When the Admissible Set Is Empty
A dose-escalation algorithm needs a defined response to a result that can occur after any cohort: none of the candidate doses meets the protocol’s overdose-control rule. Without a prespecified response, investigators might continue simply because a dose remains on the schedule, or might treat a statistical warning as an automatic command to close the entire study. Neither conclusion follows from the calculation alone.
The key distinction is between stopping escalation and deciding whether trial enrollment should pause or end. An empty admissible set means that the rule permits no candidate dose for the next allocation. A separate protocol decision determines whether to pause enrollment for review or terminate further treatment under the study. That decision should be made in advance, with patient safety, data quality, and the trial’s stopping criteria in view.
This rule is applied to the posterior distribution from the dose-toxicity model, such as the BLRM, under its specified prior and model assumptions. The analysis requires a defined candidate-dose grid, a consistent DLT definition and assessment window, and a fitted posterior that is suitable for the decision. The values \(p_T\) and \(\xi\) are protocol choices: \(p_T\) marks the toxicity level of concern, while \(\xi\) sets the tolerated posterior probability of exceeding it. They have different meanings and should not be interchanged.
If the dose set is finite, emptiness has a simple equivalent form. Every dose is inadmissible exactly when each overdose probability is at least the cutoff, or, equivalently, when the smallest of those probabilities is at least the cutoff:
The equivalence follows directly from the set definition: a dose belongs to \(\mathcal A(\mathcal D)\) only if its \(q_j\) is strictly below \(\xi\). If even the smallest \(q_j\) is not below \(\xi\), no dose can satisfy the membership condition. In particular, a dose with \(q_j=\xi\) is not admissible under this convention. If a protocol instead specifies a non-strict cutoff, that convention must be used consistently; the examples here use the strict rule above.
The Immediate Response and the Separate Trial Decision
Once the set is confirmed empty, the immediate allocation instruction is unambiguous: do not escalate to, de-escalate to, or assign a new cohort at any candidate dose under this rule. In particular, it is not valid to select the least risky dose merely because it has the smallest overdose probability. “Smallest” and “admissible” are different conditions.
Whether the study pauses or ends is a protocol-level decision, not a mathematical consequence of the empty set. A protocol can specify an immediate enrollment hold followed by a safety review. A review may be appropriate when information needed for the decision is being checked, such as whether the data were entered correctly, the intended posterior analysis was run, or the required assessment window was applied. If review confirms the result and the protocol’s stopping criteria are met, further enrollment may end. If a valid, authorized update changes the analysis, any resumption must use the newly reviewed result and the protocol’s rules—not an informal override.
A pause is not permission to keep treating participants at an inadmissible dose while waiting for a more convenient result. Management of participants already enrolled, including care and follow-up, is governed by the clinical protocol and safety oversight. Conversely, an empty set does not by itself establish that every dose is clinically unsafe in an absolute sense. It says that, given the model, prior, data, candidate doses, and chosen cutoff, none meets the prespecified rule.
Check the candidate doses, DLT definition and window, data used for the update, model fit, and protocol values for \(p_T\) and \(\xi\).
Calculate or verify \(q_j\) for every candidate and determine whether any is strictly below \(\xi\).
Do not choose a dose from the candidate list by ranking overdose probabilities alone.
Refer the result for the protocol-defined safety and data review; resume or terminate only as authorized under the protocol.
Worked Applications
Worked Example: No Candidate Meets the Cutoff
Hypothetical setting. After a cohort update, the BLRM analysis reports posterior overdose probabilities of \(0.31\), \(0.46\), and \(0.72\) at the three candidate doses \(d_1,d_2,d_3\). The protocol uses \(p_T=0.30\) and \(\xi=0.25\). Assume the data, model fit, and assessment windows have been checked and the reported probabilities are the intended decision quantities.
For each dose, compare \(q_j\) with \(\xi\):
The minimum is \(\min(q_1,q_2,q_3)=0.31\), which is also at least \(0.25\). Therefore \(\mathcal A(\mathcal D)=\varnothing\). The algorithm does not permit assigning the next cohort at any of these three doses. The result also does not mean that the posterior probability of toxicity is \(0.31\) at \(d_1\); it means the posterior probability that the DLT probability exceeds \(0.30\) is \(0.31\).
The model result triggers the protocol’s no-admissible-dose pathway. Investigators should hold new allocations and follow the specified review process. Whether that review leads to a temporary pause or permanent termination depends on the protocol and authorized safety decision, not on choosing \(d_1\) because it has the smallest \(q_j\).
Worked Example: A Dose Exactly at the Boundary
Hypothetical setting. A protocol has four candidate doses with posterior overdose probabilities \(0.12\), \(0.24\), \(0.25\), and \(0.51\). The cutoff is \(\xi=0.25\). Assume all four posterior calculations use the specified model and the same DLT threshold and assessment window.
Apply the strict membership condition \(q_j<\xi\). The first dose is admissible because \(0.12<0.25\). The second is admissible because \(0.24<0.25\). The third is not admissible because \(0.25\not<0.25\), and the fourth is not admissible because \(0.51\not<0.25\). Thus:
This set is not empty, so the no-admissible-dose stopping condition has not been met. The example illustrates why equality must be handled exactly as written in the protocol: treating \(q_j=\xi\) as admissible would change the algorithm. It also shows that the presence of one or more inadmissible doses does not imply that escalation must stop; the relevant question is whether at least one candidate is admissible.
