Re-randomization designs in oncology trials may allow improved isolation of the treatment effect of multiple phases of treatment compared to parallel trial designs, with a second randomization often following a disease assessment or completion of a clinical milestone (e.g., surgery or completion of induction therapy), in all or a subset of patients. While these designs can address multiple clinical questions in a single trial, they present unique methodological and interpretation challenges. Key challenges include potential loss of randomization in the overall population if re-randomization occurs only in a subset of patients, selection bias, determination of appropriate estimands across randomization phases, appropriate endpoint selection, and follow-up across multiple phases. Statistical analyses to address these challenges often rely on strong, unverifiable assumptions, making interpretation challenging. This open forum discussion among multi-disciplinary experts considered the rationale, potential benefits, feasibility, and methodologic and interpretation challenges when implementing re-randomization designs in cancer clinical trials for regulatory decision-making.

The speakers/panelists* for the discussion included members of the BIOP Statistical Methods in Oncology Scientific Working Group representing pharmaceutical companies, representatives from international regulatory agencies (Food and Drug Administration (FDA), Therapeutic Goods Administration (TGA, Australia), Federal Institute for Drugs and Medical Devices (BfArM, Germany), and Health Canada (HC)), clinicians, academicians, and expert statisticians. In addition, over 100 participants attended the virtual meeting. The discussions were moderated by the BIOP Statistical Methods in Oncology Scientific Working Group co-chair, Dr. Olga Marchenko from Bayer; Dr. Brittany McKelvey from LUNGevity Foundation; and Dr. Gautam Mehta from OCE, FDA.

In the introductory presentation, the OCE leadership presenter noted that clinical trial designs incorporating re-randomization typically include an initial randomization at enrollment prior to treatment initiation followed by a second randomization that may occur after an early clinical milestone, such as a disease assessment or completion of a phase of therapy (e.g., surgery or induction therapy). The second randomization may include all patients or only a subset, such as patients who are biomarker positive or who have reached a clinical milestone like pathologic complete response. The presenter emphasized that this could be a potentially efficient design allowing multiple treatment elements, such as duration of therapy or treatment sequence, to be studied within a single trial, but that additional challenges arise if the second randomization only occurs in a subset of patients, including loss of randomization in the overall population and potentially complex assumptions. Key questions posed for the panel included: what challenges exist with designing these trials, what oncology endpoints are appropriate, how to interpret the analysis after the second randomization, and whether these designs are only for hypothesis generating or can support market approval.

The first presenter, from FDA, addressed re-randomization based on early clinical milestones in oncology trials. He noted that traditional trials randomize participants at well-established milestones with fixed treatment assignments under the intent-to-treat principle, whereas re-randomization at early milestones allows investigators to address more targeted questions such as maintenance versus withdrawal, treatment de-escalation, or sequencing decisions. He discussed a hypothetical re-randomized design to isolate the timing effect (strategy effect) from the drug’s treatment effect: patients are first randomized to different switch triggers (e.g., prostate-specific antigen (PSA) rise vs. radiographic progressive disease) while receiving the same standard of care, then re-randomized to the experimental drug or standard of care upon reaching their trigger. Statistical challenges highlighted included post-baseline subsetting complicating estimand definition and multiplicity control, potential baseline imbalances, and possible dilution of power. He described methods that could be utilized such as inverse probability weighting (IPW) and propensity score adjustment, but indicated these methods rely on strong and often unverifiable assumptions. He noted that isolating the strategy effect from the conditional drug effect may require the principal stratum strategy, a causal inference approach that currently lacks clear regulatory precedent. He concluded that FDA remains open to novel trial designs and statistical methodologies.

The second presenter, from academia, discussed Sequential Multiple Assignment Randomized Trials (SMARTs), a type of multi-stage randomized design where at least some participants are randomized at least twice along critical decision points. SMARTs are motivated by understanding dynamic treatment regimens (DTRs), which are adaptive treatment guidelines that specify actions based on intermediate outcomes such as response, tolerability, or adherence. She described embedded DTRs as treatment triplets (e.g., initial treatment, action for responders, action for non-responders), emphasizing that randomization is a tool to obtain unbiased evidence for these regimens. Benefits include addressing multiple questions simultaneously, evaluating treatment interactions, and advancing personalized medicine. Challenges in oncology include financial and logistical complexity, absence of placebo arms, estimand definition with dropout, endpoint selection, and the open question of how to label or register a DTR. She highlighted the Phase 2 I-SPY 2.2 breast cancer study as an example of a SMART using covariate-adjusted Bayesian logistic regression across three treatment blocks. She concluded that pursuit of perfection in trial design should not limit progress and that broader discussion with FDA on SMARTs is needed.

The key points raised in the panel discussion following the presentations are as follows:

• Many statistical challenges associated with re-randomization designs are not unique and also arise in other multi-stage trial evaluations. Re-randomization within a unified protocol increases transparency compared with the historical practice of using separate trial protocols for different treatment phases.

• The treatment policy estimand was broadly supported as the natural target under the ITT paradigm, with IPW providing valid inference among those who do not drop out under mild assumptions. However, handling patients who do not reach the second randomization within an ITT framework remains an unresolved issue.

• Differential dropout between treatment arms prior to re-randomization impacts the second phase regardless of the endpoint type. Binary or depth-of-response endpoints following induction therapy are more straightforward but are also subject to bias if only subset of ITT is being used for the analysis. Time-to-event endpoints in re-randomized settings introduce additional complexity related to time origin and potential bias from differential dropouts prior to re-randomization.

• The choice between re-randomization and factorial designs depends on the clinical context, particularly the expected dropout rate before the second randomization. Each approach has trade-offs in efficiency, statistical power, and interpretability.

• Perspectives differed regarding the role of re-randomized designs in drug development. Industry participants favored their use in Phase 2 to inform simpler Phase 3 designs, whereas many National Cancer Institute (NCI) multi-stage studies are conducted as Phase 3 trials.

• From a regulatory perspective, a key unresolved issue is whether these trials can support approval of an individual drug or the treatment strategy as a whole. The FDA reiterated its openness to innovative trial designs incorporating re-randomization.

This forum provided an opportunity to have open scientific discussion among a diverse multidisciplinary stakeholder group—clinicians, statisticians from academia, government, and pharmaceutical companies, patient advocates, and international regulators—focused on emerging statistical issues in cancer drug development.

Acknowledgement: Authors thank Joan Todd (FDA) for technical support.

* Speakers/Panelists:

Dr. Keaven Anderson (Merck), Dr. Somak Chatterjee (FDA), Dr. Michael Coory (TGA, Australia), Dr. Boris Freidlin (National Cancer Institute, NIH), Prof. Kelley Kidwell (University of Michigan), Prof. Franz König (Medical University of Vienna), Dr. Daeyoung Lim (FDA), Dr. Olga Marchenko (Bayer), Dr. Brittany McKelvey (LUNGevity Foundation), Dr. Gautam Mehta (FDA), Dr. Catherine Njue (Health Canada), Mr. Armin Schüler (BfArM, Germany), Dr. Amy Stark (Eli Lilly), Dr. Deepti Telaraja (FDA), Dr. Yevgen Tymofyeyev (Johnson & Johnson)

Subscribe now

Biomarkers as decision infrastructure

The Innovative Design Scientific Working Group (IDSWG) held its Oncology Biomarker KOL Roundtable and Panel Discussion on May 1, 2026, at Sanofi in Morristown, New Jersey, with a hybrid audience of approximately 200 attendees and 25 panelists from industry, academia, and regulatory agencies. The program, chaired by Gina D’Angelo of AstraZeneca and co-chaired by Hong Wang of Sanofi, covered biomarker strategy in early development, prognostic and predictive biomarker identification, biomarker-driven clinical trial design, companion diagnostics, and artificial intelligence (AI).

The FDA-NIH BEST Resource defines a biomarker as a measured characteristic that indicates normal biological processes, pathogenic processes, or responses to an exposure or intervention; importantly, it is not itself a direct measure of how a patient feels, functions, or survives [1]. A biomarker program is strongest when the context of use is explicit: proof of mechanism, target engagement, dose and schedule selection, patient enrichment, response or resistance monitoring, safety management, Phase 3 population definition, or labeling and companion diagnostic (CDx) development. Each context carries different evidentiary, statistical, operational, and regulatory requirements.

Start early

Biomarker strategy must be embedded in early phase development, not retrofitted after clinical uncertainty has accumulated. Early biomarker work should connect biology, pharmacokinetics/pharmacodynamics, assay feasibility, sampling windows, and clinical decision points. Project Optimus and recent FDA dose-optimization guidance reinforce the need to move beyond maximum tolerated dose selection toward integrated benefit-risk and exposure-response thinking in oncology [2]. In this setting, biomarkers can help measure target engagement, pharmacodynamic activity and even early clinical efficacy, but only when the assay, sampling schedule, and interpretation plan are prospectively defined.

Panelists emphasized that small early phase studies provide great opportunities to learn about biomarkers using proper statistical approaches such as regression-based approaches, graphical exploration, Bayesian modeling, high dimensional approaches, and resampling methods [3,4]. Many development programs in the discovery and early phase begin with data more suited to establishing prognostic value. Speakers highlighted the importance of aligning biomarker ambitions with the available study design, especially when teams are tempted to draw predictive conclusions from single-arm data. Several panelists stressed that robust predictive biomarker assessment generally requires comparative data, careful attention to clinical relevance, and a clear path to validation.

The discussion also explored the methodological challenges involved in biomarker discovery and evaluation. Panelists referenced established regression-based approaches for binary and time-to-event outcomes, as well as the use of penalized methods, tree-based approaches, and causal machine learning for higher dimensional settings. At the same time, speakers cautioned that methodology alone is not enough: the value of a biomarker depends on context, sample size, endpoint definition, assay performance, and whether the resulting signature is clinically interpretable and operationally feasible.

Do not overclaim predictive value

The distinction between prognostic and predictive biomarkers is a key concept in clinical trials. Prognostic biomarkers identify patients with different outcomes regardless of treatment, whereas predictive biomarkers identify patients who are more or less likely to benefit or be harmed by a particular treatment. Single-arm trials can identify prognostic biomarkers that can suggest biomarkers for further evaluation of the potential predictive feature. However, single arm trials usually cannot establish treatment-by-biomarker interaction to identify a predictive biomarker for patient selection. Robust predictive assessment generally requires comparative data from trials with different treatment arms, careful clinical interpretation, and external validation. Randomized two arm treatment trials are the gold standard to establish a biomarker signature that finds an enriched population to treat. As an alternative, single-arm trials may be matched to external controls, and real-world data (RWD) may also offer a credible approach for assessing whether a biomarker could be predictive; if the biomarker, assay, analysis plan, representativeness, and external validation are clearly established [5,6].

Continuous biomarkers illustrate the challenges between statistical approaches and practical implementation. Dichotomization can reduce power, inflate the Type-I error, and amplify bias to identify the biomarker positive group [4,7]. Yet thresholds are often necessary for trial eligibility, labeling, and companion diagnostics (CDx) development. The panel recommended keeping biomarkers continuous during exploration and using modelling and graphical displays to better understand how biomarkers relate to treatment response and efficacy. Once a clear relationship has been established, candidate cut-offs can be selected based on clinical, analytical, and operational considerations. In some settings, multi-level categories or quantitative continuous scoring may capture biological and clinical information more effectively than a single binary positive/negative rule.

Biomarker-based trial designs

Enrichment, stratified all-comer, adaptive design, basket, umbrella, and platform designs each solve different problems. These designs also introduce distinct operational, statistical, and interpretational complexity. Selecting the appropriate design requires careful alignment with phase of the trial, sample size, the strength of supporting biological evidence, and the specific clinical questions the study aims to answer. FDA enrichment and adaptive design guidance documents emphasize the importance of pre-specification, operating characteristics, and adequate information to justify the interim adaptation [8,9]. Panelists noted that Phase 3 trials should generally be designed to confirm a biomarker rather than discover one. When uncertainty remains about the biomarker signature or biomarker-based patient selection strategy to use in a Phase 3 trial, adaptive designs may offer a practical solution. However, successful implementation of adaptive/seamless designs requires that key elements be established early, including clearly defined interim decision rules, operational versus inferential seamless designs, clinical operational and regulatory buy-in, analytically validated and operationally ready assays, robust procedures to ensure data quality and integrity, and ongoing alignment with regulatory expectations.

