Shared-Control Dependence in Intermediate Outcomes

Why a Common Control Creates Dependence

In a multi-arm, multi-stage trial, several experimental arms may be compared with the same control group at an interim look. Even if participants in different randomized arms are independent, the estimated treatment–control contrasts are not: each contrast contains the same estimated control mean. A chance fluctuation in that mean moves all contrasts in the same direction.

This matters when intermediate outcomes guide which arms continue. Selection rules act on several contrasts at once, so their joint distribution—not just each contrast’s standard error—determines the chance that one or more arms cross a boundary or that multiple arms appear promising. The focus here is the dependence caused by the shared control estimate, not the relationship between an intermediate outcome and a later endpoint.

Model and Covariance Derivation

For the basic derivation, assume participants are independently randomized to one of the arms, observations are independent between participants, and each arm’s sample mean has finite variance. Write the sample sizes as \(n_i\), and the outcome variances as \(\sigma_i^2\), for \(i=0,1,\ldots,K\). Thus \(\operatorname{Var}(\bar{X}_i)=\sigma_i^2/n_i\). The key assumption is independence of the arm-specific sample means across distinct randomized groups. The derivation does not require equal sample sizes or equal outcome variances.

For distinct experimental arms \(i\) and \(j\), expand the covariance of the two contrasts. The treatment-arm means are independent of each other and of the control mean, so the only nonzero term is the variance of the shared control mean:

Each contrast has variance equal to the sum of the variances of its treatment and control means. Dividing the covariance by the product of the contrast standard deviations gives their correlation:

When the variances are known, or treated as fixed for planning, the same correlation applies to the standardized contrasts \(Z_i=D_i/SE_i\), where \(SE_i=\sqrt{\sigma_i^2/n_i+\sigma_0^2/n_0}\). With estimated variances, the formula is a plug-in approximation; it does not account for additional randomness or dependence introduced by a particular variance-estimation procedure.

The covariance is positive whenever the control variance is positive. If a control mean happens to be lower than its expected value, every treatment-minus-control contrast tends to be larger. The degree of dependence is not determined by the control alone: it also depends on how much treatment-arm variability contributes to each contrast. As \(n_0\) grows, the control-mean variance shrinks and the induced correlation generally decreases. For equal arm sizes and equal variances, the correlation between any two contrasts is \(1/2\): the shared control contributes half of each contrast’s variance.

How Dependence Changes Selection

A selection rule might retain an arm when its intermediate-outcome statistic exceeds a prespecified threshold, or retain the arm with the largest statistic. In either case, dependence affects the joint chance of the possible selection outcomes. For example, it affects the probability that at least one arm crosses a boundary and the probability that two arms cross it together. Calculating each arm’s marginal crossing probability does not, by itself, give these joint probabilities.

If the vector of standardized contrasts is jointly normal, its covariance matrix has ones on the diagonal and the shared-control correlations off the diagonal. Under that model, the conditional distribution of one standardized contrast given another is also normal. For two standardized contrasts with correlation \(\rho\), unit variances, and means \(\mu_1,\mu_2\), the conditional mean and variance are:

Thus, observing an unusually high value in one arm tends to make a high value in another arm more plausible when \(\rho>0\). This is a statement about the joint sampling distribution, not evidence that one treatment causes the other arm’s result.

Arm Ranking and Threshold Decisions Are Different

A shared control affects both contrasts, but it affects their difference in a special way. Subtracting one contrast from another cancels the common control mean:

Consequently, the ordering of the raw contrasts is exactly the ordering of the experimental-arm means. Under independent arm samples, the variance of their difference is \(\sigma_i^2/n_i+\sigma_j^2/n_j\), with no control-mean variance component. This cancellation does not mean that the shared control can be ignored for all selection decisions. A rule that compares each contrast with a threshold still depends on the control estimate, and a rule that uses standardized rather than raw contrasts need not rank arms in the same way as their raw means when standard errors differ.

Practical Calculation, Alternatives, and Failure Modes

For planning or analysis, construct the joint covariance matrix rather than supplying only separate standard errors. Its diagonal entries are the variances of the contrasts, and its off-diagonal entries are \(\sigma_0^2/n_0\) under the independent-arm model. A joint-normal approximation can then be used to evaluate selection probabilities, with the approximation checked against the design and outcome distribution where appropriate.

The standard formula assumes independent randomized groups and correctly specified variances. If observations within an arm are clustered, repeated, or otherwise dependent, \(\sigma_i^2/n_i\) may not be the correct variance for that arm’s mean. Use the variance of the actual arm-specific estimator and account for any cross-arm dependence in the design or analysis; the shared-control formula alone is not a substitute for that covariance structure. Missing outcomes, unequal follow-up, or model-based adjusted means can likewise change the relevant variances and covariances.

A practical alternative is to estimate the full covariance matrix from a model or by simulation that reproduces the randomization, outcome model, and interim analysis. This can be preferable when the simple independent-means model is implausible. By contrast, treating each comparison as if it had its own independent control is appropriate only when the design truly supplies independent control groups. It is not a harmless computational shortcut for a shared-control design.

Common diagnostics include checking that the covariance matrix is symmetric and positive semidefinite, that its diagonal agrees with the reported contrast variances, and that its off-diagonal entries reflect the shared control. A simulation under the assumed model can compare empirical correlations and selection frequencies with the analytic calculations. Large discrepancies suggest a coding error or that the model assumptions do not represent the implemented design.

The central limitation is that this derivation addresses dependence arising from a common control mean under a specified mean-variance model. It does not establish that an intermediate outcome predicts the final endpoint, nor does a high correlation between selection statistics establish that the outcome is a valid surrogate. Those are separate scientific and statistical questions.