Why a DLT at a High Dose Changes the Next Decision
A dose-limiting toxicity (DLT) in a high-dose cohort is a signal to reassess safety, not a rule that automatically identifies the next dose. The decision depends on the accumulated evidence across dose levels and on the posterior probability that each candidate dose exceeds the protocol’s overdose boundary. A Bayesian logistic regression model (BLRM) represents the dose–toxicity relationship and updates that probability after the new cohort’s outcomes are observed.
The practical task here is narrower than selecting a final recommended dose: after a concerning high-dose cohort, decide whether a lower dose satisfies the prespecified overdose-control rule, or whether escalation should stop because no dose does. The examples use hypothetical, fully followed cohorts. Their posterior calculations are illustrative outputs from the specified model, not results from an actual study.
Model, Assumptions, and Updating
For these examples, suppose the protocol specifies the BLRM \( \operatorname{logit}\{p(d;\theta)\}=\eta+\beta\log(d/d_{\mathrm{ref}}) \), where \(\operatorname{logit}(q)=\log\{q/(1-q)\}\), \(d_{\mathrm{ref}}\) is a fixed reference dose, and \(\beta>0\) ensures that the modeled DLT probability increases with dose. For illustration, use independent priors \(\eta\sim N(-2,1.5^2)\) and \(\gamma=\log(\beta)\sim N(0,0.5^2)\), so \(\beta=\exp(\gamma)\). A real trial would use priors and model checks appropriate to its protocol and available evidence; these illustrative priors are not universal recommendations.
If cohort \(j\) receives dose \(d_j\), has \(n_j\) participants with complete DLT follow-up, and \(y_j\) DLTs, assume \(Y_j\mid\theta\sim\operatorname{Binomial}\{n_j,p(d_j;\theta)\}\). The likelihood treats participants’ DLT outcomes as conditionally independent given the model parameters. The update also assumes that the recorded cohort counts and dose levels are correct, the DLT definition and observation window are applied consistently, and the model and prior are appropriate enough for the intended decision.
Each observed cohort contributes a binomial likelihood term. A DLT-heavy cohort at a high dose therefore changes the posterior distribution over the dose-response curve, which in turn changes overdose probabilities at both the treated dose and other doses. The model’s monotonicity links those levels, but it does not mean that an observed DLT rate at one dose is itself the posterior probability of overdose at another.
In practice, posterior draws \(\theta^{(1)},\ldots,\theta^{(M)}\) can be used to estimate the overdose probability at a candidate dose:
Here \(\mathbb{I}\{\cdot\}\) equals 1 when the condition is true and 0 otherwise. The estimate is the fraction of posterior draws in which the dose is above the overdose boundary. It is not the fraction of participants who experienced a DLT.
A De-Escalation Decision in Three Examples
Throughout the examples, let \(\phi=0.30\), so overdose means a modeled DLT probability greater than 30%, and let \(\xi=0.25\). A dose is admissible only when its posterior overdose probability is strictly less than 0.25. For the immediate next-dose decision, consider doses below the dose just administered, and choose the highest admissible lower dose if one exists. If none exists, stop dosing under this rule. A protocol may impose additional restrictions, which remain in force.
Worked Example: Two DLTs at the Highest Dose
A hypothetical trial has dose levels 5, 10, 20, and 40 mg. After complete follow-up, the cumulative cohort data are 0/3 DLTs at 5 mg, 1/3 at 10 mg, 0/3 at 20 mg, and 2/3 at 40 mg. The current cohort was treated at 40 mg. Using the stated BLRM and safety rule, suppose 10,000 retained posterior draws give 200, 800, 1,800, and 6,100 draws above the overdose boundary at 5, 10, 20, and 40 mg, respectively.
At 20 mg, for example, the estimated posterior overdose probability is \(1{,}800/10{,}000=0.180\), or 18.0%. The other dose-specific estimates are \(200/10{,}000=0.0200\), \(800/10{,}000=0.0800\), and \(6{,}100/10{,}000=0.610\). These are posterior probabilities conditional on the data and model.
The 5, 10, and 20 mg doses each meet the strict rule because their estimates are below 0.25. The 40 mg dose does not: 0.610 is greater than 0.25. Among admissible doses below the current 40 mg dose, 20 mg is the highest. The next dose under this example’s rule is therefore 20 mg, subject to any additional protocol constraints.
The raw DLT proportion at 40 mg is \(2/3=0.667\), or 66.7%. That proportion is important evidence, but it is not the 61.0% posterior overdose probability. Nor does the 0/3 observed at 20 mg establish zero toxicity risk: the model combines all cohorts with the prior to estimate uncertainty at that dose.
Worked Example: No Lower Dose Meets the Cutoff
In another hypothetical trial, the dose levels are 2.5, 5, 10, and 20 mg. The observed counts are 1/3, 2/3, 2/3, and 3/3 DLTs, respectively; the current cohort received 20 mg and has completed follow-up. Suppose the posterior computation gives 2,900, 4,800, 7,400, and 9,300 of 10,000 draws above the 30% overdose boundary at those doses.
The estimated overdose probabilities are \(2{,}900/10{,}000=0.290\) at 2.5 mg, \(0.480\) at 5 mg, \(0.740\) at 10 mg, and \(0.930\) at 20 mg. Even the lowest dose has an estimated overdose probability of 0.290, which is not less than the cutoff of 0.25. Thus, there is no admissible dose in the tested range under this rule. The action is to stop dosing rather than to assume that the lowest dose must be safe.
