Platform Trials Explained

This allows interim boundaries to be defined according to statistical information rather than simply calendar time.

Interim Analyses Need Prespecified Rules

A platform can have many possible interim analyses.

For example:

At each interim analysis, a treatment can be evaluated according to the prespecified decision rules.

The statistical design must account for repeated opportunities to stop for success.

Group Sequential Methods Within a Platform

A platform can incorporate group-sequential methods.

For example, an efficacy boundary may be specified as:

where \(c(t)\) changes according to the information fraction.

An O'Brien-Fleming-like approach may use a more stringent early boundary and a less stringent final boundary.

A Pocock-like approach may use more similar boundaries across analyses.

The exact boundary construction depends on the platform's multiplicity and interim-analysis structure.

Multiplicity Exists in Two Dimensions

Platform trials can face multiplicity across both:

• Treatments
• Interim analyses

For example, three treatment arms evaluated at three interim analyses create multiple opportunities to make efficacy decisions.

The design therefore needs to consider the joint probability of false positive decisions across the relevant treatment and time dimensions.

Familywise Error in a Platform

Suppose the confirmatory family contains three treatment hypotheses:

The familywise error rate can be written as:

A confirmatory design may target:

The exact calculation can require multivariate distributions, simulation, closed testing, graphical procedures, or other methods depending on the design.

Why Simulation Is Often Essential

For a simple two-stage single-arm trial, exact binomial enumeration may be enough.

Platform trials can be substantially more complicated.

For example, a platform may include:

• Multiple treatment arms
• Multiple interim analyses
• Shared controls
• Unequal randomization
• Adaptive randomization
• Arms entering at different times
• Arms dropping at different times
• Covariate adjustment
• Missing outcomes
• Delayed outcomes
• Different treatment effects

In such settings, Monte Carlo simulation is often the most practical way to evaluate operating characteristics.

Simulation-Based Operating Characteristics

Important quantities include:

• Familywise type I error
• Power for each treatment
• Probability of correctly dropping futile treatments
• Probability of incorrectly dropping active treatments
• Probability of early success
• Expected sample size
• Maximum sample size
• Probability each arm enters the platform
• Probability each arm reaches final analysis
• Bias and precision of treatment-effect estimates
• Average trial duration

A platform design should be evaluated under many plausible scenarios rather than only one assumed treatment-effect configuration.

Scenario-Based Simulation

Suppose a platform has three experimental treatments.

A useful simulation program might include:

These scenarios reveal how the platform behaves when different combinations of treatments are actually effective.

Global Null Scenario

The global null is particularly important for evaluating familywise type I error.

Suppose:

The simulation asks:

This should satisfy the relevant error-control requirement.

Power Is Not a Single Number in a Platform

In a conventional two-arm trial, people often refer to "the power" of the study.

In a platform, several power quantities may be relevant.

For treatment A:

Similarly:

may be evaluated separately.

Investigators may also consider:

• Marginal power
• Joint power
• Power to identify at least one effective treatment
• Power to identify all effective treatments
• Power under different combinations of active and inactive treatments

Marginal vs. Joint Power

Suppose A and B are both truly effective.

Marginal power asks whether A is detected:

while joint power asks whether both are detected:

Joint power is generally lower than either marginal probability.

This distinction becomes important when the scientific objective is to identify a complete set of effective interventions rather than simply detect individual effects.

Conditional Power

Frequentist platform designs can also use conditional power.

Conceptually:

A treatment may be dropped if the conditional power falls below a prespecified futility threshold.

Conditional power can therefore provide an evidence-based futility decision without requiring the interim estimate itself to cross a fixed efficacy boundary.

Predictive Probability vs. Conditional Power

These concepts are related but not identical.

These quantities should not be treated as interchangeable.

Delayed Outcomes

Platform trials frequently encounter delayed endpoints.

For example, progression-free survival or overall survival may take months or years to mature.

The platform may therefore use:

• Interim information based on partially mature data
• Time-to-event methods
• Information-based stopping rules
• Predictive probabilities
• Event-driven interim analyses

The timing of decisions should be aligned with the actual information available rather than simply the number of patients enrolled.

Time-to-Event Platform Trials

For a time-to-event endpoint, treatment effects may be expressed through a hazard ratio:

where \(HR<1\) may indicate benefit when lower hazard is favorable.

A platform might test:

Interim decisions can then be based on information measured through the number of observed events.

Sample Size in Platform Trials

Sample-size planning is more complicated than simply multiplying the sample size for one two-arm trial by the number of treatments.

