Master Protocols' Borrowing Question Hits the SAP Before It Hits E20

A single issue of Statistics in Biopharmaceutical Research (Vol. 18, Issue 1) now carries eight master-protocol methodology papers — covering non-concurrent controls, ridge-shrunk shared controls, power-prior baskets, hierarchical mixtures, BOP2 for baskets, latent-subgroup survival baskets, and a dose-ranging basket with enrichment — and the May 2026 Statistics in Medicine paper on interim analyses and NCCs lands the punchline: every one of these methods is, at root, an answer to the same question. When can you legitimately share control information across arms, time, populations, or external datasets? The fact that ICH E20 on adaptive designs is being finalised with explicit attention to “borrowing of control group information in Master Protocols” makes the timing more than coincidental.

The control-borrowing problem is now the master-protocol problem

The Statistics in Medicine result is the most uncomfortable of the batch. The authors show that performing an interim analysis on Arm 1 of a platform trial biases the treatment-effect estimator for a later Arm 2 — even when the standard step-function time adjustment for non-concurrent controls is applied. Their proposed estimator substantially reduces both bias and Type I error inflation while preserving the power gains over a concurrent-controls-only analysis. The combination they break (NCC borrowing plus group-sequential interim) is not exotic: it is the modal oncology platform design. If your SAP currently leans on a step-function regression to justify NCC inclusion alongside interims, the default is now demonstrably insufficient, not merely contestable.

The rest of the cluster is best read as the field reaching for the same lever from different angles. The SBR ridge-estimation paper offers a frequentist regularisation route between full pooling and full separation of shared controls. The power-prior basket paper calibrates Bayesian borrowing to unequal basket sizes — the realistic case, not the textbook one. Hierarchical mixture models, weighted-sum and order-statistic dynamic borrowing, and order-restricted borrowing all attack the same homogeneous/heterogeneous tension with progressively more structure. BPED extends the logic into adult-to-pediatric extrapolation under ICH E11A. Pick your favourite; the operating-characteristic comparisons are not yet head-to-head, and the tuning-parameter sensitivity that has long dogged robust mixture priors applies to all of them.

Three live trials are stress-testing the same machinery

While the methods journals converge, three real platforms are publishing what the operational edge actually looks like. ACT-GLOBAL — the first multinational adaptive platform in acute stroke, launched August 2024 — has had to publish an explicit ethical justification for enrolling under deferred consent across jurisdictions. RECOVER-VITAL is attempting a platform protocol for Long COVID with no validated endpoints and contested pathophysiology, which is a more honest stress test of estimand specification than most simulations will produce. CanTreatCOVID publishes lessons learned on shared controls and adaptive decision rules in community primary care. The two-part Journal of Clinical Epidemiology planning series frankly concedes that “planning APTs is complex, and limited guidance is available” — which is the academic register for “you are on your own until E20 lands.”

What the regulatory window actually closes on

The reason to take this batch seriously now, rather than at next year’s JSM, is the E20 timing. The methodological choices sponsors lock into 2026 SAPs — which estimator for NCC adjustment, which prior for cross-basket borrowing, which sensitivity analyses for exchangeability — are the ones reviewers will benchmark against once the guidance is final. None of these papers is yet a regulatory standard, and the cross-beat work on synthetic priors and robust mixture-prior tuning makes clear that Type I error control under borrowing is still being argued out. But the direction of travel is clear enough: “we used non-concurrent controls and adjusted for time” will not survive a serious review by 2027 without a named estimator, a pre-specified borrowing strength, and sensitivity analyses that actually probe the exchangeability assumption rather than ratify it.

Protocol read: The control-borrowing question has graduated from research topic to design-level decision with regulatory consequences; teams designing platforms in 2026 should pick a borrowing method deliberately and document why, because the default will be audited.

What to do now:

• Audit any active or draft platform SAP that combines non-concurrent controls with an interim analysis against the new Statistics in Medicine estimator; the step-function regression is no longer a defensible default.
• Pre-specify the borrowing strength and exchangeability sensitivity analyses for any basket or shared-control design, rather than treating them as analysis-time choices.
• Track ICH E20 draft updates on control-group borrowing and flag any 2026 SAP commitments that would be hard to defend if E20 lands close to its current draft language.