons reuse the same control group. Their estimates are therefore dependent, even when outcomes from different participants are independent. The dependence can be calculated systematically from the group sizes, group variances, and contrast definitions. A reliable workflow is useful because sign errors, mismatched group order, and confusion between covariance and correlation can all produce plausible-looking but incorrect results.

This tutorial develops a calculation sequence: define the contrasts, calculate the variance of each group mean, assemble the contrast covariance matrix, standardize its entries to obtain correlations, and check the result. The calculations concern the sampling variation of estimates; the expected treatment effects do not determine their covariance.

Assumptions and the Calculation Formula

For the basic calculation, assume that observations are independent within and across groups, each group has finite outcome variance \(\sigma_g^2\), and group \(g\) has \(n_g\) observations. Then \(\operatorname{Var}(\bar X_g)=\sigma_g^2/n_g\), and the group means are mutually uncorrelated. A normal outcome model is not required for this covariance calculation; normality may be added when a joint normal model for inference is needed. The arms need not have equal sample sizes or equal variances.

Let \(D\) be the diagonal matrix whose \(g\)th diagonal entry is \(\sigma_g^2/n_g\). Put the contrast coefficient rows into a matrix \(A\). For the vector of contrasts \(C=A\bar X\), where \(\bar X=(\bar X_0,\ldots,\bar X_K)^\mathsf{T}\), the covariance matrix is \(ADA^\mathsf{T}\). In particular, the covariance between contrasts \(j\) and \(\ell\) is a weighted sum over the groups they both involve.

If the variances are estimated from data, replace \(\sigma_g^2\) by the corresponding within-group sample variance \(s_g^2\). The resulting covariance matrix is a plug-in estimate, rather than the exact covariance matrix based on known population variances.

Deriving the Formula and Applying the Workflow

The formula follows by expanding the covariance of two linear combinations. Covariance is bilinear, so each pair of group means contributes the product of its two coefficients times the covariance of those means. Under the independence assumption, all terms for distinct groups are zero, leaving only the terms from the same group.

For treatment-minus-control contrasts, \(C_j=\bar X_j-\bar X_0\). Each contrast has coefficient \(-1\) for control, \(+1\) for its treatment, and zero for the other treatment arms. Thus, for two distinct treatment-minus-control contrasts, the only nonzero shared contribution is the control term, \(\sigma_0^2/n_0\). The workflow below also works when contrast directions differ or coefficients are more general.

Constructing and Checking a Full Matrix

With unequal group variances and sizes, the shared-control covariance is still obtained from the control contribution, but the variances of the individual contrasts differ. These differences affect the correlations: a common covariance does not imply a common correlation.

Direction, Units, and Common Failure Modes

A covariance’s sign depends on the signs of the coefficients, not simply on whether two contrasts use the same participants. Reversing one contrast changes the signs of its covariances with all other contrasts but does not change its variance. Correlation changes sign in the same way, while remaining dimensionless.

Several checks catch common calculation errors. Confirm that all coefficient rows use the same group order and sum to zero; check that each diagonal entry is calculated as a variance, using squared coefficients; and ensure the covariance matrix is symmetric. The absolute value of every correlation must be at most one, since \(|\operatorname{Cov}(C_j,C_\ell)|\leq\sqrt{\operatorname{Var}(C_j)\operatorname{Var}(C_\ell)}\). A negative variance or a correlation outside \([-1,1]\) indicates an error or an invalid covariance estimate.

A common mistake is to treat every off-diagonal covariance as positive merely because two comparisons use a shared control. That is true for treatment-minus-control contrasts with the same orientation, but not in general: the shared group’s contribution is the product \(a_{j0}a_{\ell0}\sigma_0^2/n_0\), which is negative if the control coefficients have opposite signs. Another mistake is to report a covariance as a correlation. Covariance has squared outcome units; correlation divides by the two contrast standard deviations and has no units.

Choosing and Interpreting the Calculation

The coefficient-matrix method is efficient when there are many contrasts: construct \(A\), form the diagonal group-mean covariance matrix \(D\), and calculate \(ADA^\mathsf{T}\). The pairwise formula is often easier to audit by hand because it shows exactly which shared groups contribute. A direct simulation or empirical estimate can be useful as a check under a specified data-generating model, but it does not replace verifying that the contrast definitions and assumptions are appropriate.

The calculation relies on the stated group variances and independence of group means. If participants are clustered, observations are repeated, or groups are otherwise dependent, \(D\) is not diagonal and cross-group covariance terms must be included. If variances are estimated, plug-in correlations also carry estimation uncertainty. In that setting, the displayed numerical correlation is an estimate, not a fixed population quantity. Finally, this workflow quantifies dependence among the specified contrasts; it does not by itself define how any contrast will be used in a later decision rule.