Your control group may be entirely simulated

Under a draft guideline now out for comment, a clinical trial could in future be approved with a control arm that contains no living participants at all. The comparator would be assembled from historical records: a model trained on how patients with the same condition actually progressed, generating the arm that would otherwise have to be recruited, consented and given nothing.

The mechanics would be undramatic. Every person who enrolled would receive the investigational treatment. Their outcomes would be compared against a modelled trajectory of what the same population would have done untreated, with the model locked and registered before the first patient is dosed. The trial would still have two arms on paper. Only one of them would have a waiting room.

The ethical appeal is obvious and, in rare diseases, close to unanswerable. Assigning half of a very small population to a placebo is difficult to justify to the families and difficult to justify to the statistics, since a trial too small to recruit is also a trial too small to prove anything. A comparator that does not require anyone to be untreated removes both problems at once, and introduces a third.

“The control group would not be imaginary,” a spokesperson for the drafting body would be expected to say. “It would be derived. Every patient behind that model existed and was recorded. They simply did not exist at the same time as your trial.”

That is where the difficulty starts. A model of untreated disease learns from whoever was historically written down — which has never been everybody in equal measure. A synthetic control inherits the demographics of the archive, including the people the archive never reached, and it inherits them invisibly, expressed as a smooth curve rather than as a list of who was missing.

Statisticians would raise the structural objection: a trial run against a model is partly a test of the model. If the modelled trajectory is optimistic, the treatment looks weak; if pessimistic, it looks miraculous. The difference being measured would no longer be between a drug and nothing, but between a drug and our current beliefs about the disease — which is a respectable thing to measure, provided everyone agrees to call it that.

Nothing here is adopted. But externally controlled trials are already a regulatory topic, model-based analysis is already used to sharpen trial results, and the hardest question in the draft is not statistical at all. It is who ends up standing in for you, in the arm you were never enrolled in.