Master protocols and Bayesian borrowing were discussed as promising tools for tumor-agnostic or biomarker-defined development. Basket trials can improve trial efficiency when a molecular alteration plausibly defines a shared biology across tumor types. Bayesian hierarchical models can improve statistical efficiency by borrowing information across tumor types. To mitigate the risk of over-borrowing, the borrowing must be clinically justified and transparent [10].

CDx Development

The panel emphasized that CDx are an essential component of precision oncology and should be developed in parallel with therapeutic agents. A key distinction highlighted during the discussion is that the biomarker itself and the assay used to measure it are fundamentally different. While a biomarker may have strong biological relevance, successful biomarker-driven drug development ultimately depends on the analytical performance of the diagnostic assay used to identify eligible patients. Therefore, assay selection and validation should occur early in clinical development, with careful evaluation of sensitivity, specificity, reproducibility, and biomarker prevalence based on the intended testing platform.

From a regulatory perspective, panelists stressed that companion diagnostic readiness should be aligned with drug development. Sponsors should clearly define the biomarker’s intended use, establish assay concordance between local and central testing when appropriate, and engage regulatory agencies early to ensure that both therapeutic and diagnostic development proceed in parallel. Such planning is particularly important for pivotal trials and accelerated development programs, where delays in diagnostic readiness may affect drug approval.

Operational considerations, including specimen quality, assay turnaround time, standardized testing procedures, and collaboration with diagnostic partners, were also recognized as critical for successful implementation. Furthermore, collecting comprehensive biomarker data during pivotal trials can improve the interpretation of biomarker thresholds and support evidence-based treatment decisions after approval. Overall, the discussion reinforced that robust companion diagnostics are indispensable for accurate patient selection, efficient clinical trial design, regulatory success, and the broader implementation of precision medicine.

Case discussions, including MET-driven resistance in EGFR-mutant lung cancer and ctDNA dynamics in cholangiocarcinoma, showed how specimen timing, assay platform, tissue availability, analytical validation, and cutoff refinement can materially affect a program’s trajectory. A third presentation showcased AI-enabled computational pathology, where quantitative continuous scoring of immunohistochemistry images was used to derive features associated with clinical outcomes and to support biomarker discovery for targeted therapies. FDA CDx guidance emphasizes that, in most circumstances, a companion diagnostic and its corresponding therapeutic product should be approved or cleared contemporaneously when the CDx can demonstrate the product’s safe and effective use [11]. Oncology group-labeling guidance further describes when a CDx may support broader labeling across a class of oncology therapeutics [11].

AI is an accelerator, not a substitute for validation

AI and multimodal biomarkers are major frontier topics. Speakers described opportunities to integrate pathology images, genomics, transcriptomics, radiomics, clinical data, and real-world data. Additional opportunities mentioned were to derive quantitative continuous pathology scores; and to use foundation models or virtual twins to generate actionable signals. At the same time, the analytical and statistical challenges remain recognizable: dimensionality, batch effects, missingness, measurement error, transportability, calibration, confounding, and interpretability. Good AI practice principles were emphasized such as multidisciplinary expertise, representative data, independent test datasets, model design and development practices, risk-based performance assessment, and life cycle management [12].

Real-world data (RWD) can possibly broaden representation, support hypothesis generation, enable model pretraining, and contextualize biomarker prevalence or outcomes. However, real-world evidence (RWE) is not automatically fit for regulatory decision-making. FDA’s RWD/RWE guidance documents clarify the need to assess data relevance and reliability and to plan study design, data sources, and analyses before relying on such evidence [13]. For AI with respect to the biomarker realm, the same principle applies, discovery is valuable, but clinical translation requires a locked model or classifier, independent validation, and clinically meaningful performance.

Statistician as a Strategic Partner

For biostatisticians and quantitative scientists, the panel highlighted that their expertise is essential to separate prognostic from predictive claims, design early phase studies and analysis plans, quantify uncertainty around cutoffs, evaluate longitudinal biomarkers, calibrate adaptive and Bayesian designs, and stress-test AI-derived insights. Equally important, statisticians can collaborate with cross-functional teams to assess evidence for adequate internal decision-making and regulatory labeling.

The IDSWG biomarker panel ultimately called for a shift in mindset. The question is no longer simply, ‘Which biomarker should we measure?’ Rather, the more practical and meaningful questions are: What decision will the biomarker inform along the drug development? What level of evidence is required for that decision? How will the assay perform in the intended population? How should uncertainty be quantified and communicated? And how will the biomarker be implemented in a clinical trial, a regulatory label, or routine patient care?

Acknowledgment

We extend sincere gratitude to the panelists, attendees, organizers, and sponsors, including NJSTAT, the ASA Biopharmaceutical Section, AstraZeneca, Sanofi, CIMS Global, InventiveMatrix, and the Society of Biopharmaceutical Sciences, whose support made this event possible.

We thank all the panelists: Cong Chen, Cheng Cui, Gina D’Angelo, Jean Fan, Antonio Tito Fojo, Margaret Gamalo, Liz Garrett-Mayer, Ryan Hartmaier, Antreas Hindoyan, Shuguang Huang, Etai Jacob, Olga Kholmanskikh, Paul Newcombe, Jai Pandey, Mohini Rajasagi, Hari Sankaram, Kui Shen, Kaveri Suryanarayan, Pedro Torres-Saavedra, Danielle Townsley, Jonathon Vallejo, Hongfang Wang, Yaji Xu, David Zhang, Heng Zhou.

We thank all the committee members: Gina D’Angelo, Hong Wang, Philip He, Revathi Ananthakrishnan, Erica (Xin) Tong, Yashvi Bhandari, Kevin Dobbin, Cindy Lu, Di Ran, Yujun Wu, Manjari Narayan, Li Liu, Jingxiao Chen, Yuqian Shen, Guannan Chen, Pat Mitchell, Jun Yin, Shuguang Huang, Heng Zhou, Judong Shen, Xiaowen Tian, Sutan Wu, Haijun Ma, Luke Ouma, Nicole Li.

Competing interests

GD is an employee of AstraZeneca and may own its stocks. HW is an employee of Sanofi and may own its stocks. PH is an employee of Daiichi Sankyo and may own its stocks. Contributions by the authors are solely their own and are not intended to express the views of their employers.

References

1. FDA-NIH Biomarker Working Group. BEST (Biomarkers, EndpointS, and other Tools) Resource: Harmonizing Biomarker Terminology. U.S. Food and Drug Administration and National Institutes of Health; 2016.

2. U.S. Food and Drug Administration. Optimizing the Dosage of Human Prescription Drugs and Biological Products for the Treatment of Oncologic Diseases: Guidance for Industry. August 2024.

3. D’Angelo G, Ran D, Yu B. Evaluation of Optimal Cut-Offs and Dichotomous Combinations for Two Biomarkers to Improve Patient Selection. Ther Innov Regul Sci. 2025. doi: 10.1007/s43441-025-00829-4. Epub ahead of print. PMID: 40581696.

4. D’Angelo G, Tian X, Deng C, Zhou X. Predictive Biomarker Graphical Approach (PRIME) for Precision Medicine. Pharm Stat. 2026; 25(3):e70094. doi: 10.1002/pst.70094. PMID: 42050979.

5. Simon RM, Paik S, Hayes DF. Use of archived specimens in evaluation of prognostic and predictive biomarkers. J Natl Cancer Inst. 2009;101(21):1446-1452. doi:10.1093/jnci/djp335.

6. Mandrekar SJ, Sargent DJ. Clinical trial designs for predictive biomarker validation: theoretical considerations and practical challenges. J Clin Oncol. 2009;27(24):4027-4034. doi:10.1200/JCO.2009.22.3701.

7. Royston P, Altman DG, Sauerbrei W. Dichotomizing continuous predictors in multiple regression: a bad idea. Stat Med. 2006;25(1):127-141. doi:10.1002/sim.2331.

8. U.S. Food and Drug Administration. Enrichment Strategies for Clinical Trials to Support Demonstration of Effectiveness of Human Drugs and Biological Products: Guidance for Industry. March 2019.

9. U.S. Food and Drug Administration. Adaptive Designs for Clinical Trials of Drugs and Biologics: Guidance for Industry. November 2019.

10. Park JJH, Siden E, Zoratti MJ, et al. Systematic review of basket trials, umbrella trials, and platform trials: a landscape analysis of master protocols. Trials. 2019;20:572. doi:10.1186/s13063-019-3664-1.9. U.S. Food and Drug Administration. In Vitro Companion Diagnostic Devices: Guidance for Industry and Food and Drug Administration Staff. August 2014.

11. U.S. Food and Drug Administration. Developing and Labeling In Vitro Companion Diagnostic Devices for a Specific Group of Oncology Therapeutic Products: Guidance for Industry. April 2020.

12. U.S. Food and Drug Administration, Health Canada, and Medicines and Healthcare products Regulatory Agency. Good Machine Learning Practice for Medical Device Development: Guiding Principles. October 2021.

13. U.S. Food and Drug Administration. Considerations for the Use of Real-World Data and Real-World Evidence to Support Regulatory Decision-Making for Drug and Biological Products: Guidance for Industry. August 2023.

Subscribe now

The origins of the organization date back to October 2003, when leaders in nonclinical biostatistics gathered for an initial meeting hosted by Schering-Plough (see Table 1). Representatives from Schering-Plough, Pfizer, Merck, Johnson & Johnson, and Bristol-Myers Squibb discussed common scientific and organizational challenges facing statisticians in nonclinical development. Participants were true pioneers who recognized the need for a dedicated professional forum for nonclinical statisticians.

Table 1. Participants of first nonclinical meeting at Schering Plough

Additional companies soon joined the effort, including GSK, Amgen, Wyeth, Genentech, and AstraZeneca. These early meetings focused on topics ranging from toxicology support and assay validation to organizational structure, software infrastructure, staffing, and regulatory impact on statistical practice. Although the participating organizations differed considerably in scope and responsibilities, a common professional identity began to emerge among statisticians working outside traditional clinical trials support.

In 2007, the organization formally adopted the name Nonclinical Biostatistics Leaders’ Forum (NCBLF), reflecting its growing structure and mission. A steering committee was established, with Jim Colaianne serving as the first chair. Another major milestone occurred in 2009 with the inaugural Nonclinical Biostatistics Conference at Harvard University. The conference demonstrated strong demand for a scientific forum dedicated specifically to nonclinical statistical applications. Over time, the conference evolved into a flagship event for statisticians working across nonclinical areas, fostering scientific exchange and cross-company precompetitive collaboration.

While these early accomplishments established a strong foundation, the most significant period of growth and evolution began after the organization joined ASA-BIOP in 2016. This article is written by the NCBWG chairs from both before and after the organization joined BIOP. See Table 2 for a full listing of NCBLF/NCBWG chairs.

Table 2. NCBLF/NCBWG Chairs

Joining ASA-BIOP: A Transformational Milestone

The integration of the NCBLF into the ASA Biopharmaceutical Section represented a major turning point for the organization. Following the transition, the group became the Nonclinical Biostatistics Working Group (NCBWG), formally aligning it with the working group structure of BIOP.

This partnership substantially expanded the visibility and influence of the nonclinical statistics community. ASA-BIOP assumed sponsorship of the biennial Nonclinical Biostatistics Conference beginning in 2017, and the NCBWG gained representation within the BIOP Executive Committee through dedicated leadership positions for the NCBWG chair and the conference chair.

