What a Platform Trial Organizes
A platform trial can evaluate several treatments within one continuing trial structure. That structure matters because the set of treatments, the control group, and the patients eligible for each comparison may change over time. A clear architecture makes these changes explicit before they affect randomization or analysis.
The platform is governed by a master protocol: the overarching document that specifies the trial framework, including common procedures and the rules for adding, conducting, and stopping treatment comparisons. Treatment-specific details may be set out in appendices or subprotocols. A platform may have multiple treatment arms active at once, and an arm can enter or leave while the overall trial continues.
A platform is not defined simply by having several arms. Its defining architectural feature is the possibility of changes to the active set within a continuing framework. It is also useful to distinguish common master-protocol terms: an umbrella trial examines multiple treatments for a disease or condition, while a basket trial evaluates a treatment across multiple diseases or populations. These labels describe different organizing dimensions; either framework could have features of a platform if arms or comparisons change over time.
Representing the Randomization Structure
For each patient, the protocol must determine which arms are available. Availability can depend on the enrollment date, the patient’s eligibility, and whether an arm is open at the relevant site. Let \(E_i(t)\) denote the set of arms available to patient \(i\) at time \(t\). If the control is denoted by \(C\) and treatments by \(A\) and \(B\), a patient eligible for all three during a period might have \(E_i(t)=\{C,A,B\}\). A patient who does not meet the criteria for \(B\) might instead have \(E_i(t)=\{C,A\}\).
For an available arm \(j\), let \(\pi_{ij}(t)\) be the probability that patient \(i\) is assigned to that arm at time \(t\), given the protocol’s randomization procedure and the patient’s eligibility. For each patient, probabilities for available arms must sum to one; an unavailable arm has assignment probability zero. The probabilities can be fixed or can change according to a prespecified adaptation. The architecture specifies the arms and comparisons the procedure can use; it is distinct from the particular rule used to update probabilities.
A defensible design also states its assumptions. Randomization protects comparisons within the group of patients eligible for those arms when assignment is implemented as specified. A shared-control comparison additionally requires that the control observations used are appropriate for the treatment comparison in question. In particular, sharing a label such as “control” does not by itself make controls enrolled in different calendar periods exchangeable. Eligibility criteria, background care, outcome measurement, and the clinical context must be considered.
Worked Example: A Fixed Shared-Control Allocation
Suppose a hypothetical platform is enrolling patients eligible for control \(C\), treatment \(A\), and treatment \(B\). During an initial period, the protocol assigns probabilities \(0.40\), \(0.30\), and \(0.30\), respectively. Assume each patient in this period meets the eligibility criteria for all three arms and that the randomization procedure uses these probabilities as specified.
For \(120\) such patients, the expected number assigned to each arm is the total number multiplied by the corresponding probability:
These are expected counts, not guaranteed realized counts: randomization can produce different totals. Both treatment arms use the same control group, so their comparisons can draw on the 48 control patients randomized during this period. This is a shared-control architecture; it does not make the two treatment-versus-control estimates independent.
Why Shared Controls Change the Comparison
A shared control can avoid assigning a separate control group to each treatment comparison, and it can make more efficient use of control participants. But comparisons that use the same control observations are statistically related. This dependence matters when describing joint evidence or uncertainty across several treatment comparisons.
The relationship can be seen under a simple model. Suppose the outcome is continuous, patients’ outcomes are independent across randomized groups, and all groups have the same outcome variance \(\sigma^2\). Let \(\bar{Y}_A\), \(\bar{Y}_B\), and \(\bar{Y}_C\) be the observed group means for treatments \(A\), \(B\), and the shared control. Define treatment-minus-control contrasts \(D_A=\bar{Y}_A-\bar{Y}_C\) and \(D_B=\bar{Y}_B-\bar{Y}_C\). Since the treatment groups are distinct and independent of the control group, the only random term they share is \(\bar{Y}_C\).
The positive covariance arises because a control mean that is lower than expected makes both treatment-minus-control contrasts larger. The result depends on the stated independent-group, common-variance setup; more complex outcomes or models require the corresponding covariance calculation. The architectural lesson remains: shared-control contrasts must be analyzed as related, not as though each had a separate independent control group.
Worked Example: Quantifying Shared-Control Dependence
In a hypothetical continuous-outcome trial, suppose \(n_A=40\), \(n_B=40\), and the two treatments share \(n_C=60\) controls. Assume independent patients and a common outcome standard deviation of \(10\) units, so \(\sigma^2=100\) square units.
The variance of the first treatment-control difference is \(\operatorname{Var}(D_A)=\sigma^2/n_A+\sigma^2/n_C\):
The same calculation gives \(\operatorname{Var}(D_B)=4.167\) square units. Their covariance is the variance of the shared control mean:
Thus, their correlation is \(1.667/\sqrt{4.167(4.167)}=0.400\). Under these assumptions, the two estimated treatment-control differences are positively correlated. Treating them as independent would misrepresent their joint uncertainty.
Decision Points That Change the Structure
The master protocol should state when an arm can open, when it stops accepting new patients, and what happens to patients already assigned to it. These decisions change the set of available assignments. They can also change which patients contribute to each treatment-control comparison.
Specify the date or criteria for activation, the sites and patient groups eligible for the arm, and whether a concurrent control is available for its comparison.
State when probabilities may change and which patients are affected. If response evidence can influence allocation, make clear that this is a prespecified change to the randomization structure, not an informal adjustment.
