STAMPEDE's Control Arm Across Calendar Time

Why the Control Arm Has a Calendar-Time History

In a long-running platform trial, “standard care” is a clinical strategy defined by the protocol and the care available when patients enter. It need not be the same package of treatment throughout the trial. STAMPEDE began in an era when long-term androgen-deprivation therapy (ADT) was the central systemic treatment for advanced prostate cancer. As evidence and practice changed, the platform added comparisons, and the appropriate background care could differ by disease setting, protocol version, and recruitment period.

This matters because an arm’s result is a comparison with the control patients who were eligible and randomized alongside it—not with an abstract, timeless control group. STAMPEDE’s control history therefore has two linked parts: what standard care meant at a particular time, and which patients were eligible for each comparison at that time.

The ADT backbone is an important point of continuity, but it does not make every control group identical. Local treatment, such as radiotherapy, could be part of care according to disease setting and protocol requirements. Later, as docetaxel became an established option for some men with metastatic hormone-sensitive disease, protocol versions and comparisons could differ in whether it was available or included in background care. It would be incorrect to assume that every STAMPEDE control participant received the same combination—or that a treatment tested in one comparison automatically became part of every other comparison’s control.

Eligibility Changed Alongside the Platform

STAMPEDE was designed to evaluate multiple treatment comparisons in men with prostate cancer at substantial risk of progression. The original platform included patients with advanced disease, but there was not one permanently fixed eligibility definition that applied identically to every arm. New comparisons could focus on particular disease states or risk groups, and eligibility could be specified or amended for the question being tested.

For example, a comparison in metastatic disease and a comparison restricted to nonmetastatic disease do not answer the same clinical question, even if both use ADT as a backbone. Within nonmetastatic disease, whether a patient had nodal involvement and what local treatment was planned could also matter to the treatment strategy. The STAMPEDE comparisons of prostate radiotherapy in low-burden metastatic disease and metformin in nonmetastatic disease illustrate why the setting named in a comparison is part of its interpretation, not a descriptive footnote.

A useful record for interpreting any reported arm result therefore includes the comparison’s recruitment dates, eligibility rules, control treatment, and relevant background care. These details help establish which patients and care context the result actually represents. They also prevent a common error: treating “the STAMPEDE control” as one homogeneous group assembled across the entire life of the platform.

What Randomization Does—and Does Not—Protect

Within an eligible comparison and recruitment period, randomization assigns participants to treatment groups by chance. Under the trial’s allocation and follow-up procedures, this supports a causal comparison of the assigned strategies for that eligible population. It does not make the control strategy identical across periods, nor does it guarantee that patients eligible for one comparison represent patients eligible for another.

A compact way to state the target is to let \(Y(1)\) and \(Y(0)\) denote the outcome that a participant would have under the experimental and control strategies, respectively. For the eligible population in a particular comparison and period, a risk-difference estimand is \(E\{Y(1)-Y(0)\}\). Changing the eligibility rules changes the population over which that expectation is taken; changing the control strategy changes what \(Y(0)\) means.

This is why a result from an earlier comparison cannot simply be carried over as the control experience for a later one. Even if the endpoint name is the same, the patient mix, treatment options, follow-up practices, or standard-care components may differ. Conversely, if an experimental arm and control arm were randomized concurrently under the same rules, their contrast remains the primary comparison for that trial question.

Worked Examples: Period, Eligibility, and Interpretation

Justification: Why Period Must Be Considered

The problem in the first example can be seen directly by writing each pooled risk as a weighted average of period-specific risks. Let \(w_{g,t}\) be the proportion from period \(t\) in group \(g\), and let \(r_{g,t}\) be that group’s event risk in that period. Then the pooled risk is \(\sum_t w_{g,t}r_{g,t}\). The difference between pooled groups is \(\sum_t w_{1,t}r_{1,t}-\sum_t w_{0,t}r_{0,t}\).

If the weights differ by group, this is generally not a weighted average of the within-period treatment differences \(r_{1,t}-r_{0,t}\). Only when the groups have the same period weights, or when special conditions make the composition irrelevant, does the crude pooled contrast directly reflect a common weighted within-period contrast. In the first example the early and later weights differed, so the period mix contributed to the apparent 22-point difference despite a 10-point difference within each period.

This is a justification for examining recruitment period and eligibility, not a claim that statistical adjustment alone creates comparability. The most secure basis for an arm comparison remains concurrent randomization. If a summary across periods is needed, report the period-specific estimates and state the target weights; consider whether treatment effects differ by period rather than assuming they are constant.

Reading a Control History Carefully

Diagnostics include checking whether the arms being compared overlap in recruitment dates, whether their eligibility criteria match, and whether the background control strategy is documented. Large shifts in patient characteristics across periods, a change in local treatment, or different follow-up and outcome definitions are warnings against an unqualified pooled comparison. A regression model with a period indicator may describe measured differences, but it cannot guarantee that all time-related differences have been controlled.

Alternatives and Limitations

The main alternative to a concurrent randomized comparison is a historical or nonconcurrent control comparison. It may be useful descriptively when no concurrent control is available, but it is vulnerable to calendar-time confounding and changes in eligible patients. A second approach is to present period-stratified randomized effects and, if a combined summary is useful, standardize them to a stated target period mix. This preserves visibility of heterogeneity better than a single unqualified pooled estimate, but it requires adequate data in each period and does not remove bias from changes that were not measured or randomized.

A further limitation is that “standard care” can vary among patients even within a protocol period because clinical decisions and local practice differ. A protocol definition tells the reader what was intended or permitted; it does not by itself prove that every participant received identical background treatment. Interpretation should therefore distinguish assigned strategy from actual care received.

The key practical lesson is not that results from different periods cannot be compared. It is that a comparison must say what is being held constant and what is allowed to change. A valid concurrent result can be strong evidence for its defined population and care context without being a universal estimate for every later standard-care setting.