BLRM Escalation Under Excessive Overdose Probability

From BLRM Risk Estimates to an Escalation Decision

A BLRM describes how the probability of a dose-limiting toxicity (DLT) changes with dose. An escalation decision needs more than a point estimate of that probability: it must account for uncertainty about the dose-risk curve. In particular, a dose with an estimated DLT risk below a safety limit may still have a substantial posterior probability of exceeding that limit.

This tutorial uses the dose-risk model and posterior risk calculations established in the earlier material on logistic dose-toxicity models and posterior risk intervals. The decision focus is how to compare those posterior probabilities with prespecified acceptable and excessive risk bands. The bands and cutoffs are design choices specified in a protocol, not quantities discovered from the observed data.

Definition: Let $p_j(\theta)$ be the DLT probability at candidate dose $d_j$ under BLRM parameters $\theta$. Let $[p_L,p_U]$ be the prespecified target toxicity interval, and let $\gamma$ be the maximum allowed posterior probability of excessive toxicity. Define $$ q_j^{T}=P\!\left(p_L\leq p_j(\theta)\leq p_U\mid D\right),\qquad q_j^{E}=P\!\left(p_j(\theta)>p_U\mid D\right). $$ Here $q_j^{T}$ is the posterior probability that the dose risk is in the target band, and $q_j^{E}$ is the posterior probability that it exceeds the band’s upper limit. A dose passes the overdose-control condition when $q_j^{E}\leq\gamma$.

The event in $q_j^{E}$ is about the unknown dose-level DLT probability, not about whether the next patient will experience a DLT. For a fixed value of $\theta$, $p_j(\theta)$ is that patient’s modeled DLT probability; the posterior distribution expresses uncertainty about this probability after the observed data $D$ are considered.

Model, Assumptions, and the Decision Rule

For example, under the two-parameter BLRM parameterization, the dose-level risk can be written as

\operatorname{logit}\{p_j(\theta)\} =\log(\alpha)+\beta\log(d_j/d_{\mathrm{ref}}), \qquad \theta=(\alpha,\beta),

where $d_{\mathrm{ref}}$ is a reference dose and the parameter restrictions and prior are those specified for the model. With $\alpha>0$ and $\beta>0$, the modeled risk increases with dose. For fully assessed binary DLT outcomes, a basic likelihood is $Y_j\mid\theta\sim\operatorname{Binomial}(n_j,p_j(\theta))$, where $n_j$ patients received dose $d_j$ and $Y_j$ had a DLT. This likelihood assumes that, conditional on the model parameters, outcomes are Bernoulli with the dose-specific risk and that the model adequately represents the dose-risk relationship. It does not by itself address incomplete follow-up or delayed DLT assessment.

A protocol must specify the target interval, the excessive-risk cutoff $\gamma$, eligible doses, and how to choose among eligible doses. One transparent escalation rule is to call a dose eligible if $q_j^{E}\leq\gamma$, then recommend the highest eligible dose among those allowed by the protocol. Some protocols instead choose the eligible dose with the greatest $q_j^{T}$. These objectives are not interchangeable: the first favors the highest dose that satisfies the safety constraint, while the second favors the dose most likely to lie in the target band.

Decision principle: Apply the posterior overdose cutoff to each protocol-eligible candidate dose. A passing dose is not automatically the recommendation: the protocol’s selection objective and any escalation restrictions determine which passing dose is chosen. The cutoff controls a posterior probability under the fitted model; it is not a guarantee that a DLT will not occur.

The calculation follows directly from the definition of posterior probability. If $\theta^{(1)},\ldots,\theta^{(M)}$ are draws from the posterior distribution, transform each draw into a risk at dose $d_j$. For each draw, record whether that risk is in the target band, above the upper limit, or below the target band. The proportions of draws in the respective events estimate the required posterior probabilities:

\widehat q_j^{T} =\frac{1}{M}\sum_{m=1}^{M} I\!\left\{p_L\leq p_j(\theta^{(m)})\leq p_U\right\}, \qquad \widehat q_j^{E} =\frac{1}{M}\sum_{m=1}^{M} I\!\left\{p_j(\theta^{(m)})>p_U\right\}.

