Shared-Control Efficiency for Continuous Outcomes

Why Compare Standard Errors?

For a continuous outcome, a treatment effect is often summarised as the difference between the treatment-group and control-group means. The standard error describes the sampling variability of that estimated difference: smaller standard error means greater precision, all else equal. In a multi-arm trial, each treatment can be compared with one common control group, or each treatment can be studied in a separate two-arm trial with its own control group. The variance formula makes the recruitment trade-off explicit.

This comparison is meaningful only when the designs target the same treatment-control contrasts and use compatible assumptions about outcome variability. A shared control can avoid recruiting repeated control groups, but does not make the control observations available in unlimited quantity: the precision of every contrast still depends on the size and variability of the common control group.

Variance of a Continuous Treatment-Control Difference

The formula assumes equal population variances in the two groups. The variance calculation itself does not require normally distributed outcomes. Normality is needed if we want the difference in means to have an exact normal sampling distribution at finite sample sizes; with sufficiently large samples, a normal approximation may be reasonable under standard conditions. Random allocation supports the comparison, while independence of participants and a consistent outcome measurement process are also important.

The standard error formula follows from two elementary facts: the variance of a sample mean of independent observations with variance \(\sigma^2\) is \(\sigma^2/n\), and the variance of the difference of independent random variables is the sum of their variances. In particular,

If the group variances are not assumed equal, the corresponding variance is \(\sigma_T^2/n_T+\sigma_C^2/n_C\). In an analysis, the unknown variances are generally estimated. A common estimated standard error that does not require equal variances is \(SE=\sqrt{s_T^2/n_T+s_C^2/n_C}\), where \(s_T\) and \(s_C\) are the sample standard deviations. In a design calculation, \(\sigma\) is often a planning value based on relevant prior information or a stated assumption.

Shared Control Versus Separate Controls

Consider \(K\) treatment groups, each compared with a control group. If all treatment groups have \(n_T\) participants and the shared control has \(n_C\), the standard error for each treatment-control contrast is \(\sigma\sqrt{1/n_T+1/n_C}\). If instead each treatment has its own separate control group of size \(n_C\), the standard error of each contrast is the same, provided the outcome variance and allocation sizes are the same. But the separate-trial approach recruits \(K n_C\) control participants rather than \(n_C\).

Thus, at matched per-comparison group sizes and under these assumptions, using one common control avoids recruiting \((K-1)n_C\) additional control participants. This arithmetic compares the recruitment needed for the same precision in each contrast; it does not establish that sharing a control is always the best design. The control and treatment groups must be sufficiently comparable for their mean differences to answer the intended scientific questions.

Comparing Designs at the Same Total Recruitment

The same total number of participants can be distributed differently across designs, so a recruitment comparison should state what is held fixed. For example, suppose there are two treatments, the total recruitment is 120, and allocation is equal across the three groups in the shared-control design. That gives 40 participants in each treatment group and 40 in the common control. In two separate trials with the same total, equal allocation within each trial gives 30 participants per group in each trial.

Using Sample Standard Deviations

Once data have been collected, the population standard deviations are unknown. If equal variances are not assumed, the estimated standard error for a treatment-control mean difference is \(\sqrt{s_T^2/n_T+s_C^2/n_C}\). For example, suppose a hypothetical study has 25 participants per group, a treatment sample standard deviation of 8 points, and a control sample standard deviation of 7 points. The estimated standard error is

This estimate uses the observed variability in each group. It is distinct from a planning standard error, which uses assumed or previously estimated population variability to anticipate precision before recruitment. If observations are clustered, repeatedly measured, or otherwise dependent, the simple formula may not apply: the variance calculation must reflect that dependence.

Diagnostics, Limitations, and Common Pitfalls

Before relying on a continuous-outcome standard error calculation, check that the intended groups and sample sizes are correctly identified, that the outcome is measured on the same scale, and that the variability assumption is plausible. Extreme skewness, outliers, unequal variances, missing outcomes, or clustering can make a simple planning model a poor description of the data. With unequal variances, use group-specific variances rather than forcing a common \(\sigma^2\); with dependence, use a model that accounts for it.

A common mistake is to count each treatment-control comparison as if it required a distinct control group, even when the design uses one shared control. That overstates recruitment for the shared-control design. The opposite mistake is to count the common control only once in the recruitment total but then act as though each contrast had a different, independent control mean. Each individual contrast uses the same control observations, so the contrasts share sampling variation.

Another pitfall is to claim that shared control always improves precision. At the same treatment and control group sizes, the standard error of an individual comparison is unchanged whether that control group is dedicated or shared. The recruitment saving occurs because the shared design reuses the same control participants across comparisons. If total recruitment is fixed, the allocation determines each contrast's standard error; the designs must be compared using the actual group sizes.

A shared control is most defensible when the treatment groups are studied in a sufficiently comparable setting and period, use compatible eligibility criteria and outcome measurement, and the control condition is suitable for every comparison. Separate trials may be preferable if those conditions do not hold, or if distinct control conditions are scientifically necessary. These considerations are not captured by the variance formula alone.