Outcome Maturity at Selection Looks

Why Maturity Matters at a Selection Look

At an interim look, participants may have been randomized at very different times. Some have had enough follow-up for the intermediate outcome to be observed according to the protocol; others have not. Counting only the available outcome values without checking how they became available can make an arm appear more promising simply because its participants have had more time to contribute information. Conversely, treating every participant without a recorded outcome as if they had no event confuses incomplete observation with an observed outcome.

The first task is therefore to state what “mature” means for the intermediate outcome being used. Maturity is about whether that outcome can be ascertained under a prespecified rule at the selection look. It is not, by itself, a claim that the outcome is favorable, that an event has occurred, or that the resulting comparison is unbiased.

For participant \(i\), let \(E_i\) be the enrollment time, \(t\) the calendar time of the selection look, and \(F_i(t)=t-E_i\) the available follow-up, when \(E_i\le t\). For a rule requiring at least \(\tau\) time units of follow-up, maturity can be represented by \(M_i(t)=1\) when the rule is met and \(M_i(t)=0\) otherwise. The required rule depends on the endpoint: a fixed-landmark measurement, for example, may require a measurement in a specified window rather than simply \(F_i(t)\ge\tau\).

Among \(n_a\) participants assigned to arm \(a\) by the look, let \(m_a\) be the number whose outcomes are mature. The observed mature fraction in that arm is

The denominator should be stated explicitly. Here it is all participants assigned to the arm by the look, including those whose outcomes are not yet mature. If the intended quantity is instead the fraction of participants who have reached a particular follow-up age, specify that definition and its denominator. Either way, report the number of mature outcomes as well as the fraction: the same fraction can represent very different amounts of information in a small and a large arm.

Accrual, Follow-Up, and a Simple Maturity Calculation

A simple planning calculation illustrates how a fixed follow-up requirement limits maturity at a calendar-time look. Suppose participants are enrolled uniformly over an interval of length \(A\), and the look occurs at the end of that interval. Suppose also that an outcome requires at least \(\tau\) time units of follow-up, with \(0<\tau<A\), and that every participant enrolled at least \(\tau\) time units before the look has an ascertainable outcome. Ignoring enrollment exactly at the look and other losses to follow-up, the mature fraction is

To see this, take the look as time \(A\) and an enrollment time \(u\) uniformly distributed from 0 to \(A\). Follow-up is \(A-u\). Maturity requires \(A-u\ge\tau\), or \(u\le A-\tau\). The eligible portion of the enrollment interval has length \(A-\tau\), out of a total length \(A\), giving the stated fraction. If \(A\le\tau\), no participant can meet this full follow-up requirement under these assumptions, so \(q=0\).

This calculation is a planning approximation, not a substitute for checking actual follow-up at the look. Nonuniform accrual, pauses in recruitment, missed assessments, withdrawal, and different enrollment timing across arms can all change the observed maturity fraction. An event observed before \(\tau\) may also make a fixed-landmark outcome ascertainable earlier, depending on its definition; the simple calculation assumes that everyone needs the full follow-up period.

Maturity Is Not the Same as an Outcome Value

For a binary intermediate outcome, an outcome proportion among mature participants can be calculated as \(\hat p_a=x_a/m_a\), where \(x_a\) is the number with the outcome of interest among the \(m_a\) mature participants. This is a description of the observed mature subset. It is not automatically an unbiased estimate for all participants assigned to arm \(a\). In particular, if maturity differs by arm because the arms have different enrollment timing, comparing these subsets may compare participants with different calendar-time or follow-up profiles.

For a fixed-landmark binary outcome, a participant with an observed event by the landmark has a known event status; a participant observed event-free through the landmark also has a known status. A participant lost before the landmark without an event does not have a known binary status at that landmark. Treating that participant as event-free would assign an unobserved outcome. For a time-to-event analysis, right-censored observations can contribute information under the relevant survival-analysis assumptions, but that is a different analysis from simply counting fully ascertainable fixed-landmark outcomes.

A useful first precision check is to calculate the standard error of the difference between two independent mature-subset proportions. If \(x_a\) outcomes are observed among \(m_a\) mature participants in arm \(a\), and \(x_b\) among \(m_b\) in arm \(b\), then under independent Bernoulli outcomes within each arm, the estimated standard error is

This calculation measures sampling uncertainty for the observed mature-subset contrast under the stated model. It does not correct bias caused by selective or unequal maturity, and it does not account for a design that induces dependence between outcomes. A narrow or wide standard error alone cannot establish that the mature participants represent all randomized participants.

Why Incomplete Maturity Can Destabilize Arm Comparisons

When fewer outcomes are mature, the available outcome proportions usually have greater sampling variability. Greater variability makes the observed ordering of arms more sensitive to chance: an arm with a slightly larger observed proportion can have a lower underlying proportion than its apparent competitor. The effect is especially important when estimated differences are small relative to their standard errors. Maturity fractions help describe how much follow-up is available, while mature counts show the actual denominators contributing to a comparison.

This precision calculation is only one part of a maturity assessment. When maturity is related to prognosis, treatment, or calendar time, incomplete outcomes may affect the comparison in ways that a larger standard error cannot capture. Useful checks include:

• Report, by arm, the number assigned, the number mature, and the maturity fraction, using the same look date and ascertainment rule.
• Summarize follow-up duration and enrollment timing by arm, rather than relying only on the overall mature fraction.
• Check why outcomes are immature: insufficient time since enrollment, missed visits, withdrawal, or another reason may have different implications.
• Compare maturity across arms and across enrollment periods. Differences can reveal unequal opportunity to contribute an outcome.
• Verify that an observed event, an observed event-free landmark, and censoring before the landmark have not been coded as interchangeable states.

Choosing and Interpreting a Maturity Summary

A fixed calendar-time look and an event-driven look answer different operational questions. At a calendar-time look, participants can have a broad range of follow-up; maturity should be reported directly rather than inferred from the date of the look. At an event-driven look, a target number of events may have occurred, but that does not imply that all participants have mature fixed-landmark outcomes. The event target and the intermediate-outcome ascertainment rule are separate quantities.

The mature fraction is also not the only possible summary. Report the maturity count and fraction by arm, and, when relevant, describe the distribution of follow-up time or the number of outcomes known through the landmark. A single pooled fraction can conceal an arm-specific imbalance. A single count can conceal that one arm is much smaller than another. These summaries help characterize the available data, but they do not by themselves resolve missing-outcome bias or determine a selection decision.

The key assumptions behind the calculations should be made explicit. The uniform-accrual formula assumes enrollment is uniform over the stated interval and that maturity depends only on meeting the specified follow-up requirement. The binomial standard-error formula assumes independent outcome indicators within and across arms for the comparison shown. In a trial with dependence, clustering, or a different analysis model, its uncertainty calculation may not apply. Likewise, interpreting a mature-subset contrast for the full randomized arms requires additional justification that the observed subset is suitable for that interpretation; randomization alone does not ensure it when post-randomization maturity determines inclusion.

Common Pitfall