Joint Normality of Shared-Control Contrasts

Why Joint Normality Matters

In a multi-arm trial, each treatment is compared with the same control group. The resulting treatment-control contrasts are therefore not independent: they all include the control-group mean. When the analysis concerns several contrasts together, their joint distribution matters. For example, the probability that all contrasts fall below specified thresholds cannot generally be found by treating their marginal distributions as independent.

The key result is straightforward but important: if the vector of group means is jointly normal, then any fixed linear transformation of that vector is also jointly normal. Treatment-minus-control contrasts are such a transformation. The shared control creates off-diagonal covariance terms, but it does not prevent joint normality.

Definition: A random vector is jointly multivariate normal if every fixed linear combination of its components has a normal distribution. If the vector of group means is jointly multivariate normal, then a vector formed by fixed linear combinations of those means is also jointly multivariate normal.

Model and Contrast Vector

Consider one control group, indexed by \(0\), and \(K\) treatment groups, indexed by \(1,\ldots,K\). Let \(Y_{ij}\) be the outcome for participant \(i\) in group \(j\), and let \(\bar X_j\) be the sample mean in group \(j\), which has sample size \(n_j\). Assume that outcomes are independent between participants and groups and that, within each group, outcomes follow a normal distribution with mean \(\mu_j\). For the shared-variance model used below, assume a common variance \(\sigma^2\) in all groups, with positive sample sizes.

Under these assumptions, the group means are independent normal random variables: \(\bar X_j\sim N(\mu_j,\sigma^2/n_j)\). Consequently, the vector of group means is jointly multivariate normal. Define the treatment-control contrasts by \(D_j=\bar X_j-\bar X_0\), and collect them in \(\mathbf D=(D_1,\ldots,D_K)^\mathsf{T}\). Their expected values are \(\delta_j=\mu_j-\mu_0\).

The contrasts can be written as a matrix multiplication. Let \(\mathbf M=(\bar X_0,\bar X_1,\ldots,\bar X_K)^\mathsf{T}\), and form a matrix \(A\) whose row for treatment \(j\) has \(-1\) in the control column, \(1\) in treatment \(j\)'s column, and zeros elsewhere. Then \(\mathbf D=A\mathbf M\). This representation makes the joint distribution and covariance calculation explicit.

Theorem: If \(\mathbf M\) is jointly multivariate normal with mean vector \(\mathbf m\) and covariance matrix \(\Sigma\), then \(\mathbf D=A\mathbf M\) is jointly multivariate normal with mean \(A\mathbf m\) and covariance matrix \(A\Sigma A^\mathsf{T}\). For independent group means with common outcome variance \(\sigma^2\), the treatment-control contrasts have \(\operatorname{Var}(D_j)=\sigma^2/n_j+\sigma^2/n_0\), and, for distinct treatments \(j\ne \ell\), \(\operatorname{Cov}(D_j,D_\ell)=\sigma^2/n_0\).

Why the Result Holds

The linear-transformation result follows directly from the definition of joint multivariate normality. Take any fixed vector of coefficients \(\mathbf c=(c_1,\ldots,c_K)^\mathsf{T}\). The linear combination of contrasts \(\mathbf c^\mathsf{T}\mathbf D\) equals \(\mathbf c^\mathsf{T}A\mathbf M=(A^\mathsf{T}\mathbf c)^\mathsf{T}\mathbf M\). Because \(\mathbf M\) is jointly normal, this is a normal random variable for every fixed \(\mathbf c\). That is precisely the defining condition for \(\mathbf D\) to be jointly multivariate normal.

The mean follows by linearity of expectation: \(E(\mathbf D)=A E(\mathbf M)=A\mathbf m\). The covariance follows by applying the definition of covariance to the same transformation:

\operatorname{Cov}(\mathbf D) =E\left[(A\mathbf M-A\mathbf m)(A\mathbf M-A\mathbf m)^\mathsf{T}\right] =A\Sigma A^\mathsf{T}.

For the shared-control model, each contrast contains its own treatment mean and the same control mean. Since different group means are independent, the variance of \(D_j=\bar X_j-\bar X_0\) is the sum of the treatment-mean variance and control-mean variance. For two distinct treatments, the only random component shared by \(D_j\) and \(D_\ell\) is \(-\bar X_0\). Their covariance is therefore \(\operatorname{Var}(\bar X_0)=\sigma^2/n_0\): the two minus signs produce a positive covariance.

