In this episode of "In the Interim…", Dr. Scott Berry and NIH’s Dr. Michael Proschan conduct a detailed discussion from opposing viewpoints on response-adaptive randomization (RAR) in clinical trials. The discussion focuses on where they agree – on the positives and negatives of RAR, and where they disagree on its scientific use.

Key Highlights

• Potential issues of using RAR: Potential temporal trends, unblinding, reduction in statistical efficiency in 2-arm trials
• Potential benefits include improved statistical efficiency in multi-arm trials  depending on the goals (e.g. dose-finding trials).
• Potential unblinding of results in non-blinded trials and the need for operational excellence.
• Ethical and Bayesian perspectives are considered, but emphasis remains empirical.

For more visit: https://www.berryconsultants.com/

In this episode of "In the Interim…", Dr. Scott Berry, Dr. Elizabeth Lorenzi, and Dr. Amy Crawford discuss the STEP platform trial’s statistical methodology for evaluating which acute stroke patients benefit and which do not from endovascular therapy (EVT). The discussion critiques the inadequacy of traditional clinical trials powered for a single population to show benefit, as the goal of the trial is to identify who benefits, not if the entire population has a net benefit. The team walks through the development and simulation of a Bayesian change point model, addressing heterogeneous treatment responses across the NIH Stroke Scale. The adaptive platform design leverages scheduled interim analyses to draw timely, data-driven conclusions about patient subgroups, improving trial efficiency and relevance. The episode also previews scaling to two-dimensional modeling, incorporating both stroke severity and time since last known well, and emphasizes ongoing clinical trial simulation and close integration between clinicians and statisticians throughout trial design and execution.

Key Highlights

• STEP platform master protocol and the NIH StrokeNet collaborative infrastructure
• Clinical rationale for Bayesian change point modeling of the effect of EVT across the patients
• Shift from single to dual change point models to reflect regions of equivalence
• Development of custom C code and MCMC samplers due to limits of standard tools
• Interim analyses direct adaptive enrollment and define actionable conclusions
• Future extensions to multidimensional change point curves modeling 

This episode of "In the Interim…" features Dr. Scott Berry, Dr. Kert Viele, and Dr. Melanie Quintana of Berry Consultants dissecting the technical and operational landscape of Bayesian statistics in clinical trial design. The episode discussed what is Bayesian statistics, the impact of informative and non-informative priors, and clarifies when and why Bayesian approaches surpass frequentist analyses—especially in adaptive, platform, and rare disease trial settings. The discussion directly challenges the misconception that Bayesian methods “lower the bar," presenting evidence that they often require broader data synthesis and can raise evidentiary standards.

Key regulatory developments at FDA and EMA are reviewed, with attention to updated guidance and increased adoption. Case studies illustrate Bayesian methods in practice, including the prospectively combined phase 2 and 3 analysis for REBYOTA approval; hierarchical modeling in GNE myopathy; shared controls and endpoint integration in the HEALEY ALS Platform Trial; and robust subgroup borrowing in the ROAR basket trial. The team also addresses technical challenges such as multiplicity, subgroup analysis, complexity in endpoint modeling, and appropriate strategies for blending Bayesian and frequentist approaches for maximum regulatory and scientific clarity.

Key Highlights

• Clear explanation and real-world examples of Bayesian analysis in clinical trials.
• Theoretical and practical distinctions from frequentist methods
• Practical breakdown of control sharing, endpoint integration, and subgroup borrowing.
• Regulatory position and the increasing acceptance of Bayesian trial designs and analyses.
• Case examples: REBYOTA, GNE myopathy, HEALY ALS Platform Trial, ROAR basket trial.

Dr. Scott Berry and Dr. Kert Viele discuss the origins and implementation of the “time machine” modeling approach, beginning with sports analytics and progressing to adaptive platform clinical trials. The episode focuses on how techniques for comparing athletes across eras translate into methodology for platform trials.

Key Highlights

• Sports analytics as foundation: Early work of modelling athlete comparisons across eras using bridging methodologies.
• Platform trial application: The time machine model in I-SPY 2 enabled efficient control allocation through overlapping arms over extended trial periods.
• Core modeling principles: Additive treatment effect assumptions and the necessity of sufficient temporal overlap for reliable era comparisons.
• Statistical implementation: Approaches include categorical era adjustment and Bayesian smoothing splines for modeling change over time.
• Limitations and disease specificity: In conditions with rapid clinical or epidemiologic change, such as COVID-19, non-concurrent controls are avoided due to high risk of era by treatment interaction.
• Regulatory and methodological distinction: The model leverages within-trial overlapping data collected under a unified protocol, contrasting sharply with external or historical controls.

