When the Eligible-Dose Set Is Empty
A posterior overdose-control rule can identify doses whose estimated risk is acceptable under a prespecified criterion. Sometimes it identifies none. This is not simply an awkward ranking problem: if every candidate fails the rule, choosing the “best” candidate does not make that dose compliant. The trial team must decide whether to stop, pause and retain the current dose as an administrative reference, or consider a lower dose under an explicitly authorized fallback.
The distinction matters because the posterior overdose probability is a measure of uncertainty about a dose’s toxicity probability. It is not itself the probability that the next participant will experience toxicity. A failed criterion therefore signals that the planned rule does not permit that dose; it does not, by itself, establish that toxicity is certain or that every participant would be harmed.
The inequality is inclusive: a dose with posterior overdose probability exactly equal to \(\xi\) passes. The toxicity limit \(p_T\), the posterior threshold \(\xi\), the dose candidates, and the model must be specified before applying the rule. Changing one of them after seeing that no dose passes changes the decision rule; it is not merely a calculation correction.
Why the Rule Has No Dose to Return
Suppose the dose-selection task is to choose an eligible dose that also has a desirable property, such as being close to a target toxicity probability. That task is a constrained selection problem: the chosen dose must belong to \(\mathcal{A}(D)\). If \(\mathcal{A}(D)\) is empty, there is no dose that meets the constraint. A ranking of the candidates can still be calculated, but it cannot turn the top-ranked candidate into a passing dose.
This follows directly from the definition. A compliant recommendation \(d^*\) would have to satisfy \(d^*\in\mathcal{A}(D)\). But if \(\mathcal{A}(D)=\varnothing\), no such \(d^*\) exists. The decision must therefore change something other than the conclusion that the current eligible set is empty: stop or pause, collect and verify information before reassessment, alter the candidate dose set under an authorized procedure, or explicitly invoke a fallback that relaxes the rule.
The result is about the specified model, data, dose grid, and cutoff. Before treating it as a clinical decision, verify that those inputs are correct. For example, confirm the toxicity endpoint and assessment window, the dose labels, the toxicity-limit value and units, and the value of \(\xi\). Check that the posterior calculation used the intended data and that the model’s computation is reliable. If posterior draws are used, check convergence and Monte Carlo precision, especially when an estimate is near \(\xi\). A computationally unstable estimate should prompt diagnosis and a justified reassessment, not an unannounced change to the threshold.
Three Responses, Three Different Meanings
| Response | What it means | Relationship to the failed rule |
|---|---|---|
| Stop the trial | End enrollment and do not assign additional participants under the current design. | Consistent with having no eligible candidate; appropriate when a safe path forward is not justified. |
| Retain the current dose and pause | Keep the current level as a reference while enrollment or escalation is suspended and the evidence is reviewed. | A pause is not permission to treat more participants at a dose that failed control. |
| Select a lower dose | Choose a lower level because an authorized fallback permits it, or because the protocol is formally revised. | If that dose also fails the prespecified criterion, it is not a passing dose; the choice is an explicit exception, not a compliant selection. |
“Retain the current dose” is especially easy to misread. It can mean retaining the dose as the current reference while no new participants are assigned, or it can mean continuing to treat new participants at that level. The first is a hold; the second is continued dosing. If the current dose fails the overdose-control criterion, continued dosing is not justified by the criterion itself. It requires a separate, prespecified rule or a documented decision through the trial’s safety and governance process.
Likewise, “select a lower dose” does not guarantee that the lower dose passes. If the trial’s candidate set includes lower levels and none passes, choosing the lowest posterior risk among them is a least-risk fallback, not a finding of acceptable risk. A protocol may instead define a lower dose outside the original grid, but that is a change to the candidate set and should be governed and documented accordingly.
Worked Example: Stop When Every Candidate Fails
Hypothetical setting. A trial has three candidate doses. Its prespecified toxicity limit is \(p_T=0.30\), and its posterior overdose-control threshold is \(\xi=0.25\). Posterior calculations from 10,000 draws of the fitted dose-toxicity model give 3,100, 4,200, and 5,700 draws above \(p_T\) at the low, middle, and high doses, respectively.
For the low dose, $$ P(p_1>0.30\mid D)\approx\frac{3100}{10000}=0.310. $$ For the middle and high doses, the corresponding estimates are \(4200/10000=0.420\) and \(5700/10000=0.570\). Each is greater than \(0.25\), so all three doses fail and \(\mathcal{A}(D)=\varnothing\).
Decision. If the protocol does not define a lower-dose exception and the safety review does not support continuing, stopping enrollment is a clear response. The trial does not select the low dose and describe it as safe: its estimated posterior overdose probability is \(0.310\), above the allowed \(0.25\). The conclusion is limited to the prespecified control rule and the model update; it is not a claim that toxicity will occur in every patient.
Pausing and Retaining the Current Dose
A pause can be useful when the no-pass result is unexpected, when data or model calculations need verification, or when the responsible review group must decide whether an authorized next step exists. During the pause, the current dose can remain a reference point for interpreting the evidence. The pause should specify what is suspended, what information or review is pending, and what conditions would permit enrollment to resume.
A pause is not a substitute for the safety rule. If the current dose fails, enrolling another cohort at that dose while calling the action “retention” still exposes participants at a dose that the rule excludes. If the protocol explicitly authorizes continued treatment in this situation, that is a distinct fallback policy and its rationale and limits should be stated as such. Do not assume that an earlier dose assignment, or a desire to avoid interrupting the trial, overrides the control criterion.