The transition also marked an important cultural shift. What had originally been a leadership-focused networking forum evolved into a broad scientific and professional community integrated within one of the largest statistical organizations in the pharmaceutical industry.

Over the past decade, NCBWG members have become increasingly active throughout ASA-BIOP leadership and operations. This growing engagement reflects both the maturation of the organization and the expanding recognition of nonclinical statistics within the broader biopharmaceutical community.

NCBWG members contribute extensively across BIOP committees and operational activities. Ji Young Kim serves as BIOP Distance Learning Chair, helping expand educational programming and webinar activities. Katie Brickey serves as BIOP Webmaster, supporting communication and member engagement across the section. Olga Yee contributes on the BIOP Funding Committee. Francis Rogan serves as the nonclinical associate editor for the Biopharmaceutical Report, helping ensure visibility for nonclinical statistical topics within BIOP communications. Lina Niu contributes to the BIOP Mentoring Committee, supporting professional development and outreach to early-career statisticians. Several members have also sat on the executive committee, including the current BIOP chair, Steve Novick.

These contributions illustrate how the NCBWG has evolved from a relatively specialized industry forum into a deeply integrated and highly engaged component of ASA-BIOP leadership.

Expansion of Scientific Activities

Since joining BIOP, the NCBWG has experienced substantial growth in scientific programming, professional initiatives, and member engagement. The organization now supports multiple scientific workstreams addressing emerging challenges in pharmaceutical development, including:

• Bayesian methods in CMC and preclinical research

• Manufacturing comparability

• Virtual control groups

• Medical device CMC

• Student outreach and mentorship

These activities have created opportunities for statisticians from industry, academia, and regulatory agencies to collaborate on conference sessions, white papers, webinars, publications, and educational programming.

The Bayesian CMC workstream provides one example of the group’s scientific evolution. Established after integration into BIOP, the workstream was created to advance modern Bayesian methods in pharmaceutical CMC applications. Since 2020, it has published three journal articles and the book “Case Studies in Bayesian Methods for Biopharmaceutical CMC” as well as organize several JSM sessions, and a presentation on nonclinical opportunities in the pharmaceutical industry to the Committee of Academic Representatives.

The NCBWG has also strengthened collaboration across disciplines. Integration within BIOP has facilitated greater interaction between statisticians working in nonclinical research, clinical development, regulatory science, manufacturing, and data science. These connections have expanded scientific exchange and strengthened awareness of the critical role statisticians play across the entire pharmaceutical development lifecycle.

Building Community and Supporting the Future

Another major area of growth since joining BIOP has been mentorship, outreach, and professional development. The NCBWG has increasingly emphasized support for students and early-career statisticians through poster competitions, conference networking opportunities, a scholarship award, and mentorship activities.

These efforts are particularly important because many graduate statistics programs provide limited exposure to nonclinical pharmaceutical applications. The NCBWG helps introduce students and young professionals to career paths spanning discovery research, toxicology, manufacturing sciences, and analytical development.

The biennial Nonclinical Biostatistics Conference continues to serve as the scientific centerpiece of the community. The conference now brings together statisticians, data scientists, pharmaceutical scientists, and regulators to discuss methodological innovation and practical applications across discovery and CMC research.

Equally important has been the strong sense of professional community that has emerged. Statisticians working in nonclinical settings are often embedded within highly interdisciplinary scientific organizations and may have relatively few peers within their own companies. The NCBWG provides an important network for collaboration, mentorship, and exchange of ideas across organizations and disciplines.

Looking Forward

The evolution of the NCBWG reflects the broader transformation occurring throughout the pharmaceutical industry. Modern drug development increasingly relies on quantitative methods across the full product lifecycle, from early discovery through manufacturing and commercialization. Advances in biologics, cell and gene therapies, artificial intelligence, and high-dimensional analytical technologies continue to create new opportunities and new statistical challenges.

In large part, due to its association with BIOP, NCBWG is well positioned to help address these challenges through scientific collaboration, methodological innovation, and community-building. What began in 2003 as a small leadership forum has evolved into a thriving scientific community within ASA-BIOP, with active engagement across industry, academia, and regulatory science.

The success of the NCBWG demonstrates the value of ASA-BIOP, the long-term value of collaboration, shared scientific purpose, and professional engagement. As nonclinical biostatistics continues to expand in importance, the partnership between the NCBWG and ASA-BIOP provides a strong foundation for continued growth and impact across the pharmaceutical sciences.

Subscribe now

Hybrid control arms leverage external or historical data to augment the control arm of a randomized controlled trial and can be particularly useful in oncology settings where conducting a fully powered randomized trial may be infeasible or unethical. Hybrid trial designs require additional considerations at the study-design stage, including the borrowing strategy for external data and methods for assessing similarity between the randomized trial population and the external population. It is important to evaluate consistency across both internal and external control populations; concordance with the randomized internal control arm provides reassurance that the external dataset has been appropriately selected. This open forum of multidisciplinary experts reviewed methodologies for constructing hybrid control arms in cancer clinical trials, discussed statistical challenges inherent to hybrid designs, and explored practical approaches to address these issues.

The speakers/panelists* for the discussion included members of the BIOP Statistical Methods in Oncology Scientific Working Group representing pharmaceutical companies, representatives from international regulatory agencies (Food and Drug Administration (FDA), Medicines and Healthcare products Regulatory Agency (MHRA, UK), and Austrian Agency for Health and Food Safety), clinicians, academicians, patient advocates, and expert statisticians. In addition, over 100 participants attended the virtual meeting, including representatives from other international regulatory agencies (European Medicines Agency (EMA), Health Canada (HC), Medicines Evaluation Board (MEB, Netherlands)), Paul-Ehrlich-Institute (PEI, Germany), Therapeutic Goods Administration (TGA, Australia), Brazilian Health Regulatory Agency (ANVIS, Brazil), Health Sciences Authority (HAS, Singapore), Ministry of Health (MOH, Israel), Pharmaceuticals and Medical Devices Agency (PMDA, Japan)). The discussions were moderated by the BIOP Statistical Methods in Oncology Scientific Working Group co-chair, Dr. Olga Marchenko from Bayer, Brittany McKelvey from LUNGevity Foundation, and Dr. Rajeshwari Sridhara, consultant from OCE, FDA.

In the introductory presentation, the OCE presenter noted that hybrid control arms may be appropriate for selecting cancer indications to reduce the required sample size of a randomized clinical trial when fully powered randomization is not feasible. He emphasized that differences between randomized trial and external populations can introduce bias and complicate interpretation of trial results. Panelists and presenters were invited to discuss suitable scenarios for employing hybrid control arms, optimal data-borrowing strategies, and approaches to mitigating bias arising from both measured and unmeasured population differences.

The first academic presenter addressed the challenge of pre-specification in hybrid control trials, noting that while augmenting randomized control arms with historical control arm data may improve efficiency, it also increases the risk of bias when historical and concurrent controls differ. He contrasted static borrowing methods, which can lead to the unbounded type I error inflation, with dynamic borrowing approaches that adapt to observed outcomes and better constrain error rates. The speaker emphasized that valid inference depends on pre-specifying design choices, data sources, and analytic methods independently of outcome data. Such pre-specification can, however, be difficult to implement and validate when data from historical trials have already been published and are available to investigators. Simulation studies showed that outcome-dependent selection of historical controls may further inflate the type I error rate and lead to biased treatment effect estimates.

The second academic speaker introduced a partial bias correction (PBC) method to address unmeasured confounding when borrowing synthetic (external) control data in hybrid control trials. The approach uses propensity score weighting to balance observed covariates and step functions to partially correct bias, with an optional step to adjust for large, systematic outcome differences when scientifically justified. The PBC method requires a user-specified parameter to define the maximum tolerable bias. Simulation and real-data examples showed that PBC effectively balances bias and precision between randomized-only and fully pooled analyses, yielding estimates closer to randomized trial results with modest efficiency gains, while remaining computationally simple and broadly applicable across endpoints.

The key points raised in the panel discussion following the presentations were: 

• Well-powered RCTs remain the gold standard. Hybrid control arms may be considered when fully powered RCTs are infeasible or unethical (e.g., rare diseases, pediatrics, early oncology development). Early regulatory engagement is strongly encouraged.

• Hybrid designs are more reliable when external and internal controls are highly comparable in eligibility criteria, treatment context, endpoints, timing, and outcome measurement.

• Pre-specification in the design stage, including data sources, borrowing strategies, priors, discounting rules, and decision criteria are key to the success of the trial. Trials should avoid outcome-driven selection of external controls.

• Hybrid control arm designs may improve precision, reduce required sample size, and assign more participants to novel treatment by borrowing external control data. However, any power gain should be outweighed against potential risk of bias and the type I error inflation and this should be investigated at the design stage. Considerations should also be given to the reliability of the externa data both in terms of quality and fitness for purpose of the external data source.

• Static borrowing is straightforward but may be severely underpowered when the external data are down-weighted. Dynamic borrowing is less transparent and increases complexity. Both methods remain vulnerable to bias and inflated type I error rate. This is especially true in precision-oncology applications, where reliable external controls are scarce and randomized control arms are often too small to adequately assess how much to borrow from external data.

• Sensitivity analysis (e.g., tipping-point analyses) and quantitative bias analyses are useful for assessing robustness of the results to unmeasured confounding and other departures from key assumptions for hybrid control arm designs.

This forum provided an opportunity to have open scientific discussion among a diverse multidisciplinary stakeholder group – clinicians, epidemiologists, and statisticians from academia and pharmaceutical companies, patient advocates, and international regulators- focused on emerging statistical issues in cancer drug development.  

Acknowledgement: Authors thank Joan Todd (FDA) and Syed Shah (FDA) for technical support.

* Speakers/ Panelists:  

Dr. Anup Amatya (FDA), Dr. Ruthanna Davi (Medidata Solutions), Dr. Nicole Drezner (FDA), Dr. Alfredo Farjat (Bayer), Dr. Boris Freidlin (National Cancer Institute, NIH), Dr. Rima Izem (Novartis), Dr. Thomas Jemielita (Merck), Dr. Florian Klinglmueller (Austrian Agency for Health and Food Safety), Dr. Li Liang (University of Texas, MD Anderson Cancer Center), Prof. Bo Lu (Ohio State University), Dr. Olga Marchenko (Bayer), Dr. Brittany McKelvey (LUNGevity Foundation), Dr. Gautam Mehta (FDA), Dr. Richard Pazdur (FDA), Dr. Martin Posch (Medical University of Vienna), Dr. Khadija Rerhou Rantell (MHRA, UK), Dr. Rajeshwari Sridhara (FDA), Ms. Yinghua Su (Health Canada), Dr. Huan Wang (FDA)

Subscribe now

This note is long overdue, but the Midwest Biopharmaceutical Statistics Workshop (MBSW) Executive Committee wants to thank BIOP for its generous donation supporting student registration at the 49th MBSW in May 2026. Your support made a real difference in the experience we were able to offer students this year in our program on Shaping the Future of Medicine Through Data and AI.

Here is a summary of what the student program (found at https://mbswonline.com/program/) included.

Student Session Highlights

• Registration support. A $3,000 grant from BIOP covered registration and meals for 16 students attending MBSW.

• Group coordination. A dedicated group text chat helped students coordinate arrivals and meals, and it fostered collaboration throughout the conference.

• CV feedback. Students had the option to submit their CVs for review by two industry professionals. Several went on to submit their revised CVs to the JSM 2026 Career Service.

• Mock interview session. One committee member volunteered to be interviewed by a colleague, offering both weak and strong sample responses for students to critique. Students then submitted their own responses to interview questions tailored to their CVs, and the group worked through them together. The session ran well past its allotted time thanks to how engaged everyone was.

• Poster awards. Every student poster presentation was reviewed by several judges, and three awards (pictured below) were given for best posters.

• Community building. Students shared catered meals and spent time exploring the local area together as a group.

In Their Own Words

Students shared the following feedback after the conference.

