Partitioning Patients Into Stagewise Cohorts

Why Cohort Boundaries Need an Operational Rule

In a seamless phase 2/3 trial, the phase label attached to a participant can affect how the study team summarizes stage-specific information and documents which data were available at a transition. A calendar description such as “phase 2 patients were enrolled before the transition” is not sufficient on its own: it leaves open whether enrollment means consent, randomization, first dose, or some other event. It also says nothing about a participant whose assigned endpoint is observed only after the transition.

The core distinction is between which enrollment cohort a participant belongs to and which of that participant’s data are available at a particular cut. Define the cohort using an event and boundary rule that can be applied consistently to every participant. Record follow-up and data availability separately. An endpoint that crosses the transition does not, by itself, move a participant to a different cohort.

Specify the Event, Boundary, and Data Cut

Let \(t_i\) be the timestamp at which participant \(i\) is randomized, and let \(\tau\) be the effective transition timestamp. Use a single documented time standard, such as UTC, and define the boundary explicitly. One simple convention assigns randomizations strictly before \(\tau\) to stage 1 and randomizations at or after \(\tau\) to stage 2:

This half-open boundary convention makes the cohorts disjoint and exhaustive for randomized participants: a timestamp cannot be both before and at-or-after \(\tau\). A protocol could specify a different convention, but it must settle the exact-boundary case before enrollment. If a trial uses a different event as the enrollment marker, it must say so and apply it consistently. In particular, signing consent before \(\tau\) does not make a participant stage 1 if the defined enrollment event is randomization and randomization occurs at or after \(\tau\).

A data cut is a separate timestamp \(d\) at which the study freezes or extracts specified data. For a fixed-horizon endpoint requiring \(L\) days of follow-up, the nominal endpoint window for participant \(i\) ends at \(t_i+L\). Whether the endpoint is complete at cut \(d\) depends on the endpoint’s ascertainment rules and the actual available records, not just on cohort membership. For example, reaching \(t_i+L\) may be necessary but not sufficient if records have not yet been validated.

A practical data record therefore keeps at least the randomization timestamp, assigned stage, endpoint window, endpoint status at each relevant cut, and the date that status was established. “Not yet complete at the cut” should be distinguishable from “complete but missing,” “not collected,” or “not applicable.” The precise endpoint status categories depend on the endpoint and its prespecified ascertainment rules.

Follow-Up Can Cross the Transition

For a participant randomized before \(\tau\), a fixed endpoint window may end after \(\tau\). The participant remains in the stage 1 enrollment cohort under the rule above. If the data cut occurs before the endpoint window is complete, the endpoint is not yet mature at that cut. If the endpoint is later observed, the record can be updated as available while retaining the original stage label.

This distinction matters for endpoints whose measurement requires a specified follow-up duration. It also matters for event-time outcomes: an event may occur after the transition, while a participant who has not yet had an event may still be under observation. The date of an outcome event, or the date on which its status becomes known, does not replace the randomization date as the cohort-defining event. The endpoint’s analysis and censoring rules must still be followed; cohort assignment alone does not determine how an endpoint is analyzed.

Why the Partition Works—and What It Does Not Claim

The set-based definition makes the basic partition property explicit. Let \(S_1=\{i:t_i<\tau\}\) and \(S_2=\{i:t_i\geq\tau\}\), for all randomized participants with valid timestamps. Every such \(t_i\) is either less than \(\tau\) or greater than or equal to \(\tau\), so each participant is in at least one set. No timestamp can satisfy both conditions, so no participant is in both sets.

where \(S\) is the set of randomized participants covered by the rule. A data cut can change which endpoint records are available, but it cannot change either inequality defining \(S_1\) and \(S_2\). This is the justification for maintaining a stable cohort label while updating endpoint-status fields over time.

The partition is an operational classification, not a claim that stages are statistically independent, that participants have completed follow-up, or that data from one stage can be used for a particular inferential purpose. Those questions require their own design and analysis rules. Here, the goal is to make the participant roster and the status of its data interpretable without conflating them.

Operational Checks and Common Failure Modes

Before a transition, the protocol and data-management plan should identify the effective timestamp, the event used for cohort assignment, the boundary convention, and the data cut or cuts. At each cut, reconcile the participant roster against the randomization system and check that every record has one stage label. The endpoint-status fields should then be checked against the endpoint window and the cut date.

• Ambiguous date fields. Consent, randomization, first dose, and first endpoint assessment can fall on different dates. Confirm which field defines enrollment; do not choose a convenient date after seeing the data.
• Unresolved boundary timestamps. Local time zones, daylight-saving changes, rounded timestamps, or inconsistent system clocks can put records on the wrong side of \(\tau\). Use a documented time standard and investigate records whose recorded precision cannot resolve the boundary.
• Outcome-dependent membership. Classifying a participant according to when an outcome becomes available can make the cohort depend on follow-up or outcome behavior. Keep cohort assignment based on the specified enrollment event; record availability separately.
• Incomplete follow-up mislabeled as missing. A participant whose endpoint window has not ended at the cut is not automatically a participant with a missing completed endpoint. Distinguish “not yet mature” from the endpoint’s defined missing-data statuses.
• Duplicate or repeated records. Reconcile participant identifiers and randomization records. If the protocol permits re-randomization or multiple randomization episodes, specify whether the classification unit is the participant or the episode rather than silently counting one person twice.

The main alternative is to define stages by observation period or by which database cut contains an outcome. These can be useful labels for describing when measurements were made or became available, but they answer different questions from “which enrollment cohort does this participant belong to?” A single participant can have an enrollment-stage label and later observations collected after the transition. A data-cut-based label can also change from one extraction to the next, making the participant roster unstable. For a stagewise enrollment cohort, the randomization-time rule is usually clearer; use additional variables to describe observation timing and data availability.

A common mistake is to reason, “The endpoint was observed after the transition, so this person is a stage 2 participant.” That conclusion confuses outcome timing with the cohort-defining event. Under an enrollment-defined rule, the correction is direct: use the randomization timestamp to assign the stage, then determine separately whether the endpoint was complete and available at the specified cut. The limitation is equally important: this classification rule does not resolve how incomplete data should be analyzed, nor does it establish whether cross-transition observations are eligible for a particular analysis.