Target-Trial Specification for an Added Platform Arm

Why Specify a Target Trial?

When a platform trial opens a new arm, the phrase “compare the new treatment with control” is not yet a complete causal question. Which patients qualify when the arm opens? What exactly is the control strategy? When does follow-up begin? Which outcome is measured, and at what time? Different answers can define different populations and different treatment effects—even when the treatment name stays the same.

A target-trial specification makes those choices explicit by describing the randomized trial that would answer the question of interest. For an added platform arm, it describes the hypothetical randomized comparison among people eligible for that arm during its relevant enrollment period. This specification is useful even when the available analysis later includes nonrandomized or historical control observations: it defines the comparison those data would need to represent. It does not, by itself, establish that they do represent it.

Definition: A target trial is the precisely specified hypothetical randomized trial that defines a causal question. For an added platform arm, its core specifications are eligibility criteria, treatment strategies, assignment procedure and time zero, follow-up, outcomes, and the causal contrast. The target population is the set of people who meet the stated eligibility criteria in the stated enrollment period.

Specify the Trial Before Comparing Outcomes

The specification should be detailed enough that two analysts could identify the same eligible participants, assign the same treatment strategies, start follow-up at the same point, and calculate the same outcome contrast. In a platform setting, calendar time is part of that description: an arm that opens later may enroll people under different eligibility rules or background care than earlier arms. State the relevant calendar window rather than treating “eligible patients” as a timeless group.

Eligibility.
List inclusion and exclusion criteria and when they must be satisfied. Include the disease subtype, relevant clinical features, and the calendar-time window in which the new arm is open.

Treatment strategies.
Define the new-arm strategy and the control strategy. State treatment versions, co-interventions, and what happens after discontinuation or progression when these affect interpretation.

Assignment and time zero.
Specify how eligible participants are assigned in the hypothetical trial and identify the shared start of treatment assignment, eligibility, and outcome follow-up.

Follow-up and outcome.
State the follow-up horizon, outcome definition, measurement schedule, and how death, loss to follow-up, or other intercurrent events are handled.

Causal contrast.
State the target population, treatment strategies being compared, outcome summary, time horizon, and effect measure, such as a risk difference at 12 months.

A useful notation makes the contrast explicit. Let \(Z=1\) denote assignment to the new-arm strategy and \(Z=0\) assignment to control. For an eligible participant, let \(Y^z(\tau)\) be the outcome that would be observed by the prespecified horizon \(\tau\) under assignment \(z\). For a binary outcome, one possible target is the risk difference:

\Delta(\tau) = P\{Y^1(\tau)=1\mid S=1\} - P\{Y^0(\tau)=1\mid S=1\},

where \(S=1\) indicates membership in the target population. The sign and units need to be stated in words. For example, a risk difference of \(-0.10\) is a 10-percentage-point lower risk under assignment to the new strategy than under assignment to control by time \(\tau\). It is not a 10% relative reduction.

Worked Example: Defining an Added Oncology Arm

Suppose a hypothetical platform opens a targeted therapy arm for adults with a specified cancer subtype. Write a target-trial specification for a 12-month binary outcome.

Eligibility. Include adults with the specified subtype, the biomarker required for the therapy, measurable disease, and no prior exposure to the study drug. Require these criteria at the time the patient could be randomized. Define the target enrollment window as the period when the new arm is open, for example January through December of a stated year. The target population is not all patients ever enrolled in the platform; it is those meeting these criteria in that window.

Strategies and assignment. Compare assignment to the named targeted therapy plus protocol-defined background care with assignment to the platform’s protocol-defined control care plus the same background-care policy. Assume the hypothetical trial randomizes eligible participants 1:1. The strategies should state whether assigned treatment is continued after progression or stopped for toxicity; otherwise “targeted therapy” may conceal materially different treatment policies.

Time zero, follow-up, and outcome. Set time zero at randomization, when eligibility is confirmed and assignment occurs. Follow each participant for 12 months. Define the outcome as death from any cause by 12 months, with death counted as an event and outcome status sought even if assigned therapy is discontinued.

Contrast. The estimand is the 12-month risk difference for death under assignment to targeted therapy versus control among the eligible patients entering during the specified opening period. If the hypothetical randomized trial had 18 deaths among 100 assigned to therapy and 25 among 100 assigned to control, the observed risks would be \(18/100=0.18\) and \(25/100=0.25\). The estimated difference is \(0.18-0.25=-0.07\), or 7 percentage points lower 12-month mortality under therapy assignment in this hypothetical sample. This calculation illustrates the chosen contrast; it does not establish that the true effect is exactly \(-0.07\).

Why Time Zero Must Be Shared

Eligibility, assignment, and the start of follow-up must refer to the same point in the target trial. If follow-up begins earlier for controls than for new-arm participants, participants may have to remain alive or event-free until they receive treatment to enter the treated group. That creates a period in which the groups are defined differently. A clearly stated time zero avoids this design mismatch in the hypothetical trial.

In a platform, “time zero” is not the date the arm opened for everyone. It is the individual participant’s assignment date (or the corresponding date at which that participant would have been assigned in the target trial). The calendar date still matters for defining the enrollment window and the care environment, but it is distinct from the participant-level start of follow-up.

Worked Example: Matching Eligibility and Follow-Up Dates

In a hypothetical added arm, two eligible participants are assessed and assigned on March 3 and March 20. The protocol asks whether each participant dies within 12 months after assignment.

For the March 3 participant, time zero is March 3 and the outcome window ends 12 months later. For the March 20 participant, time zero is March 20 and the outcome window ends 12 months after that date. Each person’s eligibility is assessed at their own assignment, and follow-up begins at that same assignment.

