Adding New Arms to an Ongoing Platform

Why Arm Entry Needs Its Own Design

An ongoing platform trial may open a treatment arm after other arms have already enrolled participants. That change is not simply the addition of a treatment label: it changes which assignments are possible for future participants, when evidence about the new treatment begins to accumulate, and which control participants can provide a contemporaneous comparison. A sound entry plan makes these details explicit before the first participant can be assigned to the new arm.

The key distinction is between the platform’s calendar time and a participant’s randomization time. A participant randomized before an arm opens could not have received that treatment. Such participants do not become members of the new arm retrospectively, and their outcomes cannot be treated as randomized comparisons with it. The new arm’s randomized evidence starts at its activation.

Specify the Entry Rule Before Activation

The protocol should state the activation date or operational condition, the initial randomization probabilities, the applicable participant eligibility criteria, and the control assignment available during the new arm’s enrollment. It should also specify whether assignment is stratified, whether allocation is blocked or otherwise constrained, and what minimum information is needed before the new arm may enter an allocation update.

These are linked decisions. If an arm opens but its allocation probability is too small to produce usable enrollment, it may remain nominally active without generating timely evidence. If its probability is increased using unstable early outcomes, however, the first randomization period can have disproportionate influence on later assignments. An initial allocation rule should therefore be defined in advance, including any minimum probability or minimum control allocation required by the design.

For example, suppose control \(C\) and treatment \(A\) are open before treatment \(B\) is added. Before \(B\) activates, only \(C\) and \(A\) can be assigned. At and after activation, eligible participants may be assigned among \(C\), \(A\), and \(B\), according to the new rule. The probabilities for the earlier period remain what they were; the opening does not rewrite past assignments.

Concurrent Control and Calendar Time

An ongoing trial may change over calendar time: supportive care can improve, eligibility practices can shift, or outcome measurement can change. Concurrent control helps protect a new-arm comparison from changes that affect participants enrolled in the same period. It does not guarantee that every time-related bias is eliminated, but randomized assignment among eligible participants during overlapping enrollment makes the control comparison more credible than an unadjusted comparison with participants recruited earlier.

A short justification makes the role of overlap clear. Suppose the mean outcome in calendar period \(k\) for arm \(j\) is represented as \(E(Y\mid j,k)=\mu+\delta_k+\lambda_j\), where \(\delta_k\) is a period effect common to arms and \(\lambda_j\) is the arm effect. Comparing new treatment \(B\) with control \(C\) within the same period gives \(E(Y\mid B,k)-E(Y\mid C,k)=\lambda_B-\lambda_C\). The shared period effect cancels. If instead the new arm’s mean in period \(k\) is compared with a historical control mean from period \(h\), the difference is \((\lambda_B-\lambda_C)+(\delta_k-\delta_h)\). Unless the period effects are equal or adequately addressed, the historical contrast includes the time difference as well as the treatment contrast.

This argument assumes that the period effect is meaningfully common across arms and that trial conduct, outcome definitions, and eligibility remain sufficiently compatible for the comparison. Concurrent controls are not a remedy for changes that affect arms differently, unrecorded changes in eligibility, or a control regimen that has itself changed. Such changes may require separate design or analysis provisions.

When Can the New Arm Join Adaptation?

Opening an arm and allowing it to influence later allocation are separate decisions. An arm can begin with a fixed, protected allocation rule and become eligible for response-adaptive updates only after a prespecified interim condition is met. This condition should be defined in terms of information that is available and relevant to the update, such as a minimum number of observed outcomes or a planned review after enough outcomes have matured. Participants with pending outcomes should not be treated as if their outcomes were already observed.

The entry policy should identify whether the new arm is included in an update at the first scheduled review after activation, only after it meets a minimum information threshold, or not until a later review. It should also say how the control arm is treated, how the update uses the new arm’s evidence, and whether probability floors or other allocation constraints apply. These details keep the adaptation rule reproducible and prevent an arm from being rewarded or penalized on a handful of unusually early responses.

A Bayesian update is one possible basis for an allocation score, but the score-to-probability rule is a design choice, not a consequence of having a posterior distribution. The following example uses posterior means solely to illustrate the arithmetic of a prespecified rule. It does not prescribe posterior-mean allocation as the generally preferred policy.

Checks at the Transition

Before and after activation, the trial team should verify that the randomization system and operating procedures implement the planned transition. Useful checks include confirming the activation timestamp, the eligible arms shown for each participant, the probabilities actually used, and the number of participants assigned to each arm by calendar period. A mismatch between the protocol and the system—for example, a new arm appearing before its opening date or remaining unavailable after activation—can undermine the intended randomization.

Review the evidence available at each adaptation time as well. If outcomes are delayed, a count of enrolled participants is not necessarily a count of observed outcomes. The rule should state how pending outcomes are handled and apply the same definition consistently. Monitor whether the new arm’s enrollment is sufficient for planned comparisons and whether control enrollment during its open period is adequate. These checks assess whether the design is functioning; they do not justify changing the rule informally in response to emerging results.

Design Choices and Limitations

A fixed initial allocation after activation is transparent and can protect the new arm from being affected by very sparse early outcomes. Immediate inclusion in an adaptation rule can respond more quickly to accumulating evidence, but the resulting probabilities may be unstable when little information is available. A minimum-information gate offers a middle course: it delays influence on allocation until a prespecified threshold is reached, at the cost of keeping the initial rule in place longer.

Whatever the choice, an allocation probability is not a guarantee of a particular sample size. Random variation, changing participant eligibility, and the timing of activation all affect realized enrollment. Nor does a concurrent control automatically solve every comparability problem: eligibility or outcome measurement changes can still complicate interpretation. The entry plan should therefore be considered together with the trial’s estimand, calendar-time practices, and planned analyses, without assuming that a later statistical adjustment can repair an unclear assignment process.

The most reliable approach is to document four items before opening: when the arm becomes available, which participants can receive it, what probability it receives initially, and what explicit evidence makes it eligible for later adaptation. A clear concurrent-control plan completes that specification.