Why Correlation Diagnostics Matter
A shared-control correlation is not a free input: it follows from the contrasts, allocation, and variance assumptions. An incorrect sign or magnitude can distort joint calculations even when each contrast’s standard error looks plausible. A disciplined diagnosis asks two different questions: Is the value mathematically possible? and Is it credible under the stated design and model?
The checks here concern the dependence among estimated treatment-control contrasts, not the correlation between individual outcomes in different arms. They assume the usual model in which participants in distinct arms are independent and each contrast uses the same control group. If the design or analysis departs from those assumptions, the corresponding covariance needs to be justified from that model rather than presumed.
Reference Model and Basic Checks
These expressions require independent samples across arms, finite outcome variances, and a control group that is genuinely shared by the contrasts. The variance expressions also assume the stated arm-specific variances and sample sizes correctly describe the sampling variation of the means. They need modification if, for example, observations are clustered or participants overlap across arms.
The covariance formula follows by expanding the covariance and using independence of the arm means:
The same control mean is subtracted in both contrasts, so its sampling error moves them together. If one contrast is instead control-minus-treatment, its sign is reversed and its covariance with a treatment-minus-control contrast is \(-v_0\). More generally, multiplying contrast \(i\) by a direction sign \(s_i\in\{-1,1\}\) multiplies its covariance with contrast \(j\) by \(s_i s_j\). This sign check is often the fastest way to find a coding or contrast-orientation error.
For nondegenerate arms with positive variances, \(v_0>0\), so two contrasts oriented treatment-minus-control have positive correlation, less than 1. The covariance matrix must also be positive semidefinite: for every vector \(a\), \(a^\mathsf{T}\Sigma a\geq 0\), where \(\Sigma\) is the covariance matrix. Equivalently, no linear combination of the contrast estimates can have negative variance. Correlation values must lie between \(-1\) and \(1\), but that pairwise condition alone does not guarantee that a whole matrix is valid.
Worked Diagnostics
Worked Example: A Sign and Magnitude Check
Suppose a hypothetical trial plans a shared control group of \(n_0=40\) participants and two treatment groups of \(n_1=20\) and \(n_2=30\). Assume independent participants across arms and a common outcome variance \(\sigma^2=100\) square units in all three arms. Both contrasts are treatment-minus-control. A spreadsheet reports their correlation as \(-0.38\). Check it.
The common-variance model gives \(v_0=100/40=2.5\), \(v_1=100/20=5\), and \(v_2=100/30=3.333\) square units. Therefore,
The reported negative sign conflicts with the contrast directions, and the calculation gives approximately \(0.378\), not \(-0.38\). A likely explanation is a reversed contrast sign or a sign error in the covariance. This check does not depend on the observed treatment effects: effect estimates affect the observed contrast values, not this model-based correlation.
Worked Example: Auditing Unequal Variance Inputs
Consider a different hypothetical design with \(n_0=25\), \(n_1=20\), and \(n_2=40\). The assumed outcome variances are \(\sigma_0^2=64\), \(\sigma_1^2=100\), and \(\sigma_2^2=36\) square units. The treatment-minus-control contrasts share the same control sample. Calculate their correlation and assess a reported value of \(0.70\).
The arm-mean variances are \(v_0=64/25=2.56\), \(v_1=100/20=5\), and \(v_2=36/40=0.90\) square units. Hence the contrast variances are \(2.56+5=7.56\) and \(2.56+0.90=3.46\) square units; their covariance is \(2.56\) square units.
Thus \(0.70\) is not the correlation implied by these inputs. It is within the generic range from \(-1\) to \(1\), and has the expected positive sign, but it is still suspicious under the specified model. Replacing the arm-specific variances with a common-variance shortcut, or using the wrong sample size in \(v_0\), could explain the discrepancy. The audit should trace each variance term to its arm and sample size before treating the reported value as usable.
Worked Example: Pairwise Values That Cannot Form One Matrix
Suppose three hypothetical, standardized contrast estimates have reported pairwise correlations \(r_{12}=0.8\), \(r_{13}=0.8\), and \(r_{23}=-0.8\). Each reported value individually lies in the allowed range. Determine whether all three can belong to one joint correlation matrix.
The proposed matrix is
For a three-variable correlation matrix, positive semidefiniteness requires a nonnegative determinant. Expanding the determinant for diagonal entries equal to 1 gives \(1+2r_{12}r_{13}r_{23}-r_{12}^2-r_{13}^2-r_{23}^2\). Substitution yields
The determinant is negative, so \(R\) is not positive semidefinite and cannot be a correlation matrix. The problem is joint inconsistency, not an individual pair outside the usual range. This pattern also cannot be explained by simply reversing the direction of one contrast: such a reversal changes two pairwise signs, while preserving the matrix’s validity.