Worked Example: Checking Monte Carlo Precision Before Declaring the Set Empty
Hypothetical setting. Posterior simulation gives estimated overdose probabilities of \(0.28\), \(0.43\), and \(0.66\) for three candidate doses. Each estimate is based on \(10{,}000\) effectively independent posterior draws. The cutoff is \(\xi=0.25\). Assume the chain diagnostics support using this effective sample size and the dose-toxicity model is otherwise appropriate.
For the first dose, suppose 2,800 of the 10,000 draws have \(p(d_1)>p_T\). Then \(\hat q_1=2{,}800/10{,}000=0.28\). The corresponding Monte Carlo standard error is approximately
An approximate 95% Monte Carlo interval is \(0.28\pm1.96(0.00449)\), or approximately \(0.271\) to \(0.289\). This interval describes simulation error in estimating \(q_1\), not uncertainty about the dose’s true toxicity probability. It lies above the cutoff \(0.25\). The other estimates, \(0.43\) and \(0.66\), are also above the cutoff. On these results, the estimated admissible set is empty, and the protocol’s stopping response applies.
The check matters because an estimate close to \(\xi\) can be affected by simulation error. If the smallest estimated probability were \(0.251\) but its Monte Carlo error were substantial, the analysis should not treat the third decimal place as decisive. The analyst should assess simulation precision using appropriate diagnostics and obtain sufficiently precise estimates under the prespecified analysis procedure before confirming the decision. Monte Carlo uncertainty must not be confused with a license to choose a dose that the decision rule classifies as inadmissible.
Diagnostics, Alternatives, and Limits
Before acting, confirm that the reported \(q_j\) values are posterior probabilities of the correct event \(p(d_j)>p_T\), rather than posterior mean toxicity probabilities or probabilities of falling outside a target interval. Also check that every candidate dose in the protocol-defined set was evaluated and that the same cutoff was applied to each. Model convergence, implausible posterior behavior, data errors, or inadequate simulation precision can undermine the calculation; they call for a valid analysis review, not an ad hoc change to the cutoff.
Observed DLT rates and posterior mean toxicity are not substitutes for this rule. An observed rate is based on the participants treated at a dose and may be unstable in a small cohort. A posterior mean, \(E[p(d_j)\mid\mathcal D]\), summarizes the center of the posterior for toxicity probability, whereas \(q_j\) measures posterior probability above the specified overdose threshold. A mean below \(p_T\) does not guarantee \(q_j<\xi\), and a mean above \(p_T\) does not by itself determine \(q_j\). The rule concerns posterior uncertainty about exceeding a threshold.
The cutoff is a decision criterion, not a biological boundary that makes a dose safe below it and unsafe above it. Its result depends on the model, prior, observed data, dose grid, \(p_T\), and \(\xi\). It does not measure efficacy, replace clinical judgment, or establish that participants already treated at a dose will experience toxicity. These limitations are reasons to interpret the result in context, not reasons to ignore a prespecified safety rule.
- Mistake: “The least risky candidate must be acceptable.” Ranking \(q_j\) identifies the smallest posterior overdose probability, but admissibility requires \(q_j<\xi\). If the minimum is at or above the cutoff, the set is empty.
- Mistake: “An empty set proves every candidate is unsafe.” The result is conditional on the model, prior, data, and threshold. The correct statement is that no candidate meets the protocol’s posterior overdose criterion at this decision point.
- Mistake: “An empty set automatically ends the whole trial.” It stops new assignments under the dose rule. Whether enrollment pauses or ends follows the protocol’s separate stopping and review provisions.
Check Your Understanding
Answer each question using the strict admissibility rule \(q_j<\xi\).
- With \(\xi=0.20\), candidate probabilities are \(0.19\), \(0.20\), and \(0.37\). Is the set empty?
- With \(\xi=0.25\), the minimum posterior overdose probability is \(0.28\). What does the result instruct the allocation algorithm to do?
- Why is it incorrect to say that an empty set proves all candidate doses are clinically unsafe?
- A team pauses enrollment after an empty set is identified. Does the statistical result alone require permanent termination? Explain.
- For a dose, the posterior mean of \(p(d)\) is below \(p_T\). Does that establish that the dose is admissible? Why or why not?
Solutions. (1) The set is not empty: \(0.19<0.20\), so the first dose is admissible; the dose exactly at \(0.20\) is not. (2) Since every \(q_j\) is at least the minimum \(0.28\), every candidate is at or above \(0.25\); the algorithm must stop new assignments at all candidates under this rule. (3) The conclusion is conditional on the fitted model, prior, data, candidate set, and chosen cutoff; it establishes failure to meet the prespecified criterion, not absolute clinical unsafety. (4) No. A pause or termination pathway is specified by the protocol and safety oversight; the empty set itself requires stopping allocation under the rule but does not dictate permanent trial closure. (5) No. Admissibility depends on \(P(p(d)>p_T\mid\mathcal D)\), not solely on the posterior mean.