Worked Example: Retain the Dose as a Reference, Not an Assignment
Hypothetical setting. The current dose is \(d_2\). With \(p_T=0.25\) and \(\xi=0.20\), a model update based on 10,000 posterior draws estimates overdose probabilities of \(0.270\) at \(d_1\), \(0.350\) at \(d_2\), and \(0.520\) at \(d_3\). For example, at \(d_2\), 3,500 of the 10,000 draws have \(p_2>0.25\), giving \(3500/10000=0.350\).
All three estimates exceed \(0.20\), including the estimate \(0.270\) at the lowest candidate dose. The current dose fails by \(0.350-0.200=0.150\) on the posterior-probability scale. The low dose also fails by \(0.270-0.200=0.070\).
Decision. The team may retain \(d_2\) as the reference dose while pausing new enrollment, checking the data and model update, and obtaining the required safety review. It may not infer that \(d_2\) remains eligible merely because it was the current dose. If the review confirms the inputs and no authorized exception applies, enrollment cannot resume at \(d_2\) under this control rule.
Choosing a Lower Dose Only as an Explicit Fallback
A lower-dose fallback may be considered when the protocol or a properly authorized amendment defines how to proceed despite an empty eligible set. The choice should use a stated rationale, such as choosing the lower candidate with the smallest posterior overdose probability, and should report that the dose still fails the original criterion. Other requirements, including any review or consent procedures, remain governed by the trial protocol. Lower is not synonymous with safe, and a posterior ranking is not an authorization to treat.
If a dose is chosen under an exception, report both the original criterion and the residual evidence against it. Relevant summaries include \(P(p_j>p_T\mid D)\) at the selected dose, the posterior mean of \(p_j\) when useful, and the difference from the current dose. These summaries answer different questions: a posterior overdose probability measures uncertainty that \(p_j\) exceeds \(p_T\), while the posterior mean describes the average of \(p_j\) over its posterior distribution. Neither should be described as the observed toxicity rate among future participants.
Worked Example: A Lower-Dose Exception Is Not a Passing Result
Hypothetical setting. The current dose is \(d_2\), and \(p_T=0.25\), with \(\xi=0.20\). From 10,000 posterior draws, the estimated probabilities of exceeding the limit are \(0.340\) at \(d_1\), \(0.460\) at \(d_2\), and \(0.640\) at \(d_3\). The posterior means of \(p_1\) and \(p_2\) are \(0.240\) and \(0.300\), respectively.
The lowest candidate dose has the smallest estimated overdose probability, but \(0.340>0.20\), so it fails the original criterion by \(0.340-0.200=0.140\). Its estimated overdose probability is \(0.460-0.340=0.120\) lower than at the current dose. Its posterior mean toxicity probability is \(0.300-0.240=0.060\) lower than the current dose’s posterior mean.
Decision. If an authorized fallback permits selecting the lowest-risk candidate, the team could consider \(d_1\) under that fallback. The conclusion must remain precise: the summaries favor \(d_1\) over \(d_2\) on these measures, but \(d_1\) still fails the prespecified overdose-control threshold. The \(0.120\) difference is a difference between posterior probabilities of the parameter exceeding the limit; it is not a 12-percentage-point reduction in an individual patient’s probability of toxicity.
Common Pitfalls and Practical Checks
- Calling the least risky dose safe. A dose can rank first and still fail. Report whether it meets the actual inequality \(P(p_j>p_T\mid D)\leq\xi\).
- Treating a pause as continued assignment. Holding a dose as a reference does not permit new enrollment at that dose unless a separate, authorized policy allows it.
- Changing the cutoff after seeing the result. Raising \(\xi\) until a dose passes weakens the original control rule. Any change requires a justified, governed revision, not an ad hoc calculation.
- Confusing posterior uncertainty with patient-level risk. \(P(p_j>p_T\mid D)\) is a probability about the unknown dose-specific toxicity probability, conditional on the data and model.
- Ignoring model and computation checks. Verify the endpoint, data, dose coding, model fit, posterior computation, and Monte Carlo precision. A diagnosed error can justify correcting the analysis; it does not justify quietly changing the decision criterion.
Stopping, holding, and selecting a lower dose are not interchangeable ways to express the same action. Stopping ends the current trial path; holding suspends assignments while preserving a reference point; selecting a lower dose proceeds under an exception if that dose does not pass. A defensible decision records which action is taken, why it is authorized, what evidence supports it, and whether the original safety criterion is satisfied.
Check Your Understanding
Answer each question using the definitions and decision distinctions above. Short solutions follow the questions.
- For \(\xi=0.20\), three candidate doses have posterior overdose probabilities \(0.18\), \(0.20\), and \(0.43\). Is the eligible-dose set empty?
- All doses fail the threshold. What is the difference between retaining the current dose during a pause and continuing to enroll at it?
- A lower dose has the smallest posterior overdose probability, but that probability remains above \(\xi\). Can it be described as passing the control rule?
- In a posterior-draw calculation, 2,450 of 10,000 draws at a dose exceed \(p_T\). Calculate the estimated posterior overdose probability and determine whether the dose passes when \(\xi=0.24\).
- Why is \(P(p_j>p_T\mid D)\) not the same quantity as the probability that the next participant experiences toxicity?
Solutions. (1) No. The dose at \(0.20\) passes because the criterion uses \(\leq\xi\), so the eligible set is not empty. (2) Retaining the dose during a pause means it remains a reference while assignments are suspended; continuing enrollment assigns new participants at a dose that failed the rule and needs separate authorization. (3) No. It can be considered only under an explicit fallback or revision, and must still be reported as failing the original criterion. (4) The estimate is \(2450/10000=0.245\). Since \(0.245>0.24\), the dose fails. (5) The posterior quantity is the probability, given the data and model, that the unknown toxicity probability \(p_j\) exceeds the limit; the next participant’s toxicity is an individual outcome whose probability, conditional on \(p_j\), is \(p_j\).