Transitioning From Escalation To Dose Expansion

Why the Transition Needs Its Own Rule

Escalation is designed to learn about doses while limiting exposure to unacceptable toxicity. Expansion has a different purpose: to collect more evidence at a dose, or a small set of doses, that may merit further development. A dose that is permissible to explore during escalation is not automatically ready for expansion. The transition should therefore be governed by a rule written before outcomes are reviewed, rather than by an informal judgment that a dose “looks promising.”

The rule must use the joint efficacy–toxicity evidence relevant to the treatment decision. In particular, a dose may have an encouraging response rate but also an unacceptable toxicity rate; another may be safe but show little evidence of benefit. A transition rule makes these tradeoffs explicit and records what evidence is sufficient to stop escalating, which dose or doses may expand, and what new data will be used to revisit the recommendation.

Define the Evidence Before Applying a Gate

Suppose each patient treated at dose \(d_j\) is assessed for binary efficacy \(E\) and binary toxicity \(T\) over protocol-defined assessment windows. Let \(E=1\) denote the specified efficacy outcome and \(T=1\) the specified toxicity outcome. The four possible joint outcomes are efficacy without toxicity, efficacy with toxicity, no efficacy without toxicity, and neither efficacy nor freedom from toxicity. The windows and outcome definitions must be fixed in the protocol; changing them during the transition changes the evidence being summarized.

For \(n_j\) evaluable patients at dose \(d_j\), let \(a_j\) be the number with efficacy and no toxicity, \(b_j\) the number with both efficacy and toxicity, \(c_j\) the number with neither efficacy nor toxicity, and \(d_j^\ast\) the number with toxicity but no efficacy. The symbol \(d_j^\ast\) is a count, distinct from the dose label \(d_j\). These counts sum to \(n_j\).

Here \(\hat q_j\) estimates the probability of the joint favorable outcome, efficacy without the specified toxicity, while \(\hat p_{T,j}\) estimates the toxicity probability. These are different summaries: \(\hat q_j\) is not generally the product of the marginal efficacy rate and the no-toxicity rate. Patients’ efficacy and toxicity outcomes may be dependent, so the four-category data should be retained when applying a joint rule.

A Practical Gate and Its Justification

One transparent rule is to require (1) a minimum number of evaluable patients, (2) a toxicity rate no greater than a prespecified limit, and (3) a joint favorable-outcome rate at least as large as a prespecified target. For illustration, take \(n_{\min}=12\), a toxicity limit of \(0.30\), and a joint favorable-outcome target of \(0.35\). These are hypothetical design values, not universal clinical thresholds. The values must be justified for the disease, treatment, assessment windows, and intended use of the dose.

The rule can also require that an earlier safety stopping rule has not been triggered and that the earlier efficacy-futility screen has been cleared. Those are distinct gates: passing an efficacy-futility screen means there is not enough evidence to stop for lack of efficacy under that screen; it does not, by itself, establish acceptable toxicity or justify expansion.

The basic justification for using the proportions is direct. Under the stated independent, stable multinomial outcome model, each indicator for a joint favorable outcome has expectation \(q_j=P(E=1,T=0\mid d_j)\). The average of \(n_j\) such indicators is \(\hat q_j\), so its expectation is \(q_j\). The same reasoning applies to the toxicity indicators and \(\hat p_{T,j}\). Thus the summaries target the probabilities named by the gate. They can still be imprecise at small \(n_j\); the minimum sample-size condition and subsequent expansion data matter precisely because a threshold crossing is not certainty.

If more than one dose passes, a protocol needs a tie-breaking or selection rule. For example, it may select the eligible dose with the largest \(\hat q_j\), subject to the toxicity limit, while preferring a lower dose when joint outcomes are similar. A model-based constrained selection rule is another option. The essential point is not which rule is universally best—there is no universal rule—but that eligibility and selection are separately specified. A dose can pass the gate without being the preferred expansion dose.

What Expansion Adds to the Recommendation

Expansion should be designed to answer a stated question. If one dose is clearly eligible and the objective is to characterize that dose, additional patients may be treated there under the protocol’s monitoring plan. If two eligible doses remain plausible and choosing between them matters, expansion can include both doses, with allocation fixed or randomized in advance. A comparison is more interpretable when patients are enrolled concurrently under the same eligibility, assessment, and follow-up procedures; a simple historical comparison can be confounded by changes over time.

As outcomes accrue, add the new patients’ four-category counts to the corresponding dose’s escalation counts and recompute the joint summaries. The final recommendation should apply the prespecified selection rule to the updated evidence, not automatically retain the dose chosen at the transition. Conversely, expansion data should not be used to invent a more favorable threshold after seeing the results. If the updated data no longer satisfy a gate, the protocol should specify whether to pause enrollment, stop that dose, consider another eligible dose, or declare that no dose meets the criteria.

Diagnostics, Alternatives, and Limitations

Before acting on a transition result, verify the data and the rule’s implementation. Check that the four counts sum to the stated evaluable total, that toxicity is counted across the specified window, and that efficacy is assessed at the same dose-specific time point for all patients. Review whether the conclusion depends on a single outcome or a small change in a count. A threshold barely crossed is still a pass under a deterministic rule, but it should not be described as strong evidence; the planned expansion is meant to add information, not to certify the first estimate.

A toxicity-only escalation rule is simpler, but it may advance a dose with little efficacy evidence. A joint rule addresses that gap, at the cost of requiring clear definitions for both endpoints and an explicit way to trade off benefit and harm. A model-based rule can borrow structure across doses and express a decision in terms of posterior probabilities, as in the earlier material on posterior probabilities of acceptable doses. It depends on its model and prior assumptions. A transparent empirical gate is easier to audit but can be unstable when sample sizes are small and does not automatically account for uncertainty. Neither approach removes the need to prespecify thresholds and evaluate the design before use.

Other limitations remain. A fixed threshold may not capture the clinical consequences of different toxicities, response durations, or delayed outcomes. A dose-specific rate can also be misleading if patient mix or outcome assessment changes between escalation and expansion. Finally, expansion is not a substitute for later confirmatory evidence: the cohort may be too small or too selected to establish effectiveness or characterize uncommon harms.

Common Mistake

• Wrong reasoning: “The response rate is high and the observed toxicity rate is under the limit, so the dose should expand.”
• Correction: Apply every prespecified gate, including the joint favorable-outcome criterion, minimum evaluable sample size, and any safety or futility conditions. Then use the stated selection rule among eligible doses. A transition decision is a protocol action based on limited evidence, not a declaration that the dose is best.