The indicator $I\{\cdot\}$ equals 1 when its condition is true and 0 otherwise. This is a Monte Carlo estimate: with a finite number of posterior draws, the calculated proportions have simulation error. A decision close to $\gamma$ therefore warrants checking that the posterior computation is sufficiently stable for the intended decision.

Worked Escalation Decisions

Worked Example: Highest Dose Passing the Overdose Cutoff

Hypothetical setup. Suppose the protocol defines the target interval as 20% to 30%, sets $\gamma=0.25$, and uses the highest eligible dose that passes the overdose cutoff. For 1,000 posterior draws, the risk categories at three ordered doses are:

Dose	Below 20%	Within 20%–30%	Above 30%
Dose 1	420	450	130
Dose 2	240	540	220
Dose 3	170	510	320

At Dose 1, $\widehat q_1^{E}=130/1000=0.130$, so it passes. At Dose 2, $\widehat q_2^{E}=220/1000=0.220$, so it also passes. At Dose 3, $\widehat q_3^{E}=320/1000=0.320$, which is greater than $0.25$, so it fails. The highest passing dose is Dose 2. Its estimated target-band probability is $\widehat q_2^{T}=540/1000=0.540$.

The decision is not that Dose 2 is certain to be safe or certain to be in the target range. Rather, 22.0% of the posterior draws place its risk above 30%, which is below the protocol’s maximum of 25.0%. Dose 3 does not meet that condition.

Worked Example: A Dose Just Above the Cutoff

Hypothetical setup. Keep the same target interval and $\gamma=0.25$, but suppose the protocol chooses the passing dose with the greatest target-band probability. For a candidate dose, 1,000 posterior draws produce 250 risks above 30%, 600 within 20%–30%, and 150 below 20%.

The estimated probabilities are

\widehat q^{E}=\frac{250}{1000}=0.250,\qquad \widehat q^{T}=\frac{600}{1000}=0.600.

Because the stated rule is $\widehat q^{E}\leq\gamma$, this dose passes exactly at the cutoff. If instead 251 of the 1,000 draws were above 30%, then $\widehat q^{E}=251/1000=0.251$, and the dose would fail. The difference is small, so in practice the posterior computation should be checked for adequate precision rather than treating a rounded display as decisive. For instance, a reported value of 0.25 might represent an underlying value slightly above or below the cutoff.

Passing makes the dose eligible; it does not establish that it should be selected. Under the stated target-probability objective, it is selected only if its $q^{T}$ is greater than that of the other eligible doses, subject to the protocol’s other constraints.

Worked Example: Why the Posterior Mean Risk Is Not the Cutoff Test

Hypothetical setup. The target interval is 20% to 30%, and $\gamma=0.25$. At one dose, 700 posterior draws give a risk of 24%, while 300 give a risk of 38%. The posterior mean risk is

\frac{700(0.24)+300(0.38)}{1000} =\frac{168+114}{1000} =0.282.

The mean risk, 28.2%, lies inside the target interval. But the overdose probability is the fraction of draws with risk above 30%:

\widehat q^{E}=\frac{300}{1000}=0.300.

Since $0.300>0.25$, the dose fails the overdose-control condition. Also, $\widehat q^{T}=700/1000=0.700$. A target-range posterior mean does not answer whether enough posterior probability lies above the excessive-risk limit. The decision rule is based on the probability of that event, not on whether the mean risk falls in the target band.

