Reading the brief online abstracts for the 2026 Regulatory-Industry Statistics Workshop (RISW), along with two presentations shared with me, prompted a good deal of reflection on where drug development is heading. What follows is my own methodological extrapolation rather than a recap of the presentations, but one theme stands out: the January 2026 publication of the FDA's PDUFA VII draft guidance on Bayesian methodology represents an important milestone in the ongoing evolution of drug development. Alongside the ICH M15 guideline, the RISW 2026 abstracts suggest that Bayesian methods are moving beyond exploratory signal-seeking or internally proprietary use into the foundation of confirmatory, registrational decision-making. With global convergence across the FDA, EMA, and PMDA, quantitative strategists now have a clearer framework for incorporating external data and dynamic borrowing into late-phase trials.
The opportunity is real, but it asks something of us: to design as Bayesians from the start, rather than translating familiar frequentist habits into Bayesian language after the fact. In my reading, that means four things: starting from external data that are fit for purpose, building borrowing into the model itself, deciding in advance how borrowing will respond to the data, and demonstrating through simulation that the design behaves as intended.
Start with Data That are Fit for Purpose
Before any prior is specified, the external data themselves must earn their place. A presentation from CBER offers a clear lens, drawn from cell and gene therapy. External controls are most credible when a disease’s natural history is well defined, when the disease is known not to improve without treatment or with available therapies, and when a large, clinically meaningful effect is expected. Zolgensma in spinal muscular atrophy (SMA) and OTARMENI in OTOF-related hearing loss illustrate this well: in both, single-arm trials were compared against a natural history in which meaningful improvement does not occur on its own. By contrast, chronic inflammatory demyelinating polyneuropathy (CIDP), with its variable course, high placebo response and several approved therapies, shows where randomization remains the better path.
The same presentation describes a discipline that any borrowing design can adopt: agree on the external data source with FDA and submit the protocol and statistical analysis plan before the trial begins, keep the design blind to external outcome data, separate design from analysis, and assess the comparability of data across arms on key dimensions such as time period, geographic region, diagnostic criteria, prognosis, treatments, follow-up, intercurrent events, outcomes, and missing data. These practices echo the modular spirit described below, and they are a reminder that randomized evidence, when feasible, remains the strongest foundation of all.
Borrow Through the Model, Not Around It
In a Bayesian design, external evidence enters through the prior and the likelihood, and both should describe how the data actually arose. Appropriate confirmatory designs therefore favor fully generative models, or modular approaches that keep any causal adjustment separate from the Bayesian update, over ad-hoc weighted pseudo-likelihoods that can distort posterior variance.
One RISW 2026 abstract illustrates a modular path. Propensity scores are first used as weights to make the borrowed population comparable to the target population of the new study, and the weight-adjusted treatment effects on two endpoints are derived. Their joint distribution becomes the prior, and the design then borrows dynamically, according to how consistent those effects are with the new study. By formally decoupling this initial causal adjustment from the subsequent Bayesian update, the design contains the pseudo-likelihood problem and helps prevent prospective trial data from retroactively distorting the historical cohort balance—a concept that can also be extended to broader robust dynamic borrowing approaches.
Make Exchangeability an Explicit Decision
Dynamic borrowing rests on a single assumption: that historical and concurrent data are exchangeable, close enough to be treated as arising from the same distribution. Formally, each study's true effect can be seen as a draw from a shared population, with the between-study heterogeneity governing how much they may differ. When that heterogeneity is small, borrowing is strong; as it grows, the historical evidence speaks with less authority until it falls silent. Continuous approaches, such as robust meta-analytic-predictive, commensurate, and power priors, let the data inform this judgment gradually. Modern frameworks can instead make it explicit and prospectively specified.
Adaptive Bayesian Borrowing (ABB) exemplifies this by establishing a deterministic congruence gate at an interim analysis, replacing the continuous dial with a clear choice: the new trial is either exchangeable with the past or it is not. If the concurrent control aligns with historical expectations, the informative prior activates to borrow data. However, if divergence is detected, historical borrowing drops to zero, and the design automatically triggers compensatory sample size re-estimation (SSR). This coupled adaptation seamlessly recruits replacement patients to preserve the trial's nominal power.
Whether through decoupling the initial covariate balancing step from the subsequent Bayesian update, settling the borrowing decision once at a pre-specified interim analysis, or considering fully generative causal frameworks,¹ designs that guard against feedback from the outset rest on a coherent inferential foundation and are ready for rigorous validation.
Show That the Design Behaves as Intended
The FDA draft guidance describes two paradigms for establishing operating characteristics: frequentist metrics for calibrated designs, or prior-referenced metrics such as Bayesian power. One RISW 2026 abstract names controlling Type I error under dynamic borrowing as a key calibration challenge. Taking this further, when a sponsor chooses to calibrate a Bayesian design using Type I error, it is worth considering a comprehensive mapping of Type I error drift surfaces across a continuum of population drift to understand potential peak inflation at boundary edges. Tipping point sensitivity analyses can vary borrowing fractions and missing data patterns simultaneously to explore the thresholds where an efficacy conclusion might change. To facilitate independent agency reproducibility, documenting MCMC audit parameters—such as warmup lengths, convergence metrics, and random seeds—in the Statistical Analysis Plan remains a helpful practice.
Ultimately, my key takeaway is that achieving regulatory success with Bayesian borrowing relies on external data that are fit for purpose and on translating mathematical hyperparameters into transparent clinical decision boundaries. By engaging early with platforms like the FDA's CID program and EMA scientific advice, ideally before trial initiation, development teams can align these innovative designs with global evidentiary standards to deliver trials that reach trustworthy answers while asking less of patients.
¹ I'm happy to share more detail on these frameworks with interested readers.
References
FDA. Use of Bayesian Methodology in Clinical Trials of Drug and Biological Products. Draft guidance; Jan 2026.
FDA. Considerations for the Design and Conduct of Externally Controlled Trials for Drug and Biological Products. Draft guidance; Feb 2023.
EMA. Concept Paper on the Development of a Reflection Paper on the Use of Bayesian Methods in Clinical Development. EMA/CHMP/1813/2026; Jan 2026.
PMDA. Points to Consider for Small Clinical Trials (Early Consideration). Mar 2026.
ICH M15. General Principles for Model-Informed Drug Development.
EC/EMA/HMA. Complex Clinical Trials – Questions and Answers. 2022.
Best N (GSK), et al. Utilization of ICH M15 to Support Bayesian Design Submissions. ASA RISW 2026 Presentation.
Lim D (FDA), et al. A Case Study in External Controls and Bayesian Inference in a Phase 3 Trial: Inferential Coherence in Hybrid Design. ASA RISW 2026 Presentation.
Mukhopadhyay S (AbbVie), et al. Adaptive Bayesian Borrowing with Prospective Specifications. ASA RISW 2026 Presentation.
Quan H (Sanofi), et al. Propensity Score Weighting and Power Prior Approach for Historical Data Borrowing in Bivariate Regression Analysis. ASA RISW 2026 Presentation.
Rabbee N (Regeneron), et al. Bridging Bayesian Trial Design from FDA Guidance to Practice. ASA RISW 2026 Presentation.
Zhou T (FDA). External Controls in Cell and Gene Therapy: From Data to Regulatory Decision. ASA RISW 2026 Presentation.
Mukhopadhyay S, Zhao Y, Chen X, et al. Prospectively specified adaptive Bayesian borrowing. Pharm Stat. 2026;25(1):e70051.
Oganisian A, Linero A. Priors and propensity scores in Bayesian causal inference. Obs Stud. 2025;11(1):47–59.