In this episode, Dr. Don Berry and Dr. Scott Berry provide an in-depth account of I-SPY 2, focusing on the trial’s use of the “time machine” methodology—a Bayesian solution allowing bridging across arms to inform ongoing analyses. The discussion details how predictive probabilities and adaptive randomization shaped pivotal decisions, including the handling of Pertuzumab’s approval and Neratinib’s subtype-specific performance. This episode also documents the technical and operational contributions of Laura Esserman, Anna Barker, Janet Woodcock, Meredith Buxton, and Ashish Sanil, clarifying the roles that enabled the platform’s success and broader impact on subsequent adaptive trials.

Key Highlights

• Introduction of the “time machine” concept, enabling valid comparison between experimental and control arms even when enrollment periods differ—a pragmatic solution originally utilized in sports examples for evolving platform trials as treatments and control arms change.
• Ongoing trial conduct driven by a Bayesian adaptive algorithm, developed and maintained by Berry Consultants statisticians, which computes predictive probabilities to guide arm graduation, futility, and real-time adjustment of randomization probabilities.
• Neratinib serves as a case study in subtype-specific adaptive randomization: the platform set randomization probability to zero in subtypes without signal, while effective subtypes increased randomization and advanced to graduation.
• I-SPY 2’s methodologies shaped subsequent adaptive platform trials (GBM AGILE, Precision Promise, COVID-19 ACTIV networks), with regulatory acceptance reflected in FDA guidance and Janet Woodcock’s public recognition of adaptive randomization as “adequate and well controlled” for registration studies.
• Specific recognition: Laura Esserman (trial leadership), Anna Barker (funding and strategic input), Janet Woodcock (FDA guidance and adaptive methods support), Meredith Buxton (logistics; GCAR leadership), and Ashish Sanil (Berry Consultants; ongoing algorithm implementation).

In Episode 20 of Berry’s "In the Interim..." Podcast, The Legend of I-SPY 2 - Part A, Dr. Don Berry and Dr. Scott Berry discuss the origins and design of the I-SPY trials. Their conversation explains the inefficiency of traditional adjuvant breast cancer trials and details the shift to the neoadjuvant approach, where tumor response can be observed prior to surgery. 

I-SPY 1 served as a proof-of-concept using MRI for probabilistic prediction of pathologic complete response (pCR). I-SPY 2 represents a major advancement in clinical trial science, introducing a multi-arm bandit methodology, integration of biomarker-driven subtypes and signatures, and a structured funding model that transitioned from philanthropy to “pay to play” industry support.

This episode of "In the Interim..." features an in-depth discussion of the StrokeNet Thrombectomy Endovascular Platform (STEP), a multi-domain, multi-factorial, adaptive platform trial for acute stroke, anchored in the NIH StrokeNet network. Guests Dr. Eva Mistry (University of Cincinnati) and Dr. Jordan Elm (Medical University of South Carolina) join us to explain how STEP enables simultaneous investigation of multiple treatment strategies in patients with acute ischemic stroke. The conversation details the use of a master protocol, the integration of industry partners through the Other Transactional Authority (OTA) NIH mechanism, and innovative statistical designs to efficiently identify improved treatment strategies.

Key Highlights:

• STEP utilizes a master protocol within NIH StrokeNet, unifying eligibility, procedures, and data collection across all study domains.
• The platform supports multiple research questions.
• In an initial domain STEP applies a statistical change-point model to empirically estimate the thresholds where EVT is effective, neutral, or potentially deleterious based on medium vessel occlusions and baseline clinical status. 
• Protocols may be adapted in response to new external data, including pausing and revising enrollment in specific subpopulations when emerging science warrants.
• Shared control groups are used wherever applicable, improving trial efficiency by reducing the number of patients allocated to control arms and allowing eligible patients to contribute to multiple domains when protocol and scientific rationale permit.

Dr. Scott Berry applies a statistician’s review of a random trial result published in JAMA – the FAIR-HF2 clinical trial.  Interrogating the frequentist paradigm and the focus on the binary outcome of the primary hypothesis test. He scrutinizes the Hochberg multiplicity adjustment, challenges the prevailing disregard for accumulated scientific evidence, and contrasts the limitations of black/white view of clinical trial of over 1000 patients and 6 years of enrollment. A contrast is made to what a potential Bayesian approach, grounded in practical trial interpretation and evidence integration would look like. The episode argues how current norms, created by dogmatic statistical views, in clinical trial analysis can obscure or perhaps mislead from meaningful findings and limit the utility of costly, complex studies.