The required information depends on:

• Number of treatment arms
• Control allocation
• Treatment allocation
• Endpoint variance or event rate
• Target treatment effect
• Multiplicity adjustment
• Interim analyses
• Shared-control correlations
• Expected treatment-arm entry and dropout
• Accrual rate
• Information timing

A Simple Continuous-Endpoint Approximation

For a single treatment-control comparison with equal allocation and a continuous endpoint, a rough two-sided sample-size relationship is:

A platform with multiple treatment arms cannot simply use this equation without modification.

The allocation ratio, multiplicity strategy, shared control, and adaptive structure all matter.

Control Allocation in a Multi-Arm Trial

Suppose there are \(K\) experimental arms and a control.

If each experimental arm receives \(n_T\) patients and the control receives \(n_C\), then the allocation ratio is:

A larger control group improves precision for every treatment-control comparison, but it also means fewer patients are allocated to experimental treatments for a fixed total sample size.

Thus, control allocation is a design optimization problem rather than a purely statistical afterthought.

Operational Efficiency

Platform trials can potentially reduce more than patient numbers.

A common platform can also share:

• Clinical sites
• Data-management infrastructure
• Central laboratories
• Imaging procedures
• Randomization systems
• Statistical programming
• Safety monitoring infrastructure
• Vendor relationships
• Protocol training

These efficiencies can be particularly important when multiple treatments address the same disease population.

Operational Complexity

Efficiency does not mean simplicity.

Platform trials can be operationally complex because:

• Treatment arms may have different dosing schedules.
• Arms may have different safety profiles.
• Arms may enter at different times.
• Randomization probabilities can change.
• Eligibility criteria may evolve.
• Control treatment may change over time.
• Different arms may have different follow-up requirements.

The platform therefore requires strong governance and careful version control.

Protocol Amendments

Adding or removing an arm generally requires a prespecified amendment and appropriate regulatory and operational processes.

The statistical framework should specify:

• How a new treatment becomes eligible for entry
• Which data contribute to its analysis
• How randomization is modified
• How multiplicity is handled
• How its sample size is determined
• How success and futility are defined

Blinding

Platform trials can be blinded or open-label depending on the interventions and endpoints.

Blinding can become more complicated as multiple treatments enter and leave the platform.

For example, if treatments have visibly different administration procedures, maintaining blinding may require separate matching procedures.

The statistical design should be developed together with the operational blinding strategy.

Safety Monitoring in a Platform

Each treatment arm can have its own safety profile.

The platform therefore needs rules for:

• Arm-specific safety monitoring
• Shared safety oversight
• Stopping an individual treatment for toxicity
• Stopping the platform for broader safety concerns
• Reviewing cumulative safety data

An arm-specific safety stop does not necessarily imply that other treatments must stop.

A Simple Platform Decision Matrix

Estimands in Platform Trials

A platform trial should define its estimands explicitly.

For treatment \(j\), an estimand might specify:

• Population
• Treatment condition
• Endpoint
• Summary measure
• Handling of intercurrent events

For example, a continuous endpoint estimand could target:

The exact estimand depends on the scientific question and treatment context.

Different Arms Can Address Different Questions

A platform should not be assumed to have a single universal estimand.

Treatment A might be evaluated in one biomarker-defined population while treatment B is evaluated in another.

In that situation, the platform can share infrastructure without pretending that all treatment comparisons answer exactly the same scientific question.

Common Statistical Models

Depending on the endpoint, platform analyses may use:

• Linear models
• Generalized linear models
• Logistic regression
• Cox proportional-hazards models
• Mixed-effects models
• Repeated-measures models
• Bayesian hierarchical models
• Joint models

The platform structure determines how the treatment indicators, shared control, time, site, and other factors enter the analysis.

Hierarchical Borrowing

Bayesian platform trials sometimes consider hierarchical models that allow information to be partially shared across treatment arms or populations.

Conceptually, treatment effects might follow:

where \(\tau^2\) represents between-treatment heterogeneity.

This can allow some information sharing while retaining differences between treatments.

Platform Trials With Biomarkers

Platform trials can be especially useful when treatments are associated with different biomarkers.

For example:

The platform can therefore combine treatment and subgroup adaptation.

This creates additional multiplicity and sample-size considerations.

What Happens When a Treatment Is Dropped?

Dropping a treatment has several consequences.

First, future patients are no longer randomized to that treatment.

Second, the allocation probabilities among the remaining arms may change.

Third, the analysis population for the dropped treatment is effectively closed.

Fourth, the platform's future control observations may or may not contribute to the final analysis depending on the prespecified analysis plan.

Therefore, "drop an arm" is a statistical operation as well as an operational one.

What Happens When a Treatment Enters?

