Why a CRM Recommendation May Need a Safety Restriction
A continual reassessment method (CRM) uses accumulated toxicity outcomes to estimate dose-specific toxicity probabilities and identify a dose near the target toxicity rate. That statistical recommendation is not, by itself, a complete dosing policy. A protocol can require a dose to be excluded when its posterior probability of being too toxic is unacceptably high. It can also restrict escalation for operational reasons, such as a toxicity in the most recently treated cohort or an untried dose between the current dose and a proposed higher dose.
These safeguards answer different questions. A posterior safety exclusion asks what the model implies about a dose given the data and prior. A no-escalation rule asks which transitions the trial permits, regardless of whether the model would otherwise recommend them. Both rules must be specified before their application to a particular patient or cohort; neither should be improvised in response to a preferred recommendation.
The comparison is often made to the target \(\phi\), but a protocol may instead define a separate unacceptable-toxicity boundary. The boundary and the cutoff must be stated explicitly. The distinction between \(>\) and \(\geq\) also matters: a posterior mass at exactly the boundary is not counted when the rule is strictly \(>\). A protocol should settle this convention in advance.
Model Conditions and the Logic of Excluding Higher Doses
For the one-parameter power CRM, let \(q_j\) denote the skeleton probability at dose \(j\), with \(0<q_1<q_2<\cdots<q_J<1\). The model is \(\pi_j(\alpha)=q_j^\alpha\), with \(\alpha>0\). The prior and likelihood produce a posterior distribution for \(\alpha\). This model assumes that the skeleton orders doses by toxicity and that the dose-toxicity curve is monotone over the studied grid. The safety calculations below use this model and its posterior; they do not make an exclusion rule valid if the model or data processing is inappropriate.
Monotonicity gives a useful result. For every fixed \(\alpha>0\), if \(k>j\), then \(q_k>q_j\), so \(q_k^\alpha>q_j^\alpha\). Consequently, whenever \(\pi_j(\alpha)>\phi\), it is also true that \(\pi_k(\alpha)>\phi\). The event for dose \(j\) is contained in the corresponding event for dose \(k\), and therefore
Thus, under a monotone dose-toxicity model and a common boundary and cutoff, if a dose is excluded, every higher dose is excluded as well. This is a posterior statement: it follows because the same posterior draws of \(\alpha\) generate the probabilities at all dose levels. It should not be assumed when dose-specific models do not enforce monotonicity.
Worked Example: Calculating Posterior Safety Exclusions
Worked Example: Calculating Posterior Safety Exclusions
Consider a hypothetical decision look with target \(\phi=0.25\) and safety cutoff \(c=0.25\). The CRM skeleton is \(q=(0.05,0.12,0.25,0.40,0.55)\). Suppose the posterior distribution has three possible values of \(\alpha\):
| \(\alpha\) | Posterior probability |
|---|---|
| 0.5 | 0.20 |
| 1 | 0.50 |
| 2 | 0.30 |
These are hypothetical posterior weights from the model update; the calculation below applies the stated cutoff to those weights. At dose 3, the modeled toxicity probabilities are \(0.25^{0.5}=0.5000\), \(0.25^1=0.2500\), and \(0.25^2=0.0625\). Only the first is strictly greater than \(0.25\), so
Dose 3 is not excluded. At dose 4, the corresponding probabilities are \(0.40^{0.5}=0.6325\), \(0.40^1=0.4000\), and \(0.40^2=0.1600\). The first two exceed the target, giving
Dose 4 is excluded. At dose 5, the probabilities are approximately \(0.7416\), \(0.5500\), and \(0.3025\). All exceed \(0.25\), so \(T_5=1.00\) and dose 5 is excluded. This agrees with monotonicity: once dose 4 fails, the still-higher dose 5 fails too.
No-Escalation Rules Are Operational Constraints
A no-escalation rule must say exactly which transition it prohibits. Common protocol choices include forbidding the next cohort from skipping an untried dose, prohibiting escalation immediately after one or more DLTs in the current cohort, or limiting the next assignment to at most one dose level above the highest dose with adequately assessed outcomes. These are not interchangeable rules. The protocol must specify the relevant cohort, the DLT condition, the required follow-up, and what assignments remain permitted.
For example, a rule might state: “If any participant in the most recently treated cohort has a DLT, the next cohort may not receive a dose higher than the dose just administered.” This rule blocks upward movement after that cohort’s DLT; it does not, unless separately stated, require de-escalation, mandate treating again at the same level, or automatically stop the trial. A different protocol may impose one of those additional actions.
Worked Example: Applying a No-Escalation Rule After a DLT
In a hypothetical trial, the current cohort of three participants receives dose 1, and one participant experiences a DLT. Suppose the protocol’s rule is that any DLT in the most recently treated cohort prohibits escalation above that cohort’s dose at the next assignment. Assume the posterior weights and skeleton from the preceding example summarize all data available at this look, including this cohort.
