Time-to-Event 3+3 Rules for Late Toxicity

Why Follow-Up Time Matters

A conventional 3+3 decision is made from the number of dose-limiting toxicities (DLTs) observed among participants assigned to a dose after the protocol-defined DLT observation period. When some toxicities occur late in that period, waiting for every participant to complete it can slow enrollment. Deciding earlier can shorten that wait, but a participant who has not yet had a DLT may still experience one later. A time-to-event 3+3 variant makes this trade-off explicit by linking decisions to both observed DLTs and elapsed follow-up.

There is no single universal partial-follow-up gate that defines every time-to-event 3+3 variant. The rule must be specified in the protocol. This tutorial develops one transparent cohort-based version: a participant with an observed DLT is counted immediately; a participant without an observed DLT is decision-ready only after reaching a prespecified minimum follow-up time. A cohort decision waits until all its members are decision-ready, then applies the usual 3+3 count rule. This approach can reduce waiting without treating short, event-free follow-up as a completed toxicity window.

The gate does not declare that a participant who has reached \(\tau\) is free of DLT through \(W\). It means only that the protocol permits the 3+3 decision to proceed with the event information available at that minimum follow-up. Events arising later in the observation window remain relevant to ongoing safety monitoring and must be handled according to the protocol.

Specify the Rule Before Applying It

For the version used here, each cohort has three participants. At a dose, enroll three participants and assess their decision-readiness. If two DLTs occur among those three, do not escalate. If the cohort is ready and has zero DLTs, escalate to the next dose. If it is ready and has one DLT, enroll three additional participants at the same dose; do not escalate while that expansion cohort is still awaiting its gate. Once all six are decision-ready, escalate if there are zero or one DLT among the six, and do not escalate if there are two or more.

This is a cohort-based, gated rule; it does not permit rolling enrollment into the next dose while the current cohort is waiting. A protocol could specify a different enrollment schedule or gate, but it would need to state those rules precisely. The familiar count thresholds are those of the conventional 3+3 rule; the time-to-event feature here changes when the count can be used, not the count thresholds themselves.

The rule requires several practical choices to be fixed in advance: the DLT definition, \(W\), the minimum follow-up \(\tau\), how time is measured, and which events count as DLTs. Follow-up is generally measured from the participant’s first dose at that level, under the protocol’s specified calendar convention. DLTs are counted as observed events, not as fractions or weights based on time. In particular, a participant with no DLT and follow-up shorter than \(\tau\) is not counted as a DLT-free participant for a cohort decision.

Applying the Count Rule After the Gate

The time gate and the 3+3 count rule answer different questions. The gate asks whether enough follow-up has accrued to make the scheduled decision. The count rule asks what to do given the observed number of DLTs. Keep these stages separate: an event-free participant below the gate is not evidence that the cohort has passed, and reaching the gate does not change how an observed DLT is counted.

Why the Gate Does Not Eliminate Late Risk

A gate short of the full window trades some information for speed. A simple survival argument shows why an event-free participant at \(\tau\) cannot be treated as certain to remain event-free until \(W\). Let \(T\) be the time to DLT, and let \(S(t)=P(T>t)\) be the probability of remaining DLT-free beyond time \(t\), under a specified dose and population. Among participants still DLT-free at \(\tau\), the probability of a DLT after \(\tau\) but by \(W\) is

The second equality follows because the event \(\{\tau<T\leq W\}\) is the part of \(\{T>\tau\}\) removed when restricting to \(\{T>W\}\). The resulting conditional probability is positive whenever some participants who are event-free at \(\tau\) go on to have a DLT before or at \(W\). Its value depends on the timing distribution; it cannot be determined from the gate alone.

For illustration only, suppose DLT times among participants who will have a DLT within the window are uniformly distributed over the 30-day window, and suppose \(\tau=15\) days. Among that event group, half of the events occur after day 15. This does not mean the conditional probability for all participants is one half: that also depends on the overall probability of DLT by day 30. The uniform timing assumption is a simplifying illustration, not a general property of toxicity data.

Operational Checks and Common Pitfalls

A dose-escalation log should record, for each participant, the dose, start of the DLT clock, current follow-up time, whether a DLT has been observed, and whether the gate is met. Before applying a count threshold, verify that the cohort is ready under the protocol’s exact definition. Recalculate readiness when follow-up changes, rather than assuming that all participants have accrued the same time.

• Counting a short event-free follow-up as a completed observation. A participant followed for 10 days with no DLT does not meet a 14-day gate. The absence of an event so far is not a completed negative observation.
• Confusing the gate with the observation window. If \(W=28\) days and \(\tau=14\) days, reaching day 14 makes a participant decision-ready under this variant; it does not complete the 28-day window.
• Delaying recognition of an observed DLT. A DLT is counted when it occurs, even if the participant has not yet reached \(\tau\). A minimum follow-up gate is not a reason to ignore an event already observed.
• Escalating while a cohort member has not met the gate. In the cohort-based rule specified here, all members must be decision-ready before the scheduled cohort decision, unless a stopping condition such as two DLTs has already been reached.
• Assuming every time-to-event 3+3 variant has the same gate. The gate and enrollment schedule are protocol choices. Applying an informal or remembered threshold instead of the written rule can change trial decisions.

A particularly common error is to say that a cohort has passed because most members meet the gate. Under the version defined here, all DLT-free participants in the cohort must meet it. For example, with a 15-day gate, follow-up times of 30, 16, and 14 days do not pass: only the first two meet the threshold. The correction is to check each participant’s comparison separately before applying the count rule.

How This Variant Compares With Other Approaches

The gated variant is most useful when the protocol can state a credible minimum follow-up threshold and when reducing waiting time is important. Its limitations are equally important: a shorter gate leaves less observed time for detecting DLTs; a fixed count rule does not model the exact timing of events; and decisions may be sensitive to the chosen threshold. If many DLTs occur late, a gate that appears operationally convenient may allow escalation with substantial unobserved toxicity risk. The gate should therefore be justified for the treatment and observation window, not selected solely to speed enrollment.