Escalation After a Cohort With No DLTs

What a Zero-DLT Cohort Does—and Does Not—Show

After a cohort completes its dose-limiting toxicity (DLT) assessment window, observing no DLTs is evidence about toxicity at the administered dose. It is not proof that the dose has zero risk, and it does not by itself establish that a higher dose is acceptable. In a Bayesian logistic regression model (BLRM), the cohort reweights the posterior distribution of the dose-toxicity curve. The escalation decision then depends on the posterior overdose probability at the proposed higher dose and the overdose cutoff specified by the protocol.

The cohort is informative in a particular way: parameter values predicting lower toxicity at the dose given receive greater likelihood than values predicting higher toxicity. But a higher dose depends on the joint posterior for the BLRM intercept and slope. If posterior configurations with lower risk at the observed dose also have steep slopes, risk at a higher dose need not fall as much as risk at the current dose.

Definition: In the BLRM, the DLT probability at dose $d$ is $p(d)$, with $$ \operatorname{logit}\{p(d)\}=\alpha+\beta\log(d/d_{\mathrm{ref}}), \qquad \beta>0, $$ where $\operatorname{logit}(u)=\log\{u/(1-u)\}$, $d_{\mathrm{ref}}$ is a reference dose, and $\alpha,\beta$ are uncertain model parameters. For a cohort of $m$ patients treated at dose $d_c$, let $Y$ be the number with a DLT. Conditional on $\alpha,\beta$, the cohort likelihood is binomial: $$ Y\mid \alpha,\beta \sim \operatorname{Binomial}\{m,p(d_c)\}. $$ For a completed cohort with no DLTs, $Y=0$, so its likelihood contribution is $\{1-p(d_c)\}^{m}$.

This update presumes that the cohort has completed the protocol-defined DLT assessment window, DLT status is recorded as a binary outcome, and patients’ DLT outcomes are conditionally independent given the BLRM parameters. It also relies on the specified BLRM and prior being suitable for the dose range being considered. If follow-up is incomplete, treating a patient as DLT-free is not justified; incomplete follow-up requires separate handling.

Why Zero DLTs Reweight the Posterior

Let $\theta=(\alpha,\beta)$, and let $\pi(\theta\mid\mathcal D)$ be the posterior density before the new cohort, given existing data $\mathcal D$. Bayes’ rule gives the updated posterior density after the zero-DLT cohort:

\pi(\theta\mid\mathcal D,Y=0) \propto \{1-p(d_c\mid\theta)\}^{m}\pi(\theta\mid\mathcal D).

The likelihood multiplier is larger when $p(d_c\mid\theta)$ is smaller. For example, if two possible parameter values predict current-dose risks $p_1$ and $p_2$, their relative likelihood contributions from the zero-DLT cohort are in the ratio

\frac{L_0(p_1)}{L_0(p_2)} = \left(\frac{1-p_1}{1-p_2}\right)^m.

Thus, a larger cohort generally makes the zero-DLT observation more discriminating between predictions at the treated dose. It does not eliminate uncertainty: parameter values with nonzero toxicity risk still have positive likelihood whenever that risk is less than one.

Update the joint parameter distribution.
Multiply each prior or existing-posterior weight by $\{1-p(d_c)\}^{m}$, then normalize the weights.

Evaluate risk at each dose of interest.
For each posterior parameter value, calculate $p(d)$ at the current dose and at the proposed higher dose.

Apply the overdose rule to the higher dose.
Calculate the posterior probability that its DLT risk exceeds the protocol’s overdose threshold and compare that probability with the prespecified cutoff.

For clarity, the worked examples use a finite set of posterior parameter configurations. This makes posterior reweighting exact for the stated discrete distribution and keeps the arithmetic visible. In a continuous BLRM analysis, the same quantities are typically estimated from posterior draws or numerical integration.