Thank you again for such an amazing workshop. This has been the best experience of my PhD so far.

MBSW was honestly one of my favorite conferences so far. The people, the conversations, and the insights made it truly special. Thank you, Ann, for organizing everything so beautifully and bringing this amazing group together.

It was an amazing experience. It was great to meet all of you, and I learned a lot from everyone. Looking forward to seeing some of you at JSM.

MBSW was a wonderful conference, and it was so great to meet you all this week 🙂. Thank you, Ann Marie, for bringing all of us together before, during, and after MBSW, and for creating such a PhD student centered space for us. We will definitely see each other again and stay in touch.

Pictures

A few highlights from the student program are included for your enjoyment.

[Pictured: Students enjoying food and listening to engaging speakers.]

[Pictured: Best student poster award winners. Top right (Asteria Herbert Chilambo); bottom left (Carly Middleton); bottom right (Quynh Long Khuong).]

Thank you again for helping us create such an engaging and welcoming environment for the students. We, and they, are deeply grateful for your generosity.

With appreciation,

The MBSW Executive Committee

Subscribe now

I will start by acknowledging that we all have different career journeys and there is no “one size fits all”. My own career began as an entry-level graduate statistician and has involved both upwards and sideways moves. I started working in early phase Oncology, moved to late phase Oncology, took on a blended role that is accountable for both biometrics input to a program as well as line management, then moved therapy areas to focus on Respiratory and Immunology. Along the way, I added being a randomization process owner, moved countries (from UK to USA) and had a family. Now I’m a Senior Director at AstraZeneca, and the company I joined back as a graduate is very different from the one it is today.

These days, I think less of myself as a statistician and more as a drug developer. I intend to cover, in this article, some key themes in a pharmaceutical statistician’s career journey, and what helps us (as statisticians) broaden our impact on the whole development path of a project.

Let us start with who we are as statisticians. We’re often seen as adding common sense and being logical. Our purpose is to ensure that the relevant information is available and considered when decisions are made. We may not be the final decision makers when it comes to development plans, but we can (and should) have a strong voice at the table. We should not underestimate the impact of that.

If our voice is heard, that’s great, and even better if our opinions are requested. As statisticians, every topic in drug development can benefit from our opinion, and we can measure success when we’re being asked for our opinion on non-statistical topics.

The following ingredients help broaden our impact: (1) communication and translation, (2) stepping outside our lane, (3) problem solving by providing logical solutions, and (4) taking accountability.

Communication and translation: I believe a statistician’s super strength is often to make logical arguments and be data-driven. However, colleagues may just want the key messages. As statisticians, we love details/data, so how do we communicate effectively? What is the key message and who are we communicating to? To have the message land, think about describing information in ways that resonate (e.g., using sports metaphors to frame the issue in a common language), be concise, consider using visuals if needed. Listen to the other person to ensure you each understand the others’ view. Impactful feedback I received early in my career was to be clear in the first line what the ask is. Be concise, and if detail is needed, add it as an attachment, keeping only the key messages in the email body (along with your recommendation). It was a gift, and I learnt that for critical communication, it pays to invest time in crafting the message.

Step outside your swim lane: Don’t be afraid to ask questions of other functions, to learn to understand the issue at hand. Be participative in meetings and offer your opinion – know you don’t have to be 100% sure to chime in. I appreciate this may be daunting at first, but experiment with it.

Problem solve and provide logical solutions: We are logical thinkers and like to solve problems, so apply this skill to statistical and non-statistical challenges alike. Treat obstacles as challenges and find ways to remove or navigate them or offer alternative pathways. This also applies to those situations when we receive new information that feels like a deviation to the plan, and we need to adapt to those new assumptions. I have seen many examples where a problem is seen as insurmountable, but then assumptions are probed and a way around or “over” the obstacle is found.

Accountability and ownership: I can’t emphasize the importance of this behavior enough – we build trust and strong relationships with our study/project team colleagues when our behaviors reflect that we’re on the team so own the decisions/recommendations made. Think about high-performing teams you know and I’m sure as well as strong communication, you see a strong level of trust between colleagues.

There are some tangible actions that can be taken to go from having the ingredients to making a complete dish. The list below is not exhaustive; these are simply suggestions to consider:

• Continue learning, be it technical, disease area, soft skills, cultural awareness, or AI.

• Elicit feedback in the moment and remember that receiving feedback is a gift. Consider pro-actively soliciting feedback ahead of important activities so the “giver” knows to pay attention. Ask probing questions – remember we each have blind spots so receiving feedback can allow us to adjust as needed.

• Lean in and say yes to opportunities (even if they don’t sound like an opportunity). Push yourself outside your comfort zone – you really will learn more. You may make mistakes, but the key is to learn from them. For me, taking on a process owner role gave me growth opportunities and connections I had not foreseen.

• Continue to flex or adapt. Work at different levels (sometimes we need to be in the details and sometimes we need to work at a higher level). If you spend all your time at the helicopter level, you could end up giving the wrong advice, and if you spend all your time in the weeds, you may miss the bigger picture/impact.

• Take decisions. We gather info but sometimes it feels hard to then take a decision, despite us maybe being best placed due to the knowledge gained. Learn to take them as sometimes more harm is done through a lack of decision than by making the wrong decision.

• Be present. Be camera on for video calls, which allows others to see that you are paying attention. When joining new teams, work to build key relationships early – it will make it much easier to walk through challenging issues when needed.

• Have difficult conversations. They are uncomfortable and we dread them, so invest time in preparing for them as changes often happen because of them, so they’re to everyone’s benefit.

These final three are more focused on self-care and maintaining a work life balance.

• Find your village – find those people who are sounding boards for advice.

• Find an outlet – I took up running several years ago and I found it really helps me handle challenges and often gives me perspective. Last year, at my daughter’s softball practice, the phrase “flush it away” was used after every bad catch/bat – what a great way to learn to take responsibility but not to let it eat you up.

• Spend time wisely – block your time, recognize when input is good enough, recognize what’s in your sphere of control and be kind to yourself. I block two afternoons a week in my calendar to ensure I have some stretches where I can focus (and not be in meetings). I also block time needed for key school events (or key activities outside of school), and conversely, let my family know in advance when I know there is going to be an intense period. This is key to the two-way connection of work-life balance.

Hopefully the above provides a few new nuggets to think about. If I were to think of the three pieces of advice to take away, it’s (1) Say yes to opportunities, (2) Take ownership (of your work, your development, be present) and (3) continue to seek feedback.

I hope you’ll see that the result is that your voice is sought out more, and beyond just statistical input.

Why I Chose to Step Away and Look Inward

As a scientific leader, I am accustomed to solving problems through data, models, and structured reasoning. Yet as my work has expanded to include guiding people and shaping organizational direction, I have come to appreciate that leadership draws not only on analytical discipline, but also on perspective, empathy, and influence. These responsibilities often require navigating ambiguity, balancing priorities, and bringing others along—not simply to reach a sound decision, but to build shared commitment behind it.

Over time, I became increasingly curious about questions at the heart of effective leadership: how my instincts shape the way I lead, how to make sound decisions under pressure and uncertainty, and how to inspire people across an organization during periods of change. Those questions drew me to High-Potential Leaders: Accelerating Your Impact at the Wharton School of University of Pennsylvania—an intensive five-day program designed to help experienced professionals deepen self-awareness, strengthen team leadership, and translate reflection into practical impact.

Like many organizations, ours is evolving through a period of meaningful transformation. This evolution is not only structural; it also reflects shifts in mindset, ways of working, and the role of advancing technologies. As I observed these changes and contributed to them, I realized the importance of preparing myself—and helping prepare the broader organization—to adapt with clarity, confidence, and purpose.

What the Week Taught Me

The program was less about memorizing frameworks than about looking inward and outward at the same time. Five themes stayed with me, each anchored in a session that connected personal reflection with practical leadership application.

1. Adaptability starts with self-awareness

The week opened with “The Adaptable Leader” and a deeply personal Wharton Personality Profile, which mapped traits such as conscientiousness, openness, extraversion, and emotional reactivity against a cohort of peers. Rather than revealing something entirely unexpected, the assessment gave me a clearer language for understanding patterns I already recognized in myself. Seeing those patterns reflected through data—and considering how they may be experienced by others—was a powerful reminder that adaptability begins with self-awareness.

What stayed with me most was the idea that adaptability is not about changing who we are, but about becoming more intentional in how we apply our strengths. Sometimes that means choosing environments where our natural traits can be most effective; at other times, it means adjusting our approach to meet the needs of the situation. The value of the exercise was not in labeling personality, but in helping me better understand when to lean into my instincts, when to stretch beyond them, and how to lead with greater flexibility and purpose.

2. Defining my leadership identity

In the “Leader Identity” sessions, we were asked to articulate the qualities that matter most to us and how we want to be perceived as leaders. Leadership was framed not as a title or position, but as something expressed through our actions and relationships—grounded in authenticity, compassion, and accountability. It was an invitation to lead more deliberately from values rather than by default.

When I reflected on the qualities I value most in a leader, several early-career moments came immediately to mind. In one, I observed a leader respond to a crisis with calm focus, moving quickly from the pressure of the moment to the actions needed to address it. In another, I saw a leader shield the team from unnecessary organizational noise, taking on that burden personally so others could stay focused on the work. Together, these experiences clarified the leadership values that matter most to me: steadiness under pressure, accountability, care for the team, and purposeful action.

3. Decision-making under pressure

Professor Marissa King’s “Decision Making for Leaders” introduced the Vroom-Yetton model and a personalized report showing how much I involve others in decisions—from deciding independently to consulting, facilitating, or delegating—compared with peers and expert models. The core insight was both practical and freeing: no single style is best. Effectiveness depends on matching the approach to the situation, balancing decision quality, buy-in, time, and team development.

The personalized profile helped me see my own decision-making tendencies more clearly. In scientific work, I naturally value rigor, consultation, and multiple perspectives; the session reminded me that involving others is most effective when it is intentional rather than automatic. Some decisions benefit from broad input and shared ownership, while others require timely judgment and clear direction. The session reinforced the importance of reading each situation carefully—understanding where alignment is essential, where expertise should be brought in, and where a timely decision can create momentum for the team.

4. Driving change

Deputy Dean Nancy Rothbard’s “Leading & Managing Innovative Change” brought this lesson to life through a hands-on simulation, where we advised a struggling company and worked to build organizational buy-in within limited time and budget. What made the exercise powerful was how closely it mirrored real change efforts: the pull to move quickly to a vision before fully understanding the problem, the importance of communicating more than feels necessary, and the value of urgency, social proof, and careful framing of gains and losses.

The experience strongly echoed the organizational changes I have observed and experienced in my own work. It reminded me that successful implementation starts well before execution—with understanding the landscape, engaging the right people, and helping others see both the need for change and their role in shaping it. I also gained a deeper appreciation for the balance between motivation and communication: people need to understand why change matters, and they need clear, credible, and repeated messages that help them move forward with confidence. It also helped me recognize why some change efforts felt more effective than others: much depends on how thoughtfully each step is taken.

5. Collaboration in action

The Leadership Navigation Challenge, led by Jeff Klein and John Kanengieter, took us out of the classroom and across the Penn campus to locate waypoints as a team—assigning roles, setting strategy, and debriefing with honest after-action reviews. Paired with the sessions on high-performing teams, it offered a vivid reminder that trust, clear roles, and open communication are what fuel collaborative success.

For me, the most memorable part of the challenge was how quickly the quality of our teamwork became visible. When we paused to clarify roles, listen to different perspectives, and adjust our path together, we moved with more confidence and purpose. It reinforced a lesson I have often seen in professional settings: collaboration does not happen simply because capable people are placed on the same team. It requires leaders to create the conditions for trust, alignment, and shared ownership so that individual strengths can become collective impact.