Key Highlights

• FAIR-HF2 randomized 1,105 patients with heart failure and iron deficiency to intravenous ferric carboxymaltose or placebo across 70 sites, with three pre-specified co-primary analyses.
• The study relied on the Hochberg procedure to control family-wise error across analyses: (1) time to first cardiovascular death or heart failure hospitalization; (2) total heart failure hospitalizations; (3) time to first event in a highly iron-deficient subgroup.
• Results showed a favorable hazard ratio (0.79) and a p-value below 0.05 for primary composite 1, but statistical significance was nullified under Hochberg multiplicity criteria as other endpoints failed threshold requirements.
• Berry challenges the reduction of trial outcomes to discrete “significant” or “not significant” designations—critiquing the scientific and statistical culture that ignores gradient evidence in favor of only black-and-white outcomes.
• He details the likelihood principle and Bayesian analysis as superior frameworks, quantifying a 98% posterior probability of benefit; he contextualizes findings with prior evidence from the HEART-FID, IRONMAN, and AFFIRM-AHF trials and published meta-analyses—arguing that isolated, negative conclusions defy cumulative data.
• The discussion extends to the inefficiency of fixed trial designs, the missed value in adaptive methodologies, and the inefficacy of requiring full-scale repeat trials all analyzed in isolation, when evidence already points strongly to a beneficial effect.

Scott Berry convenes co-authors Kert Viele, Joe Marion, and Lindsay Berry to discuss the statistical and developmental nuances of inferentially seamless phase 2/3 clinical trial designs. The group dissects the simple method for distributing alpha when including stage 1 data, whether it is a good idea to distribute alpha, and the optimal allocation of sample size when Stage 1 data are carried forward, all referencing their recently published work in Pharmaceutical Statistics.

Key Highlights:

• Systematic definition of seamless phase 2/3 trial designs, contrasting fixed, separate-phase models with integrated, inferentially seamless approaches.
• Detailed explanation of the required alpha adjustment when selecting doses partway through—leveraging group sequential theory, normal approximations, and quadrature for explicit formula derivation; R code and calculation procedure are made available for practitioners.
• Exploration of the information fraction curve for adjusted alpha, emphasizing that initial adjustment is numerically significant but does not inherently reduce statistical power.
• Findings indicate that power is always higher when including stage 1 data – and outperforms a closed testing procedure.
• Discussion of when seamless trials may not be advantageous: operational and statistical limitations: insufficient endpoint/regulatory understanding for phase 3, differences in manufacturing readiness, need for public phase 2 results for funding, and proof of concept hurdles; identifies real scenarios where seamless 2/3 designs are challenging.
• Considerations for operational bias and blinding, with technical commentary on the boundaries of unblinding within company roles.
• Provision of practical R code and explicit analytic guidance for calculating adjusted alpha in seamless design protocols.

In this solo episode of "In the Interim...", Scott Berry, President and Senior Statistical Scientist at Berry Consultants, addresses deep-rooted confusion in the field of adaptive clinical trial design surrounding the concept of “spending alpha.” Drawing on practical experience and rigorous statistical foundations, Berry addresses the prevailing language and myths that conflate interim analysis with loss of type I error. He clarifies that, with planned and transparent allocation of alpha, interim analyses enable more power with more efficient design, and robust clinical trials—without sacrificing statistical validity. This is a precise and fact-driven examination for those demanding technical clarity, not marketing gloss.

Key Highlights

• Explains the basics of hypothesis testing in superiority trials, highlighting why a one-sided 2.5% alpha is the operational standard despite persistent use of two-sided 5% language in clinical protocols.
• Refutes the widespread belief that reviewing interim data costs available alpha, making clear that statistical error is not “penalized”—it is allocated, with potential efficiencies in average sample size and, in thoughtfully extended designs, gains in operating characteristics such as power.
• Describes real-world examples, including the SEPSIS-ACT (selepressin) trial sponsored by Ferring Pharmaceuticals, which incorporated more than 20 interim analyses while maintaining a pre-specified final alpha of 0.025; underscores the necessity of transparent, prospective design and explicit documentation for regulatory acceptance.
• Distinguishes between interim actions—such as futility analyses or response-adaptive randomization, which require no alpha adjustment, and early efficacy analyses, which must be precisely modeled to preserve type I error.
• Challenges terminology like “penalty” and “spending alpha,” asserting that imprecise language fosters misunderstanding and leads to missed opportunities in adaptive trial efficiency.
• Emphasizes the crucial role of prospective, simulation-based planning and clear protocol definition at every interim, anchoring statistical practice in measured evidence, not historical convention.