Reparameterizing Shared-Control Comparisons

Why Reparameterize Shared-Control Comparisons?

A set of treatment-control contrasts is often a convenient starting point: each contrast describes how one treatment arm compares with the same control arm. But a scientific question may instead concern an average across treatments, the difference between two treatment effects, or a sequence of comparisons between neighboring treatments. These are alternative linear summaries of the original contrasts. Changing the summaries also changes their covariance, so the new uncertainty must be calculated rather than carried over unchanged.

This tutorial focuses on that calculation. It does not require the transformed comparisons to be independent, nor does it require the transformation to preserve all the information in the original vector. The key is to state the transformation explicitly and apply it to both the estimates and their covariance matrix.

In a shared-control setting, a common starting vector is \(\mathbf{D}=(D_1,\ldots,D_k)^{\mathsf T}\), where \(D_i=\bar{X}_i-\bar{X}_0\) compares treatment arm \(i\) with control. If arm means are independent across randomized groups and have finite variances, then \(\operatorname{Cov}(D_i,D_j)=\operatorname{Var}(\bar{X}_0)\) for distinct treatment arms \(i\) and \(j\): the shared control mean enters both contrasts with a minus sign. The diagonal entries additionally include each treatment-arm mean variance. The transformation rule itself, however, applies to any vector with a valid covariance matrix; the shared-control structure is one way to obtain \(\boldsymbol{\Sigma}\).

Deriving the Covariance Transformation

The rule follows directly from bilinearity of covariance. Write the \(r\)th transformed comparison as \(T_r=\sum_i L_{ri}D_i\), where \(L_{ri}\) is the entry in row \(r\), column \(i\) of \(\mathbf{L}\). Then for transformed comparisons \(T_r\) and \(T_s\),

The entries of \(\mathbf{L}\) are fixed coefficients, not random quantities. In particular, the formula accounts for every covariance between original contrasts; it is not correct to transform only the diagonal variances unless the original contrasts are in fact uncorrelated. The transformed estimates themselves are obtained by the same linear operation, \(\mathbf{T}=\mathbf{L}\mathbf{D}\).

The general rule needs finite second moments and a correctly specified covariance matrix, but it does not require normality. If \(\mathbf{D}\) is jointly normal, a fixed linear transformation is also jointly normal, with mean \(\mathbf{L}E(\mathbf{D})\) and covariance \(\mathbf{L}\boldsymbol{\Sigma}\mathbf{L}^{\mathsf T}\). A singular covariance matrix is allowed; it can arise, for example, when the transformed comparisons are linearly dependent.

Changing Treatment-Control Contrasts Into Treatment Comparisons

Subtracting two treatment-control contrasts removes the shared control term algebraically: \((\bar{X}_i-\bar{X}_0)-(\bar{X}_j-\bar{X}_0)=\bar{X}_i-\bar{X}_j\). This can make a transformed contrast easier to interpret as a direct treatment-to-treatment comparison. It does not make its uncertainty disappear: its variance still depends on the covariance between the original contrasts.

Weighted Summaries and Their Uncertainty

A row of \(\mathbf{L}\) can also define a weighted summary. For instance, \(T=w_1D_1+w_2D_2\) has variance \(w_1^2\operatorname{Var}(D_1)+w_2^2\operatorname{Var}(D_2)+2w_1w_2\operatorname{Cov}(D_1,D_2)\). The cross-product term matters whenever the original contrasts are correlated, as they commonly are with a shared control.

Choosing and Checking a Reparameterization

A transformation should be chosen to match the comparisons that need interpretation, estimation, or later use. When \(\mathbf{L}\) is square and nonsingular, the transformation is invertible, so no information in the original contrast vector is lost. When \(\mathbf{L}\) has fewer rows than columns, it produces fewer summaries and generally discards information about directions not represented by those summaries. A rectangular transformation may still be exactly what is wanted when the target is a selected set of comparisons.

If a linear hypothesis is expressed in the original coordinates, it must be translated consistently when the coordinates change. For an invertible \(\mathbf{L}\), \(\mathbf{D}=\mathbf{L}^{-1}\mathbf{T}\), so the original mean vector can be recovered from the transformed one. For a dimension-reducing transformation, that recovery is generally impossible; one should not claim that the reduced vector answers every question that the original vector could answer.

Several simple checks catch common mistakes:

• Check dimensions. If \(\mathbf{D}\) has \(k\) entries and the new vector has \(m\), then \(\mathbf{L}\) must be \(m\) by \(k\), and the result must be \(m\) by \(m\).
• Check symmetry. The transformed covariance matrix must be symmetric, apart from minor rounding error.
• Check units. If each original contrast is measured in outcome units and the coefficients in \(\mathbf{L}\) are dimensionless, transformed variances and covariances are in squared outcome units.
• Check the coefficients and signs. A difference \(D_i-D_j\) uses \(+1\) and \(-1\); reversing the order changes the estimate's sign and can change its covariance with other transformed comparisons.
• Check variances. A negative diagonal entry signals an incorrect calculation or an invalid input covariance matrix, not a meaningful negative uncertainty.

A frequent error is to compute the variance of a sum or difference by adding the two variances alone. That is correct only when the covariance is zero. The correct expressions are \(\operatorname{Var}(D_i+D_j)=\operatorname{Var}(D_i)+\operatorname{Var}(D_j)+2\operatorname{Cov}(D_i,D_j)\) and \(\operatorname{Var}(D_i-D_j)=\operatorname{Var}(D_i)+\operatorname{Var}(D_j)-2\operatorname{Cov}(D_i,D_j)\). Shared-control dependence makes the omitted term especially consequential.