Adding an arm introduces a new treatment-control comparison.

The platform should therefore specify:

• Entry criteria
• Required sample size
• Initial randomization probability
• Interim timing
• Futility rule
• Efficacy rule
• Final analysis rule
• Multiplicity framework
• Control eligibility

The fact that a new treatment entered later does not eliminate the need to define its statistical operating characteristics.

Platform Trial vs. Standard Multi-Arm Trial

Platform Trial vs. Group Sequential Trial

A group sequential trial generally has one primary treatment comparison and multiple interim analyses.

A platform can have:

Therefore, a platform can incorporate group-sequential methods without being equivalent to a conventional group sequential trial.

Platform Trial vs. Adaptive Trial

The terms are also not synonymous.

An adaptive trial is any trial that prospectively plans changes based on accumulating data while maintaining the specified validity of inference.

A platform trial is a particular type of adaptive infrastructure designed to evaluate multiple interventions over time.

Thus:

conceptually, although the exact terminology used in the literature can vary.

Simulation in R

A simplified platform simulation can illustrate the basic mechanics.

set.seed(123)

nsim <- 10000

n_control <- 100
n_A <- 100
n_B <- 100

mu_control <- 0
mu_A <- 0.5
mu_B <- 0

sigma <- 1

results <- data.frame(
  A_success = logical(nsim),
  B_success = logical(nsim)
)

for (i in seq_len(nsim)) {

  y_control <- rnorm(
    n_control,
    mean = mu_control,
    sd = sigma
  )

  y_A <- rnorm(
    n_A,
    mean = mu_A,
    sd = sigma
  )

  y_B <- rnorm(
    n_B,
    mean = mu_B,
    sd = sigma
  )

  z_A <- (
    mean(y_A) - mean(y_control)
  ) / sqrt(
    sigma^2/n_A +
    sigma^2/n_control
  )

  z_B <- (
    mean(y_B) - mean(y_control)
  ) / sqrt(
    sigma^2/n_B +
    sigma^2/n_control
  )

  results$A_success[i] <- z_A > 2.5
  results$B_success[i] <- z_B > 2.5
}

This is intentionally a simplified illustration.

A production platform-trial simulation should reproduce the actual randomization algorithm, interim analyses, stopping rules, enrollment timing, missingness, endpoint distribution, multiplicity procedure, and analysis model.

Estimating Operating Characteristics

The simulated probability of success for treatment A is:

mean(results$A_success)

Similarly:

mean(results$B_success)

If A is truly effective and B is not, these quantities help characterize the marginal behavior of the platform.

The simulation can be extended to calculate the probability of at least one false positive:

mean(
  results$A_success |
  results$B_success
)

The actual expression should use the intended logical structure carefully; for example:

mean(
  results$A_success |
  results$B_success
)

would estimate the proportion of simulations in which at least one of the two decisions was positive.

Adding an Interim Analysis to the Simulation

A more realistic simulation would divide enrollment into stages.

n_stage1 <- 50
n_stage2 <- 50

for (i in seq_len(nsim)) {

  # Stage 1 enrollment

  # Calculate interim treatment effect

  # Apply futility/efficacy boundaries

  # If treatment continues:
  # enroll Stage 2 patients

  # Perform final analysis
}

This structure allows the simulation to estimate:

• Probability of early success
• Probability of early futility
• Probability of reaching final analysis
• Expected sample size
• Final power
• Familywise type I error

Expected Sample Size in a Platform

For a fixed two-stage trial, expected sample size can often be written simply.

For a platform, the calculation can become substantially more complicated.

Suppose there are \(K\) treatment arms.

The total sample size can be represented as:

where:

• \(N_C\) is the number of control patients.
• \(N_j\) is the number of patients assigned to treatment \(j\).

But \(N_j\) is itself random when treatment arms can enter, stop, or change allocation.

Therefore:

with each expectation depending on the platform's adaptive rules.

Expected Number of Patients by Treatment

A useful operating characteristic is:

for every treatment arm.

This can reveal how much exposure is expected under different treatment-effect scenarios.

For example, a simulation might produce:

These values are illustrative.

They demonstrate why reporting only total sample size can hide important features of an adaptive platform.

Probability of Dropping an Effective Treatment

Futility rules create a potential risk:

This is sometimes called a false-futility probability.

A good platform evaluation should examine this probability under clinically meaningful effect sizes.

For example:

Probability of Early Success

Similarly, investigators can evaluate:

This helps determine how often highly effective treatments can leave the platform before reaching the maximum planned information.

Trial Duration

Patient numbers are not the only efficiency metric.