Locating the Source of a Suspicious Value
A value can be mathematically possible but inconsistent with the assumed shared-control model. For the treatment-minus-control orientation, compare the reported value with the formula using the stated \(v_0,v_i,v_j\). In particular, the covariance is the variance of the shared control mean, not the variance of an individual outcome and not a correlation between individual outcomes. The denominator uses both full contrast variances.
Allocation errors often leave a recognizable trace. The control sample size determines \(v_0=\sigma_0^2/n_0\), which appears in both the covariance and the two contrast variances. The treatment sample sizes determine \(v_i\) and \(v_j\). Increasing the control sample size reduces \(v_0\); under fixed positive treatment-arm variances, the shared-control correlation therefore decreases. A claimed value that changes in the opposite direction warrants a check of how the allocation was entered. Verify that each sample size belongs to the correct arm and that the control is counted once as a shared group, not as separate independent control samples.
Variance assumptions matter too. Under a common-variance assumption, the same \(\sigma^2\) is used for each arm. Under unequal variances, each \(v_g\) must use that arm’s variance. A correlation computed with a common variance can be a poor match if the design or analysis instead assumes substantially different arm variances. Conversely, unequal variance inputs should not be mixed with a formula derived under equal variances without checking that the formula still applies.
A useful distinction is between an exact model-based covariance and an estimated covariance. Plugging estimated variances into the formula gives an estimated correlation; it can differ from the value based on planning assumptions. Separately estimated pairwise correlations may also fail to produce a positive-semidefinite matrix, especially with limited data. That failure signals incompatibility among the estimates as a joint matrix; it does not establish that any single pairwise estimate is necessarily wrong. Rounding can produce tiny apparent violations, so compare unrounded values before diagnosing a substantive problem.
Comparisons, Diagnostics, and Limits
| Approach | What it checks | Main caution |
|---|---|---|
| Model-based calculation | Whether the correlation follows from arm variances, sample sizes, contrast directions, and shared-control sampling variation. | Its answer depends on the model assumptions and inputs being appropriate. |
| Randomization-based dependence assessment | Whether dependence under the randomization scheme agrees with the design-based analysis. | It need not match a calculation based on independent-arm sampling assumptions. |
| Empirical pairwise estimates | Whether observed joint variation supports particular pairwise correlations. | Pairwise estimates can be noisy and may not form a valid joint matrix. |
These approaches answer related but not identical questions. A discrepancy between a model-based value and a randomization-based or empirical assessment should prompt a review of the estimand, design, and assumptions—not automatic replacement of one value with another. In particular, the standard formula is not justified merely because a study has a control group used in more than one comparison. Dependence from clustering, repeated measurements, covariate adjustment, or other design features may require a different covariance calculation.
Common warning signs include the wrong sign for the contrast orientation, a correlation outside \([-1,1]\), a value inconsistent with the variance-and-allocation formula, or a correlation matrix that is not positive semidefinite. A plausible-looking value is not proof of a correct calculation. Check the contrast definitions and the arm-level variance inputs, then verify the covariance matrix as a whole. These are diagnostics, not a substitute for specifying the correct probability model.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- Under a common-variance model, \(n_0=50\), \(n_1=25\), and \(n_2=50\). What correlation is implied for two treatment-minus-control contrasts?
- What covariance sign is expected if one contrast is treatment-minus-control and the other is control-minus-treatment?
- Why is it not enough to check that every reported pairwise correlation lies between \(-1\) and \(1\)?
- Name one design feature that can make the standard shared-control covariance formula inapplicable without modification.
Solutions. (1) Set \(\sigma^2=1\), since its scale cancels: \(v_0=1/50=0.02\), \(v_1=1/25=0.04\), and \(v_2=1/50=0.02\). The correlation is \(0.02/\sqrt{(0.06)(0.04)}\approx 0.408\). (2) The covariance is negative: reversing one contrast direction multiplies the shared-control covariance by \(-1\). (3) Pairwise-valid values can still form a matrix with a negative eigenvalue or determinant, violating positive semidefiniteness and implying a negative variance for some linear combination. (4) Clustering, overlapping participants, or repeated measurements can induce dependence not represented by independent arm means; the covariance must then reflect that structure.