An external comparator can provide useful context when a concurrent randomized control is difficult or unavailable. It also introduces design challenges that cannot be solved by presenting a matched table alone. The central question is whether the comparison can support the intended inference under credible assumptions.
The FDA’s February 2023 external-control guidance remains draft guidance with nonbinding recommendations. It is a relevant discussion of design considerations, not evidence that any particular external-control analysis will be accepted.
1. Align the clinical question
Define the eligible population, treatment conditions, endpoint, start of follow-up, and observation period for both groups. Differences in time origin can create bias before any model is fitted. An external cohort assembled from people who survived long enough to meet a later criterion is not automatically comparable to a trial cohort enrolled earlier.
For a hypothetical disease study, examine whether eligibility can be applied using information available at the same point in time. Document criteria that cannot be measured equivalently and explain their potential consequence.
2. Examine measurement compatibility
Outcome definitions, assessment schedules, clinical practice, and data completeness may differ. A response measured by scheduled imaging and a response inferred from routine clinical notes need careful consideration. A shared label does not establish equivalent measurement.
The FDA’s real-world data guidance is useful additional reading for external sources. Review the provenance of key variables and whether the source captures the information required by the proposed design.
3. Explain what adjustment can address
Matching or weighting may improve balance on measured covariates. It does not by itself resolve unmeasured confounding, incompatible endpoints, or selection processes that are poorly understood. Assess overlap and the consequences of excluding participants without suitable comparators.
Prespecify the analytical approach where feasible and report the populations to which the result applies. A weighted estimate may target a different population from a simple unadjusted comparison; reviewers should understand that distinction.
4. Challenge the conclusion
Evaluate plausible biases and alternative specifications. Sensitivity analyses should be motivated by the design’s vulnerabilities rather than selected merely to display numerical stability. Explain which assumptions cannot be empirically verified.
For rare-disease development, the FDA’s rare-disease guidance provides broader context. Limited recruitment can motivate alternative designs, but it does not remove the need for a defensible evidence strategy.
A credible external-control package combines clinical alignment, compatible measurement, transparent cohort construction, justified adjustment, and explicit limitations. The comparison should be understandable before the effect estimate becomes the center of the conversation.