Worked Example: A Zero-DLT Cohort and the Next Dose

Suppose the existing posterior is represented by four equally weighted BLRM configurations. At the current dose their DLT probabilities are $0.10, 0.20, 0.30,$ and $0.40$. At the next candidate dose, their probabilities are $0.1818, 0.3333, 0.4615,$ and $0.5714$, respectively. These dose pairs are consistent with a positive slope: each next-dose log-odds is the current-dose log-odds plus $\log 2$. A cohort of $m=3$ completes follow-up with $Y=0$. The protocol defines overdose as $p(d)>0.30$ and permits a dose only when its posterior overdose probability is below $0.25$.

The four unnormalized posterior weights are the prior weights, each $0.25$, multiplied by the zero-DLT likelihood. Since the equal prior factor cancels when normalizing, use the likelihoods directly:

(1-0.10)^3=0.729,\quad (1-0.20)^3=0.512,\quad (1-0.30)^3=0.343,\quad (1-0.40)^3=0.216.

Their sum is $0.729+0.512+0.343+0.216=1.800$. The normalized posterior weights are therefore $0.729/1.800=0.4050$, $0.512/1.800=0.2844$, $0.343/1.800=0.1906$, and $0.216/1.800=0.1200$.

At the current dose, only the configuration with risk $0.40$ exceeds $0.30$. Its posterior overdose probability is $0.1200$, below the $0.25$ cutoff. At the next dose, the last three configurations exceed $0.30$, so the posterior overdose probability is

$$ 0.2844+0.1906+0.1200=0.5950. $$

The next dose fails the overdose rule because $0.5950$ is not below $0.25$. This cohort makes the current dose look safer under the rule, but it does not make the next dose admissible. The decision is about posterior probability of exceeding the overdose threshold, not merely whether the cohort had any DLTs.

When the Rule Permits Escalation

A no-DLT cohort can reduce the posterior probability of overdose at a higher dose enough to permit escalation. The result depends on the previous posterior, the cohort size, the treated dose, the relationship between doses encoded by the BLRM, and the rule’s threshold and cutoff. The following example shows that a zero-DLT cohort is not automatically a prohibition on escalation either.

Worked Example: A Higher Dose Falls Below the Cutoff

Consider four equally weighted posterior configurations. Their current-dose risks are $0.10, 0.15, 0.20,$ and $0.25$. At the next dose the risks are $0.1818, 0.2609, 0.3333,$ and $0.4000$; each pair has a positive dose slope and a doubled odds at the next dose. Ten patients complete follow-up at the current dose with no DLTs. Keep the illustrative overdose threshold $0.30$ and cutoff $0.25$.

The zero-DLT likelihoods, in configuration order, are

0.90^{10}=0.3486784,\quad 0.85^{10}=0.1968744,\quad 0.80^{10}=0.1073742,\quad 0.75^{10}=0.05631351.

The sum is $0.7092405$, giving normalized posterior weights approximately $0.4916, 0.2775, 0.1514,$ and $0.07940$. At the next dose, the final two configurations exceed $0.30$. Hence

P\{p(d_{\mathrm{next}})>0.30\mid\mathcal D,Y=0\} = \frac{0.1073742+0.05631351}{0.7092405} = 0.2308.

The next dose meets the overdose rule because $0.2308<0.25$. On the stated overdose criterion, escalation is permitted. This is not a general conclusion that ten DLT-free patients make escalation safe: it follows from this particular posterior representation and the stated protocol cutoff. Other protocol requirements, such as cohort timing or constraints on dose assignment, must also be satisfied.

Why Risk at the Higher Dose Can Behave Differently

A positive BLRM slope ensures that each modelled dose-toxicity curve increases with dose. It does not guarantee that posterior overdose probability at every higher dose decreases after a DLT-free cohort. The cohort likelihood directly reweights predictions at the dose administered; it does not observe toxicity at the proposed higher dose. The posterior relationship between intercept and slope matters.

Worked Example: Lower Current Risk, Higher Posterior Risk at the Next Dose

Suppose four equally weighted posterior configurations have current-dose and next-dose risks, respectively: $(0.25,0.29)$, $(0.10,0.40)$, $(0.20,0.28)$, and $(0.30,0.45)$. Each next-dose risk is greater than its paired current-dose risk, as required by a positive slope. Before observing a new cohort, two of the four configurations put next-dose risk above the $0.30$ overdose threshold, so the prior-posterior overdose probability there is $0.5000$.

