What an Overdose-Probability Threshold Controls
A CRM model represents uncertainty about toxicity at each dose using a posterior distribution. An escalation rule can use that distribution to screen doses: a dose is considered sufficiently safe for escalation only when its posterior probability of exceeding a prespecified toxicity limit is not too large. The threshold sets how much posterior overdose risk the rule is willing to tolerate.
This threshold is a decision parameter, not an estimate learned automatically from the current patients. A smaller threshold applies a stricter safety screen. It can exclude doses that a more permissive threshold would allow, potentially limiting exploration of higher doses. A larger threshold may allow more doses to be considered, but accepts a greater posterior probability that an eligible dose is above the toxicity limit.
The inequality is inclusive: a dose with \(q_j=\delta\) passes. Passing the screen means that, under the fitted model and current data, the posterior probability of \(p_j>\phi\) is at most \(\delta\). It does not by itself say which dose the trial must assign next. The screen is one component of a dose-selection rule; other prespecified rules may determine which passing dose is selected.
Choosing a Threshold Through Decision Loss
A useful way to understand a threshold is to state the consequences being balanced. Consider a simplified choice for one dose: escalate to it, or withhold escalation to it. Let \(O_j\) denote the event \(p_j>\phi\). Suppose escalating when \(O_j\) is true has loss \(L>0\), while withholding escalation when \(O_j\) is false has loss \(B>0\). For simplicity, suppose the other two action-event combinations have zero loss. These losses represent the decision maker’s relative consequences, not measured clinical outcomes.
Under this loss specification, the posterior expected loss of escalation is \(Lq_j\), because the loss \(L\) occurs on the event \(O_j\), whose posterior probability is \(q_j\). The posterior expected loss of withholding escalation is \(B(1-q_j)\), because \(p_j\leq\phi\) has posterior probability \(1-q_j\). Escalation has no greater expected loss when
Thus this particular loss model implies the threshold \(\delta=B/(L+B)\). The boundary matters: if \(q_j=\delta\), the two expected losses are equal. Escalation has strictly lower expected loss only when \(Lq_j<B(1-q_j)\), equivalently \(q_j<\delta\). At equality, the actions tie; the decision maker may use a prespecified tie-breaking rule.
This derivation is a justification for interpreting a threshold, not a claim that every trial must specify losses in this form. A clinical team may set \(\delta\) directly after considering the seriousness of toxicity, the consequences of delaying access to a potentially useful dose, the treatment setting, and the population. If the team uses the simplified loss interpretation, the implied relative loss is \(L/B=(1-\delta)/\delta\). A small \(\delta\) corresponds to assigning relatively greater loss to escalating when the dose is truly above the limit.
Worked Example: Deriving a Threshold From Relative Losses
Suppose, hypothetically, a trial’s decision makers assign loss \(L=9\) to escalating to a dose whose toxicity probability exceeds \(\phi\), and loss \(B=1\) to withholding escalation when the dose is at or below \(\phi\). Use these values to derive the threshold and evaluate a dose with posterior overdose probability \(q_j=0.10\).
The threshold implied by the stated losses is $$ \delta=\frac{B}{L+B}=\frac{1}{9+1}=0.10. $$ At \(q_j=0.10\), the expected loss of escalation is \(Lq_j=9(0.10)=0.90\). The expected loss of withholding escalation is \(B(1-q_j)=1(0.90)=0.90\). The losses are equal, so this dose is exactly on the boundary. It passes the inclusive screen \(q_j\leq\delta\), but escalation does not have strictly lower expected loss at this value.
For comparison, if \(q_j=0.08\), escalation has expected loss \(9(0.08)=0.72\), while withholding has expected loss \(1(0.92)=0.92\); escalation has strictly lower expected loss. If \(q_j=0.12\), the losses are \(1.08\) and \(0.88\), respectively, so withholding has lower expected loss.
How Strictness Changes the Eligible Set
For fixed posterior probabilities, a threshold defines a set of doses that pass the safety screen. If \(\delta_1\leq\delta_2\), then every dose passing at \(\delta_1\) also passes at \(\delta_2\), because \(q_j\leq\delta_1\) implies \(q_j\leq\delta_2\). A stricter threshold can therefore only leave the eligible set unchanged or make it smaller. It cannot make an additional dose pass.
