Why a Shared Control Matters
In a multiarm trial, several experimental treatments may be compared with the same standard-of-care group. This can make better use of participants assigned to control: a control participant may contribute information to more than one treatment comparison. The gain is not free, however. The comparisons now share data, so their estimates are statistically dependent, and conclusions across the set of treatments require attention to multiplicity.
STAMPEDE is a multiarm, multistage platform trial evaluating treatment strategies for prostate cancer. Its platform structure allowed experimental treatment comparisons to be added or closed over time, while standard of care provided a common comparator where protocol eligibility and timing made the comparison appropriate. A shared control does not mean that control participants received several treatments, nor that every control participant in the platform automatically belongs in every comparison. It means that the same eligible control observations can contribute to more than one specified treatment-versus-control contrast.
The mathematical results below use a deliberately simple setting to make the dependence visible. Assume that each participant contributes one outcome, participants are assigned to only one arm, and outcomes from different participants are independent. Let the outcome variance in arm \(j\) be \(\sigma_j^2\), and in control be \(\sigma_0^2\). Assume the treatment and control samples are independent, and that the same \(n_0\) control participants are used in each contrast. These assumptions give the basic variance formulas; actual trial analyses may instead use time-to-event, binary, covariate-adjusted, or stratified models, for which the relevant covariance must be obtained from that model and design.
Precision and Dependence
For a treatment-versus-control comparison, independence between the treatment and control samples gives
The control contribution, \(\sigma_0^2/n_0\), appears in every comparison that uses those controls. For two distinct treatment arms \(j\) and \(k\), assume their participant samples are disjoint and independent. The covariance is
The covariance is strictly positive when the control outcome variance is positive; if control outcomes are constant, it is zero. The shared control mean shifts both treatment-minus-control estimates in the same direction: a higher-than-expected control mean lowers both estimates, while a lower-than-expected mean raises both. This is the source of their nonnegative covariance under the assumptions above.
For equal treatment-arm variances and equal treatment sample sizes, the correlation between two estimates is the control variance contribution divided by the total variance of either estimate. More generally, it is \((\sigma_0^2/n_0)/\sqrt{\operatorname{Var}(\hat{\Delta}_j)\operatorname{Var}(\hat{\Delta}_k)}\). The amount of dependence therefore depends on the control sample size and variability as well as the treatment-arm information.
Worked Example: Calculating Shared-Control Standard Errors
Suppose, purely hypothetically, that a continuous outcome is lower when a patient does better. A shared control group has \(n_0=100\) participants, mean 10.0 outcome units, and standard deviation 4.0 units. Treatment A has \(n_A=80\), mean 8.8, and standard deviation 4.0; treatment B has \(n_B=80\), mean 9.2, and standard deviation 4.0. For this calculation, treat the common outcome variance as 16 square units.
The estimated contrasts are \(\hat{\Delta}_A=8.8-10.0=-1.2\) units and \(\hat{\Delta}_B=9.2-10.0=-0.8\) units. Negative values favor treatment under the stated outcome direction. The variance estimate for either contrast is
Thus each standard error is \(\sqrt{0.36}=0.600\) units. The covariance between the two estimates is \(16/100=0.160\) square units, and their correlation is \(0.160/\sqrt{0.36\times0.36}=0.444\). The estimates are positively correlated because of their shared variable control mean. This is an illustrative calculation, not a result from STAMPEDE.
What Efficiency Does—and Does Not—Mean
Sharing controls can improve precision relative to dividing a fixed control sample among separate comparisons. It does not necessarily mean that a platform trial needs fewer participants under every design or achieves more power for every arm. The result depends on allocation, outcome variability, recruitment over time, and the number and timing of comparisons. In particular, adding experimental arms without increasing control enrollment can reduce the amount of control information available per comparison.
Worked Example: Reusing Versus Splitting the Same Controls
Continue with a hypothetical setting in which each arm has 80 participants and all three groups have outcome variance 16 square units. Compare two designs for two treatment comparisons. Design 1 uses one shared control group of 100 participants. Design 2 splits the same total of 100 control participants into two separate groups of 50, one for each comparison. Assume no other design differences.
For each comparison in Design 1, the variance is \(16/80+16/100=0.36\), so the standard error is \(0.600\) units. In Design 2, each separate control group contributes variance \(16/50=0.32\). Each contrast therefore has variance \(16/80+16/50=0.20+0.32=0.52\), and standard error \(\sqrt{0.52}=0.721\) units.
The variance reduction for either contrast in the shared-control design is \((0.52-0.36)/0.52=0.308\), or about 30.8% relative to splitting the controls; the standard error falls by about 16.8%. This comparison isolates the effect of control reuse under the stated assumptions. It does not account for the positive dependence between the two shared-control estimates, nor does it prove that a shared-control design is preferable once recruitment, calendar time, safety, and multiplicity are considered.
A short justification clarifies the trade-off. The variance of each contrast includes one control-mean variance term. A shared group of 100 controls makes that term \(16/100\) for both contrasts. Splitting the controls leaves only 50 observations in each comparison, making the term \(16/50\). Reuse improves each marginal contrast’s precision, but the common control mean also makes the two estimates move together. The gain in precision for individual comparisons and the dependence across comparisons are two consequences of the same design choice.
