Arm Dropping in a Multiarm Platform Trial

Why Drop an Arm?

A multiarm platform trial can evaluate several experimental treatments against a control while sharing trial infrastructure and, in many designs, a concurrent control group. If an experimental arm has little posterior support for a clinically worthwhile effect, continuing to assign participants to it may use trial resources without a reasonable prospect of benefit. A prespecified arm-dropping rule turns that concern into a reproducible interim decision.

The rule is not a declaration that a treatment has no effect. It is a decision to stop allocating new participants to that arm under a specified design, based on the evidence available at an interim analysis. The decision depends on the outcome definition, the clinical-benefit threshold, the prior distributions, the analysis timing, and the probability cutoff. These elements must be stated before applying the rule.

For the examples, use \(\delta=0.05\), meaning a five-percentage-point absolute increase in response probability, and \(\gamma=0.10\). Thus an arm is dropped if its posterior probability of exceeding control by more than five percentage points is below 10%. These values are illustrative design choices, not universal standards. A different clinical setting may require a different benefit margin or cutoff.

Model and Posterior Calculation

Suppose each participant has a binary response by a specified assessment time. For arm \(j\), let \(Y_j\) be the number of responses among \(n_j\) participants, and assume \(Y_j\mid p_j\sim\operatorname{Binomial}(n_j,p_j)\). Use independent beta priors \(p_j\sim\operatorname{Beta}(a_j,b_j)\) for the experimental arms and \(p_0\sim\operatorname{Beta}(a_0,b_0)\) for control. The analysis assumes a well-defined response outcome, correctly recorded response counts and denominators, and conditionally independent participant outcomes within each arm given its response probability. It also assumes no borrowing across arms in this simple model.

With \(y_j\) observed responses, the likelihood is proportional to \(p_j^{y_j}(1-p_j)^{n_j-y_j}\). Multiplication by the beta prior density gives a beta posterior:

The corresponding control posterior is obtained by the same update. Under independent priors and arm-specific binomial likelihoods, the posterior distributions for \(p_j\) and \(p_0\) are independent. The probability \(Q_j\) can be evaluated by numerical integration or simulation. A useful approximation, when both posterior distributions are reasonably concentrated and not dominated by boundary behavior, treats their difference as normal:

where \(\mu_j\) and \(v_j\) are the posterior mean and variance for \(p_j\), \(\mu_0\) and \(v_0\) are those for control, and \(\Phi\) is the standard normal cumulative distribution function. For a \(\operatorname{Beta}(a,b)\) distribution, the posterior mean is \(a/(a+b)\) and the variance is \(ab/[(a+b)^2(a+b+1)]\). The approximation should be checked against direct beta integration or posterior simulation when sample sizes are small or posterior mass lies near zero or one.

Applying the Rule to a Platform Snapshot

At an interim analysis, calculate the posterior distribution for each experimental arm and the common control using the data available at that analysis. Then calculate \(Q_j\) separately for each arm and compare it with \(\gamma\). The following examples use hypothetical data and the illustrative rule \(\delta=0.05,\ \gamma=0.10\).

Operational Decisions and Continuing Comparisons

Dropping an arm means ending future enrollment or randomization to that treatment according to the trial protocol. It does not erase the data already collected, change the control observations, or automatically stop other arms. Each remaining arm is assessed against its relevant control comparison using the rule and analysis plan specified for the platform.

When several arms share a control, each contrast \(p_j-p_0\) contains the same uncertain \(p_0\). Even if the arm-specific response probabilities are posterior independent, contrasts for two arms are not generally independent:

under the independent arm-posterior model. Thus arm-specific \(Q_j\) values answer individual comparison questions; they should not be treated as independent probabilities or combined as though the shared control were known without uncertainty. Dropping one arm does not, by itself, increase the posterior probability of benefit for another.

A posterior-probability futility rule is also distinct from predictive futility monitoring. Here the decision is based on the current posterior probability that the treatment effect exceeds a fixed clinical margin. Predictive monitoring instead concerns the probability of a future trial outcome under continued enrollment. The two criteria answer different questions and should not be substituted for one another.

Diagnostics, Limitations, and Common Mistakes

Before acting on the calculation, verify that the interim data are complete to the planned assessment time and that response definitions, denominators, and analysis populations are applied consistently across arms. Check that participants have not been counted twice, that pending outcomes are handled as specified, and that the control data correspond to the platform period and population relevant to the comparison. A change in control response rate over calendar time or a change in eligibility can undermine the interpretation of a pooled concurrent control unless the design and model address it.

The beta-binomial model assumes a single response probability within each arm for the analyzed population and period. Meaningful patient heterogeneity, differing follow-up, noncompliance, or a changing standard of care may make this simple model inadequate. Posterior probabilities are also sensitive to the prior when data are sparse. Examine prior choices and the sensitivity of \(Q_j\) to plausible alternatives, especially when the result is near \(\gamma\). If the normal approximation is questionable, evaluate \(Q_j\) from the beta posteriors directly, for example by numerical integration or posterior draws.

The posterior cutoff is not automatically a frequentist error-rate guarantee. Its operating characteristics depend on the design, including the number of arms, allocation, interim timing, prior distributions, and the response rates that might occur. A rule that is sensible for one platform may be too aggressive or too conservative for another. The cutoff and margin should be justified for the trial context rather than selected after seeing the interim results.

Common mistake: “The posterior mean difference is less than five percentage points, so the arm must be futile.” This is incorrect because the decision rule is based on \(P(p_j-p_0>0.05\mid\text{data})\), not on whether the posterior mean exceeds 0.05. An arm with a mean below the margin may still have enough posterior probability above it to continue, as arm B illustrates. The correction is to calculate the stated posterior probability and compare that probability with the prespecified cutoff.

Another error is to describe a dropped arm as “proven ineffective” or to interpret \(Q_j\) as the probability that the null hypothesis is true. Neither conclusion follows. \(Q_j\) is a posterior probability for a defined effect region under the stated model, priors, and observed data. The operational conclusion is limited to the rule: the arm does or does not meet the prespecified criterion for dropping.