Carrying the Learning Forward

What made the experience unforgettable was not only the faculty, but also the cohort itself—a group of accomplished leaders from diverse industries and geographies whose perspectives enriched every discussion and whose connections will last well beyond the week. The program closed with a simple but resonant charge: be a student of leadership, surround yourself with coaches and trusted colleagues, and keep seeking stretch experiences.

It is early to know exactly how these lessons will shape my day-to-day work, and that is precisely the point: leadership development is a practice, not an event. My intention is to lead with sharper self-awareness, choose my decision-making style more consciously, communicate change with more repetition and empathy than feels natural, and invest deliberately in the teams and networks around me.

Looking ahead, I see this program influencing not only how I lead, but how I continue to learn as a leader. It has encouraged me to pause more intentionally, listen more deeply, and choose my approach with greater awareness of both the situation and the people involved. For colleagues considering the program, my advice is to enter it with openness and curiosity rather than a narrow expectation of acquiring tools. The most valuable learning may come from the questions it prompts, the perspectives it opens, and the honest reflection it invites about the leader you are becoming. For me, it was a reminder that leadership growth begins with looking inward and continues through the choices we make every day.

Bio:

Dr. Wenqiong Xue is Global Head of Biostatistics and Data Sciences Clinical Development in Mental Health and Eye Health at Boehringer Ingelheim.

Over more than a decade, Dr. Xue has developed broad experience across global drug development, beginning as a trial statistician in respiratory and progressing into leadership roles across multiple therapeutic areas, including respiratory, oncology, ophthalmology, and programs spanning early- to late-phase development. As BDS TA Head, she leads the strategic and operational contributions of Biostatistics and Data Sciences to the therapeutic area portfolio and shapes the broader BDS strategy within the global organization.

Dr. Xue is passionate about advancing innovative statistical methodologies within the medicine community and leveraging AI/ML approaches to accelerate and strengthen drug development. Her leadership philosophy centers on trust, authentic relationships, and open communication, reflecting her belief that strong collaboration leads to stronger outcomes. She has served as a Steering Committee member of the Duke Industry Statistics Symposium since 2023.

Subscribe now

Hosted from May 19–21, 2026, at the Busch Student Center at Rutgers University, this year’s conference highlighted how classical experimental design principles are evolving to meet modern computational realities. As fields like artificial intelligence, machine learning, and high-dimensional modeling reshape data science, experimental design remains foundational, not merely for efficiency, but as a strategic driver of scientific rigor, causal validity, and optimization across industry and academia.

Welcome Address and Technical Sessions

The conference opened on May 19th with a welcome address by Thu Nguyen, Dean of Mathematical and Physical Sciences at Rutgers University. Following the opening remarks, the scientific program unfolded across 10 invited technical sessions that bridged classic design theory with modern computational and pharmaceutical challenges. Presenters explored robust frameworks for design-based causal inference, dynamic nonlinear kinetic modeling for bioassays, and nonclinical pharmaceutical sciences. A major focal point of the scientific program was the integration of Artificial Intelligence (AI) and Machine Learning (ML) into experimental design. Researchers demonstrated how large language models can boost the power of randomized experiments, and how novel optimal designs can streamline complex AI hyperparameter tuning while improving model explainability. The sessions also showcased methodological breakthroughs in uncertainty quantification using Gaussian process surrogate models, particularly for computationally expensive computer experiments. Furthermore, the program highlighted advancements in classical areas, introducing new space-filling criteria, parallel flats designs for split-plot experiments, and robust solutions for constrained order-of-addition configurations and adaptive Neyman allocations.

Mentoring Session and Panel Discussion

A standout feature of the first day was the dedicated Mentoring Session held on May 19 from 3:00–4:00 PM. Aimed at equipping Ph.D. students and early-career researchers with guidance on potential career paths and strategies for advancing their research, the session was structured into an intimate roundtable-discussion format. Participants chose between two academia-focused tables hosted by senior researchers Robert Mee (University of Tennessee), Xinwei Deng (Virginia Tech), Hongquan Xu (UCLA), and Simon Mak (Duke University), and an industry-focused table hosted by Stan Altan (Johnson & Johnson) and Ryan Lekivetz (JMP), fostering rich, direct dialogue on career development.

Academia-focused tables hosted by Robert Mee (University of Tennessee) and Xinwei Deng (Virginia Tech).

Academia-focused tables hosted by Hongquan Xu (UCLA), and Simon Mak (Duke University).

Industry-focused tables hosted by by Stan Altan (Johnson & Johnson) and Ryan Lekivetz (JMP).

Another major highlight was Wednesday afternoon’s interactive Panel Discussion, “Design of Experiments for academia and industry in the era of AI: challenges and opportunities,” moderated by Matteo Bonvini (Rutgers) and featuring panelists Ryan Lekivetz (JMP), Sam Gardner (Eli Lilly), Dennis Lin (Purdue), and John Stufken (George Mason). The conversation examined the shift toward AI-augmented experimentation while addressing the gap between academia’s focus on statistical theory and industry’s focus on speed and deployability. Panelists also highlighted emerging methodological opportunities, key practical bottlenecks, and the urgent need to modernize university curricula so future statisticians can effectively navigate the intersection of experimental design and AI.

Poster Sessions

This year’s program featured two active poster sessions, hosting 15 contributions across applied and theoretical statistics. Emerging scholars presented innovations in subdata selection for massive datasets, variance estimation for spatial modeling, sliding window discrepancy metrics, and Gaussian copula-based sensitivity analysis for causal outcomes.

Organizational Leadership and Looking Ahead

The success of DAE 2026 was made possible by the dedication of the Steering Committee (Ryan Lekivetz, Hongquan Xu, Wei Zheng, Haiying Wang, and Xinwei Deng), Organizing Committee Co-Chairs Tirthankar Dasgupta and Ying Hung, and Conference Coordinators Amanda Velcheck, Eileen Sharkey, and Sherrae Thomas.

As data environments grow increasingly complex, the role of experimental design remains vital. By uniting methodological innovation, real-world translation, and peer networking, DAE 2026 strengthened our community’s ability to drive statistical excellence. We look forward to building on this momentum at future DAE gatherings!

Subscribe now

Source: BBSW AI Committee report, “AI in Biotech and Pharma R&D Survey Report.” https://www.bbsw.org/ai-report-2026

Highlights

• AI adoption has hit a tipping point — a BBSW survey of 60 senior data leaders shows AI is now part of daily work in biotech and pharma R&D, yet only 1 in 3 organizations is scaling beyond pilots.

• The bottleneck isn’t ideas — it’s the delivery system. Privacy, integration, talent, and regulatory uncertainty are what keep AI stuck in pilot mode.

• Readiness predicts scaling. Organizations with governance and data foundations in place are nearly 10x more likely to reach enterprise-scale adoption.

• Teams want automation with guardrails: routine analytics automation and real-time data quality monitoring top the wish list, paving the way toward automated submissions and discovery.

• Biometrics is the wedge. Where outputs must be defensible and reproducible, organizations are forced to build the governance muscle that scales across all of R&D.

• The winners will compete on “trust at scale” — not the newest model, but AI that is safe, repeatable, and audit-ready under regulation.

1. Executive Summary

Based on the survey of 60 senior data experts in the biotech and pharma industry, AI adoption is clearly well underway and has reached a tipping point. Teams are using AI widely for writing, searching, and coding; but when work crosses organizational boundaries — where privacy, validation, auditability, and system integration become essential — progress slows. This creates a familiar pattern: strong local productivity wins, but limited enterprise-scale transformation.

Three signals in the survey data. First, tool usage is already mainstream. The most common tools teams report using are general-purpose GenAI assistants (82% of respondents), literature review or knowledge extraction (60%), and coding assistants (52%). These are “low friction” entry points which deliver benefits without deep integration into regulated data flows.

Second, organizational scale is still the exception. At the organizational level, only 22 of 60 respondents (37%) describe their organizations as either “scaling across functions” or treating AI as a “core strategic priority with widespread adoption.” Most are still exploring or running pilots in isolated areas.

Third, the blockers are not “lack of ideas,” but “lack of a delivery system.” The top barriers concentrate on the classic enterprise gap: data privacy or security (45%), integration with existing systems or workflows (42%), lack of skilled AI personnel (40%), regulatory uncertainty (38%), and insufficient data quality or availability (30%). In other words, what limits scaling is the ability to make AI safe, repeatable, and governable, not the ability to run a demo.

The most important practical takeaway is where to start. Near-term unmet needs cluster around routine analytics automation (67%) and real-time data quality monitoring (48%). In the Biometrics / Clinical Data Science subgroup (n=30), those priorities intensify: 77% cite routine analytics automation, and 50% cite real-time data quality monitoring. This function is therefore a natural “wedge” for scale: it has large value pools, strong constraints that force governance discipline, and deliverables that benefit from standardization and reuse.

The path forward: start where ROI is fastest as pilot for the short-term gains (1–2 years) — routine analyses and real-time data quality as top choices — then build the governance infrastructure that turns a successful pilot into repeatable, GxP-compliant workflows for enterprise-level long-term transformation (3–5+ years).

2. Survey scope, sample, methods, and how to read results

The AI in Biotech & Pharma Survey is conducted by BBSW, a non-profit community dedicated to Biotech and Pharma data leaders (www.bbsw.org), from December 2025 through February 2026. It is built from 60 survey responses across pharma (60%), biotech (15%), Clinical Research Organizations (CROs) (7%), MedTech (7%), academe (5%), and other related sectors (6%), with strong representation from biometrics and data science-adjacent roles in the regulated R&D environments.

The respondents represented a broad range of biometrics leadership and practitioners with meaningful input from Director/manager (25%), and executives at VP level or C-suite (15%). Most of the survey questions are multi-choice. Because participation is voluntary, the sample is best interpreted as directional for identifying themes, priorities, and capability gaps.

This report uses a value-chain lens — from early discovery, clinical development, and regulatory submission — as well as a function lens, with special focus on Biometrics and Clinical Data Science (CDS). The report also looks at a simple maturity segmentation to show how constraints shift as organizations move from experimentation to scale.

3. External context: why “auditable AI” is becoming the standard

The regulatory environment is shifting from abstract debate (“is AI acceptable?”) toward practical frameworks (“under what use, with what evidence, at what level of confidence?”). FDA’s draft guidance on AI supporting regulatory decision-making emphasizes context-of-use and credibility considerations. EMA similarly highlights transparency, bias, human oversight, and governance across the medicinal product lifecycle. These signals do not simply increase scrutiny; they also clarify the path to scale by making “credibility and control” explicit design requirements.

A second signal is that regulators themselves are adopting AI internally (e.g., FDA’s “Elsa” initiative and later agentic AI announcements). This matters because it reinforces a realistic stance: AI is increasingly treated as a capability that can be deployed safely within controlled environments — provided the operating model, security posture, and auditability are sound.

The survey results align with this context. Respondents cite regulatory uncertainty as a major barrier (38%; 23 of 60), but at the same time they cite validation and quality/reliability of outputs as key drivers of confidence. This combination suggests organizations are not waiting for “permission”; they are looking for a defensible way to implement AI under compliance constraints.

4. Where the industry stands: adoption is broad, scaling is uneven

Similar trend in AI adoption and readiness is reported at the team level and at the organizational level (Figure 1). At the team level, AI has moved beyond curiosity. Most teams are either piloting AI or actively using it for specific tasks (78%; 47 of 60). Only a small fraction report that AI is integrated into regular workflows (8%) or central to most activities (5%). The gap between “active use” and “workflow integration” is the key: teams can benefit without changing how work is produced, reviewed, and governed; integration requires that change.

Cross-functional adoption is even earlier in the curve. The dominant state is still piloting (28 of 60), with fewer reporting active use (14 of 60). Notably, 7 respondents describe cross-functional teams as not using AI tools at all. This matters because value in clinical development often depends on handoffs and shared definitions; if AI cannot travel across those boundaries, it remains local productivity rather than enterprise capability.

At the organizational level, only 22 of 60 (37%) are in “scaling/core priority” for enterprise adoption, while 30 of 60 are still exploring or piloting. This indicates a market that has proven useful but is still building the institutional capacity to deploy safely and repeatedly on a scale.

