Why Confirm a Selected Dose?
An escalation procedure identifies a dose for further consideration using the outcomes observed during escalation. That selection is a rule-based decision, not a guarantee that the dose's DLT rate is known precisely or that additional participants will have the same experience. A prespecified confirmation cohort can add DLT experience at the selected dose before the protocol recommends retaining it for subsequent study.
The confirmation step must be specified in advance: the number of participants to enroll, the DLT observation window, how an unevaluable participant is handled, and the count-based conditions for retaining or changing the recommendation. It is not an opportunity to invent a more favorable boundary after seeing the outcomes. The purpose is to apply a defined safety rule to additional experience at the same dose, while preserving the distinction between the original escalation decision and the later confirmation decision.
A participant belongs in a denominator only according to the protocol's evaluability definition. For example, a participant who has not completed the DLT window cannot automatically be counted as having no DLT. The protocol may specify replacement, continued follow-up, or a time-to-event rule; the confirmation calculation must use that rule rather than silently changing the denominator. Similarly, toxicity classification and attribution must follow the protocol's existing DLT definitions.
Prespecify the Recommendation Rule
There is no single confirmation boundary that is correct for every escalation design. One protocol might require a maximum number of DLTs in the confirmation cohort; another might use the cumulative count at the selected dose, or both. The rule should also state what follows a failure: for example, recommend the next lower dose, suspend dose use pending review, or stop dose assignment. The recommendation must respect the dose levels and actions that the protocol allows.
For illustration throughout this tutorial, consider a hypothetical protocol with the following rule. After a candidate dose has been validly selected by its escalation procedure, enroll six evaluable confirmation participants at that dose. Retain the dose only if the confirmation cohort has at most one DLT and the combined escalation-plus-confirmation count is at most two. In symbols, the retention conditions are \(y\leq 1\) and \(x+y\leq 2\). If either condition fails, do not retain the candidate under this rule; follow the protocol's specified alternative recommendation. These numbers are illustrative, not universal clinical standards.
The candidate must first have been selected validly. The confirmation rule does not erase or revise the escalation procedure's eligibility conditions. For instance, under the conventional 3+3 escalation procedure, a dose with two or more DLTs among three participants triggers de-escalation rather than selection of that dose as the recommended MTD. A confirmation rule cannot retroactively make such a dose an eligible candidate.
Worked Example: Retaining a Candidate With Limited Additional DLT Experience
Suppose a hypothetical escalation procedure has validly selected 60 mg as the candidate dose. Six evaluable participants were treated at 60 mg during escalation, with \(x=1\) DLT. The protocol enrolls \(m=6\) confirmation participants, among whom \(y=1\) has a DLT. Apply the illustrative retention rule.
First check the confirmation-only condition: \(y=1\), so \(y\leq 1\). Then check the cumulative condition: \(x+y=1+1=2\), and \(2\leq 2\). Both conditions hold, so the rule allows retaining 60 mg as the recommendation.
Across the two cohorts, the observed DLT proportion is \(\frac{x+y}{n+m}=\frac{1+1}{6+6}=\frac{2}{12}=0.1667\), or about 16.7% of the 12 evaluable participants. This descriptive proportion summarizes the combined observations; the decision above follows from the two prespecified count conditions, not from a claim that the true DLT probability is exactly 16.7%.
What the Combined Count Estimates
When the escalation and confirmation outcomes are combined descriptively, the relevant estimate is the observed DLT proportion \((x+y)/(n+m)\). Interpreting it as an estimate of one common DLT probability assumes that the participants are independent and have the same probability \(p\) of a DLT at this dose under the protocol-defined window and treatment conditions. In particular, the dose, DLT definition, follow-up window, and relevant administration conditions must be sufficiently consistent. The combined proportion is not automatically appropriate if those conditions differ.
A short likelihood argument explains the pooled estimate under those assumptions. Treat each participant's DLT indicator as a Bernoulli observation with probability \(p\). For \(N=n+m\) participants and \(k=x+y\) observed DLTs, the likelihood, up to a factor that does not depend on \(p\), is \(p^k(1-p)^{N-k}\). For \(0<k<N\), the log likelihood is \(k\log(p)+(N-k)\log(1-p)\). Differentiating and setting the derivative to zero gives the interior maximum:
The endpoint cases need separate treatment. If \(k=0\), then the likelihood is \((1-p)^N\), which is maximized at \(p=0\); if \(k=N\), then the likelihood is \(p^N\), maximized at \(p=1\). Thus the same pooled proportion is the maximum-likelihood estimate in those cases as well, although there is no interior derivative solution. This estimate describes the observed data under a common-probability model; it does not replace the protocol's count-based recommendation rule.
Worked Example: A Confirmation Rule Is Not Met
Suppose a candidate dose was validly selected, with \(n=6\) evaluable escalation participants and \(x=1\) DLT. In the six-person confirmation cohort, \(y=2\) participants experience DLTs. Apply the illustrative rule and calculate the pooled proportion.
The confirmation condition fails because \(y=2>1\). The cumulative count is \(x+y=1+2=3\), which also exceeds the permitted maximum of two. Therefore, the protocol's rule does not allow the candidate dose to be retained. The protocol's prespecified alternative action, such as recommending the next lower dose if available, determines what happens next; the count alone should not be used to invent an action absent from the protocol.
The pooled observed proportion is \(\frac{1+2}{6+6}=\frac{3}{12}=0.2500\), or 25.0%. It provides a concise summary of the observed DLT experience if pooling assumptions are reasonable. The rejection of the candidate here follows from the prespecified count boundaries, not from comparing 25.0% with an unstated target.
