Bayesian synthesis in stroke medicine

The problem

Stroke trials answer yes or no at the 0.05 gate. Clinicians do not make yes-or-no decisions; they weigh graded probabilities for one patient at a time, and the gap between what the trials report and what the bedside needs is where much of the field's confusion lives.

A trial that "fails" at p = 0.06 and a trial that "succeeds" at p = 0.04 may carry nearly identical information about whether a treatment helps — and the binary reading of them drives treatment fashions, premature abandonments, and a decade of arguments that better questions would have dissolved.

What the method requires

Three disciplines make Bayesian synthesis worth trusting, and each is a practice rather than a slogan. Transparent priors: every analysis declares what was believed before the data arrived — enthusiastic, neutral and sceptical — and the declaration precedes the analysis, in writing. Reproducible engines: the same data and the same prior return the same posterior, byte for byte, on any machine, or the analysis does not ship. Pre-registration: the question, the priors and the analysis plan are registered before results exist, so that the answer cannot quietly become the question.

Proof the approach works, from the published literature

The approach is not speculative; the literature already carries its demonstrations, at both ends. DAWN (Nogueira RG, et al. N Engl J Med 2018; doi:10.1056/NEJMoa1706442; NCT02142283) showed a Bayesian design running a pivotal trial: adaptive enrichment, a posterior stopping rule, credible intervals — and a result the whole field acted on. The Bayesian reanalysis of EOLIA (Goligher EC, et al. JAMA 2018; PMID 30347031) showed the interpretive power of the same machinery pointed backwards: a trial the significance test called negative carried a posterior probability of mortality benefit (relative risk below 1) of 88% to 99% under stated priors, from strongly sceptical to enthusiastic. Design and interpretation are the two halves of this kind of work, and both already have their existence proofs in print.

Where our interest stands

Two kinds of work come out of this. The Readings, published monthly on the Bayesian Centre page, are translations: one published paper, one clinical question, one posterior probability against a decision line fixed before the arithmetic runs, computed under a single declared reference prior. A Reading is not pre-registered and is not a re-analysis — it adds no data and rebuilds none of the original model — and what stands in place of pre-registration is that it is checkable: the counts, the code, the seed and the execution record are published with it, and anyone is invited to reproduce the number and say so in writing.

The re-analyses are the other kind, and they carry the discipline described above in full: pre-registered before results exist, with the sceptical prior as the one output that adjudicates and every other prior demoted to a labelled sensitivity annex. Nothing of that kind is published yet. When something is, it will have been pre-registered before it was anything else, and it will appear here.

Collaborate

If graded probability in cerebrovascular medicine is your problem too — as a trialist, a methodologist, or a clinician who has felt the 0.05 gate misdescribe a result you cared about — write: eurostrokes@gmail.com

The method, stated precisely — and the objection to it

The discipline above has one refinement worth stating before anything of ours is published, because it decides what a re-analysis is allowed to claim. Adjudicate by subtraction: declare one primary output — the posterior probability that the effect clears a pre-specified region of practical equivalence under the sceptical prior — and demote every other prior to a labelled sensitivity annex, which can strengthen or caveat that verdict but is never itself a competing headline. The sceptical prior is the adversarial one: it asks whether the data would move a reasonable sceptic, which is the question a re-analysis exists to answer.

The refinement exists because of a published objection from inside the community that uses these methods. de Grooth and Cremer argue that re-analysing a trial under many priors "does not contribute to consensus" and can appear "vacuous" — every reader keeps the posterior that matches their own prior, and nobody moves (Am J Respir Crit Care Med 2024; PMID 37922492). The objection is fair, and against a many-prior re-analysis with no decision rule it is right: relabelling the same estimate five ways adjudicates nothing. The answer is not another prior; it is the subtraction above — one adjudicating output, declared before the data are touched, with everything else explicitly sensitivity. A re-analysis that will not say in advance which lens decides has not answered the critique. One that does, has.

Source: the programme's own methods brief, 2026, and the published critique it answers.