Platform simulations should often estimate:

• Calendar time to first treatment decision
• Calendar time to final treatment decision
• Time required to introduce a new treatment
• Time spent waiting for endpoint maturation
• Time to platform completion

A design that saves patients but requires substantially longer follow-up may have a different operational profile from one that uses more patients but produces faster decisions.

Accrual Rate Matters

Suppose a platform enrolls 10 patients per month.

If treatment A requires 100 evaluable patients, recruitment alone requires approximately:

months, before accounting for endpoint maturation.

If three arms are recruiting simultaneously, the total platform accrual rate may be considerably higher, depending on eligibility and randomization.

Eligibility and Arm-Specific Eligibility

A platform can have common eligibility criteria plus treatment-specific criteria.

For example:

This means a patient can be eligible for the platform but not eligible for every treatment arm.

Randomization must therefore occur only among the treatments for which that patient is eligible.

Platform Randomization Is Not Always a Simple \(1:1:1:1\)

Suppose a patient is eligible for A and B but not C.

A randomization algorithm could assign among:

rather than:

The platform therefore needs an eligibility-aware randomization system.

Master Protocol Governance

Long-running platforms require governance structures that can manage:

• Treatment-arm entry
• Treatment-arm closure
• Protocol amendments
• Statistical changes
• Safety signals
• Data-access rules
• Blinding
• Independent decision-making

An independent data monitoring or decision-making body may be involved depending on the trial.

Operational Blinding and Information Leakage

Adaptive platforms can create opportunities for information leakage.

For example, investigators may infer that a treatment is performing well if its enrollment probability suddenly increases.

The platform should therefore determine which adaptation information is visible to:

• Investigators
• Patients
• Sponsors
• Statistical teams
• Operational teams

and which information remains restricted.

Statistical Team Structure

Complex platforms often benefit from clearly separated responsibilities for:

• Trial statisticians
• Independent statistical monitoring
• Data management
• Randomization programming
• Simulation programming
• Final analysis programming

The exact structure depends on the platform's design and blinding requirements.

Common Mistakes

1. Thinking a platform is simply several trials sharing a website. The shared-control and adaptive statistical structure must be explicitly designed.
2. Ignoring calendar time. A treatment entering years later may not have a directly exchangeable control population.
3. Using an unadjusted 0.05 threshold repeatedly. Multiple treatment arms and interim looks can inflate false-positive error.
4. Assuming all arms have equal information. Arms can enter and leave at different times.
5. Automatically using response-adaptive randomization. Adaptive randomization adds substantial statistical and operational complexity and is not required for a platform.
6. Failing to simulate the actual platform. Simple closed-form calculations often do not capture the full adaptive structure.
7. Ignoring treatment-arm entry. Adding an arm creates a new statistical comparison.
8. Ignoring treatment-arm exit. Dropping an arm changes future randomization and information accumulation.
9. Failing to distinguish concurrent from non-concurrent controls. This can create important interpretational and statistical issues.
10. Reporting only overall power. Treatment-specific marginal and joint operating characteristics may also be important.
11. Confusing posterior probability with frequentist type I error. These are different statistical concepts.
12. Assuming a platform must be Bayesian. Platform trials can use Bayesian, frequentist, or hybrid methods.

A Practical Platform-Trial Design Workflow

What Should Be Specified in the Statistical Analysis Plan?

The statistical documentation should be sufficiently detailed that the platform's decisions can be reproduced.

At minimum, specify:

• Estimands
• Primary endpoints
• Analysis populations
• Treatment-arm definitions
• Control definition
• Randomization procedure
• Shared-control rules
• Concurrent-control rules
• Interim-analysis timing
• Futility boundaries
• Efficacy boundaries
• Multiplicity strategy
• Type I error control
• Power definitions
• Sample-size assumptions
• Arm-entry criteria
• Arm-removal criteria
• Missing-data handling
• Safety stopping procedures
• Simulation assumptions

Example Operating-Characteristic Table

A platform design might summarize its simulated operating characteristics as follows.

Actual numerical values should come from the final platform simulation rather than being inferred from a simple two-arm sample-size calculation.

Platform Trials and Regulatory Considerations

A platform trial can be attractive from a development perspective, but its statistical complexity means that the design should be justified carefully.

Important topics include:

• Definition of each treatment comparison
• Control comparability
• Multiplicity
• Interim analyses
• Type I error
• Statistical adaptations
• Integrity of treatment comparisons
• Data access and blinding
• Protocol amendments

For confirmatory development, the platform's statistical operating characteristics should be demonstrated under realistic scenarios before the trial begins.

Confirmatory vs. Exploratory Platform Objectives

Not every platform comparison has the same evidentiary purpose.

