The Protocol

A new paper in Statistics in Biopharmaceutical Research argues that mainstream meta-analytic methods applied to clinical trials produce strongly inconsistent estimators — meaning they don’t converge to the true parameter even as data accumulate, a structural flaw that more data cannot fix. If the argument holds broadly, it implicates the evidentiary foundation of systematic reviews, HTA submissions, and regulatory benefit-risk assessments simultaneously.

Three recent papers in JBS and SBR refine the multiplicity toolkit for co-primary endpoints: a Holm-based k-of-n procedure with proven optimality and confidence regions, a “free” stepwise extension bridging Holm and max-p without alpha cost, and improved trimmed weighted Hochberg procedures with integrated sample size optimization for the two-endpoint case. None are yet standard practice, but all three arrive with the validity proofs and confidence-region machinery that make a regulatory justification argument tractable — worth flagging to your SAP leads before the next co-primary design discussion.

A systematic review in Clinical Trials compares MI against complete case analysis across published trials to ask how often the switch materially changes treatment effect estimates or inferential conclusions — empirical evidence, not simulation. The full findings are behind a paywall and the abstract is truncated, so precise conclusions remain unconfirmed; but if MI rarely shifts results, expect uncomfortable questions about SAP boilerplate and the resource overhead teams routinely absorb to implement it.

The MHRA’s AI Airlock has secured multi-year funding and is now in Phase 2, with pilot-phase reports (October 2025) covering hallucination evaluation, explainability, and post-market surveillance for AIaMD — all non-binding, but likely the clearest preview available of where formal MHRA guidance is heading. Teams responsible for algorithmic performance monitoring and drift detection should treat the surveillance and hallucination outputs as early reading material, not background noise.

A new paper in Statistics in Biopharmaceutical Research shows that “on-treatment” analyses in RCTs are systematically biased when the two arms have different dosing frequencies — daily oral versus monthly injectable, for instance — because asymmetric exposure windows create differential censoring structures that distort comparative estimates. The practical reach is wide: any SAP that defines on-treatment follow-up without accounting for dosing interval differences, particularly under ICH E9(R1) “while on treatment” intercurrent-event strategies, may be quietly introducing bias that no sensitivity analysis will catch if it shares the same flawed definition.

A cross-sectional analysis in Clinical Trials characterizes the trial design and outcome transparency of EMA drug approvals between 2020 and 2023 — a window that includes conditional marketing authorizations, rolling reviews, and pandemic-era flexibility. Expect it to quantify the share of approvals resting on single pivotal trials, surrogate endpoints, uncontrolled or open-label designs, and imperfect disclosure: the numbers that quietly anchor every internal conversation about how much EMA will actually accept. For biometrics teams calibrating design choices against regulatory reality rather than regulatory aspiration, this is a more useful input than another guidance document.