Why Control-Group Structure Matters
When several treatments are compared with a control, using one control group or a separate control group for each treatment changes more than the study’s logistics. With a shared group, every treatment-control estimate contains the same observed control mean. This creates dependence between estimates. With independent control groups, the comparisons do not share that sample source, but the number of control observations assigned to each comparison may differ.
The choice should therefore be assessed using both the marginal standard errors and the full covariance matrix. The comparison also depends on what is held fixed: the number of controls available to each comparison, or the total number of control participants. These are different design questions.
Model and Covariance Matrices
Suppose there are \(K\) treatment groups, indexed by \(i=1,\ldots,K\), and a control group indexed by 0. Write \(\bar{X}_i\) for the sample mean in treatment group \(i\), and \(\bar{X}_0\) for the control sample mean. The treatment-control estimate is \(D_i=\bar{X}_i-\bar{X}_0\) when one control group is shared. Let group \(i\) have sample size \(n_i\) and outcome variance \(\sigma_i^2\); the shared control has size \(n_0\) and variance \(\sigma_0^2\).
Assume observations are independent within and between groups, observations within each group have a common finite outcome variance, and the sample means are calculated from the stated group observations. Equal variances across groups are not required. The covariance results below follow from independence and the variance of a sample mean; normality is not needed for them.
The earlier material on covariance of independent arm means and deriving covariance from the shared control gives the key calculation. In this setting, each \(D_i\) contains its own treatment mean and the same control mean with a minus sign. The two minus signs in a covariance multiply to a positive contribution from the shared control.
The first three terms are zero because the treatment samples and control sample are independent, and distinct treatment samples are independent. The remaining term is the variance of the shared control mean. For separate control groups, every sample mean used in one comparison is independent of the sample means used in another, so the covariance between comparisons is zero.
Worked Covariance-Matrix Comparisons
Worked Example: Same Marginal Variances, Different Dependence
Consider two treatment groups with \(\sigma_1^2=24\), \(n_1=5\), \(\sigma_2^2=8.1\), and \(n_2=5\). The control outcome variance is \(\sigma_0^2=10\), and a control sample of size \(n_0=4\) is used. Compare a shared control with two independent control groups, each of size 4 and with variance 10.
The treatment-mean variance components are \(24/5=4.80\) and \(8.1/5=1.62\). The control-mean variance component is \(10/4=2.50\). With a shared control, the two variances are \(4.80+2.50=7.30\) and \(1.62+2.50=4.12\), while their covariance is 2.50. Hence
The standard errors are \(\sqrt{7.30}=2.702\) and \(\sqrt{4.12}=2.030\), in the outcome’s units. The correlation is
Thus the shared-control estimates are positively correlated, with correlation about 0.456. With separate controls of the same sizes and variances, the marginal variances remain 7.30 and 4.12, but the covariance matrix is \(\begin{pmatrix}7.30&0\\0&4.12\end{pmatrix}\). This separate-control design uses eight control observations rather than four.
Worked Example: Comparing Designs With a Fixed Total Control Sample
Suppose the total number of control observations is fixed at 8. Let \(\sigma_0^2=16\), with treatment-mean variance components \(\sigma_1^2/n_1=4\) and \(\sigma_2^2/n_2=3\). Compare one shared control group of size 8 with separate control groups of size 4 each. Assume independent observations and the same control variance in either design.
For the shared design, the control-mean variance is \(16/8=2\). The covariance matrix and standard errors are
For the separate design, each control mean has variance \(16/4=4\). The comparison variances are \(4+4=8\) and \(3+4=7\), and the off-diagonal covariance is zero:
Under this allocation and variance model, pooling the same total of eight controls gives a smaller standard error for each individual treatment-control estimate. It also induces covariance between the estimates. This is a comparison of marginal precision, not a claim that the shared design is better for every possible analysis.
Worked Example: Holding Per-Comparison Control Size Fixed
Now let the control variance be \(\sigma_0^2=9\), and use three control observations per comparison. The treatment-mean variance components are 1 and 4. Compare one shared control group of size 3 with separate control groups of size 3 each. Assume all groups are independent except for the deliberate reuse of the shared group.
