Recruitment Decisions After Interim Signals

Why Recruitment Can Change After an Interim Signal

Interim findings can affect a trial even when no one changes the treatment protocol. A clinician who suspects that one treatment is performing better may screen more people thought likely to benefit, describe potential benefits with different emphasis, or devote more time to recruiting at particular sites. These decisions can change who enters the trial and, in some circumstances, can also make the treatment groups incomparable.

The distinction matters. In a properly randomized trial, recruitment decisions made before assignment do not, by themselves, let staff choose which treatment an individual receives. Randomization can therefore preserve treatment-group comparability among those enrolled, even if the enrolled population changes. But recruitment can still change the population to which the result applies, the mix of treatment effects being averaged, and the reliability of conclusions drawn across recruitment periods. If selection can influence treatment assignment, or if assignment is known or predictable, the concern is more direct: treatment groups may differ for reasons other than randomized assignment.

Where the Recruitment Process Can Be Influenced

Recruitment is a sequence of decisions, not just a count of consented participants. An interim signal can affect several stages, and different stages leave different evidence in the trial records.

• Screening and referral: Staff may refer more patients who seem likely to respond, or may stop looking for patients with characteristics thought to be associated with poorer outcomes.
• Eligibility assessment: Subjective criteria may be applied more strictly to some candidates than to others. Even without changing written criteria, ambiguous cases can be handled differently.
• Invitation and consent: Staff may approach some eligible people more often, describe expected benefits more enthusiastically, or spend less time explaining uncertainty. Consent must remain voluntary and based on accurate, balanced information.
• Site effort: Recruitment support may move toward sites believed to have participants who will do well, or toward sites whose staff are more enthusiastic about a suspected favorable treatment.
• Timing: The characteristics of people available later in recruitment may differ from those enrolled earlier, even if staff behavior does not change. A shift around an interim review therefore needs investigation, not automatic attribution to it.

Earlier material on operational bias describes how information can change trial conduct. Here the practical question is whether that change altered the pathway into the enrolled sample, and whether it affected treatment assignment or only the population being studied.

What Randomization Protects—and What It Does Not

Suppose eligibility, consent, and enrollment are completed before a concealed random assignment. If assignment is implemented as planned, then recruitment decisions cannot systematically direct a particular enrolled person to a preferred treatment. Conditional on the people enrolled, randomization balances their characteristics between groups in expectation. With a finite sample, chance imbalances remain possible.

That protection does not make the enrolled sample immune to selection. If recruitment shifts toward a subgroup that responds differently to treatment, the average treatment effect among participants recruited later can differ from the average effect among those recruited earlier. This is a change in the composition of the target population for the estimate, not necessarily a failure of randomization. In contrast, if recruiters can foresee assignments, choose when to assign, or enroll people differently by treatment group, the randomized comparison itself may be compromised.

A Short Derivation: Changing Mix, Changing Average Effect

Let \(Z\) denote a measured participant subgroup, such as baseline disease severity, and let \(S=1\) mean that a person is enrolled. Let \(\tau(z)\) be the average treatment effect for enrolled participants in subgroup \(z\), and let \(\pi_t(z)=P(Z=z\mid S=1,\text{ recruitment period }t)\) be that subgroup's share of enrollees in period \(t\). Under random assignment within the enrolled sample, the period-specific average treatment effect is the subgroup-weighted average:

This follows by partitioning the enrolled population into the mutually exclusive subgroups \(z\). The mean treatment effect across all enrollees is the sum of each subgroup's share multiplied by its mean treatment effect. If subgroup effects differ, changing the shares can change \(\theta_t\), even when each \(\tau(z)\) stays constant. If subgroup effects are identical, changing the subgroup mix alone does not change this average treatment effect, though it can still affect generalizability and precision.

For a trial that combines recruitment periods, the overall average will also depend on how many participants were enrolled in each period. A reported pooled effect can therefore be a weighted average over both subgroup composition and period-specific enrollment. This is one reason to define the target population and the intended treatment-effect summary before interpreting a changing recruitment profile.

Worked Patterns in Recruitment Data

Monitoring and Diagnosing Recruitment Changes

Useful monitoring separates the stages of recruitment and uses stable denominators. A rise in enrollment alone cannot show whether the change arose from more referrals, more screening, altered eligibility decisions, greater willingness to consent, or different site effort. Where relevant and permitted, compare counts and proportions by site and recruitment period, and examine pre-specified characteristics such as severity, age, referral source, and reasons for exclusion.

A baseline imbalance between randomized groups is not, by itself, evidence that recruitment was manipulated; chance imbalances occur. Conversely, similar measured baseline summaries do not prove that no selection occurred, because unmeasured characteristics or undocumented decisions may matter. Interpretation should use the design, process records, timing, and the full pattern of evidence rather than one statistical test.

Prevention, Alternatives, and Limits

The strongest safeguards address the pathway before a concerning pattern appears: keep interim treatment information away from people making recruitment decisions when feasible, use objective eligibility criteria, standardize balanced consent materials and staff training, document screening outcomes, and maintain consistent recruitment support across sites unless changes are justified independently of treatment-specific expectations. These safeguards complement, rather than replace, the trial's planned information controls.

When recruitment composition changes, investigators may describe or standardize results to a clearly defined target population if the needed characteristics are measured and the assumptions are credible. That is not a substitute for intact randomization. Statistical adjustment cannot reliably repair unrecorded eligibility discretion, unmeasured selection factors, or treatment assignment that was influenced by recruitment decisions. Nor does adjustment make a changed consent process ethically acceptable.

Common Mistake

Wrong reasoning: “The treatment effect changed after the interim review, so recruitment bias must have made the randomized groups incomparable.” This conclusion skips two steps. First, a change in the enrolled population can alter the average effect when subgroup effects differ, even if randomization remains intact. Second, an observed difference over time can also reflect chance or other changes, rather than the interim signal.

Correction: Separate the question of who enrolled from the question of how treatment was assigned. Examine composition and recruitment stages, then verify whether allocation remained concealed and correctly implemented. Describe a recruitment pattern as evidence to investigate unless the process records support a stronger conclusion.