Synthetic control arms: from pioneering to emerging regulatory precedent
Externally-controlled and synthetic-arm designs have moved from pioneering to structured regulatory openness. What the guidance actually asks for, and where the method still does not hold.
Insight · · 4 min
In rare and paediatric disease, the control arm is often the hardest part of the trial. There are too few patients to randomise twice over, there are real ethical objections to allocating a child with a progressive condition to placebo, and the natural history against which any effect must be judged is frequently uncharacterised. Synthetic and external control arms, built from historical trial data, disease registries or routinely collected real-world data, offer a way through. The regulatory posture toward them has shifted, and it is worth being precise about what has shifted and what has not.
Externally controlled trials are not new, and it is a mistake to present them as a recent invention. ICH E10, adopted in 2000, already described the external control as a legitimate design choice and set out, in unusually blunt language, why it is the weakest of the control types: the comparison is vulnerable to differences in patient population, in supportive care, in diagnostic criteria and in outcome ascertainment, none of which randomisation is there to absorb. Everything written since builds on that assessment rather than overturning it.
What changed is the volume and structure of the data available, and the regulators' willingness to say formally how they will assess it. The 21st Century Cures Act obliged the FDA to publish a framework for real-world evidence, which it did in December 2018. In February 2023 the agency issued draft guidance specifically on the design and conduct of externally controlled trials for drug and biological products. In Europe, EMA established DARWIN EU in 2022 to generate real-world evidence from a federated network of data sources, and its methodological guidance on registry-based studies has been in place since 2021. The direction of travel is consistent: not a relaxation of the evidentiary standard, but a published account of what the agencies will look at.
There is precedent in approvals, too, and it predates the current wave of interest. Blinatumomab received accelerated approval in 2014 for relapsed or refractory acute lymphoblastic leukaemia with a historical comparison drawn from pooled patient-level data across prior studies. That case is instructive precisely because it was not simple: the comparison required patient-level access, careful alignment of response definitions, and a disease where the untreated course is well enough understood to make the counterfactual credible. Those three conditions are the real gate, and most programmes fail at least one of them.
Set out plainly, the design asks four questions, and the answers have to be prepared before any data is pulled. Is the source population comparable on the variables that drive outcome, and can you demonstrate it rather than assert it? Is the outcome measured the same way in both arms, with the same definition, the same timing and the same ascertainment intensity? Is the confounding structure understood well enough to be adjusted for, and are the variables needed for that adjustment actually recorded in the source? And is the treatment effect you expect large enough to survive the bias you cannot rule out? A modest effect in a design with unquantifiable bias is not a finding; it is a hypothesis with a p-value attached.
That last point deserves emphasis, because it is where most enthusiasm collapses. External controls work best where the effect is unambiguous and the natural history is grim and well documented. They work poorly where the effect is incremental, where standard of care has moved between the historical period and the present, or where the endpoint is subjective and the historical arm was assessed by different people under different expectations. Time is itself a confounder. A control cohort assembled from data five years old carries five years of drift in diagnosis, imaging, supportive care and coding practice.
Our position, stated so it can be quoted back to us: a synthetic control arm is not equivalent to a concurrent randomised arm, and anyone who tells you otherwise is overselling. What is true is that a well-constructed external arm, with transparent data provenance, a pre-specified analysis plan, pre-specified sensitivity analyses and a governance trail a reviewer can audit, is increasingly accepted as supportive evidence. Emerging regulatory precedent, not settled standard. The distinction matters when a programme is being planned around the assumption that the agency will accept it.
The practical work sits earlier than most sponsors expect. Data provenance has to be established before the cohort is defined, not reconstructed afterwards to explain a result. The estimand has to be written down while both arms are still hypothetical. The sensitivity analyses have to be pre-specified, including the ones you expect to be unfavourable, because the credibility of the whole exercise rests on having declared them in advance. Where the source is a registry, its data dictionary, completeness profile and curation history become part of the submission, not background material.
Where this leads is a narrower but more honest use case. For a rare condition with a documented and uniformly poor natural history, a hard endpoint, and a candidate expected to produce a large effect, an external control can make an otherwise infeasible study possible, and can spare a small population an allocation nobody wanted to defend. Outside those conditions, the design usually costs more credibility than it saves in recruitment. Applied responsibly, in line with emerging regulatory precedent, it is a real instrument. Applied to avoid the difficulty of running a randomised trial, it produces evidence nobody can use.
Sources. ICH E10, Choice of Control Group and Related Issues in Clinical Trials, Step 4, 2000. FDA, Considerations for the Design and Conduct of Externally Controlled Trials for Drug and Biological Products, draft guidance, February 2023. FDA, Framework for FDA's Real-World Evidence Programme, December 2018. EMA, Guideline on registry-based studies, 2021. EMA, DARWIN EU, established 2022.
Written by the Evidilya scientific team. For interviews, references or a full publication list, use the contact page.
Talk to our team.
Bring us the evidence question you're wrestling with, we'll tell you the shortest defensible path.