That mathematical relationship explains the basic trade-off. A smaller threshold is more restrictive with respect to posterior overdose probability, but may reduce the range of doses available for exploration. A larger threshold may make more doses eligible, but its safety screen permits a larger posterior probability of exceeding \(\phi\). Neither direction is universally preferable; the appropriate choice depends on the trial’s risk tolerance and objectives.
Worked Example: Comparing Thresholds Across a Dose Grid
In a hypothetical CRM update, the toxicity limit is \(\phi=0.30\). The fitted model gives the following posterior overdose probabilities for four ordered doses. Assume the model and data analysis used to obtain these posterior summaries are appropriate for the trial.
| Dose | Posterior overdose probability \(q_j\) |
|---|---|
| 1 | 0.02 |
| 2 | 0.07 |
| 3 | 0.16 |
| 4 | 0.34 |
With \(\delta=0.05\), only dose 1 passes: \(0.02\leq0.05\), while \(0.07\), \(0.16\), and \(0.34\) all exceed \(0.05\). With \(\delta=0.10\), doses 1 and 2 pass because \(0.02\leq0.10\) and \(0.07\leq0.10\), but doses 3 and 4 do not. With \(\delta=0.20\), doses 1, 2, and 3 pass; dose 4 remains excluded because \(0.34>0.20\).
The calculation shows how a stricter threshold narrows the set of doses that can be considered for escalation. It does not establish that dose 3 should be selected when \(\delta=0.20\), nor that dose 1 should be selected when \(\delta=0.05\). Selection among doses that pass requires the trial’s other prespecified dose-selection rules.
Thresholds Are Not Toxicity Targets or Error Guarantees
A threshold \(\delta\) and a toxicity limit \(\phi\) answer different questions. The limit \(\phi\) defines what counts as overdose: the event is \(p_j>\phi\). The threshold \(\delta\) sets the maximum posterior probability of that event permitted by the screen. For example, \(\phi=0.30\) and \(\delta=0.10\) mean that a dose passes only if the posterior probability that its true toxicity probability exceeds 0.30 is no greater than 0.10. They do not mean that the dose’s toxicity probability is targeted to 0.10.
Nor is \(\delta\) automatically a frequentist guarantee that no more than a fraction \(\delta\) of trials, patients, or assigned doses will be overdoses. It is a posterior probability conditional on the model, prior, and observed data. Across repeated trials, the actual frequency of unsafe recommendations depends on the design and its operating characteristics, including the dose-toxicity relationship, sample size, accrual pattern, and model specification. A posterior threshold alone does not ensure a particular repeated-sampling error rate.
The posterior probabilities also depend on the model and prior. In CRM, the model structure and prior encode assumptions about how toxicity changes over the dose levels. A threshold screen can be internally applied correctly while the underlying posterior is poorly calibrated because those assumptions do not adequately describe the data-generating process. Threshold choice should therefore be considered together with model-based design evaluation, rather than treated as a substitute for it.
Worked Example: Interpreting a Threshold and Its Implicit Trade-Off
Suppose a hypothetical trial uses \(\phi=0.25\) and chooses \(\delta=0.20\). For a candidate dose, the current model gives \(q_j=0.18\). Interpret the decision screen and, under the simplified loss model, calculate the implied relative losses.
Because \(0.18\leq0.20\), the dose passes the posterior overdose screen. In context, the fitted posterior assigns an 18% probability to the event that the dose’s true toxicity probability exceeds 25%, conditional on the current data and model. This is not an 18% probability that a particular patient will experience toxicity.
Under the loss derivation, \(\delta=B/(L+B)\). Rearranging gives \(L/B=(1-\delta)/\delta\), so $$ \frac{L}{B}=\frac{1-0.20}{0.20}=\frac{0.80}{0.20}=4. $$ Thus \(\delta=0.20\) corresponds, in that simplified decision model, to assigning four times as much loss to escalating when the dose is above the toxicity limit as to withholding escalation when it is not. The dose’s posterior expected escalation loss, in units of \(B\), is \(4(0.18)=0.72\); the expected withholding loss is \(1(1-0.18)=0.82\). Escalation has strictly lower expected loss for these inputs, although passing the screen does not by itself determine the full trial action.