For the shared design, the control-mean variance is \(9/3=3\). The covariance matrix is
The standard errors are \(\sqrt{4}=2.000\) and \(\sqrt{7}=2.646\). In the separate design, each comparison retains the same control sample size and hence the same marginal variance components, but it requires six control observations in total. Its covariance matrix is \(\begin{pmatrix}4&0\\0&7\end{pmatrix}\), with the same two standard errors. The distinction here is dependence and total control use, not marginal precision.
What the Comparison Does—and Does Not—Establish
The examples show why a statement such as “shared controls are more precise” needs a design condition. If a fixed total of control observations is divided among separate groups, a shared group can provide more control observations to each contrast and thereby reduce each marginal variance. If each comparison instead gets the same number of control observations under both designs, marginal variances can match while the shared design has nonzero covariance and uses fewer controls overall.
For a shared group, the standard error for comparison \(i\) is \(\sqrt{\sigma_i^2/n_i+\sigma_0^2/n_0}\). For separate controls, it is \(\sqrt{\sigma_i^2/n_i+\sigma_{0i}^2/n_{0i}}\). These formulas make the allocation effect explicit. A covariance matrix is not summarized by its diagonal alone: off-diagonal entries matter whenever more than one comparison is considered jointly.
Under the assumptions above, shared-control covariances are nonnegative because all treatment-control estimates subtract the same control mean, and are strictly positive when the shared control variance is positive. The precise correlation depends on the variance components as well as sample sizes; unequal treatment variances or allocation can produce different correlations across comparisons. If group observations are not independent, or if the control data are reused in a more complicated way than described, the stated matrices need not apply. Likewise, an incorrect sample size or variance assumption changes the standard errors and covariances.
A common mistake is to treat shared-control comparisons as independent because the treatment groups themselves are independent. That reasoning omits the control mean: the estimates \(D_i\) and \(D_j\) both include \(-\bar{X}_0\), so their covariance includes \(\operatorname{Var}(\bar{X}_0)\). The correction is to identify every sample mean used in each estimate before applying independence. Another mistake is to compare standard errors without stating how many controls each design uses. Specify whether the comparison fixes the total control sample size or the per-comparison control sample size.
These calculations describe sampling covariance under the stated group-sampling model. They do not by themselves determine which design is preferable: that depends on the design objective, resource constraints, and how the collection of comparisons will be used. Nor does a zero covariance under independent controls mean that all features of a study are otherwise equivalent.
Record which treatment and control sample means enter each treatment-control estimate.
For each group, divide its outcome variance by its sample size to obtain the variance of its sample mean.
Add the relevant components on the diagonal; include the shared control-mean variance off the diagonal only when that control sample is reused.
Calculate standard errors from the diagonal and state whether total controls or controls per comparison are being held fixed.
Check Your Understanding
Try each question, then check the concise solutions below.
- Two treatment means have variance components 2 and 5. A shared control mean has variance 1. Write the covariance matrix and give both standard errors.
- For the matrix in Question 1, calculate the correlation between the estimates.
- Why does an independent-control design have zero covariance between comparisons under the model described here?
- With control variance 12 and a total of 6 controls, compare the control-mean variance for one shared group of size 6 and for two separate groups of size 3.
Solutions. (1) The diagonal entries are \(2+1=3\) and \(5+1=6\); the shared-control covariance is 1. The matrix is \(\begin{pmatrix}3&1\\1&6\end{pmatrix}\), and the standard errors are \(\sqrt{3}=1.732\) and \(\sqrt{6}=2.449\). (2) The correlation is \(1/(\sqrt{3}\sqrt{6})=1/\sqrt{18}\approx0.2357\). (3) Each comparison uses its own treatment sample and its own independent control sample, so no sample mean is shared between the comparisons; independence makes their covariance zero. (4) The shared control-mean variance is \(12/6=2\). Each separate control-mean variance is \(12/3=4\). The separate design has the same total control count but allocates fewer observations to each comparison’s control mean.