The posterior mean toxicity probability at each dose is obtained by weighting the probability at each \(\alpha\) value by its posterior probability. At dose 3, for example,
The corresponding posterior means at doses 1 through 5 are approximately \(0.0705,\ 0.1336,\ 0.2438,\ 0.3745,\ 0.5141\). With target \(0.25\), dose 3 is closest in posterior mean to target, so it is the unconstrained candidate under a rule that selects the dose whose posterior mean toxicity is nearest to \(\phi\). But the latest cohort had a DLT at dose 1. Under the stated no-escalation rule, the next cohort cannot be assigned above dose 1. The candidate dose 3 is therefore not assignable at this look.
Dose 1 is not excluded by the posterior safety rule calculated in the preceding example, so it remains among the doses allowed by the two rules as stated. The protocol’s final action could still be a hold, de-escalation, or stopping if another protocol rule requires it. The no-escalation restriction alone does not choose among those actions.
Combining Safety Exclusions and Transition Limits
The final decision should be made in a defined sequence. First apply the posterior safety rule to identify excluded levels. Then apply operational restrictions, such as prohibiting a dose skip or an escalation after a recent DLT. Among the doses still permitted, use the protocol’s CRM selection rule. If the permitted set is empty, follow the prespecified hold or stopping instruction rather than selecting an excluded dose by default.
Use outcomes with the protocol-required toxicity follow-up to obtain the posterior for \(\alpha\).
Calculate \(T_j\) for each relevant dose and exclude every dose that fails the stated criterion.
Remove choices that violate the no-escalation, no-skip, or other prespecified operational rules for this decision.
Apply the CRM selection rule to the remaining choices; if none remain, use the protocol’s specified hold or stopping action.
Worked Example: Combining Both Restrictions at One Decision Look
Suppose a hypothetical trial has the skeleton, target, posterior weights, and cutoff used in the first example. The highest dose with completed assessment is dose 2. The most recent cohort at dose 2 has no DLT, and the protocol allows escalation by only one level at a time. Assume the dose-selection rule chooses the posterior mean toxicity probability closest to \(0.25\).
From the weighted calculations, the posterior mean probabilities are approximately \(0.0705,\ 0.1336,\ 0.2438,\ 0.3745,\) and \(0.5141\). Dose 3 is closest to target, since \(|0.2438-0.25|=0.00625\). Dose 4 and dose 5 are excluded because their posterior overdose probabilities are \(0.70\) and \(1.00\), respectively. Dose 3 is exactly one level above the highest assessed dose, and the most recent cohort had no DLT, so the stated transition rule permits moving to dose 3. The final allowable recommendation is therefore dose 3.
If the most recent cohort at dose 2 had instead had a DLT and the protocol barred escalation after any such DLT, dose 3 would be blocked even though it passed the posterior safety cutoff. The outcome would depend on the remaining permitted options and the protocol’s instructions, not on changing the posterior calculation.
Diagnostics, Alternatives, and Limitations
At each look, verify that the posterior probabilities use the same toxicity definition, target, boundary, and cutoff as the protocol. Check that the likelihood includes only outcomes whose assessment is sufficiently complete under the trial’s follow-up rules; treating participants without adequate follow-up as if they had no DLT can make a safety calculation misleading. Confirm that the posterior weights sum to one, that the modeled curve is monotone if higher-dose exclusions are propagated upward, and that the strict or inclusive boundary convention has been implemented correctly.
A common implementation failure is to apply the overdose cutoff to posterior mean toxicity rather than to posterior probability. In general, \(E\{\pi_j(\alpha)\mid\text{data}\}\) and \(P\{\pi_j(\alpha)>\phi\mid\text{data}\}\) are different summaries and can lead to different decisions. Another failure is to treat a no-escalation rule as though it were evidence that the higher dose is unsafe. It is an assignment restriction, not a posterior update.
The main alternative is to use only a model-based target-dose rule, with no additional posterior exclusion or transition restriction. That is simpler, but it does not enforce an explicit posterior tolerance for overdose or a conservative operational constraint. Conversely, adding stringent restrictions can slow dose finding, leave few permissible doses, or prevent a dose that appears acceptable under the model from being tested. Their effect depends on the chosen target, cutoff, prior, skeleton, cohort size, and follow-up policy. These safeguards do not guarantee that every treated participant avoids severe toxicity, nor do they correct model misspecification.
Check Your Understanding
Answer each question using the definitions and decision rules above. Brief solutions follow the questions.
- If a dose has posterior overdose probability \(0.25\) and the protocol excludes only when \(T_j>c\), with \(c=0.25\), is it excluded?
- Under a monotone CRM curve, if dose 3 is excluded by a common posterior overdose cutoff, what can be concluded about doses 4 and higher?
- A cohort has a DLT at its assigned dose. Does a rule forbidding escalation after any cohort DLT automatically require stopping the trial?
- Why is a posterior mean toxicity below target insufficient to establish that a dose passes a posterior overdose criterion?
Solutions. (1) No. The rule is strict, and \(0.25\not>0.25\); an inclusive rule would give a different result. (2) All higher doses are excluded under the stated monotonicity and common-threshold assumptions, because their posterior overdose probabilities are at least as large. (3) No. The rule blocks escalation as specified; stopping, holding, or de-escalating requires an additional protocol instruction. (4) The mean does not determine the posterior tail probability. The dose can have a mean below target while a substantial fraction of posterior probability lies above target, so the tail probability must be calculated directly.