Multiplicity and Interpretation Across Arms
If a trial tests several treatment-versus-control hypotheses, the chance of at least one false positive across the family can exceed the nominal significance level when each hypothesis is tested separately at that level. Sharing controls changes the joint distribution of the test statistics; it does not make the multiplicity problem disappear. A familywise procedure should be prespecified and should reflect the comparisons, direction of testing, stages, and analysis model. For multiple treatments compared with one control, a Dunnett-type procedure can use the known dependence among contrasts; Bonferroni adjustment is a simpler alternative that controls familywise error without relying on that dependence, but can be more conservative.
The choice of which hypotheses form the family matters. A platform protocol should specify which treatment comparisons are confirmatory and how the overall error rate is controlled, including any interim decisions that affect testing. In a multistage design, early stopping or arm selection can also affect the distribution of the final evidence; the analysis must follow the prespecified design rather than treating a selected arm as though it had been the only comparison all along.
Worked Example: Two Comparisons and a Simple Multiplicity Adjustment
Use the hypothetical means and standard errors from the first example. For a transparent illustration, assume normally distributed outcomes with known standard deviation 4.0 units in all groups, a two-sided test for each treatment difference, and no interim selection. The test statistics are \(z_A=-1.2/0.600=-2.00\) and \(z_B=-0.8/0.600=-1.33\). Their approximate unadjusted two-sided p-values are 0.0455 and 0.182, respectively.
For two comparisons, a Bonferroni procedure tests each at \(\alpha/2=0.025\), which corresponds to a two-sided standard normal critical value of about 2.241. Since \(|z_A|=2.00<2.241\) and \(|z_B|=1.33<2.241\), neither comparison meets this adjusted threshold. The Bonferroni-adjusted p-values are \(\min(2\times0.0455,1)=0.0910\) for A and \(\min(2\times0.182,1)=0.364\) for B.
This conclusion is not a claim that the treatments have no effect. It says that neither comparison provides sufficient evidence to reject its null hypothesis under this particular familywise procedure and model. A Dunnett-type analysis could use the positive dependence between these contrasts and may have a different critical value; its exact adjustment should be calculated for the specified design, not guessed from the pairwise p-values.
Case-Study Interpretation and Design Checks
The shared-control idea is useful in STAMPEDE because a platform can evaluate more than one experimental strategy against standard of care without constructing a wholly separate control group for every comparison. But interpretation must remain comparison-specific. Each contrast estimates the effect of a defined treatment strategy relative to its specified comparator, in the eligible population and period represented by that comparison. The shared control does not imply that different experimental treatments have been compared directly with one another: that would require a separate contrast and its own estimand and uncertainty.
When arms enter or leave a platform at different times, control observations should not be pooled indiscriminately just because they were assigned to standard care. Calendar changes in prognosis, supportive care, diagnostic practice, or the standard treatment itself can make an earlier control group unlike the concurrent comparator for a later arm. Eligibility criteria may also change. Valid analyses must follow the protocol’s rules for which controls are relevant to each comparison and account for any design features such as calendar period or stratification. A shared-control estimator is not a remedy for systematic differences between nonconcurrent groups.
The principal alternatives have different trade-offs:
| Approach | Potential advantage | Important caution |
|---|---|---|
| One shared concurrent control | Uses control participants in several comparisons, often improving marginal precision. | Creates dependence among contrasts; multiplicity and control eligibility must be handled explicitly. |
| Separate control group for each comparison | Each treatment comparison has its own comparator sample. | May require more control recruitment or split a fixed control sample, reducing precision per comparison. |
| Use nonconcurrent controls from different periods | May allow use of earlier platform data when a new arm begins. | Calendar-time and eligibility differences can bias a comparison if not adequately addressed by design and analysis. |
Diagnostics should therefore include more than the standard error. Check the allocation and eligibility records for each arm, confirm the control participants included in each contrast, examine recruitment timing and relevant baseline or care differences, and verify that the analysis uses the prespecified multiplicity method. If the assumed independence across participant samples is doubtful—for example because of clustering—or if the analysis uses adjusted or time-to-event outcomes, the simple covariance formula above is not sufficient. Use the covariance implied by the actual analysis model and design.
Common Pitfall
The practical lesson is to report what is shared and what is not: the control group used for each contrast, the treatment-specific estimate and uncertainty, and the method used to control error across the planned comparisons. This makes clear both the efficiency gained from reuse and the limits on interpreting a collection of results.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- Under the independent-sample model, what is the covariance of two treatment-minus-control estimates when the shared control variance is \(\sigma_0^2=9\) square units and \(n_0=36\)?
- If \(\sigma_0^2=0\), is the covariance strictly positive? Explain.
- Why does a shared control not eliminate the need to adjust for multiple treatment comparisons?
- Name one reason that control participants from an earlier period may not be appropriate for a later treatment-arm comparison.
Solutions. (1) The covariance is \(\sigma_0^2/n_0=9/36=0.250\) square units. (2) No. It is \(0/36=0\); the covariance is nonnegative and is strictly positive only when the control outcome variance is positive. (3) Several hypotheses are still being tested, so testing each at the nominal level can inflate the probability of one or more false rejections. Dependence changes the joint distribution but does not remove the need for a prespecified familywise procedure. (4) Standard of care, supportive care, eligibility, or other features of the enrolled population may have changed over calendar time, making earlier controls noncomparable to later participants.