What an Arm-Level Enrollment Rule Decides
A platform trial can stop enrolling participants to one experimental arm without stopping enrollment to its control arm or to other experimental arms. This distinction matters because an arm-level rule governs where future participants may be assigned; it does not, by itself, determine whether participants already enrolled stop treatment, whether their outcomes continue to be collected, or whether the platform as a whole closes.
A stopping rule should state the decision, the evidence used, and when the decision is evaluated. For example, a protocol might close enrollment to arm \(a\) at a scheduled interim analysis if the posterior probability that its response rate exceeds the concurrent control response rate by a clinically important amount is sufficiently small. The control and other eligible arms may continue enrolling. The platform-level decision is separate and requires its own rule.
The rule must also distinguish permanent closure from temporary suspension. If a temporary pause is allowed, the protocol needs to say what evidence permits reopening. If enrollment is permanently closed, the arm is not reopened under that rule. Neither choice follows automatically from a posterior probability threshold.
A Posterior Rule for One Experimental Arm
For a concrete formulation, suppose the endpoint is a binary response, with response preferred to nonresponse. Let \(\theta_a\) be the response probability on experimental arm \(a\), and let \(\theta_0\) be the response probability on the concurrent control. At an interim analysis, let \(x_a\) of \(n_a\) participants on arm \(a\), and \(x_0\) of \(n_0\) control participants, have observed responses.
Assume, for this basic model, that outcomes are Bernoulli within each arm, participants’ outcomes are conditionally independent given the arm-specific response probabilities, and the outcome definition and follow-up window are the same across arms. Assign independent beta priors, \(\theta_a\sim\mathrm{Beta}(u_a,v_a)\) and \(\theta_0\sim\mathrm{Beta}(u_0,v_0)\). These priors and the binomial likelihood give independent beta posteriors within this two-arm comparison:
Choose a clinically important absolute response-rate difference \(\delta_E>0\). Define the posterior probability of clinically important benefit as \(q_a=P(\theta_a-\theta_0>\delta_E\mid\text{data})\). One possible prespecified rule closes enrollment to arm \(a\) for insufficient evidence of benefit if \(q_a\leq c_F\), and closes enrollment for sufficient evidence of benefit if \(q_a\geq c_E\), where \(c_F<c_E\). Values between the boundaries imply that the arm remains open, subject to other rules in the protocol. A separate harm boundary can use an adverse-event probability and a clinically important increase in harm.
These cutoffs are design choices, not universal constants. They need to be selected for the trial’s purpose and assessed before the trial using operating-characteristic simulations. In a platform with a shared control, each arm’s comparison uses control information, but the resulting arm-level decisions are statistically dependent through that shared information.
Why the Posterior Calculation Has This Form
The rule uses a posterior probability about the difference between two response probabilities. The beta-binomial update follows directly from multiplying the prior density by the binomial likelihood. For one arm, omitting factors that do not depend on \(\theta_a\), the prior contributes \(\theta_a^{u_a-1}(1-\theta_a)^{v_a-1}\), and the likelihood contributes \(\theta_a^{x_a}(1-\theta_a)^{n_a-x_a}\). Therefore:
which is the kernel of \(\mathrm{Beta}(u_a+x_a,v_a+n_a-x_a)\). The same argument applies to control. The probability \(q_a\) is then obtained from the joint posterior for \((\theta_a,\theta_0)\), for example by numerical integration or posterior simulation. A normal approximation can be useful for illustration, but it is not the definition of the rule.
For a beta distribution with shape parameters \(A,B\), the mean is \(A/(A+B)\) and the variance is \(AB/[(A+B)^2(A+B+1)]\). When both posterior distributions are reasonably well approximated by normal distributions, independence implies that the posterior difference has approximate mean equal to the difference of posterior means and variance equal to the sum of posterior variances. Thus an approximate calculation is:
where \(m_a,m_0\) and \(s_a^2,s_0^2\) are the two posterior means and variances, and \(\Phi\) is the standard normal cumulative distribution function. This approximation should be checked against the beta-posterior calculation, especially with sparse data or posterior mass near zero or one.