Figure 1: Distribution of AI adoption and readiness at organizational and team levels.

5. The important linkage: readiness strongly correlates with organizational scaling

A recurring mistake in AI adoption and practice is treating “readiness” as a soft concept. The data suggests they have strong correlation (Figure 2). Among organizations that describe themselves as “very ready,” 9 out of 10 are already at the organizational “scaling/core priority” stage. Among those that are “not ready,” only 1 out of 13 has reached that stage. This hints at a meaningful relationship: scaling is less about intent, and more about whether the enabling foundations are in place.

This linkage helps explain why many organizations remain stuck in pilots. Without governance, secure data access, integration pathways, and evaluation discipline, AI stays in “assistive” modes that are hard to operationalize, especially for work products that carry regulatory or patient-safety implications.

Figure 2: AI readiness strongly correlates with organizational scaling.

6. What teams are doing today: the “tool-layer” is mature; the “system-layer” is not

What is mostly used today reflects practical starting points that can accelerate individual productivity without touching regulated systems. Teams overwhelmingly use tools that require minimal integration: general GenAI assistants (82%), literature review/knowledge extraction (60%), and coding assistants (52%) are the top 3 use cases. These tools deliver immediate benefits because they operate on text/LLM and personal workflows.

The challenge is that in clinical development, critical outputs are not personal artifacts. They are shared, reviewed, and versioned — most of the time between teams — and often traceable back to source data and controlled processes. However, when the tool shifts from “help me draft” to “help me decide” or “help me submit,” the constraints change.

This is reflected in the usage levels cited for more regulated or system-dependent tasks — regulatory document preparation (26%), clinical trial design/optimization (22%), and data quality/anomaly detection (18%) — all sit well below the top three. The interpretation is not that these areas lack opportunity; rather, they demand a system to deliver that most organizations are still building.

Measured impact shows a similar trend. Many respondents either don’t know (30%) or say it is too early to measure (15%). Only 5 report significant gains >30% of impact, and 6 report moderate gains 15–30%. This suggests that benefits are real but not yet consistently tracked, and that large-scale impact likely appears only when AI is embedded into repeatable workflows with clear metrics.

7. Where AI is needed across the R&D value chain: near-term efficiency, long-term reinvention

Respondents’ near-term unmet needs point to a pragmatic agenda. The top 1 need is automating routine data analysis tasks (67%). The #2 is real-time data quality monitoring (48%). These are not independent requests — automation increases throughput, but without strong data quality controls it also increases the throughput of errors, eroding trust and forcing teams back to manual verification. In practice, this means the fastest path to sustainable value is to pair automation with “quality rails.”

Other near-term priorities — regulatory document generation (35%), accelerating drug discovery timelines (35%), literature synthesis (32%), and cross-source data integration (28%) — reinforce a common theme: teams want AI to reduce cycle time, but they need it to do so in a way that can be governed and integrated.

Long-term aspirations (3–5+ years) shift from efficiency to transformation. The top long-term capability is fully automated regulatory submissions (48%), followed by end-to-end drug discovery automation (40%), and AI-powered decision support (33%). These ambitions are coherent and plausible, but they are built on foundations that organizations must first establish: reliable data, controlled pipelines, traceable systems, and validation approaches that are aligned to intended use become the backbone of how AI is delivered, measured, and governed.

8. Deep dive: Biometrics / Clinical Data Science as the scaling “wedge”

The Biometrics and Clinical Data Science (CDS) subgroup (50%; 30 of 60) provides the clearest picture of what it takes to scale if under high requirements of compliance. Their priorities are sharply focused on production work: routine analytics automation (23 of 30) and real-time data quality monitoring (15 of 30) lead by a wide margin. Unlike general productivity use cases, these are directly tied to deliverables that influence trial decisions, oversight, and ultimately submission packages to regulatory agencies for product approvals and surveillance.

Just as important is what drives their confidence. For this group, the strongest confidence driver is quality and reliability of AI outputs (23 of 30), followed by understanding AI capabilities and limitations (21 of 30). Validation and regulatory considerations (16 of 30) and data privacy/security concerns (16 of 30) also rank very high. This profile is telling: for Biometrics/CDS, the question is not whether AI can generate something useful; it is whether the output can be trusted, reproduced, and defended.

Barriers reinforce the same point. Privacy/security (14 of 30) and integration (13 of 30) are top obstacles, but data quality/availability and lack of AI expertise are close behind (10 of 30 each). This indicates that in the highly compliant environments, capability gaps are coupled: even strong models will fail to scale if data access is not controlled, if integration is brittle, or if evaluation and operational ownership are unclear.

A practical implication follows. Biometrics/CDS can be used as a “first domain to standardize,” because the domain naturally forces organizations to build the mechanisms required for enterprise scale: controlled data access, versioned artifacts, review workflows, regression testing for analyses, and auditable provenance from source data to outputs. Once those mechanisms exist, they become reusable across adjacent R&D functions (medical writing, clinical operations analytics, safety analytics), accelerating scale beyond the initial domain.

9. The scaling playbook: build the delivery system, not a collection of pilots

When asked what capabilities are needed to improve AI readiness, the top choices from the survey respondents form a clear blueprint: a clear AI strategy and roadmap (52%), data infrastructure and governance (47%), AI talent acquisition/development (35%), and a regulatory/compliance framework (33%). These are precisely the components that convert experimentation into a managed capability.

This blueprint becomes even more actionable when paired with respondents’ preferred ways of receiving AI solutions. The top delivery preferences are customizable/configurable platforms (48%), API integrations with existing systems (43%), and open-source tools with internal support (43%). Cloud SaaS is still meaningful (33%), but the preference signals a desire for AI that can be embedded into existing enterprise workflows rather than replacing them wholesale.

Vendor selection criteria make the same point from another angle. The top three criteria are domain expertise in biotech/biopharma (57%), regulatory compliance and validation capabilities (50%), and data security/privacy standards (47%). These are not “model-first” criteria. They are “deployment-first” criteria, reflecting that the key risk lies in operationalization and governance, not in generating a plausible answer.

In operating-model terms, organizations that scale tend to do three things early. They define a small number of high-value, high-constraint/compliance domains where governance must be solved (Biometrics/CDS as an archetype). They build shared platform capabilities that make safe deployment repeatable (identity/access, data controls, logging, evaluation, versioning). And they institutionalize accountability through cross-functional ownership — often via an AI team or Center of Excellence (CoE) structure. The survey responses suggest organizations are at different stages and may hold overlapping views, but the direction is clear nonetheless: many are actively building governance structures to enable scale.

10. A potential 12–18 months roadmap

A common failure mode in AI programs is pushing too quickly into “ambitious automation” before the foundations exist to control risk and prove value. The survey provides helpful hints for a potential roadmap with a more reliable sequence.

In the first 0–3 months, the priority should be to turn pilots into reusable assets. This is where organizations define intended-use categories and risk levels, establish secure data access and de-identification patterns, implement logging and audit trails, and put in place evaluation approaches (including regression testing for analytics workflows). Without these pieces, every pilot becomes a bespoke deployment, and scaling costs increase rather than decrease.

In the 4–9 months, the focus should be to operationalize 2–3 “hard value” capability domains that map to the highest unmet needs and naturally require governance discipline. For many organizations, the strongest trio is: real-time data quality monitoring (to protect trust), routine analytics automation (to capture the largest near-term efficiency pool), and document consistency/evidence traceability (to reduce downstream review and submission risk). This combination is powerful because it creates a reinforcing loop: better quality reduces rework, automation increases throughput, and traceability maintains defensibility.

In months 10–18, the work shifts from “build” to “embed and scale.” Here the determinant factor is integration: connecting AI capabilities to the data systems where work happens — EDC, CTMS, eTMF, RIMS, statistical programming platforms, document repositories — and making adoption measurable through standardized metrics. The outcome is not “more pilots”; it is “fewer manual handoffs” and a repeatable production line from data to evidence to deliverable.

11. Closing perspective: winners will differentiate on “trust at scale”

The survey suggests an industry in transition. AI is already part of daily work, but enterprise-level transformation requires governance, compliance, integration, and evaluation disciplines. In this environment, competitive advantage will come less from adopting the newest model, and more from building a delivery system that makes AI safe and repeatable with compliance under regulation.

Biometrics and Clinical Data Science stand out as the most effective starting point because it sits at the intersection of high value and high constraint. If an organization can make AI credible there — where outputs must be defensible, reproducible, and reviewable — it can reuse the same governance and platform components across broader R&D workflows. In practice, that is how AI moves from local productivity to enterprise capability.

References

FDA draft guidance on AI for regulatory decision-making (January 2025). https://www.fda.gov/regulatory-information/search-fda-guidance-documents/considerations-use-artificial-intelligence-support-regulatory-decision-making-drug-and-biological

FDA “Elsa” press release (June 2025). https://www.fda.gov/news-events/press-announcements/fda-launches-agency-wide-ai-tool-optimize-performance-american-people

FDA Agentic AI press release (December 2025). https://www.fda.gov/news-events/press-announcements/fda-expands-artificial-intelligence-capabilities-agentic-ai-deployment

EMA reflection paper on AI in the medicinal product lifecycle (September 2024). https://www.ema.europa.eu/en/use-artificial-intelligence-ai-medicinal-product-lifecycle

ICH E6(R3) GCP information on FDA website. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/e6r3-good-clinical-practice-gcp

McKinsey on scaling GenAI in life sciences (January 2025). https://www.mckinsey.com/industries/life-sciences/our-insights/scaling-gen-ai-in-the-life-sciences-industry

Subscribe now

STATGEN 2026, the third annual conference of the American Statistical Association (ASA) Section on Statistics in Genomics and Genetics (SSGG), was held May 18-20, 2026, at Emory University in Atlanta, Georgia. The conference was organized by ASA SSGG together with the Department of Biostatistics and Bioinformatics at Emory University’s Rollins School of Public Health. Since its launch in 2024, STATGEN has grown into the flagship annual gathering of the statistical genomics and genetics community, bringing together methodologists, applied statisticians, computational biologists, and trainees from academia, industry, and government to present new work, exchange ideas, and build collaborations [1, 2].

STATGEN 2026 drew more than 290 attendees, including over 160 trainees. The strong trainee participation was one of the defining features of the meeting and reflects the section’s sustained investment in the next generation of statistical geneticists and genomicists.

Scientific Program

The three-day program featured three keynote lectures, delivered by Dr. Gina D’Angelo, Dr. Michael P. Epstein, and Dr. Nancy Zhang, spanning statistical methods for genomics, genetic association analysis, and the translation of genomic data into biomedical and therapeutic insight. Dr. Michael Wu delivered the banquet address.

Beyond the keynotes, the program included 16 invited talk sessions, two invited poster sessions, and two panel discussions. The invited sessions covered a broad methodological landscape, from single-cell and spatial omics to genetic epidemiology, multi-omics data integration, and the application of statistical genomics in drug development. The panel discussions gave attendees, and trainees in particular, a candid forum to discuss career paths across academia, industry, and government, as well as emerging directions for the field.

The two poster sessions were a highlight for early-career participants, offering an accessible setting in which students and postdoctoral fellows could present their research and receive direct feedback from senior investigators.

2026 SSGG Founder’s Award

During the conference banquet, Dr. Michael Wu received the 2026 SSGG Founder’s Award in recognition of his significant contributions to statistical genomics and genetics and his leadership and service to the broader community. The Founder’s Award is the section’s highest honor, and the banquet presentation was a fitting occasion for the community to celebrate a colleague whose methodological work and mentorship have shaped the field [3].

Community and Networking

In keeping with STATGEN tradition, the conference again hosted its annual morning jogging and hiking outing, which drew approximately 20 participants to exercise and mingle before the scientific sessions began. Informal activities such as this one, along with the poster receptions and the banquet, are a deliberate part of the STATGEN design: they lower the barrier for trainees to meet established researchers and help sustain the collegial culture that has characterized the meeting since its inception.

