De-Escalation and Trial Stopping Rules in Model-Based Escalation

Why De-Escalation and Stopping Need Separate Rules

Model-based escalation produces a posterior distribution for toxicity risk at each dose, but a dose recommendation is not determined by that distribution alone. The trial protocol must specify what posterior evidence counts as too much risk, which dose moves are permitted, and what action follows when a dose fails the safety rule. These choices are related, but they are not interchangeable: a dose may be unsuitable for escalation while remaining suitable to continue treating patients, and a lack of safe escalation options does not by itself establish that every dose is unsafe.

The focus here is a dose-specific posterior safety criterion and the decisions built from it. The criterion can be applied to posterior risk estimates from a CRM, a Bayesian logistic regression model, or a time-to-event model. The data model and its assumptions must be appropriate for the outcome information available; for example, a probability of DLT by the end of the assessment window must be evaluated using a model that accounts for the timing and follow-up represented in the data.

Define the Posterior Safety Criterion First

Suppose the trial has ordered dose levels \(d_1 < d_2 < \cdots < d_K\), and let \(D\) denote the accumulated trial data. For model parameter \(\theta\), let \(p_j(\theta)\) be the probability of a DLT by the end of the specified assessment window at dose \(d_j\). The target toxicity probability is \(\phi\). The trial also specifies a posterior cutoff \(\xi\), between zero and one.

This is a probability about the model’s unknown dose-specific risk, conditional on the data and model. It is not the probability that a patient will experience a DLT, which is \(p_j(\theta)\) conditional on \(\theta\), and it is not a probability that the safety hypothesis is true. A small \(r_j\) means the posterior places relatively little probability on the dose’s DLT risk exceeding \(\phi\); it does not guarantee that the dose is safe.

The rule also depends on modeling and decision assumptions. The assessment window and DLT definition must match the protocol; the posterior must reflect the observed outcomes and any pending follow-up handled by the chosen model; and the dose ordering and any monotonicity assumption must be defensible. If the model assumes toxicity risk increases with dose, evidence that a dose is unsafe may support excluding higher doses as well. That implication depends on the model’s monotonicity structure; it is not a license to mark every dose unsafe when the model does not impose that structure.

From Posterior Risk to an Action

A de-escalation rule begins with the current dose and the set of permissible lower doses. If the current dose fails the safety criterion, the protocol may direct the trial to move down to an acceptable lower dose. When more than one lower dose is acceptable, choosing the highest acceptable dose below the current dose is a common rule: it makes the smallest dose reduction among those acceptable choices. A protocol can instead require a larger reduction or impose additional restrictions, but that action must be stated rather than inferred from the posterior criterion.

For an escalation decision, define the set \(E\) of doses that could be considered next under the protocol’s dose-move restrictions. Escalation is permitted only to a dose in \(E\) that meets the specified safety criterion. If no dose in \(E\) is acceptable, the rule says do not escalate to those candidates. Depending on the current dose’s status and the protocol, the appropriate action might be to remain at the current dose, de-escalate, or pause for review. It does not automatically mean that treatment must stop.

A stronger conclusion is that all dose levels are unsafe under the rule. In a model where each level is evaluated individually, this requires \(r_j > \xi\) for every dose level under consideration. If the model enforces monotonicity and the lowest dose is unsafe, then higher doses are also unsafe under that model. A statement that there is no acceptable dose among a restricted escalation set is weaker: the current dose or a lower dose may still satisfy the criterion.

Worked Example: De-Escalating From an Unsafe Current Dose

Worked Example: No Safe Escalation Candidate Is Not “All Doses Unsafe”

Worked Example: Declaring All Dose Levels Unsafe

Why the Posterior-Drawing Calculation Works

When posterior draws \(\theta^{(1)},\ldots,\theta^{(M)}\) represent the posterior distribution of \(\theta\) given \(D\), the event \(p_j(\theta)>\phi\) can be checked for each draw. Define an indicator \(I_m\) equal to 1 if \(p_j(\theta^{(m)})>\phi\), and 0 otherwise. Its posterior expectation is \(r_j\), because an indicator’s expectation is the probability of its event. The average of these indicators estimates that expectation.

For independent posterior draws, the Monte Carlo standard error of this estimate is approximately \(\sqrt{\widehat{r}_j(1-\widehat{r}_j)/M}\). For correlated simulation draws, this expression must be adjusted for the effective sample size or replaced with an appropriate simulation diagnostic. This error describes numerical uncertainty in estimating \(r_j\) from finite draws; it is not the posterior uncertainty about the dose’s toxicity risk, which is already represented by the event probability \(r_j\).

For example, in the first worked example, \(\widehat{r}_3=0.280\) from \(M=1000\) independent draws. The approximate Monte Carlo standard error is \(\sqrt{0.280(1-0.280)/1000}=\sqrt{0.0002016}\approx0.0142\). That is small relative to the difference between 0.280 and the cutoff 0.25, but it is not zero. If an estimate lies very close to the decision boundary, more simulation or a protocol-defined conservative action may be warranted.

Diagnostics, Alternatives, and Limitations

Before implementing a rule, check that the posterior risk corresponds to the protocol’s DLT definition and assessment window, that each dose is mapped correctly in the model, and that the posterior calculation is stable. Review convergence and effective sample size for posterior simulations, especially for draws near the threshold \(p_j(\theta)=\phi\). Compare results across independent chains or simulation runs when appropriate. Verify that a reported posterior probability is calculated as a tail probability over parameter draws, rather than by comparing the posterior mean risk with \(\phi\): in general, \(P\{p_j(\theta)>\phi\mid D\}\) is not equal to \(I\{E[p_j(\theta)\mid D]>\phi\}\).

Threshold choices are design decisions, not universal constants. A stricter posterior cutoff can classify more doses as unsafe; a less strict cutoff can allow more candidates. The consequences depend on the trial’s target toxicity, dose spacing, available patient data, and the prespecified acceptable level of posterior evidence. A rule should therefore be assessed as part of the design, rather than selected after seeing a particular outcome merely to obtain a preferred recommendation.

The posterior overdose rule is one approach among several. A protocol may use an overdose-control rule such as an EWOC feasibility boundary, a direct posterior interval criterion, or a simpler cohort-based rule. These approaches can yield different decisions because they summarize posterior uncertainty differently and may impose different constraints. Whatever approach is used, distinguish the rule that judges a dose from the operational action that follows. The criterion \(r_j\leq\xi\) is not itself a complete escalation algorithm.

Important failure modes include sparse data at the lowest dose, a poorly calibrated prior, model misspecification, and an unsupported monotonicity assumption. These can make the posterior appear more decisive than the information warrants or make risk estimates at unobserved doses depend heavily on modeling assumptions. If a model imposes monotonicity, a safety conclusion propagated to higher doses is conditional on that structure. If monotonicity is not imposed, evaluate the relevant doses individually and do not infer the status of unexamined doses from their rank alone.

Common Pitfall: Treating Every Failed Escalation as a Trial Stop

A common error is to see that all upward candidates exceed the posterior cutoff and conclude that the trial must terminate because “no dose is safe.” That conclusion confuses a restricted set of escalation options with the full set of trial doses. The correction is to identify the set being tested: if only \(d_3\) and \(d_4\) are eligible escalation candidates, their failure establishes only that escalation to those doses is not allowed at that time. A conclusion that all dose levels are unsafe requires the safety criterion to fail at every dose level included in the stopping rule. The protocol must separately state whether to hold, de-escalate, pause, or terminate when escalation is blocked.