Worked Example: Closing an Arm for Futility
Suppose a hypothetical platform evaluates response at a scheduled interim analysis. Use independent \(\mathrm{Beta}(1,1)\) priors, set \(\delta_E=0.05\), and specify \(c_F=0.10\). There are 8 responses among 40 participants on arm \(a\), and 18 among 40 concurrent controls. Assume the response data are complete for the participants included in this analysis.
The posteriors are \(\theta_a\mid\text{data}\sim\mathrm{Beta}(9,33)\) and \(\theta_0\mid\text{data}\sim\mathrm{Beta}(19,23)\). Their means are \(9/42=0.2143\) and \(19/42=0.4524\), so the posterior mean difference is \(0.2143-0.4524=-0.2381\).
The posterior variances are \(9(33)/(42^2\cdot43)=0.003915\) and \(19(23)/(42^2\cdot43)=0.005761\). The approximate standard deviation of the difference is \(\sqrt{0.003915+0.005761}=0.09837\). Therefore:
Because \(0.0017<0.10\), this rule closes enrollment to arm \(a\) for futility. That action does not imply that the platform stops: control and other arms may continue under their own eligibility and stopping rules. The 40 participants already enrolled on arm \(a\) remain in the trial’s follow-up unless a separate safety or treatment policy says otherwise.
Worked Example: Closing an Arm for Evidence of Benefit
Consider a separate hypothetical interim analysis with the same \(\mathrm{Beta}(1,1)\) priors, \(\delta_E=0.10\), and efficacy enrollment boundary \(c_E=0.90\). Arm \(a\) has 30 responses among 40 participants; control has 18 among 40. For illustration, assume complete outcomes and use the normal approximation to the posterior difference.
The posteriors are \(\mathrm{Beta}(31,11)\) for arm \(a\) and \(\mathrm{Beta}(19,23)\) for control. The posterior means are \(31/42=0.7381\) and \(19/42=0.4524\), giving a difference of \(0.2857\). The variances are \(31(11)/(42^2\cdot43)=0.004496\) and \(19(23)/(42^2\cdot43)=0.005761\). Thus the difference’s approximate standard deviation is \(\sqrt{0.004496+0.005761}=0.1013\).
Since \(0.967>0.90\), the specified rule closes enrollment to arm \(a\) for evidence of benefit. This is an operational enrollment decision under the stated boundary; it is not, by itself, a claim that the treatment is proven effective for every population or that the entire platform should end. A confirmatory or reporting decision may require additional criteria.
Worked Example: Closing an Arm for a Safety Signal
Suppose a separate hypothetical safety review uses a binary adverse event, for which a higher probability is worse. Use \(\mathrm{Beta}(1,1)\) priors for the adverse-event probabilities, set the important excess-harm margin to \(0.10\), and close enrollment to the experimental arm if the posterior probability of an excess above that margin is at least \(0.95\). There are 15 events among 30 participants on the experimental arm and 5 among 30 controls.
The posteriors are \(\mathrm{Beta}(16,16)\) and \(\mathrm{Beta}(6,26)\), with means \(16/32=0.5000\) and \(6/32=0.1875\). The estimated posterior mean difference is \(0.3125\). The variances are \(16(16)/(32^2\cdot33)=0.007576\) and \(6(26)/(32^2\cdot33)=0.004616\); the difference’s approximate standard deviation is \(\sqrt{0.007576+0.004616}=0.1104\).
Because \(0.973>0.95\), the arm meets this hypothetical harm boundary and enrollment to it is closed. The action is arm-specific; the platform’s other arms and control may continue if their own rules and safety oversight permit. For an actual design, the posterior probability should be computed using the prespecified analysis method rather than relying on this approximation without validation.
Operational Assumptions and Diagnostics
A posterior threshold alone is not a complete operating procedure. The protocol should specify the interim schedule, the response window, the analysis population, prior distributions, clinical margins, thresholds, minimum information requirements, and whether the decision is reversible. It should define which participants contribute at each analysis and how data corrections, missing outcomes, and delayed outcome ascertainment are handled.
In particular, participants whose outcomes are not yet observed must not silently be counted as nonresponders. One defensible rule is to calculate the boundary using only outcomes observed by the analysis cutoff, with a minimum number of evaluable participants before a decision is permitted. Another is to model pending outcomes explicitly. Either approach must be prespecified and checked by simulation, because outcome delay can make the observed cases unrepresentative of those still pending.