Acknowledgments

The organizers gratefully acknowledge the ASA Biopharmaceutical Section (BIOP) for its sponsorship of STATGEN 2026 [4]. BIOP’s support helped make the conference accessible to a large trainee cohort and strengthened the connection between the statistical genomics community and the biopharmaceutical statistics community, a link that continues to grow in importance as genomic data play an expanding role in target discovery, patient stratification, and clinical development. We also thank Emory University’s Department of Biostatistics and Bioinformatics for hosting the meeting, and the speakers, panelists, session organizers, volunteers, and attendees whose contributions made STATGEN 2026 a success.

The STATGEN 2026 Program Committee:

• Stephanie Hicks (Chair; Department of Biostatistics, Johns Hopkins Bloomberg School of Public Health)

• Eric Lock (Division of Biostatistics, University of Minnesota)

• Kui Shen (Senior Director, Clinical Statistics, Bayer)

• Zhaohui “Steve” Qin (Department of Biostatistics and Bioinformatics, Emory University)

Looking Ahead

The location and planning details for STATGEN 2027 will be announced in the summer of 2026. Colleagues in the biopharmaceutical statistics community with interests in genomics, genetics, and precision medicine are warmly encouraged to watch for the announcement and to consider submitting an abstract. Please stay tuned!

References

[1] STATGEN 2026: Conference on Statistics in Genomics and Genetics. Emory University, Atlanta, GA, May 18-20, 2026. https://statgen26.emory.edu/

[2] American Statistical Association, Section on Statistics in Genomics and Genetics (SSGG). STATGEN Conference. https://community.amstat.org/sectiononstatisticsingenomicsandgenetics/statgen-conference

[3] American Statistical Association, Section on Statistics in Genomics and Genetics (SSGG). Founder’s Award. https://community.amstat.org/sectiononstatisticsingenomicsandgenetics/founders-award

[4] American Statistical Association, Biopharmaceutical Section (BIOP). https://community.amstat.org/biop/home

One of the most important yet often under-appreciated aspects of clinical trials is the Data Monitoring Committee (DMC) or Data and Safety Monitoring Board (DSMB) as known in many NIH circles. A DMC is a group of ~3-7 individuals who review accumulating clinical trial data by treatment group in order to protect trial: (1) participants, and (2) integrity and credibility. The DMC accomplishes this by monitoring patient safety and efficacy, making a benefit/risk assessment of trial continuation, maintaining independence from study sponsors and investigators, and ensuring confidentiality of interim trial results. The DMC enhances the scientific integrity of the clinical trial, because it is the only entity with access to aggregate trial data by unblinded treatment assignment, which is required for comprehensive understanding of emerging treatment effects, whether beneficial or harmful. This allows trial sponsors and study staff to remain blinded to ongoing trial results, a key to protecting trial credibility and integrity.

Scientific and operational leadership from statisticians is crucial in the DMC process. Statisticians often have the best understanding and appreciation of many of the relevant issues: error control and associated statistical methods, protection from integrity compromises arising from operational biases induced by data leaks, and the importance of thoughtful and informative data presentation so that DMCs can understand the data signals.

I have sat on more than 100 DMCs. It is one of my favorite clinical trial activities and is a most rewarding and educational experience. I get to see many types of trial designs, in many disease areas, and discuss emerging and sometimes challenging-to-interpret data trends, with a small group of disease-area experts. I uniquely contribute to the data interpretation and the discussion given greater knowledge of the issues above and generally having greater DMC experience than the non-statistician DMC members that generally only serve in specific disease areas of expertise. It is a particularly important contributory role for academics given that people in the government and industry sectors are often unable to serve in this important role.

The importance of the DMC role cannot be overstated. Stop a trial too soon, and the trial is inconclusive and fails to obtain answers to important questions that inform clinical practice. Stop a trial too late, and participants are exposed to potentially harmful or ineffective interventions, which can be either the investigational therapy or the current standard of care, longer than necessary. The benefits of obtaining convincing and conclusive evidence and the ethical responsibility to current and future patients are weighed carefully during DMC discussions.

I have the privilege and the enjoyment of being a card-carrying member of the “DMC Mafia”, a term coined by Dave DeMets for a small group of experienced statisticians and a few clinicians that are trying to improve the integrity of the DMC process and decision-making through education and training. A group of them consisting of Dave DeMets, Susan Ellenberg, Tom Fleming, Frank Rockhold, and Janet Wittes recently discussed areas in need of improvement at the 2025 Annual Meeting of the Society for Clinical Trials. In practice, we see: DMC reports that are indigestible and disorganized, violations of independence in the form of restricting DMC access to important data, and the need for a larger pool of experienced people that can serve as DMC members. In that spirit, here is a top-ten list of areas, in no particular order, where improvement is needed.

• Decisions: stopping boundaries are guidelines

  • The decision to stop a trial is more complex than results on a single endpoint. A DMC cannot leave unanswered questions on the table. Even if a boundary is crossed, additional information may be needed on other endpoints, components of composites, or safety in order to understand the data. See a detailed discussion see: Ge, L., Hamasaki, T. & Evans, S.R. Inside the Mind of the DMC: A Review of Principles and Issues with Case Studies. Ther Innov Regul Sci 59, 234–244 (2025). https://doi.org/10.1007/s43441-024-00720-8.

• Respect for independence

  • In some cases, DMCs are not provided access to all important data. DMCs require access to all unblinded data throughout the trial to allow the most learned judgments about the benefits and harms of the interventions under study as part of the ethical and fair treatment and protection of all trial participants, regardless of assignment to an investigational therapy or control. An advantage for one therapy is a disadvantage for the other. This ensures the most informed recommendations. Restricted access exhibits control over the DMC violating independence.

  • Data should remain strictly confidential while trials are ongoing with access to interim results limited to the DMC. Formal interim testing is not justification for results to be released beyond the DMC. The DMC should not report results of tests or whether boundaries have been crossed unless an invasive recommendation is made. Pressures to provide information about interim trial results in DMC recommendations put the integrity of the clinical trial at risk and violate independence, confidentiality, and public trust.

  • Do not limit recommendations e.g., to 3-4 checkboxes to avoid leading the DMC and ensure independence.

  • Ensure that presented analyses are independent of investigator beliefs. DMC assessment needs to be objective and independent.

• Meetings

  • Return to in-person meetings when formal interim testing is planned or complex discussions are needed. The quality of DMC interactions and evaluations is optimized in focused meetings where data can be interpreted and discussed in depth rather than on just another call.

  • Closed meetings

    • Use consensus recommendations rather than voting.

    • Avoid recording closed session meetings.

• Charters

  • Simplify: often too long and legalistic.

  • Avoid restrictions regarding access to data.

  • Avoid restrictions to DMC recommendations.

  • Make charters public for transparency as with other important trial documents such as the protocol.

• Statistical representation on the DMC

  • Have at least two statisticians on DMCs with >3 members. Some sponsors and DMC chairs do not want any statisticians!

• Independent statistician

  • Be engaged and knowledgeable of the trial and the data.

  • Seek intuition regarding needs and concerns of the DMC.

  • Proactively address questions arising from the data e.g., unexpected events.

  • Be responsive to DMC requests.

• DMC reports

  • Issues

    • Often too voluminous and indigestible with lengthy tables and listings.

    • Lack of appropriate background, organization, and text.

    • Void or limited with regard to figures that are more effective at displaying data trends and outliers.

    • Inflexibility in responding to DMC requests for additional data summaries.

  • Suggestions

    • Comprehensive but comprehensible; digestible and concise.

    • Transition from TFL to FTL (figures, tables, and listings) increasing the priority and utilization of thoughtful graphical summaries as visual summaries of data trends are often easier to consume and interpret.

    • ITT focus: only ITT preserves the benefits provided by randomization regardless of whether an endpoint is characterized as one of efficacy or safety. DMCs are protecting the welfare of all randomized participants to ensure equipoise and ethical assignment of therapy, not limiting to those that adhere nor to those on investigational therapy.

    • Flexibility to adjust data summaries due to questions arising from the data or to be responsive to DMC requests.

    • Include:

      • Table of contents for navigation

      • Study synopsis

      • Minutes from prior meeting

      • Executive summary noting important results

      • Include text descriptors and summaries: facts not opinions

      • Figures

        • CONSORT disposition flow diagram

        • Forest plots risk difference or DOOR probability AEs and key variables: use an absolute scale

        • Patient profiles for trials with small N (<10-20) or for special cases of interest during review

      • Appendices for voluminous information that can be perused as needed

        • Charter

        • Long tables and listings

• Indemnify DMC members

• Simplify contracts

• Relevant expertise for DMC members and others involved with the DMC process

  • Ensure that trial integrity is protected and that those involved have an awareness to statistical and operational biases and how to prevent them.

  • Train and educate: Educational materials are available:

    • See an 8-article collection in Therapeutic Innovation & Regulatory Science: Data Monitoring Committees: Issues and Myths all Trial Sponsors and Vendors Should Know, edited by Frank Rockhold and Dave DeMets.

    • See a 6-article mini-series series in NEJM Evidence edited by Scott Evans.

  • Support an apprentice mentee to participate in DMC meetings.

  • Consider training sessions at professional meetings.

It is an ethical imperative to: (i) protect the welfare of clinical trial participants, and (ii) maximize the value of the clinical trial for patients and society by ensuring that clear and convincing evidence is obtained through high integrity clinical trials and good science. DMCs play a pivotal role in attaining these goals and advancing the evaluation of medical interventions by weighing the benefits and harms of treatments and therapeutic alternatives, in the face of complexities associated with clinical trial data. Unfortunately flawed and suboptimal practices associated with the DMC process are commonplace. DMC member training, sponsor education about important fundamental issues, and improved DMC operations will help ensure that ethical and scientific obligations are met.

Subscribe now

Disclaimer: This article is adapted from an RISW (Regulatory-Industry Statistics Workshop) 2025 talk where I shared what I have learned about career growth so far. This is a reflection on my career. Your career journey is different. Hopefully, this will be informative and add to the data needed to drive your career in the right direction.

The Career We Often See Is Not the Whole Career

When looking at someone else’s career from a distance, it can appear surprisingly orderly. Imagine pulling up LinkedIn and seeing the job title changes. Each promotion appears as if someone took a direct path from one role to the next. The bullets summarize the change in responsibilities, scope, or management. From the outside, it is easy to imagine that the person had a clear progression in what they were doing and it was as easy to follow as the voice in your car’s GPS. Most careers are more meandering than that. The distance from the start of your career to where you are today is usually not the shortest path GPS would direct you on. But that is what makes careers. Each one includes uncertainty, setbacks, changes that you didn’t want, and figuring out how to overcome the obstacles that appear. Most importantly, many of the moments that lead to growth are things that make the journey hard. If you never step into an unfamiliar or difficult situation, you might deny yourself the opportunity to find new strengths to fuel your career journey.

Don’t let comparison, the thief of joy, rob you as you are looking at your career journey. Your journey will be about you. Others’ careers will serve as opportunities for reflection and conversations around how to grow. If you discuss careers with others, talk about hard-won skills, new statistical and managerial techniques that were needed, and how each career has ups and downs that lead to new horizons. The most useful career growth often happens in the parts of the job that aren’t easily visible.

Careers are Not Professional Highlight Reels

One of the struggles of our current society is social media. While it connects many people, it often shows the highlights and not the day-to-day. You might see a beautiful picture of a single person on the Great Wall of China, walking the Champs-Élysées or in Santorini. They have the whole place to themselves. Except the crowds have mainly been eliminated by clever camera angles and AI. Don’t fall for the Instagramification of life. If you pulled up my LinkedIn profile, it would look like I went from a Senior Biostatistician role to a Vice President role without many hiccups along the way. That is the Instagram story. If you looked at my personal job satisfaction, how I have used innovation, the number of people I have managed, the impact on potential patients and workplace recognition, you would see something that was very non-linear. The graph would look like a highly variable stock moving up and down at random. So, remember to talk to people about what has made their career meaningful instead of telling yourself a story about their career based on LinkedIn.

