Why Boundary and Saturation Diagnoses Matter
A continual reassessment method (CRM) uses observed dose-limiting toxicities (DLTs) and a dose-toxicity model to recommend a dose whose toxicity risk is near a prespecified target. A recommendation at the lowest or highest available dose is not, by itself, evidence that the dose range is wrong. It may reflect the observed outcomes, the fitted curve, the target’s location relative to that curve, or restrictions on which dose is currently eligible. A recommendation can also be weakly determined when fitted risks at several doses are almost indistinguishable.
These situations call for different responses. If the fitted risks are all above the target, selecting the lowest dose follows directly from the recommendation rule; if all are below, selecting the highest follows. If risks are nearly equal, the selected dose may be sensitive to small changes in data or model assumptions. Neither diagnosis alone establishes which dose is clinically appropriate. The observed outcomes and uncertainty in the fitted curve must also be considered.
The definition gives the unconstrained recommendation. In an actual trial, the operational recommendation may be limited to doses that are eligible under the protocol, including its dose-skipping constraints. It is useful to inspect both: the unconstrained fit helps diagnose what the model is saying, while the operational rule determines what the trial may do. The posterior-mean rule is not the only possible Bayesian decision rule. For example, minimizing posterior expected squared deviation, \(E[(p_j(\theta)-\phi)^2\mid D]\), can produce a different choice because it incorporates posterior variability as well as distance of the posterior mean from the target.
Why a Target Outside the Fitted Risk Range Selects an Endpoint
Assume that the fitted risk estimates are nondecreasing across ordered doses, so \(\hat p_1\leq\hat p_2\leq\cdots\leq\hat p_J\). This ordering is typical of a monotone CRM dose-toxicity model. If \(\phi\) is below every fitted risk, the lowest dose is closest; if it is above every fitted risk, the highest dose is closest. This is a property of the decision rule, not proof that the true target dose lies outside the studied range.
The argument for a target above the fitted range is analogous: when \(\phi\geq\hat p_J\), each \(\hat p_j\leq\hat p_J\), so \(|\hat p_j-\phi|=\phi-\hat p_j\geq\phi-\hat p_J=|\hat p_J-\phi|\). If the target falls inside the range, the dose closest to it is selected, with ties requiring a prespecified tie-breaking rule. This reasoning concerns the posterior means under the specified model. It does not account for clinical constraints unless the set of eligible doses is explicitly restricted.
Compare \(\phi\) with the smallest and largest fitted risks, \(\hat p_1\) and \(\hat p_J\). This identifies whether the fitted range brackets the target.
Review cohort outcomes, posterior uncertainty, and how the fitted curve changes across doses. A boundary selection with very uncertain risks needs a different interpretation from a precise fit entirely on one side of the target.
Check sensitivity to reasonable skeleton or prior choices and confirm that operational eligibility rules have not changed the recommendation. Do not extrapolate beyond the dose range solely because its endpoint is selected.
Worked Boundary Diagnoses
Worked Example: Repeated Selection of the Lowest Dose
Hypothetical scenario. A CRM uses four ordered doses and a target DLT probability of \(\phi=0.25\). At two successive interim looks, its posterior mean DLT estimates are, respectively, \((0.28, 0.41, 0.56, 0.70)\) and \((0.31, 0.44, 0.58, 0.72)\). At the later look, the cumulative outcomes are 1 DLT among 6 patients at dose 1, 3 among 6 at dose 2, 4 among 6 at dose 3, and 5 among 6 at dose 4. For this example, treat the posterior estimates as the output of the stated CRM fit to the hypothetical data.
At the later look, the absolute distances from the target are
Thus, dose 1 is the closest fitted risk, and the target is below even the lowest fitted risk. The earlier estimates likewise place every dose above target, so the repeated lowest-dose recommendation is not an arithmetic anomaly. The substantial observed DLT counts, including 1 of 6 at the lowest dose, are consistent with concern about toxicity, although the data are sparse. The next diagnostic question is whether posterior uncertainty is small enough to support the conclusion that risks across the dose range exceed target. If uncertainty is wide, the recommendation remains conditional on limited evidence; if it is concentrated and the result is robust to reasonable model choices, the available range may indeed be too toxic for this target. The example does not justify treating an unstudied lower dose as safe.
Worked Example: Repeated Selection of the Highest Dose
Hypothetical scenario. At an interim analysis, four dose levels have each been given to 3 patients, with no observed DLTs at any level. A CRM fit gives posterior mean risks of \(0.04, 0.08, 0.14,\) and \(0.21\); the target is \(\phi=0.25\). These are hypothetical posterior summaries from a specified CRM fit, not estimates from a real trial.
The target distances are
The highest dose is recommended because its fitted risk is nearest the target, but all four risks are below target. The zero-DLT observations are compatible with this direction of fit, yet 3 patients per dose provide limited information about toxicity. The recommendation does not show that the highest dose is sufficiently safe, nor does it establish that the target is attained at a higher, unstudied dose. Review posterior uncertainty and the protocol’s overdose-control and escalation rules before interpreting the endpoint choice. If the trial cannot safely study doses beyond the current maximum, the highest-dose recommendation is a boundary signal, not a licence to extrapolate.
Diagnosing Saturation and Weak Dose Separation
Boundary selection is about the target’s position relative to the fitted risk range; saturation is about how much the fitted risks differ across doses. They may occur together, but they need not. A curve can be saturated near the target, making several doses nearly tied, or saturated entirely below target, in which case the highest dose may still be selected. Conversely, a well-separated curve can select an endpoint because all its risks lie on one side of the target.