Practical Selection, Diagnostics, and Limitations
Threshold selection should be justified before examining interim results where feasible. Otherwise, changing \(\delta\) after seeing the current posterior can make the rule difficult to interpret and may weaken the protection intended by the original design. A rationale can state the toxicity limit, the chosen threshold, the clinical consequences represented by that choice, and how the rule will be applied at the boundary.
Useful checks include confirming that posterior overdose probabilities are computed for the event \(p_j>\phi\), not for a different event such as \(p_j\geq\phi\) unless that alternative is deliberately specified; verifying that the same \(\phi\) and \(\delta\) are applied consistently; and checking how eligible doses change over plausible data patterns and model assumptions. If model averaging is used, the relevant \(q_j\) should reflect the posterior probability under the prespecified model-averaging procedure, not an unweighted average of probabilities unless that is the actual procedure.
A common mistake is to interpret a larger threshold as a higher toxicity target. For example, if \(\phi=0.30\), changing \(\delta\) from 0.05 to 0.20 does not change the toxicity limit from 30% to 20%. It changes how much posterior probability of exceeding the unchanged 30% limit the screen tolerates. Another mistake is to say that a dose with \(q_j=\delta\) has strictly lower expected loss under the threshold’s implied loss model. At equality, the expected losses tie; the weak-preference rule permits escalation but does not establish a strict advantage.
Posterior overdose control is not the only possible way to regulate escalation. A rule based on a posterior feasibility bound, such as an EWOC constraint, frames safety through posterior uncertainty about the maximum tolerated dose. A target-probability CRM rule instead focuses on proximity to a target toxicity probability. These approaches have different decision interpretations; their numerical parameters should not be treated as interchangeable. A posterior overdose threshold is especially direct when the decision question is whether the probability of exceeding a specified toxicity limit is acceptable.
Finally, a threshold is a simplifying decision rule. It compresses a posterior distribution into a pass-or-fail screen and does not capture every dimension of clinical benefit, toxicity severity, or the value of information from learning about higher doses. It also cannot repair a misspecified dose-toxicity model. Its value is that it makes one safety trade-off explicit and reproducible, provided its meaning and limitations are stated clearly.
Check Your Understanding
Answer each question using the definitions and decision rule above. Brief solutions follow the questions.
- If \(q_j=0.12\) and \(\delta=0.12\), does the dose pass the screen? Under the simplified loss interpretation, are the expected losses strictly different?
- For \(L=4\) and \(B=1\), calculate the implied threshold. State whether escalation has strictly lower expected loss at \(q_j=0.20\).
- With \(\phi=0.25\), what does \(q_j=0.08\) mean in words? What does it not mean?
- If the threshold increases from 0.05 to 0.15 while posterior probabilities remain fixed, can a dose that passed at 0.05 fail at 0.15?
- Why does a posterior overdose threshold not by itself guarantee that the same proportion of repeated trials will recommend unsafe doses?
Solutions. (1) Yes: \(0.12\leq0.12\), so it passes. At the boundary \(Lq_j=B(1-q_j)\) under the implied loss ratio, the expected losses tie; neither is strictly lower. (2) \(\delta=B/(L+B)=1/(4+1)=0.20\). At \(q_j=0.20\), escalation loss is \(4(0.20)=0.80\) and withholding loss is \(1(0.80)=0.80\), so they tie. (3) Conditional on the model and current data, the posterior probability that the dose’s true toxicity probability exceeds 25% is 8%. It is not the probability that a particular patient will experience toxicity. (4) No. Since \(0.05\leq0.15\), every dose with \(q_j\leq0.05\) also satisfies \(q_j\leq0.15\). (5) The threshold is conditional on the fitted posterior, whereas repeated-trial frequencies depend on the full design and data-generating process, including the model’s adequacy and the true toxicity probabilities.