This conclusion does not assert that 2.5 mg has a 29.0% chance of causing a DLT in an individual. It says that, given the data and model, the posterior probability that the dose’s underlying DLT probability exceeds 30% is estimated to be 29.0%.
Worked Example: A DLT at the Current Dose, but a Lower Dose Remains Admissible
Consider dose levels 1, 2, 4, 8, and 16 mg. The complete data are 0/3 DLTs at 1 mg, 0/3 at 2 mg, 0/3 at 4 mg, 1/3 at 8 mg, and 1/3 at 16 mg. The current dose is 16 mg. Suppose 10,000 posterior draws yield 100, 300, 900, 2,200, and 4,400 draws above the overdose boundary at the respective doses.
The estimated probabilities are 0.0100, 0.0300, 0.0900, 0.220, and 0.440. The 16 mg dose fails the rule because \(0.440>0.25\), while 8 mg satisfies it because \(0.220<0.25\). Since 8 mg is the highest admissible tested dose below 16 mg, it is the next dose under the stated decision rule.
A DLT at 16 mg does not mechanically require a reduction by exactly one dose level in every trial. Here the posterior evidence supports 8 mg under the cutoff; in a different dataset, the same observed count could produce a different posterior probability because the evidence at all doses and the prior both matter.
Why This Rule Is Useful—and Where It Can Fail
An observed DLT proportion and a posterior overdose probability answer different questions. The observed proportion \(y_j/n_j\) summarizes outcomes in one cohort. The posterior overdose probability evaluates uncertainty about the underlying dose-specific DLT probability relative to a prespecified boundary, using the full model and the accumulated data. A small cohort can have an extreme observed proportion while leaving substantial uncertainty, and a dose with no observed DLTs can still have a meaningful posterior probability of exceeding the boundary.
The overdose cutoff makes the safety requirement explicit. Lowering \(\xi\) makes the admissibility rule more conservative; raising it can make more doses admissible. The cutoff should be prespecified and interpreted alongside the model, not adjusted after seeing a concerning cohort merely to obtain a preferred next dose. Posterior probabilities near the cutoff deserve particular care: numerical Monte Carlo error, prior sensitivity, and model fit may affect whether the estimated value falls just above or below it.
Useful checks before acting include confirming that the high-dose cohort has complete DLT assessment, checking the cohort counts and dose assignments, reviewing posterior sampling diagnostics, and examining the fitted dose-response curve across the dose range. Posterior draws should explore the relevant parameter space adequately; convergence problems or poor effective sample sizes can make estimated tail probabilities unreliable. Prior sensitivity is also important when data are sparse, especially at untested or weakly informed doses.
The monotone BLRM assumes an increasing dose-toxicity relationship. Observed cohort rates need not increase monotonically, as in the first example, but a strongly inadequate dose-response model can still give misleading inferences. Model fit, prior plausibility, and the dose range matter. Extrapolation beyond well-supported doses should not be treated as precise merely because the model returns a probability. Incomplete follow-up is another important failure mode: a cohort with participants still at risk of a DLT should not be treated as if all outcomes were observed. The treatment of incomplete follow-up requires an explicit protocol method.
This posterior-rule approach differs from choosing the next dose from raw rates alone: raw rates do not directly encode uncertainty or borrow information across doses. It also differs from using the posterior mean DLT probability as the safety criterion. A posterior mean below 30% does not guarantee that the posterior probability of exceeding 30% is below the cutoff; the distribution’s uncertainty and upper tail matter. Finally, this immediate de-escalation rule is not a final dose-selection procedure. It specifies a safety-screened next action after the current cohort, not a complete evaluation of a dose’s overall clinical suitability.
Confirm the DLT definition, dose, cohort counts, and completion of the relevant follow-up period.
Combine the likelihood from all observed cohorts with the prespecified prior to obtain the posterior distribution.
Estimate each candidate dose’s posterior probability of exceeding the overdose boundary and compare it with the cutoff.
Choose the highest admissible lower dose if one exists; otherwise stop under the overdose-control rule.
Check Your Understanding
Apply the stated rule: a dose is admissible only when its posterior overdose probability is strictly less than the cutoff.
- With cutoff 0.25, is a dose with posterior overdose probability 0.24 admissible? What about a dose with probability exactly 0.25?
- The current dose is 30 mg. Two lower candidates, 10 mg and 20 mg, have posterior overdose probabilities 0.06 and 0.31. Which dose is selected under the rule, assuming no other constraints?
- A dose has posterior overdose probability 0.18. Is it admissible with cutoff 0.20? Would it be admissible with cutoff 0.15?
- After a fully followed cohort with 3/3 DLTs at the current dose, every tested dose has posterior overdose probability at least 0.27. What action follows if the cutoff is 0.25?
Solutions. (1) The dose at 0.24 is admissible because \(0.24<0.25\). A dose at exactly 0.25 is not, because the rule requires a value strictly below the cutoff. (2) The 10 mg dose is admissible because \(0.06<0.25\); 20 mg is not because \(0.31>0.25\). Select 10 mg, the highest admissible lower candidate. (3) It is admissible with cutoff 0.20 because \(0.18<0.20\), but not with cutoff 0.15 because \(0.18\) is not below 0.15. (4) No tested dose is admissible, so stop dosing under this rule.