Now suppose an analyst instead starts follow-up for the new-arm group on the date treatment is first received, which is March 10 for the first participant, but starts control follow-up on the assignment date, March 3. The new-arm participant must remain observable until March 10 to be classified as treated, while the control participant does not face that same requirement. The comparison then does not implement the specified assignment-based trial. The correction is to define assignment as the strategy allocation and use the same assignment time to start follow-up in both groups; treatment initiation and deviations are recorded after that point.

What a Randomized Target Trial Identifies

The target trial’s randomized assignment gives a simple justification for comparing outcome risks. Assume the trial randomizes among eligible participants, the assigned strategies are sufficiently well-defined, and the observed outcome under the received assignment equals the potential outcome for that assignment (consistency). Also assume one participant’s assignment does not change another participant’s outcome (no interference). With complete outcome ascertainment through \(\tau\), randomization gives \(Z\) independent of the potential outcomes within the eligible population. Therefore:

P\{Y(\tau)=1\mid Z=z,S=1\} = P\{Y^z(\tau)=1\mid S=1\}, \qquad z\in\{0,1\}.

The left side is the outcome risk observed among those randomized to \(z\); the right side is the risk under assignment to strategy \(z\) in the target population. Subtracting the two group risks identifies the specified risk difference. If outcomes can be missing before \(\tau\), this equality is not automatically justified by randomization alone; the specification must say how outcome ascertainment and any censoring are addressed.

This reasoning describes the hypothetical randomized comparison. A platform analysis using controls who were not randomized at the same time is not made randomized by writing down the target trial. Its causal interpretation requires additional assumptions about whether those controls can stand in for the target trial’s control outcomes. The protocol specification clarifies the question; it does not prove those assumptions.

Worked Example: A Contrast Depends on the Outcome Definition

A hypothetical target trial follows 100 participants per assignment group. By 6 months, 12 participants in the new-arm group and 20 in control have progressed. By 12 months, the corresponding counts are 30 and 40. Define progression as a binary event by the stated horizon.

For the 6-month outcome, the risks are \(12/100=0.12\) and \(20/100=0.20\), so the risk difference is \(0.12-0.20=-0.08\), an 8-percentage-point lower 6-month progression risk under assignment to the new strategy.

For the 12-month outcome, the risks are \(30/100=0.30\) and \(40/100=0.40\), so the difference is \(0.30-0.40=-0.10\), a 10-percentage-point lower 12-month progression risk. These are distinct time-specific causal contrasts, not inconsistent calculations. The outcome, horizon, population, and strategy must be specified before results are interpreted.

Diagnostics, Alternatives, and Limitations

A well-specified target trial gives practical checks for a platform comparison. Review whether the eligibility rules can be applied at a known date, whether treatment and control are described at comparable levels of detail, and whether each outcome has a clear definition and observation window. Check that time zero is identifiable for every participant and that the arm-opening calendar window is explicit. If these items cannot be established from the available records, the intended comparison may not be reproducible.

The target trial is a design framework, not an estimation method. For a genuinely randomized comparison, the randomized groups can estimate the specified assignment effect, subject to outcome ascertainment and the usual trial conduct conditions. For an observational comparison, analysts may instead compare the new arm with concurrent controls or with nonconcurrent control observations. Concurrent controls may better align calendar context, while nonconcurrent controls may provide more observations but require stronger justification for representing the target control strategy and population. Neither choice changes the target question; neither choice is guaranteed to identify it.

The specification also has limits. It cannot compensate for vague or changing eligibility criteria, irretrievably missing outcome data, or treatment strategies that combine clinically different versions without explanation. Nor does it remove uncertainty about whether a comparison population is suitable. It does make these issues visible early enough to revise the question or acknowledge the limits of the evidence.

Common mistake: treating arm opening as each participant’s time zero. Arm opening defines a calendar window; participant follow-up begins at that participant’s assignment.
Common mistake: defining the new treatment but leaving control vague. “Usual care” can change over time. Specify the control policy and its relevant version for the target period.
Common mistake: reporting an effect without a horizon. A 6-month risk difference and a 12-month risk difference answer different questions. Name the outcome and time point with the estimate.
Common mistake: assuming the target-trial description validates historical controls. It defines the hypothetical randomized comparison. Additional evidence and assumptions are needed to justify a nonrandomized substitute.

Key takeaway: For an added platform arm, specify who would be randomized, to which well-defined strategies, at what participant-level time zero, for what follow-up and outcome, and to which time-specific causal contrast. This defines the question before any available control group is judged as an answer to it.

Check Your Understanding

Answer each question using a clearly stated target-trial specification.

A new arm enrolls patients from July through December. Is July 1 automatically time zero for every participant? Give the appropriate distinction.
In a randomized target trial, 24 of 120 participants assigned to the new strategy and 36 of 120 assigned to control have the outcome by 1 year. Calculate and interpret the risk difference.
Why should discontinuation of assigned treatment be addressed in the treatment-strategy description?
What does specifying a target trial establish, and what does it not establish when the available controls are historical?

Solutions. (1) No. July through December is the arm’s calendar enrollment window. Each participant’s time zero is their own assignment date, when eligibility is confirmed and follow-up begins. (2) The risks are \(24/120=0.20\) and \(36/120=0.30\). The risk difference is \(0.20-0.30=-0.10\), meaning a 10-percentage-point lower 1-year risk under assignment to the new strategy in the hypothetical randomized comparison. (3) Discontinuation affects what strategy assignment means. A policy that continues follow-up and counts outcomes after discontinuation differs from an analysis that only considers outcomes while treatment is taken. (4) The specification defines the population, strategies, time zero, outcome, and causal contrast of the hypothetical randomized comparison. It does not establish that historical controls are comparable to the target trial’s control group or that their outcomes identify the causal contrast.