Useful diagnostics include the number enrolled and evaluable on each arm at every look, the amount of pending outcome data, sensitivity to reasonable prior choices, and the frequency and timing of arm closures under simulated scenarios. Simulations should include no benefit, clinically important benefit, harm, different accrual and delay patterns, and the possibility of multiple arms meeting a boundary at the same look. In a shared-control design, simulations should preserve the common control data; treating arm-level posterior decisions as independent can mischaracterize how often several arms stop.
The simple model also assumes that the endpoint is measured comparably across arms and that the likelihood is appropriate. Temporal changes in patient mix, nonadherence, differential missingness, or changes in outcome assessment can undermine that interpretation. If the model or data are poorly calibrated, a precise posterior probability can still be misleading. Sensitivity analyses and data-quality monitoring are therefore part of responsible use of the rule, not optional substitutes for it.
How This Rule Differs From Other Decisions
| Approach | Decision it targets | Key distinction |
|---|---|---|
| Arm-specific posterior enrollment boundary | Whether to enroll more participants to one arm | Other platform arms can continue; posterior evidence is compared with prespecified boundaries. |
| Whole-platform stopping rule | Whether the platform as a whole should stop | Requires a platform-level criterion; one arm’s boundary does not automatically trigger it. |
| Fixed enrollment target | How many participants to enroll before planned closure | Does not close an arm early in response to accumulating outcomes unless an additional rule is added. |
| Predictive-probability rule | Whether future enrollment is likely to achieve a specified end-of-trial goal | Uses predictions about future data and a stated final criterion, rather than only the current posterior probability of a treatment contrast. |
Posterior efficacy or futility rules for dropping arms provide related decision frameworks, but an enrollment-closure rule must explicitly state its operational consequence. The key distinction is not just the posterior calculation: it is whether the action closes one arm to new assignments, removes it under a broader arm-management policy, or ends the platform. These actions should not be treated as interchangeable.
Common Pitfalls and Limitations
- Confusing an arm decision with a platform decision. A low posterior probability for one arm does not, by itself, support stopping the whole trial. Apply the rule to the named arm and evaluate platform continuation separately.
- Treating a posterior threshold as a universal standard. A boundary such as 0.90 is a design choice. Its consequences depend on the margin, prior, analysis schedule, and data-generating conditions, so assess the complete design by simulation.
- Ignoring pending outcomes. Unobserved responses are not known failures. Specify how pending outcomes enter the decision and check whether delays could bias the observed interim data.
- Interpreting enrollment closure as proof. The rule defines an action under a model and prespecified criterion. It does not make a posterior probability a guarantee of benefit, nor does a futility decision prove that benefit is impossible.
- Relying uncritically on a normal approximation. Sparse or extreme binary data can yield skewed beta posteriors. Compute the posterior probability directly or validate the approximation before using it operationally.
An arm-specific rule is useful when the platform can learn efficiently from continuing control and other experimental arms while no longer assigning participants to an arm whose evidence crosses a prespecified boundary. Its limitation is that the decision is only as reliable as the endpoint, model, prior, interim-data handling, and operating-characteristic evaluation behind it.
Check Your Understanding
Answer each question, then compare with the solutions.
- If enrollment to one experimental arm is closed for futility, what can happen to participants already enrolled and to the rest of the platform?
- With \(\delta_E=0.08\), a posterior probability \(P(\theta_a-\theta_0>0.08\mid\text{data})=0.06\), and \(c_F=0.10\), what does the specified futility rule decide?
- Why should pending response outcomes not automatically be counted as nonresponses at an interim analysis?
- In a platform with a shared control, why can arm-level stopping decisions be dependent?
Solutions. (1) Enrollment to that arm stops under the rule, but follow-up of already enrolled participants can continue; control, other arms, and the platform can continue if their rules permit. (2) Since \(0.06\leq0.10\), the arm meets the futility boundary and enrollment to that arm closes. (3) A pending outcome is unobserved, not a known failure; counting it as a nonresponse can distort the interim evidence, especially when outcome timing varies. (4) The same control outcomes contribute to each arm-versus-control comparison, so posterior contrasts and decisions share information and need not be independent.