A practical first check is to calculate adjacent differences, \(\Delta_j=\hat p_{j+1}-\hat p_j\), and the full fitted range, \(\hat p_J-\hat p_1\). Small differences suggest little separation on the probability scale. But “small” is contextual: a difference that is negligible for the decision may still matter clinically, and posterior mean differences alone do not show how uncertain the comparisons are. Examine posterior uncertainty in the risks and contrasts, as well as whether modest changes in the data or reasonable model specifications change the recommended dose.
Worked Example: Nearly Tied Risks Across the Dose Range
Hypothetical scenario. A CRM fit to sparse hypothetical outcomes produces posterior mean DLT risks \((0.191, 0.196, 0.201, 0.206)\) at four increasing doses. The target is \(\phi=0.200\). The estimated standard deviations of the three adjacent posterior risk differences are \(0.012, 0.012,\) and \(0.013\), respectively.
The absolute distances from target are \(0.009, 0.004, 0.001,\) and \(0.006\). The posterior-mean matching rule therefore selects dose 3. Yet each adjacent fitted increase is only \(0.005\), and the full fitted range is \(0.206-0.191=0.015\). Each adjacent mean increase is also smaller than its stated posterior standard deviation: \(0.005/0.012=0.417\), \(0.005/0.012=0.417\), and \(0.005/0.013=0.385\). On these summaries, dose 3 is the mathematical choice under the stated rule, but the apparent advantage over dose 2 is only \(0.004-0.001=0.003\) in target-distance units.
The appropriate diagnosis is weak practical separation, not that the model has discovered a clearly superior dose. The posterior uncertainty in adjacent differences makes the small estimated increments difficult to distinguish from their uncertainty. Inspect whether the fitted risks change materially under reasonable skeleton or prior alternatives, and whether new observations would plausibly sharpen the distinctions. A monotone model can impose the ordering of fitted risks even when data provide little evidence about the size of the differences; monotonicity by itself is not evidence of strong separation.
What to Check Before Changing the Dose Range
The same endpoint recommendation can arise for different reasons. The following checks help distinguish them without treating a single fitted curve as definitive.
- Observed outcomes: Summarize patients and DLTs by dose, including the amount of exposure at each level. A few patients with no DLTs are not equivalent to extensive follow-up without DLTs.
- Target relative to fitted risks: Establish whether all posterior mean risks lie above the target, all lie below, or the target is bracketed. Report the actual distances, not just the selected dose.
- Posterior uncertainty: Determine whether the fitted endpoint risks and the relevant risk contrasts are precise enough for the proposed interpretation. Do not infer precision from a smooth curve or a single point estimate.
- Model and skeleton sensitivity: Check whether modest, clinically defensible changes to the CRM skeleton or prior materially alter the fitted risk range or recommendation. Prior predictive calibration is relevant when prior assumptions appear to dominate sparse outcomes.
- Operational restrictions: Confirm whether dose eligibility or overdose-control rules have removed a dose from consideration. A constrained recommendation may differ from the unconstrained fit and should be described accordingly.
- Data-model tension: Compare observed dose-level outcomes with the fitted curve. A monotone model may smooth nonmonotone observed rates; a discrepancy is a reason to examine sparse sampling and model fit, not to assume the observations themselves must increase at every dose.
A useful comparison is between model-based fitted risks and raw dose-level proportions. Raw proportions describe the observed cohorts directly, but can be highly variable when cohort sizes are small; the CRM borrows information across ordered doses under its model assumptions. Neither view should be substituted uncritically for the other. A decision rule based on posterior risk matching also differs from a rule that minimizes posterior expected loss or imposes a separate safety constraint. State which rule generated the recommendation before diagnosing why it selected an endpoint.
Common causes of boundary or saturation behavior include a target outside the fitted risk range, sparse or unusually uniform outcomes, a skeleton whose risks are compressed at the studied doses, and limited information about the model parameter. Model misspecification is another possibility: a one-parameter curve may be unable to represent the actual dose-risk pattern. These are diagnostic possibilities, not conclusions that follow automatically from a boundary selection. If changing the dose range is considered, it must be justified by trial design, clinical evidence, and safety oversight, rather than by mechanically extending a fitted curve.
Check Your Understanding
Answer each question, then compare with the solutions below.
- For fitted risks \(0.30, 0.42,\) and \(0.61\) with target \(0.25\), which dose is selected by posterior-mean matching, and what does that imply?
- For fitted risks \(0.05, 0.12,\) and \(0.22\) with target \(0.20\), is the recommendation a boundary selection? Calculate the selected dose’s target distance.
- Why is a dose recommendation based on fitted risks \(0.198, 0.200,\) and \(0.202\) potentially weakly determined when the target is \(0.200\)?
- Give one reason a CRM may select an endpoint repeatedly even when the data do not strongly identify the target dose.
Solutions. 1. The distances are \(0.05, 0.17,\) and \(0.36\), so the lowest dose is selected. All fitted risks exceed target, but this alone does not establish that the true risks do. 2. Distances are \(0.15, 0.08,\) and \(0.02\), so the highest dose is selected; it is a boundary recommendation, and its fitted risk is \(0.02\) above target. 3. The adjacent fitted differences are each \(0.002\), so small changes in the fit can change the closest dose; uncertainty and sensitivity should be examined before treating the selected level as clearly preferable. 4. For example, sparse outcomes can leave fitted risks strongly influenced by the skeleton or prior, placing all posterior means on one side of target and repeatedly selecting an endpoint without precise identification of the target dose.