Diagnostics, Alternatives, and Limitations

The rule is only as reliable as the posterior risk calculations and the model supporting them. Before interpreting a candidate dose’s classification, check that posterior draws are valid and that the risk transformation has been applied to the intended doses and target limits. A useful basic check is that the below-target, target-band, and excessive-risk proportions sum to one, apart from rounding, because these categories partition the possible risks. Inspect dose-specific results across the ordered dose set; an unexpected non-monotone pattern can signal a coding error, a model or parameterization mismatch, or a problem with the assumptions used for that analysis.

Also inspect how close each $\widehat q_j^{E}$ is to $\gamma$. A decision far from the cutoff is less sensitive to small Monte Carlo fluctuations than one near it. If the estimate is close, increase or otherwise assess the adequacy of posterior simulation and report enough digits to avoid making a classification based only on rounding. Posterior predictive and other model diagnostics are important as well, but assessing them in detail is a separate task.

The principal alternatives differ in what they use for the decision. A rule based on a point estimate, such as the posterior mean risk, ignores how posterior uncertainty is distributed around the upper risk limit. A rule based on $q_j^{T}$ alone ranks doses by the probability of lying in the target band but does not, by itself, impose a maximum acceptable probability of excessive risk. The overdose-constrained rule explicitly screens on $q_j^{E}$, then uses the protocol’s stated objective to choose among doses that pass. These approaches can recommend different doses, as the examples illustrate.

The posterior cutoff is conditional on the model, prior, data, and specified target band. It does not account automatically for model misspecification, an unsuitable prior, incorrect DLT coding, or incomplete outcome assessment. Nor does passing the cutoff guarantee an absence of toxicity at the next dose. The rule is a prespecified decision aid under uncertainty, not a replacement for trial oversight or other protocol requirements.

Common Mistakes

Using the mean risk as if it were the overdose probability. The correction is to count posterior draws above $p_U$ and divide by the total number of draws.
Reversing the comparison. A dose is eligible under this rule when its posterior probability of exceeding $p_U$ is no greater than $\gamma$, not when its probability of being below $p_U$ is no greater than $\gamma$.
Treating a passing dose as the automatic recommendation. Passing only satisfies the overdose screen. The protocol’s selection objective and dose-eligibility restrictions still apply.
Calling $\gamma$ the probability of a DLT in the next patient. It bounds posterior probability that the dose-level DLT risk exceeds $p_U$; it is not the individual patient’s DLT probability.
Overinterpreting a rounded result at the boundary. A displayed estimate of 0.25 may conceal a value on either side of the cutoff. Check simulation precision and the unrounded value.

Key takeaway: BLRM overdose control compares $P\{p_j(\theta)>p_U\mid D\}$ with a prespecified cutoff. First identify the doses that pass; then apply the protocol’s selection rule. The posterior probability of excessive dose risk is not interchangeable with a point estimate of risk or with the chance of a DLT in one patient.

Check Your Understanding

Use the stated cutoff and show how each classification follows from the posterior probabilities.

With $\gamma=0.20$, a dose has 180 excessive-risk draws out of 1,000. Does it pass? What is the estimated excessive-risk probability?
For $\gamma=0.25$, Dose A has $q^E=0.10$ and $q^T=0.35$; Dose B has $q^E=0.20$ and $q^T=0.50$. Which dose is chosen by the highest-target-probability rule, assuming both are protocol-eligible?
A dose’s posterior mean risk is 26%, but 28% of posterior draws put its risk above $p_U$. If $\gamma=0.25$, does the dose pass? Explain why the mean does not determine the answer.
In 2,000 posterior draws, 520 risks exceed $p_U$. Calculate $\widehat q^E$ and classify the dose when $\gamma=0.26$.

Solutions. (1) $\widehat q^E=180/1000=0.180$, and $0.180\leq0.20$, so the dose passes. (2) Both pass because their $q^E$ values are at most 0.25; Dose B is chosen because $0.50>0.35$. (3) It fails because $q^E=0.28>0.25$. The posterior mean does not give the fraction of posterior risk values above $p_U$. (4) $\widehat q^E=520/2000=0.260$, so the dose passes exactly at the stated inclusive cutoff.