Choosing an Intermediate Outcome for Arm Selection

Why the Choice of Intermediate Outcome Matters

An intermediate outcome can make arm selection possible before the final time-to-event endpoint is mature. But an outcome is not useful for this purpose merely because it is recorded early. It must be available in time for the planned decision, measured well enough to distinguish arms, clinically interpretable, and supported by evidence that it bears a meaningful relationship to the final endpoint.

These requirements answer different questions. Availability asks whether the trial will have usable observations at the selection look. Measurement quality asks whether the recorded values reflect the intended outcome. Clinical relevance asks whether differences matter to patients or clinical practice. Association with the final endpoint asks whether the intermediate outcome is informative about the endpoint of interest. Passing one check does not compensate automatically for failing another.

Suppose \(Z(t^*)\) denotes a candidate outcome measured at a planned time \(t^*\), and \(T\) denotes a participant’s time to the final event. The candidate may be a score, a laboratory measure, a response status, or an earlier event measure; its specific form changes what evidence is needed. The introductory material on defining intermediate outcomes distinguishes the interim measurement from the final endpoint. Here the focus is how to judge a proposed candidate before using it for arm selection.

Four Checks Before Using a Candidate

1. Availability at the actual decision time

Availability is not the same as eventual ascertainment. The question is whether the data will be collected, processed, quality-checked, and accessible before the decision must be made. For a time-to-event trial, a fixed-time assessment may be unavailable for participants who have not reached that time, while a visit-based measure may be delayed by scheduling or data transfer. A recorded value that arrives after the selection meeting cannot guide that meeting.

Estimate the proportion with an assessable outcome at the planned look, separately by arm where appropriate. Examine reasons for missingness, timing of data receipt, and whether outcome maturity differs across arms. A high overall completion proportion can conceal poor availability in one arm or among participants whose outcomes are most relevant to the decision. Availability should be assessed under the anticipated accrual and follow-up pattern, not only in a completed study.

2. Clinical relevance to the decision

Ask what a difference in the candidate outcome means in clinical terms. Is the scale interpretable? Is there a change that clinicians or patients would regard as meaningful? Does the candidate capture an aspect of benefit that is relevant to the final endpoint, or could it improve while the final outcome worsens? For arm selection, the direction of benefit and any decision threshold must be clear before the data are inspected.

Clinical relevance is not established just because a measure changes quickly or has a small p-value. A statistically detectable difference can be too small to matter, and a convenient biomarker can have uncertain meaning for patients. Conversely, a clinically important outcome may be unsuitable for an early look if too few participants can contribute data by the decision date.

3. Measurement quality

The outcome needs a clear measurement protocol: who assesses it, when, with what instrument, and how ambiguous or missing values are handled. Consider repeatability, assessor agreement, calibration, adherence to the protocol, and whether the measurement has the same meaning across arms and sites. If an outcome is classified using a threshold, the threshold and the handling of values near it should be specified in advance.

Measurement error matters particularly when the arm-selection rule depends on small differences. Random error can obscure a real contrast; systematic error can create or distort one. A candidate with high nominal completion but inconsistent measurement may provide less useful information than a somewhat less complete candidate measured reliably. Quality checks should therefore examine both the amount of data and how the data were generated.

4. Evidence of association with the final endpoint

A candidate intended to inform a final-endpoint decision should have a defensible relationship to that endpoint. Relevant evidence may come from prior studies, follow-up data, or a prospectively planned validation component. The evidence should match the proposed context as closely as possible: population, treatment setting, outcome definition, and measurement time all matter.

Association is a screening requirement, not a guarantee. A candidate can be associated with the final endpoint because it reflects prognosis, yet still fail to identify which treatment arm improves the final endpoint. A relationship observed among individuals does not by itself establish that treatment effects on the candidate track treatment effects on the final endpoint. Therefore, do not describe an outcome as a validated surrogate or assume that it preserves the ranking of arms solely because the two outcomes are associated.

Why Availability Changes the Information at a Look

A simple calculation illustrates why availability is part of the statistical design, not just administration. Suppose the candidate is quantitative, has variance \(\sigma_Z^2\) in each of two arms, and the observed values are independent. Let \(n_1\) and \(n_0\) be the numbers enrolled by the look and \(q_1\) and \(q_0\) the proportions with an assessable outcome. If availability is independent of the candidate values within each arm, then the numbers observed are approximately \(n_1q_1\) and \(n_0q_0\). For a difference in arm means, the variance is approximately:

This expression is a planning illustration, not a universal missing-data formula. It assumes independent observations, equal outcome variance in the two arms, and availability unrelated to the candidate values within each arm. If availability depends on outcome, prognosis, or treatment, the observed participants may not represent the intended arm population; simply replacing sample sizes by observed counts does not fix that problem.

When both arms have \(n\) participants and the same availability proportion \(q\), the standard error is \(\sqrt{2\sigma_Z^2/(nq)}\). Relative to complete availability, the standard error is multiplied by \(1/\sqrt{q}\). Thus availability affects how precisely arm differences can be assessed, while selective missingness can also affect what those differences mean.

Apply the Checks to Concrete Candidates

Make the Decision Explicit and Check for Failure

A candidate should not be accepted through an informal impression that it is “early” or “promising.” Write down what would make it suitable before examining the trial’s arm comparisons. A concise assessment can record the planned assessment time, expected availability by arm, clinical interpretation, measurement safeguards, supporting endpoint evidence, and known limitations. The purpose is to make the trade-offs visible and to identify reasons not to use the candidate.

Useful diagnostics include delayed data delivery, unexpectedly different completion across arms, missingness concentrated among participants with worsening health, disagreement between assessors, frequent movement across a response threshold on repeat assessment, and weak or inconsistent endpoint relationships in relevant evidence. These are not merely reporting details: they can undermine the information on which the selection decision depends.

The principal alternative is to wait for more mature final-endpoint information before selecting arms. That can improve direct relevance to the final objective but may delay a decision and require more follow-up. A second alternative is to use a different intermediate outcome or a combination of outcomes; that may improve coverage or relevance but can add measurement burden and complicate interpretation. There is no universally best candidate. The choice is a balance, and any proposed rule should be assessed as a whole rather than assuming that one favorable property guarantees good selection.

Common Mistake and Key Takeaway

Wrong reasoning: “The early response is strongly associated with the final endpoint in a pilot, so selecting arms by early response will retain the arms with the best final outcomes.” This confuses an individual-level association with evidence that treatment effects on the early measure preserve treatment effects on the final endpoint. It also ignores availability, measurement error, clinical importance, and whether the pilot applies to the trial setting.

Correction: Treat the association as one component of a broader suitability assessment. Verify that results will be available in time, that the measure is clinically interpretable and reliably recorded, and that the endpoint evidence is relevant. Then evaluate the planned selection rule using assumptions that reflect uncertainty and possible failure of the candidate’s relationship to the final outcome.