Eligibility, Boundaries, and a Critical Distinction
A dose can satisfy the arithmetic of a confirmation rule without being a valid candidate for confirmation. These are separate questions: (1) was the dose eligible under the escalation procedure, and (2) given a valid candidate, do the additional outcomes satisfy the confirmation rule? Always answer the first question before applying the second.
Worked Example: A Boundary Case and Candidate Eligibility
For a direct boundary check, suppose a validly selected candidate has \(n=6\), \(x=2\), and the confirmation cohort has \(m=6\), \(y=0\). Under the stated illustrative retention rule, the confirmation count satisfies \(y=0\leq1\), and the combined count satisfies \(x+y=2+0=2\leq2\). Therefore, the stated confirmation rule allows retention, assuming the candidate was validly selected.
That conditional assumption matters. Under the conventional 3+3 procedure, a dose with \(x=2\) DLTs among six participants would not ordinarily be the selected highest dose meeting the MTD rule. If that conventional procedure governs candidate selection, this particular combination of candidate status and \(x=2\) cannot arise as a valid selected candidate. It remains a useful check of the confirmation rule's arithmetic, but it must not be used to bypass escalation eligibility. If a different escalation procedure validly selected a candidate with these data, the stated confirmation conditions themselves still allow retention.
The combined observed proportion in this arithmetic example is \(\frac{2+0}{6+6}=\frac{2}{12}=0.1667\). Neither this proportion nor the zero confirmation count changes the eligibility question: the protocol's escalation rule determines whether the dose was a valid candidate in the first place.
Diagnostics, Alternatives, and Limitations
Before applying the rule, check that the dose assignment and candidate status are documented, all participants are classified using the same DLT definition, and each included participant has the required follow-up or is handled under the protocol's incomplete-follow-up rule. Verify the denominators and counts separately for escalation and confirmation before calculating \(x+y\). A disagreement between these counts and the source listings, an unresolved DLT classification, or a change in dose or observation window is a reason to resolve the data or follow the protocol's review procedure, not to proceed with a convenient denominator.
The count-based rule and the pooled estimate answer different questions. The rule gives an operational recommendation at the specified boundary. The pooled proportion summarizes observed toxicity frequency under a common-probability assumption. A separate estimate for each cohort can reveal a difference in observed experience, but small cohorts can vary substantially by chance; a difference alone does not establish that the dose's risk changed. If enrollment conditions or follow-up differ materially, report the cohort-specific outcomes and follow the protocol's decision rule rather than treating the pooled proportion as a valid common-rate estimate.
The main alternative is to make the recommendation directly from the escalation data without enrolling a confirmation cohort. That avoids delaying the decision and using additional participants at the candidate dose, but supplies no added DLT experience before the recommendation. A confirmation cohort adds observations and can expose a concerning count, while requiring more participants, time, and a clearly prespecified response. It does not guarantee a precise estimate of the DLT probability, and it does not turn a rule-based design into a model-based dose-finding method.
A confirmation cohort is also not a license to continue dosing after the protocol's safety boundary has been crossed. Rules for pausing enrollment, managing a suspected safety signal, or responding to incomplete observation remain in force. Confirmation outcomes support the protocol's dose recommendation; clinical and oversight responsibilities continue independently.
Common Mistakes
- Adding outcomes before checking candidate eligibility. The confirmation rule applies only after valid selection under the escalation procedure. It cannot repair an invalid selection.
- Changing the rule after observing \(y\). Applying a newly chosen threshold makes the recommendation outcome-dependent in a way the prespecified procedure did not authorize. Use the stated boundary and document any protocol amendment separately.
- Treating the pooled proportion as the rule. A proportion such as 2/12 is descriptive; it does not replace conditions such as \(y\leq1\) and \(x+y\leq2\).
- Assuming zero confirmation DLTs always means retention. The cumulative condition can still fail. Conversely, assess the actual written conditions rather than adding a restriction that is not in them.
- Counting incomplete observation as no DLT. Use the protocol's evaluability and follow-up rules, especially when the DLT window has not closed.
Check Your Understanding
Answer each question using the illustrative rule in this tutorial: six confirmation participants, retain only if \(y\leq1\) and \(x+y\leq2\). Then compare with the solutions.
- A valid candidate has \(x=0\) among six escalation participants. The confirmation cohort has \(y=2\). Is the dose retained?
- A valid candidate has \(x=1\) and \(y=1\), with six evaluable participants in each cohort. Calculate the pooled observed DLT proportion and state whether the rule permits retention.
- For \(x=2\) and \(y=0\), do the confirmation conditions permit retention? What additional eligibility check is required?
- Name two conditions that make pooling escalation and confirmation outcomes as observations of one common DLT probability questionable.
Solutions. (1) No. Although \(x+y=0+2=2\) meets the cumulative boundary, \(y=2>1\) fails the confirmation-only condition. (2) The pooled proportion is \((1+1)/(6+6)=2/12=0.1667\), or 16.7%. Both \(y=1\leq1\) and \(x+y=2\leq2\), so the rule permits retention, assuming valid candidate selection. (3) Yes, the written confirmation conditions are met: \(0\leq1\) and \(2+0=2\leq2\). One must still check that the escalation procedure validly selected this candidate; under conventional 3+3 selection, two DLTs among six at that dose would generally make it ineligible as the selected dose. (4) Examples include a changed DLT definition or observation window, materially different treatment administration, or incomplete follow-up handled differently. If conditions differ, use the protocol's specified rule and avoid presenting the pooled proportion as an estimate under a common-probability model without justification.