Overdose Control in a First-in-Human Escalation

Why Overdose Control Matters in Escalation

In a first-in-human escalation, each dose assignment is a safety decision made with limited information. A dose that has not produced many observed toxicities may still have a substantial probability of exceeding the trial’s acceptable toxicity target. An overdose-control rule makes this uncertainty explicit: it calculates a posterior probability of excessive toxicity and compares that probability with a cutoff specified in the protocol.

The calculation does not establish that a dose is safe, nor does it identify a dose that should be selected for later expansion. It addresses a narrower operational question: does the evidence meet the prespecified criterion for excluding this dose from further assignment? A rule can also require that all higher doses be excluded once a dose crosses the criterion. That higher-dose exclusion is a protocol safeguard; it is not, by itself, a separate posterior finding about every higher dose.

A Dose-Specific Beta-Binomial Calculation

For a transparent dose-level calculation, suppose each treated patient has a binary outcome: a DLT during a specified assessment window, or no DLT during that window. At dose \(d\), assume the outcomes are conditionally independent with a common DLT probability \(p_d\). If \(x_d\) DLTs are observed among \(n_d\) patients, then \(x_d\mid p_d\) has a binomial distribution.

For this tutorial’s calculations, use independent dose-specific priors \(p_d\sim\operatorname{Beta}(1,2)\). This is a deliberately specified illustrative prior, not a universal first-in-human prior. Its density is proportional to \(p_d^{1-1}(1-p_d)^{2-1}\). Under the binomial likelihood, the posterior is beta with parameters equal to the prior parameters updated by the observed DLT and non-DLT counts:

This model treats each dose separately: the prior distributions are independent, and the likelihood at one dose does not borrow information from another. It does not impose a monotone dose-toxicity curve. The earlier material on Bayesian Dose Selection With a Logistic Toxicity Model considers a model-based dose-toxicity relationship; the calculation here instead applies a prespecified overdose rule to dose-specific posterior probabilities.

The rule requires two protocol choices. First, \(\phi\) defines what toxicity probability counts as excessive. Second, \(\gamma\) defines how much posterior probability of exceeding \(\phi\) is sufficient for exclusion. Neither number is supplied by the data alone. They should be justified and fixed in advance, rather than chosen after inspecting the observed toxicity counts.

Computing the Posterior Overdose Probability

For a posterior distribution \(p_d\mid\text{data}\sim\operatorname{Beta}(a,b)\), the quantity needed is the upper-tail probability beyond \(\phi\). It can be written as one minus the beta cumulative distribution function evaluated at \(\phi\). For positive integer \(a\) and \(b\), an equivalent expression uses a binomial probability:

The identity follows from the relationship between the beta distribution and binomial tails: the beta cumulative probability \(F_{\operatorname{Beta}(a,b)}(\phi)\) equals the probability that a \(\operatorname{Binomial}(a+b-1,\phi)\) random variable is at least \(a\). Taking the complement gives the sum from zero through \(a-1\). In practice, software can evaluate the beta tail directly; the finite sum is useful for checking calculations when the parameters are integers.

What the Rule Does—and Does Not—Infer

The comparison is strict as defined here: exclude when the posterior probability is greater than \(\gamma\). If it equals \(\gamma\), the rule does not exclude unless the protocol explicitly specifies a different convention. A protocol should also define what happens when an estimate is unavailable, when a cohort has unresolved DLT assessment, or when no dose remains eligible.

The posterior probability and the operational exclusion set are related but not identical. The posterior calculation is specific to a dose. The protocol’s monotone exclusion safeguard may then remove higher doses from assignment even if their own posterior probabilities have not crossed the cutoff. In the first worked example, the rule excludes the dose that crossed the cutoff and doses above it; it does not assert that every excluded higher dose independently crossed the threshold.

This approach is useful because it places uncertainty and the safety threshold on a common probability scale. It is also limited. Independent dose-specific analyses do not share information across doses, so estimates can be unstable when cohorts are small. They do not enforce the clinically plausible expectation that toxicity risk tends to rise with dose. A model-based alternative, such as a logistic toxicity model, can represent a dose-toxicity relationship and borrow information across levels, but depends on the chosen model and its assumptions. The dose-specific rule is simpler to audit, but simplicity does not remove the need to justify the prior, target, and cutoff.

Assumptions, Diagnostics, and Failure Modes

Before applying the rule, check that its data and model conditions are credible. The binomial likelihood assumes a well-defined binary DLT outcome and a common underlying probability for patients treated at the same dose under the relevant trial conditions. It also assumes conditional independence of patient outcomes. The assessment window and DLT definition must be consistent, and the outcomes used for a decision should be sufficiently complete for the protocol’s timing rule.

A missing or immature safety outcome is not equivalent to a non-DLT. Treating an unresolved patient as toxicity-free can lower the posterior overdose probability and make a dose appear more acceptable than the evidence supports. The protocol should specify how delayed outcomes, withdrawals, and missing assessments are handled before the calculation is used. If dose assignment or supportive care changes the meaning or observation of DLTs across cohorts, the assumption of comparable outcomes deserves scrutiny.

Check the arithmetic as well as the assumptions: the posterior shape parameters must add the DLT count to the first prior parameter and the non-DLT count to the second. Confirm that the computed tail is \(P(p_d>\phi\mid\text{data})\), not the probability below \(\phi\), and compare it with the protocol’s actual cutoff. A tail probability unexpectedly close to zero after many observed DLTs, or close to one after none, is a prompt to check the target, parameter order, and calculation.

The posterior probability is not a frequentist error rate and is not the probability that a null hypothesis is true. It expresses uncertainty about \(p_d\), conditional on the specified prior and data. The cutoff \(\gamma\) is a decision threshold, not a guarantee of a particular rate of severe toxicity across trials. Operating characteristics of a proposed rule may need separate evaluation under plausible scenarios; this tutorial’s calculations do not establish those operating characteristics.

Common Mistake and Practical Comparison

• Wrong reasoning: “The overdose probability is 0.8059, so there is an 80.59% chance that the dose is unsafe in every meaningful sense.” Correction: It is the posterior probability that the dose-specific DLT probability exceeds the particular target \(\phi\), conditional on the model, prior, and observed data. “Unsafe” may involve additional clinical considerations not represented by that target.
• Wrong reasoning: “A dose below the cutoff has been shown safe.” Correction: It has not met this rule’s exclusion condition. The calculation does not prove safety or rule out clinically important toxicity.
• Wrong reasoning: “If one dose crosses the cutoff, the posterior calculations prove that every higher dose is also excessive.” Correction: The rule may exclude higher doses by protocol, but the dose-specific posterior calculation only quantifies the risk at the dose whose data were used.

Compared with a deterministic rule based only on the observed toxicity proportion, the posterior rule accounts for uncertainty and makes the prior contribution explicit. Compared with a model-based dose-toxicity analysis, the independent beta-binomial approach requires fewer structural assumptions but does not pool information or enforce an ordered toxicity curve. No method removes the need to define the DLT outcome, target, decision threshold, and assignment safeguards in advance.