STAMPEDE's Original Six-Arm Structure

Why the Original Arm Structure Matters

A multi-arm trial is not simply a collection of separate two-arm studies. Its treatment arms, eligible population, shared comparator, and timing together determine what a treatment contrast means. STAMPEDE’s original structure is a useful case: several systemic regimens were evaluated against a common standard-care backbone in men starting long-term hormone therapy for high-risk prostate cancer.

The trial’s six-arm structure is best read as a set of planned comparisons, not as a ranking of all six treatments. The central question for each experimental regimen was whether adding that regimen to standard care improved outcomes relative to standard care alone. It was not, by design, a direct test of every experimental regimen against every other one.

The Population and Treatment Backbone

The population relevant to these original comparisons comprised men with newly diagnosed high-risk prostate cancer for whom long-term hormone therapy was appropriate. This included men with metastatic disease and men with locally advanced disease meeting high-risk criteria. The clinical setting matters: the question was about intensifying initial systemic treatment in this population, not about treatment of every person with prostate cancer or about disease that had already been managed in a different treatment setting.

Standard care provided the common backbone. The experimental arms added one or more assigned treatments to that backbone; they did not replace the need to identify what care the control group received. This distinction is essential to the estimand. For example, the docetaxel comparison asks about the effect of adding docetaxel to standard care, not about docetaxel in isolation.

The primary outcome was overall survival: time from randomisation until death from any cause. Failure-free survival was used as an intermediate measure of disease control in the multi-stage screening process. These outcomes play different roles. Interim screening can determine whether an arm remains promising enough to continue; it does not change the definition of the final overall-survival comparison.

Reading the Six Arms as Planned Comparisons

The arm list can be translated into five experimental-versus-control contrasts. In each row below, “plus” means the experimental regimen is added to standard care. The table describes the treatment structure; it does not imply that each arm was open for the same calendar period or that every participant could have been randomised to every arm.

The primary comparisons are therefore not all pairwise contrasts among six arms. There are five planned treatment-versus-control questions. In particular, the docetaxel-plus-zoledronic-acid arm is compared with standard care to assess that combination regimen. Its contrast alone does not establish whether the combination is better than docetaxel alone, whether zoledronic acid contributes benefit when added to docetaxel, or whether the two drugs interact.

The six arms also do not make a complete factorial design. A complete \(2 \times 2\) factorial comparison of docetaxel and zoledronic acid would require all four combinations of those two factors, including a standard-care-only group, each single-treatment group, and the combined-treatment group, with assignment and analysis suited to estimating main effects and interaction. The original structure contains those four docetaxel/zoledronic-acid combinations, but the primary question in the multi-arm framework is each regimen versus the shared control; one should not substitute a factorial interpretation without the corresponding design and analysis assumptions.

Shared Control: What It Buys and What It Changes

Using a common control can make a multi-arm trial more efficient than conducting separate trials, because control participants contribute to more than one treatment-versus-control comparison. That advantage does not make the resulting contrasts statistically independent. If two estimated treatment effects use the same control observations, random variation in those observations can influence both estimates.

A simple derivation makes the dependence visible. Consider a hypothetical continuous outcome and suppose independent samples are observed in treatment arms \(A\), \(B\), and control \(C\). Let the estimated mean differences be \(\hat\theta_A=\bar Y_A-\bar Y_C\) and \(\hat\theta_B=\bar Y_B-\bar Y_C\). Under independent sampling across arms, the covariance between the two contrasts includes the variance of the shared control mean:

The first equality expands the contrasts. In the second, the treatment-arm means are independent of one another and of the control mean, so their cross-covariances are zero; the shared term contributes \(\operatorname{Var}(\bar Y_C)\). This calculation is an illustration of shared-control dependence, not the survival-analysis method used to estimate STAMPEDE’s overall-survival effects. A design or analysis that considers several comparisons must use an approach appropriate to their joint dependence and the trial’s pre-specified testing plan.

Stages, Outcomes, and Interpretation

In a multi-stage design, interim information is used to decide whether an experimental arm should continue. For STAMPEDE, failure-free survival supplied an earlier indication of disease control, while overall survival remained the primary outcome for evaluating benefit. An arm’s continuation through a screening stage is not proof of benefit; nor is stopping an arm equivalent to proving that its treatment has exactly no effect. Those decisions depend on the pre-specified rules, the available information, and the purpose of the interim assessment.

When interpreting a reported treatment estimate, keep the population, comparator, outcome, and time origin together. For example, a hazard ratio below 1 for overall survival would indicate a lower estimated instantaneous death rate in the experimental group than in the standard-care group over follow-up, subject to the model and proportional-hazards interpretation. It would not mean that the same percentage of patients avoid death, and it would not by itself give an absolute survival difference.

Diagnostics, Alternatives, and Limits

A sound reading of the design starts by checking whether the comparison is supported by the relevant randomised population and calendar period. In a platform whose arms can open or close, not every participant is necessarily eligible for every arm. Comparing a treatment group with all control participants across the entire trial, without regard to whether those controls were eligible and concurrent, can create a misleading contrast if recruitment patterns or care change over time. The comparison should reflect the trial’s randomisation and analysis plan.

Other warning signs include describing an experimental arm as “treatment versus no treatment” when both groups received standard care, treating interim screening as a final efficacy conclusion, or reading a treatment-versus-control result as a direct comparison between two active regimens. For survival outcomes, check that the reported measure is identified correctly: a hazard ratio, a survival probability at a stated time, and a difference in median survival are not interchangeable.

The principal alternative to a shared-control multi-arm platform would be separate two-arm trials, or direct randomisation between selected active regimens. Separate trials can make concurrent comparisons clearer but may use more control participants overall and can fragment evidence across studies. Direct active-arm randomisation answers a different question: which of those regimens performs better against the other? Neither alternative is universally superior. The appropriate structure depends on the clinical questions, eligibility, timing, and statistical plan.

The original arm structure also has limits. It does not justify generalising its comparisons to every prostate-cancer setting, and a regimen-versus-control contrast does not automatically identify the effect of each component in a combination. The trial’s evolving set of open arms and the use of interim screening also mean that the six-arm diagram should be understood as a design structure, not as a claim that all six options were simultaneously available to every eligible patient at every point in time.