\operatorname{Cov}(\mathbf D) =\sigma^2\left[ \begin{pmatrix} 1/n_1 & 0 & \cdots & 0\\ 0 & 1/n_2 & \cdots & 0\\ \vdots & \vdots & \ddots & \vdots\\ 0 & 0 & \cdots & 1/n_K \end{pmatrix} +\frac{1}{n_0}\mathbf 1\mathbf 1^\mathsf{T} \right], \qquad E(\mathbf D)= \begin{pmatrix} \delta_1\\ \delta_2\\ \vdots\\ \delta_K \end{pmatrix}.

Here \(\mathbf 1\) is a \(K\)-component vector of ones. The diagonal entries include both the treatment and control contributions; the off-diagonal entries are the shared-control contribution. Thus, the covariance matrix describes the dependence as well as the marginal uncertainty of each contrast.

Worked Example: Two Treatment Arms and One Shared Control

Suppose a hypothetical trial measures an outcome in milligrams per deciliter. The control and two treatment groups have sample sizes \(n_0=20\), \(n_1=n_2=15\), and the model assumes a common standard deviation of \(\sigma=6\) mg/dL. Suppose the group means are \(\mu_0=100\), \(\mu_1=104\), and \(\mu_2=97\) mg/dL. Assume independent normal outcomes within and between groups, as required by the model.

The expected contrasts are \(\delta_1=104-100=4\) mg/dL and \(\delta_2=97-100=-3\) mg/dL. The control-mean variance is \(6^2/20=36/20=1.8\) \((\text{mg/dL})^2\). Each treatment-mean variance is \(6^2/15=36/15=2.4\) \((\text{mg/dL})^2\).

Each contrast variance is \(2.4+1.8=4.2\) \((\text{mg/dL})^2\). The covariance between the two contrasts is the shared control-mean variance, \(1.8\) \((\text{mg/dL})^2\). Therefore,

\mathbf D\sim N_2\left[ \begin{pmatrix}4\\-3\end{pmatrix}, \begin{pmatrix}4.2&1.8\\1.8&4.2\end{pmatrix} \right].

The marginal standard deviation of either contrast is \(\sqrt{4.2}\approx2.049\) mg/dL. Their correlation is \(1.8/\sqrt{4.2(4.2)}=1.8/4.2\approx0.4286\). The positive correlation reflects the fact that a lower observed control mean raises both treatment-minus-control contrasts. The joint-normal statement describes their distribution together, not just these two marginal standard deviations.

Worked Example: A Linear Combination of Three Contrasts

Consider a hypothetical trial with one control and three treatments, each with \(n_j=12\), under independent normal outcomes with common standard deviation \(\sigma=4\) units. Let the group means be \(\mu_0=50\), \(\mu_1=52\), \(\mu_2=55\), and \(\mu_3=49\) units. The assumptions give a jointly normal contrast vector with mean \((2,5,-1)^\mathsf{T}\) units. Each group-mean variance is \(4^2/12=16/12=1.3333\) units squared.

Each contrast variance is \(1.3333+1.3333=2.6667\) units squared, and every off-diagonal covariance is \(1.3333\) units squared. Thus the contrast covariance matrix has diagonal entries \(2.6667\) and off-diagonal entries \(1.3333\). Consider \(L=D_1+D_2-2D_3\), corresponding to coefficient vector \(\mathbf c=(1,1,-2)^\mathsf{T}\). Its mean is \(2+5-2(-1)=9\) units.

Using the covariance structure, its variance is

\begin{aligned} \operatorname{Var}(L) &=2.6667(1^2+1^2+(-2)^2) +2(1.3333)(1\cdot1+1\cdot(-2)+1\cdot(-2))\\ &=2.6667(6)+2.6666(-3)\approx8.000\text{ units}^2. \end{aligned}

Equivalently, the coefficients sum to zero, so the common control mean cancels: \(L=(\bar X_1-\bar X_0)+(\bar X_2-\bar X_0)-2(\bar X_3-\bar X_0)=\bar X_1+\bar X_2-2\bar X_3\). Its variance is then \((1^2+1^2+(-2)^2)(16/12)=8\) units squared. Since a linear combination of a jointly normal vector is normal, \(L\sim N(9,8)\) in the stated units, with standard deviation \(\sqrt 8\approx2.828\) units.