Record the stopping decision and its effective date. New patients should no longer be assigned to the closed arm, while follow-up and analysis responsibilities for previously assigned patients remain defined.
After an arm closes, identify the remaining eligible arms and control group, the resulting randomization probabilities, and whether each planned comparison still has an appropriate concurrent comparator.
A decision to remove a treatment arm does not erase its previously randomized patients. Nor does it automatically justify pooling every control patient enrolled before, during, and after that arm’s activity. The time period and eligibility conditions for each comparison should be recorded. Where an arm’s entry or exit changes which patients could be randomized to which treatments, the analysis plan needs to preserve that structure.
Worked Example: Removing an Arm and Renormalizing Probabilities
Consider a hypothetical period in which eligible patients are randomized to \(C\), \(A\), or \(B\) with probabilities \(0.40\), \(0.30\), and \(0.30\). The protocol then closes \(B\) to new enrollment. Suppose it explicitly requires preserving the relative allocation between \(C\) and \(A\) for the remaining eligible patients.
The remaining probabilities must sum to one. The original probabilities for \(C\) and \(A\) sum to \(0.70\), so proportional renormalization gives:
The allocation probability for \(B\) becomes zero for new patients. For the next 70 eligible patients, expected assignments are \(70(4/7)=40\) to \(C\) and \(70(3/7)=30\) to \(A\). These expectations apply only after the change takes effect; they do not reassign patients already randomized to \(B\). Proportional renormalization is appropriate here only because the hypothetical protocol specified preservation of the \(C:A\) ratio. A different protocol could prescribe different probabilities.
Concurrent Controls, Diagnostics, and Alternatives
A concurrent control consists of control patients randomized during a period in which the treatment under comparison is also open to eligible patients. Concurrent controls are generally the clearest comparator because the treatment and control groups are exposed to the same broad calendar-time context. When controls from a different period are proposed for a comparison, calendar-time differences in standard care, recruitment, measurement, or patient mix may undermine comparability. Such patients should not simply be treated as interchangeable with concurrent controls without a justified design and analysis.
Useful structural checks include reviewing arm opening and closure dates, comparing eligibility criteria across periods, confirming that each patient’s randomization probabilities were valid for their actual available-arm set, and checking that allocation records match protocol changes. Investigators should also examine whether the intended shared controls are genuinely eligible for the comparisons that use them. These checks concern trial conduct and interpretability; a well-specified allocation does not by itself guarantee that every comparison will be adequately precise or free from bias.
| Approach | Architectural advantage | Main caution |
|---|---|---|
| Separate two-arm trials | Each trial can have a distinct control and comparison structure. | May require more control enrollment and does not share information by design. |
| Multi-arm trial with shared control | Several treatments can be compared within one concurrent framework. | Contrasts share control data and are statistically dependent. |
| Platform with changing arms | New treatments can enter and inactive arms can leave while the framework continues. | Changing eligibility and calendar periods require careful documentation and comparison-specific analysis. |
The platform approach is not automatically superior to separate trials. Its efficiency depends on the availability of suitable common controls and on whether the arms can share a defensible framework. Its limitations include operational complexity, the possibility of changing standards of care, and the need to track distinct enrollment periods and eligibility groups. The master protocol must make these choices transparent rather than relying on the word “platform” to settle them.
Common Mistakes and Practical Limits
- Assuming every control is shareable. A common control label is not enough. Check timing, eligibility, outcome measurement, and the assumptions needed for the specific comparison.
- Treating shared-control estimates as independent. Both contrasts include the same control mean, so their uncertainty is related.
- Applying a new allocation rule retrospectively. A probability change governs assignments after its effective time; it does not alter the randomization of patients already enrolled.
- Calling a probability change a complete design description. The protocol must also specify which arms are available, to whom, and when the change applies.
- Confusing architecture with an adaptation algorithm. Architecture determines the possible arms, comparisons, and change points. The rule for choosing updated probabilities is a separate methodological decision.
A practical limitation is that a platform’s structure can become difficult to summarize once arms enter at different times or serve different eligibility groups. A trial should retain a clear record of each patient’s available arms, assignment probabilities, randomization date, and protocol version. Without those records, even a correctly designed platform may be hard to analyze or reproduce.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- During a period, 90 patients eligible for control and two treatments are randomized with probabilities \(0.40\), \(0.30\), and \(0.30\). What are the expected assignments?
- Why are two treatment-versus-control contrasts statistically dependent when they use the same control patients?
- An arm closes on a stated date. What happens to its probability for newly randomized patients, and what does not happen to patients already assigned?
- State two reasons a control group enrolled well before a treatment arm opens may not be an appropriate comparator for that arm.
- If an arm is removed and the protocol says to preserve the relative allocation between the two remaining arms, how should their probabilities be updated?
Solutions. (1) The expected assignments are \(90(0.40)=36\) to control and \(90(0.30)=27\) to each treatment. These are expectations, not guaranteed counts. (2) Both contrasts subtract the same control mean; under independent groups, their covariance includes the positive variance of that shared control mean. (3) Its assignment probability becomes zero for new patients after closure takes effect, but patients already assigned to it remain in the trial’s follow-up and analysis as specified by the protocol. (4) Standard care, eligibility, patient mix, or outcome measurement may have changed between periods; any two are sufficient. (5) Divide each remaining arm’s original probability by the sum of the remaining probabilities. This preserves their relative allocation only because the protocol specifies that rule.