Now suppose five patients complete follow-up at the current dose with no DLTs. The likelihood contributions are $0.75^5=0.2373047$, $0.90^5=0.5904900$, $0.80^5=0.3276800$, and $0.70^5=0.1680700$. Their sum is $1.3235447$. The configurations predicting next-dose risk above $0.30$ are the second and fourth, so the updated next-dose overdose probability is

\frac{0.5904900+0.1680700}{1.3235447}=0.5731.

It has increased from $0.5000$ to $0.5731$, even though the zero-DLT observation favors configurations with lower risk at the current dose. The second configuration has relatively low current-dose risk but high next-dose risk, reflecting a steeper increase across doses. This example is a valid illustration of posterior reweighting; it is not a claim that this direction of change is inevitable.

Diagnostics, Alternatives, and Limitations

Before acting on an escalation calculation, verify that the data and model inputs match the intended update. Check the number of patients with fully observed DLT outcomes, the administered dose, the cohort size used in the likelihood, and the reference-dose and dose-slope parameterization. Confirm that posterior draws produce increasing toxicity probabilities across candidate doses when the model imposes $\beta>0$. If a computed posterior overdose probability changes implausibly, inspect the dose-specific probabilities and weights, and check that weights were normalized using the same likelihood contributions used in the numerator.

The decision quantity is the posterior probability of the overdose event, not the posterior mean DLT probability. A mean risk below the overdose threshold does not establish that the probability of exceeding that threshold is below the protocol cutoff. Likewise, the observed cohort rate $0/m=0$ is not the posterior risk at the next dose. The overdose definition and probability cutoff must come from the protocol; changing either changes the admissible decision.

The main alternative summaries answer different questions. A posterior mean or median describes a central value of the dose-specific risk, while a credible interval describes posterior uncertainty about that risk. A target-interval probability asks whether risk lies within a specified target range. The overdose rule instead asks whether the posterior probability of risk above the overdose threshold is sufficiently small. None of these summaries can be substituted for another without changing the decision criterion.

The BLRM also has limitations. Posterior conclusions depend on the model, prior calibration, observed dose range, and available cohort data. With a small cohort, a zero-DLT result may leave considerable uncertainty. With sparse dose coverage, predictions at a substantially higher dose may rely strongly on the assumed dose-response form. A posterior calculation quantifies uncertainty conditional on the model and prior; it does not remove model misspecification or guarantee clinical safety.

Common pitfall: “No patients in the cohort had a DLT, so the next dose is safe.” This treats the observed cohort rate as the true risk and ignores uncertainty, the increase in risk with dose, and the protocol’s overdose cutoff. The correction is to update the joint BLRM posterior using $\{1-p(d_c)\}^{m}$, calculate the posterior probability of overdose at the proposed dose, and compare that probability with the prespecified cutoff.

Check Your Understanding

Answer each question using the overdose definition and cutoff stated in the question.

For a cohort of $m$ patients with no DLTs at dose $d_c$, what likelihood factor updates a BLRM parameter configuration?
In the first worked example, does the next dose meet the overdose rule? Give the posterior probability and cutoff.
Why can the posterior overdose probability at a higher dose increase after a zero-DLT cohort at the current dose?
If a protocol defines overdose as $p(d)>0.30$ and uses a cutoff of $0.25$, does a calculated posterior overdose probability of $0.2500$ meet a rule requiring the probability to be below the cutoff?

Solutions. (1) The factor is $\{1-p(d_c)\}^{m}$. (2) No. The next-dose posterior overdose probability is $0.5950$, which exceeds the $0.25$ cutoff. (3) The zero-DLT cohort reweights parameter configurations according to their risk at the current dose. Configurations with lower current risk but steeper slopes can gain weight and still predict high risk at the next dose. (4) No. The rule requires a probability below $0.25$; $0.2500$ is equal to, not below, the cutoff.