Marginal Normality Is Not Enough

A normal distribution for each contrast separately does not, by itself, establish that the vector of contrasts is jointly normal. The joint-normal conclusion follows from the model for the group means, or from a direct justification of joint normality; it cannot be inferred merely by checking each marginal distribution.

Worked Example: Normal Marginals Without Joint Normality

Let \(U\sim N(0,1)\), and independently let \(B\) equal \(1\) or \(-1\), each with probability \(1/2\). Define \(X=U\) and \(Y=BU\). The marginal distribution of \(X\) is standard normal. The marginal distribution of \(Y\) is also standard normal because it is an equal mixture of \(U\) and \(-U\), both standard normal.

Their covariance is \(E(XY)=E(BU^2)=E(B)E(U^2)=0\). Yet \(Y\) always equals either \(X\) or \(-X\), so the joint distribution places probability one on the union of the two lines \(Y=X\) and \(Y=-X\). A bivariate normal distribution with these zero covariances and unit marginal variances would have a nonsingular covariance matrix and a density, and so would assign probability zero to those lines. The pair is therefore not jointly normal despite having normal marginals. This is why the joint distribution must be justified from the sampling model, not assumed from separate marginal checks.

Assumptions, Diagnostics, and Limits

The exact result above relies on normal outcomes within groups and independence of observations and groups, with a common variance for the displayed covariance matrix. These conditions make the group means independent normal variables. More generally, the linear-transformation result requires that the group-mean vector itself be jointly normal; it does not require that its covariance matrix be diagonal. The stated shared-control covariance entries use the particular independent, common-variance sampling model.

In an application, check that the allocation and data structure match the assumed groups, that participants contribute independent outcomes, and that the outcome distributions are not seriously inconsistent with the normal model. Plots of within-group outcomes or model residuals can reveal strong skewness, heavy tails, or influential observations. Such checks are informative but cannot prove joint normality. Dependence between participants, repeated measurements treated as independent, or departures from the assumed variance model can make the stated covariance matrix inappropriate.

With sufficiently large group sizes, a central limit approximation may make group means approximately normal even when individual outcomes are not normal. That is an approximation, not the exact finite-sample conclusion established under normal outcomes. Likewise, knowing each contrast’s marginal normal distribution does not supply the joint probabilities needed for simultaneous events; the covariance among contrasts is essential.

The main alternative to deriving the joint distribution from the group means is to work contrast by contrast using marginal distributions. That can be adequate for a question about one prespecified contrast, but it discards dependence when the question concerns several contrasts together. A numerical simulation under a justified sampling model can also illustrate joint behavior, but it does not replace checking that model or deriving the covariance structure. These distinctions are especially important when interpreting joint or simultaneous procedures: the exact normal model is a useful foundation, but its assumptions and the shared-control dependence must remain visible.

Common mistake: “Each contrast is normally distributed, so the contrast vector is jointly normal and the contrasts are independent.” Neither conclusion follows from marginal normality. Joint normality must be established from the group-mean model or another valid joint argument, and even jointly normal contrasts are generally correlated because they share the control mean.

Check Your Understanding

Answer each question, then compare with the brief solutions.

Under the independent, common-variance model, what is the covariance between two distinct treatment-control contrasts?
Suppose the variance of each group mean is \(2\) units squared and there are two treatment arms. Give the covariance matrix of the two contrasts.
Why does joint normality of the group-mean vector imply joint normality of the contrast vector?
Is normality of every contrast considered separately sufficient to establish joint normality? Explain.

Solutions. (1) The covariance is the variance of the shared control mean, \(\sigma^2/n_0\). (2) Each contrast variance is \(2+2=4\) units squared, and their covariance is \(2\) units squared, so the matrix is \(\begin{pmatrix}4&2\\2&4\end{pmatrix}\). (3) Every fixed linear combination of contrasts is a fixed linear combination of the jointly normal group means, hence is normal; this is the definition of joint multivariate normality. (4) No. Normal marginals alone do not determine the joint distribution; the example with \(X=U\) and \(Y=BU\) has standard-normal marginals but is not jointly normal.