The Prior Problem Is Now the Regulatory Problem

The Clinical Trials narrative review puts the central tension plainly: Bayesian confirmatory designs remain the exception in drug development despite being unremarkable in early-phase and adaptive work. The bottleneck is evidentiary. Agencies want sponsors to demonstrate that prior choices do not corrupt frequentist error rates, and the methods community has spent the last eighteen months producing tools to do exactly that.

Four instruments are worth knowing. First, a new Bayesian equipoise calibration framework in JBS provides a structured approach to aligning trial design operating characteristics with what investigators actually believe before a trial begins — connecting the ethical requirement of equipoise to the quantitative design process in a way that directly addresses FDA/EMA expectations on prior justification. Second, on borrowing: the constrained power prior paper imposes explicit limits on the borrowing parameter to preserve type I error guarantees, while SPx in Statistics in Medicine takes a practically distinctive angle — it requires only trial-level summary statistics rather than individual patient data, uses Bayesian model averaging to discount non-commensurate historical controls, and demonstrated meaningful control-group size reduction with preserved frequentist properties in a rheumatoid arthritis trial. The IPD-free requirement alone broadens real-world applicability considerably. Third, a Contemporary Clinical Trials paper formally proves that conditional power and predictive probability of success are analytically equivalent under a non-informative prior — diverging only when an informative prior enters. The result is not new in spirit, but the peer-reviewed derivation provides the citable anchor that DSMB charters have been missing. Fourth, a Bayesian decision-theoretic approach for basket trials borrows at the decision layer rather than the estimation layer: adaptive loss functions make a basket more likely to be declared promising when peer baskets show activity, while keeping effect estimates independent and tuning parameters calibrated to pre-specified frequentist error rates — which makes the framework more auditable than hierarchical borrowing alternatives.

No regulator has yet qualified a specific Bayesian borrowing method for pivotal use, and none of these papers changes that — the toolkit is methodological scaffolding for sponsor–reviewer conversations, not a cleared submission route. What it does change: a sponsor proposing Bayesian confirmatory analysis in 2026–2027 now has named methods to cite when an FDA reviewer asks how the prior was justified, instead of inventing the justification on the call.

Protocol read: The vocabulary for justifying Bayesian priors to regulators is now considerably richer, but no specific borrowing method has been formally qualified for pivotal use — sponsors planning Bayesian confirmatory designs are still building the precedent rather than citing it.

What to do now:

• Use the equipoise calibration framework as the structuring document for prior-justification conversations with FDA/EMA in pre-IND meetings.
• For external-data borrowing, evaluate constrained power prior vs. SPx by IPD availability — SPx’s summary-statistic-only requirement is a practical unlock for cross-study borrowing.
• Cite the CP/PPS equivalence proof in DSMB charters where the analytical anchor was previously missing; deploy basket-trial decision-theoretic borrowing only where pre-specified frequentist error rates are auditable.