Some treatment arms may be intended for exploratory screening.

Others may have confirmatory objectives.

The distinction matters because the required error control, multiplicity strategy, and interpretation of results can differ substantially.

Platform Trials Can Be Perpetual

One of the most interesting possibilities is the perpetual platform.

Instead of designing the platform to end after treatment A, B, and C have been evaluated, the infrastructure can continue as new interventions become available.

Conceptually:

The platform becomes a continuing mechanism for evaluating treatments within a disease area.

Why Perpetual Platforms Are Attractive

A successful perpetual platform can potentially avoid repeatedly rebuilding:

• Clinical infrastructure
• Site contracts
• Randomization systems
• Data standards
• Central laboratory procedures
• Statistical programming infrastructure
• Safety monitoring processes

This can make the development ecosystem more responsive to new candidate treatments.

Why Perpetual Platforms Are Difficult

A long-running platform also creates accumulating complexity.

Over time:

• Standard of care can change.
• Eligibility criteria can evolve.
• New treatments can have different mechanisms.
• Control outcomes can change.
• Statistical assumptions can become outdated.
• Operational processes can change.

The longer a platform runs, the more important careful governance and statistical monitoring become.

A Mental Model for Platform Trials

A useful way to remember the concept is:

The Most Important Statistical Insight

The defining statistical challenge of a platform trial is not simply that there are many treatments.

It is that the treatment comparisons are embedded in a shared, adaptive system.

The system creates dependencies through:

• Shared controls
• Shared patients and eligibility criteria
• Common calendar time
• Repeated interim analyses
• Adaptive randomization
• Arm entry and exit
• Common decision rules

Consequently, the platform should be designed as one statistical system rather than as a collection of independent analyses.

Platform Trial Checklist

Summary of the Worked Platform

The hypothetical platform began with:

Treatment B was dropped for futility.

Treatment C subsequently entered.

Treatment A eventually met its success criterion and left the platform.

The resulting active platform became:

The key feature is that the platform infrastructure remained operational throughout these transitions.

What a Platform Trial Is Really Buying You

The central value proposition of a platform is not simply "fewer patients."

It is the ability to create a reusable statistical and operational infrastructure for answering multiple treatment questions.

Potential advantages include:

• Shared controls
• Shared infrastructure
• Faster evaluation of new treatments
• Early discontinuation of futile treatments
• Potentially earlier identification of successful treatments
• Efficient use of accumulated control information
• Continuous learning within a disease area

These advantages come with corresponding costs in statistical planning, governance, simulation, operational complexity, and regulatory coordination.

Key Takeaways

• A platform trial is a persistent trial framework, not merely a multi-arm trial.
• A master protocol defines the overarching operational and statistical framework.
• Multiple treatment arms can be evaluated simultaneously.
• Shared controls can improve efficiency but create statistical dependence and require careful consideration of concurrent control and calendar-time effects.
• Treatment arms can be added or dropped according to prespecified rules.
• Interim analyses can be used to identify futile or successful treatments before maximum enrollment.
• Adaptive randomization is optional.
• Bayesian statistics are optional. Platform trials can be frequentist, Bayesian, or hybrid.
• Multiplicity must be addressed when multiple confirmatory hypotheses are evaluated.
• Simulation is often essential for demonstrating type I error, power, sample size, arm-selection properties, and trial duration.

References

Berry, S.M., Carlin, B.P., Lee, J.J. & Muller, P. (2010). Bayesian Adaptive Methods for Clinical Trials. CRC Press.

Park, J.W., Liu, M.C., Yee, D., et al. (2016). Adaptive randomization of neratinib in early breast cancer. Clinical Cancer Research.

Berry, D.A. (2015). The brave new world of clinical cancer trials. New England Journal of Medicine, 373, 1753–1755.

Saville, B.R. & Berry, S.M. (2016). Efficiencies of platform clinical trials: a vision of the future. Clinical Trials.

Woodcock, J. & LaVange, L.M. (2017). Master protocols to study multiple therapies, multiple diseases, or both. New England Journal of Medicine, 377, 62–70.

Angus, D.C. (2019). Fewer patients, more trials: a new paradigm for clinical trials. Critical Care Medicine.

Park, J.W., Liu, M.C., Yee, D., et al. (2020). Adaptive randomization of neratinib in early breast cancer. Clinical Cancer Research.

Hobbs, B.P., Landin, R. & Wang, L. (2021). Statistical considerations for platform trials. Clinical Trials.

Greenstreet, A., et al. (2023). Platform trial designs and their statistical considerations. Contemporary Clinical Trials.