Writing an Unambiguous Escalation Rule Set

From a Design Description to an Operational Rule

A rule-based escalation design is not protocol-ready merely because its general logic is familiar. Terms such as “complete the cohort,” “evaluate toxicity,” or “escalate if acceptable” can conceal choices about timing, denominators, pending outcomes, and authority to act. If those choices are left unstated, two teams can follow the same nominal design and make different decisions from the same participant outcomes.

The task here is to express a proposed rule as a sequence of observable states and actions. The examples use a hypothetical three-participant cohort design: escalate after 0 of 3 dose-limiting toxicities (DLTs); expand to 6 after 1 of 3; escalate after no more than 1 of 6; and stop escalation at a dose after at least 2 of 3 or 2 of 6. These numbers illustrate how to write rules, not a recommendation for a particular trial. The protocol must define its own DLT criteria, review process, dose levels, and intended dose-selection statement.

The details should be fixed before enrollment, not improvised after an inconvenient outcome. In particular, the protocol should identify the dose levels and starting dose; cohort size; DLT observation window; evaluability and replacement rules; decision thresholds; handling of pending participants; pause triggers and restart authority; and stopping and dose-selection language. Earlier material on handling unevaluable participants and safety pauses provides the relevant design concepts; the focus here is making their operational consequences explicit in one coherent rule set.

Write the Decision Sequence in Order

A useful rule set follows the actual order of operations. First determine whether enrollment may proceed. Then enroll the specified cohort at the authorized dose, collect outcomes for the stated evaluation period, decide which participants are evaluable, and calculate the DLT count using the stated denominator. Only after the required outcomes are available may the escalation table be applied. This ordering prevents a recommendation based on a partly observed cohort.

For a hypothetical design with an evaluation window of 28 days from first administration, suppose a participant is evaluable if they either complete day 28 with adequate toxicity assessment or experience a DLT before day 28. A participant who withdraws early for a reason unrelated to toxicity and has no DLT is not evaluable and is replaced. A participant who has a DLT and then withdraws remains evaluable and contributes that DLT. These definitions specify both the window and how early events affect the denominator; another protocol could choose different definitions, but it must state them.

A decision table can expose missing cases. For the illustrative rules, the relevant states are the first cohort of 3 evaluable participants and, only if exactly 1 of those 3 has a DLT, the expanded total of 6 evaluable participants. The table uses “stop escalation at this dose” for a dose that crosses the specified DLT boundary. The protocol must separately define what stopping means for the study and how a dose for further study is selected.

Check That the Rules Are Complete and Consistent

The decision table must cover every reachable state, and its conditions must not overlap. Here, after 3 evaluable participants, the possible integer DLT counts are 0, 1, 2, and 3. The conditions “0,” “1,” and “2 or 3” partition those possibilities: each count has exactly one action. After expansion, counts range from 0 to 6; “0 or 1” and “2 to 6” likewise cover all possible counts without overlap. The expansion state is reachable only after 1 of the initial 3, so it cannot be confused with a separate state at the same dose.

This is a simple but important justification: if every reachable state maps to one and only one action, the rule is executable without an undocumented judgment about which branch applies. The table alone is not sufficient if a state can be delayed by incomplete follow-up, a safety pause, or a replacement. Those transitions also need rules.

Evaluability, Replacement, and Pending Outcomes

Replacement rules should not erase safety information or permit a decision from an inconsistent denominator. A protocol should say when a participant is unevaluable, whether replacement is allowed, whether replacement occurs at the same dose, and whether the participant’s observed toxicities still enter safety review. It should also specify whether enrollment continues while another participant’s window is pending.

For the illustrative cohort design, a conservative operational rule is to pause enrollment at a dose until each participant in the active cohort has either completed the DLT window or experienced a DLT. A participant with a DLT counts as evaluable immediately under the stated definition. A participant without a DLT who stops treatment early for a non-DLT reason is replaced; the replacement receives the same dose, and the cohort decision waits until the target number of evaluable participants is reached. The protocol should define whether other interruptions or missing assessments make a participant unevaluable rather than leaving that judgment implicit.

Pauses, De-escalation, and Stopping

Escalation rules describe what to do when no override applies. A pause rule must therefore state which actions are suspended, who reviews the event, and what authorizes resumption. For example, a protocol might require an immediate hold on new enrollment and dose administration after any suspected treatment-related death. It should identify who is notified, what information is reviewed, the decision-maker authorized to restart, and whether written restart authorization is required. A pause is not itself a decision to escalate, de-escalate, or stop; those outcomes follow only after the specified review.

“De-escalate” also needs a defined destination and trigger. If the rule is that an unacceptable DLT count at a dose stops escalation there, the protocol should state whether enrollment at that dose stops, whether a lower dose may be evaluated, and who makes that determination. At the lowest dose, there may be no lower level to use. The protocol should say what happens then, rather than implying that a lower dose exists. It should also distinguish stopping escalation from stopping all trial treatment or terminating the study.

Diagnostics and Common Failure Modes

Before finalizing a rule set, test it against a decision table and hypothetical edge cases. A useful audit asks whether every reachable DLT count has exactly one action; whether a dose decision can occur with pending outcomes; whether replacement changes the denominator; whether a pause overrides ordinary enrollment instructions; and whether stopping at the first or highest dose has a stated consequence. Earlier material on auditing escalation decisions is useful for checking whether actions follow the rules; this drafting step checks whether the rules themselves are complete enough to audit.

• Ambiguous denominator. “Two toxicities in the cohort” is unclear if participants have been replaced. State whether the denominator is enrolled participants or evaluable participants, and how replacements count.
• Premature decision. “Escalate after three patients” may be read as three enrolled, even when one has pending follow-up. State that the required evaluations must be complete.
• Unbounded replacement. “Replace unevaluable participants” can leave accrual open-ended. State who determines unevaluability, whether replacement is permitted, and whether enrollment is held pending a decision.
• Conflicting authorities. A general escalation rule and a safety hold may appear to require incompatible actions. State explicitly that the pause takes precedence until the designated review permits resumption.
• Unclear stopping endpoint. “Stop at this dose” might mean close that dose, stop escalation, or end the trial. Name each intended consequence separately.

These issues are not fixed by adding more complicated statistical calculations. They are fixed by defining states, precedence, and actions precisely. Compared with a fixed 3+3 rule set, the rolling six design described in earlier material allows staggered enrollment and therefore requires especially explicit rules about pending DLT windows and when a decision may be made. Different designs can use different operational rules; the writing principle is the same.