I have had a group as small as three statisticians reporting to me and one as large as 90. I currently have 30+ statisticians in my group. The least interesting thing about that experience is the number in the group. More interesting is what they have taught me, how I have learned to manage others, what I do to delegate and all the lessons my direct reports teach me on how to work better and smarter. It has also given me room to grow new skills and shown me that I need to leave room to change in the future. If you want to maximize your career, look to the lessons of life over the story of LinkedIn. Because LinkedIn is just a professional version of Instagram.

LinkedIn curates only a portion of a career. It leaves gaps in the story that you can fill in with your imagination. I often think that everyone else is doing more than me and is better than me. My inclination is to assume that it was all smooth sailing for others. But what if the gaps are the hard part? What if the stories that you would tell yourself remove what makes a career meaningful? You can easily show a promotion on LinkedIn. But how do you highlight the tradeoffs you made for that promotion and why you did it?

This distinction matters because comparison can distort how you evaluate yourself. If you compare the inside of your career to the outside of someone else’s, you will almost always be unfair to yourself. You know your own doubts, setbacks, and unfinished skills. You do not see those same details in others’ stories. That can make your path feel slower, less impressive, or less coherent than it really is. And this comparison distracts you from the most important question that you can ask yourself: Am I growing in meaningful ways? While growth could be a new title, real growth often comes from a better understanding of the business, wiser judgments, confidence in decision-making, more technical understanding, broader influence, and better communication skills. All these tools will help you become a better statistician and leader in any organization.

Let’s stop talking about a career ladder and train ourselves to think about careers as lattices with many paths to an optimal solution. There are many ways to optimize your journey. It could be the number of reports, it could be impact on project teams, it could be statistical skills. Each one of those goals will have a different path for your career and its own way to optimize it. Focus on making your career the best it can be for you with your growth centered in that conversation.

How Should We Measure Our Careers?

Measuring a career by a title tends to mask what has been learned. I like to think of other ways to measure a statistician’s career in these areas: Technical Skill, Communication, Creativity, Collaboration, and Management. Each category brings different ideas to the forefront of career growth. And this list is not exhaustive. As I look at the ways that I have grown, I can see each of these groups has given me a chance to learn in different ways. While I have not grown equally in each of these areas, I have seen how my technical expertise, ability to communicate and work as a team, the chances I have taken to put forward a new solution and what I have learned as a manager tell a more complete story of my career.

Technical excellence is the baseline for a pharmaceutical statistician. You won’t stay in a role long without that expertise, but how do you want to grow it? Some people are technical wizards who lean into the statistical side of each problem. Others might focus on a single statistical area or question. There are plenty of ways to grow your career based on technical expertise. You will need your technical skills to survive, so find ways to invest in them and grow in the areas that resonate with you.

Communication skills are essential since most pharmaceutical teams are cross-functional. You will likely be the only statistician on many teams, and some teams might not have other quantitative functions. The ability to explain technical ideas to team members or senior management becomes a clear advantage. Convincing a non-statistical audience of the importance of statistical considerations is a path to differentiation. This could be how you speak in small groups or present in larger meetings. This opens different career doors. How well a statistical idea is communicated often influences the impact of that idea because you can convince others to help bring it to life.

Creativity also plays a role in statistics. Some studies are nearly identical, but many have their own needs. This can be reflected in the experimental design, types of study arms, analyses or how conclusions will be presented. It might also be research into new statistical techniques. You can carve out a place for yourself in many organizations if you bring new ideas to the table. There isn’t a perfect development plan, but there are many ways to solve the problem. Don’t be afraid to tap into your creativity for your projects!

Collaboration is separate from communication. Real collaborators on a team understand the demands put on different disciplines. They can articulate why a statistical problem might be the key to this study and why a drug supply or clinical pharmacology problem might be more important for the next. Real collaborators understand the perspective of others, understand that clinical development is a multivariate optimization problem and advocate for the function that needs their problem solved in the current study. Every team I have been on is looking for individuals who understand different perspectives. It is a great way to impact teams.

Lastly, management and leadership add a different dimension to careers. I put this last because many think it is the apex of a career to manage others. I believe that the apex of many careers is leadership growth. You don’t have to manage a single person to lead. You can lead your teams through statistical influence, making an impact on decisions and building a cohesive culture for teams. Leadership does not equal management. Management is not for everyone. I hope that those chasing management opportunities are also chasing management skills. Do you know how to delegate and build trust with your reports? Are you willing to have hard conversations and give unwanted feedback? Are you the one who makes a tough decision and answers for any possible outcome? Do you think carefully about the individuals’ careers that you are guiding? These skills can be developed, and you can learn from those around you. Don’t be afraid to grow in this area if you want to manage others. And please make sure you are growing your leadership irrespective of managing.

Be bold in your career choices and goals. It is great to have ambition. Just make sure the goals that you are chasing have the full lens of options and opportunities and that you are building the skills to succeed in them. If you want to grow the number of people in your group, how are you growing your managerial expertise? If you want to be promoted as an individual contributor, how are you doing better at communicating and influencing a team with your technical skills? Take career growth through a multi-dimensional lens without projecting everything back to a one- or two-dimensional space.

Self-Awareness is a Force Multiplier

Your optimal career will likely be found when you are learning about yourself. Be curious in life and especially curious about yourself. What do you think about outside of work? Are you calm in stressful situations? What have you learned from your life circumstances (volunteering, family life, etc.)? What do you think about when other statisticians present? What do you learn when watching other leaders? How do you receive both positive and negative feedback? Do you listen or are you dismissive? Do you build new skills from the feedback you receive? Each one of these questions will help you understand something more about you. Who you are outside of work and where your interests lie will shape who you are inside your job and what the best options for career growth might be.

If you want to grow, become a careful student of who and what you admire. Notice the lessons you have learned as a volunteer and how those lessons rhyme with situations you might encounter at work. Notice leaders who give credit generously, handle pressure calmly, or make complicated topics understandable. Figure out which of these things are innate to you and how you can adapt them to your style. I believe this is a more direct way to find fulfillment in your career because it will be authentic and you will form your own career and leadership model.

Lessons you can apply to your career are everywhere. Asking yourself what you can learn about your career in many situations will help you be ready to expand your influence or role the next time you are faced with challenges. Engaging in life outside of work can prepare you to be better at work if you are self-aware. These experiences may not appear on a resume, but they can shape the person you are on the job.

Careers are a Journey

Your career is your own journey. Take that journey. Be multidimensional. Stay curious. Build skills that work in any role. Expand your view of the lattice in front of you. Ask others to help you. Figure out what motivates you. This will expand your career. It has expanded mine.

Highlights

• Healthcare decision-makers increasingly require timely, relevant, and actionable evidence throughout the product lifecycle, driving greater integration of diverse data sources beyond traditional clinical trials.

• Real-world data (RWD), causal inference methods, and advances in artificial intelligence (AI) offer significant opportunities to generate evidence that better reflects clinical practice, while also introducing important methodological challenges.

• Credible, decision-relevant evidence depends on rigorous study planning, including clear research questions, assessment of data fitness, appropriate study design, estimand definition, causal inference strategies, and sensitivity analyses.

• Evidence-generation approaches should be aligned with the needs of regulators, health technology assessment (HTA) bodies, payers, clinicians, and other healthcare stakeholders.

• High-quality evidence is driven by scientific rigor, sound study design, and transparent assumptions; sophisticated analytics alone are not sufficient.

• AI, machine learning, and natural language processing can improve efficiency, scalability, and the use of unstructured data, but they do not replace causal reasoning and robust statistical thinking.

• The future of biopharmaceutical statistics lies in integrating randomized trials, real-world evidence, causal inference, digital health technologies, and AI within a unified evidence-generation framework.

• Statisticians are evolving from data analysts to evidence architects, connecting scientific questions, data, methodology, technology, and stakeholder needs to support trustworthy and impactful healthcare decision-making.

  ---

1. Introduction

The practice of biopharmaceutical statistics is undergoing a profound transformation. For decades, randomized controlled trials (RCTs) have been regarded as the gold standard for generating evidence on the efficacy and safety of medical interventions. While randomized trials remain fundamental to drug development and regulatory decision-making, the increasing availability of real-world data (RWD), advances in causal inference methodologies, and the rapid emergence of artificial intelligence (AI) are expanding both the opportunities and responsibilities of statisticians. (FDA, 2018; Hernán & Robins, 2020)

Today, decision-makers across the healthcare ecosystem—including regulators, health technology assessment (HTA) bodies, payers, healthcare providers, and patients—are seeking evidence that is not only scientifically rigorous but also clinically relevant, timely, and reflective of real-world practice. As a result, statisticians are increasingly asked to move beyond traditional trial-based paradigms and integrate information from diverse data sources to support decisions throughout the product lifecycle.

In this evolving environment, the role of the statistician extends far beyond data analysis. Statisticians must ensure that evidence generated from real-world settings is fit-for-purpose, scientifically credible, and capable of informing high-stakes decisions. The emergence of AI further amplifies both the potential and the complexity of evidence generation, making rigorous statistical thinking more essential than ever.

2. The Growing Importance of Real-World Evidence

Real-world evidence (RWE) has emerged as a critical complement to evidence generated from randomized clinical trials. Derived from sources such as electronic health records, claims databases, disease registries, patient-generated data, and digital health technologies, RWD offers unique opportunities to evaluate treatment effectiveness, safety, utilization patterns, and healthcare outcomes in routine clinical practice. (FDA, 2018; Fang Y, et al., 2020)

The value of RWE stems from its ability to address questions that may be difficult or impossible to answer through traditional trials alone. Clinical trials often involve highly selected patient populations, controlled treatment settings, and relatively limited follow-up periods. In contrast, RWE can provide insights into treatment performance among broader and more heterogeneous populations, including those frequently underrepresented in clinical trials, such as elderly patients, individuals with multiple comorbidities, and diverse demographic groups.

However, the availability of data alone does not guarantee meaningful evidence. A fundamental principle is that evidence quality depends not only on the quantity of data but also on the appropriateness of the data for the research question. Large datasets may still be unsuitable if critical variables are missing, poorly measured, or inconsistently captured. Consequently, statisticians must carefully evaluate data fitness, ensuring alignment between the research objective, the available data, and the proposed analytic approach.

3. Starting with the Right Research Question

High-quality evidence generation begins with a well-defined research question. Before selecting data sources or statistical methods, investigators must clearly specify the decision problem, target population, treatment strategies, outcomes of interest, and relevant stakeholders. (Hernán & Robins, 2020; Ho et al., 2021, Fang et al., 2023)

Unfortunately, many observational studies begin with available data and subsequently search for questions that can be answered. This approach often leads to analyses of limited relevance or questionable interpretability. Instead, evidence generation should adopt a decision-focused framework that starts with understanding what decision needs to be informed and what evidence is required to support that decision.

This principle has become increasingly important in both regulatory and HTA settings. While these stakeholders may share common interests, they frequently differ in evidentiary requirements, decision criteria, and tolerance for uncertainty. Careful formulation of the research question helps ensure that generated evidence remains relevant to the intended context of use.

4. From Data to Credible Evidence

Once a research question has been established, the next challenge is determining whether available data are capable of supporting valid inference. (FDA, 2018; Berger et al., 2017; Levenson et al., 2023, He et al., 2023)

Data fitness assessments should examine multiple dimensions, focusing on data reliability and relevancy that include:

• Completeness of key variables

• Accuracy of treatment and outcome measurement

• Population representativeness

• Adequacy of follow-up

• Availability of important confounders

• Temporal alignment between exposures and outcomes

• Data provenance and quality controls

Equally important is the selection of an appropriate study design. While randomized trials benefit from random treatment allocation, observational studies require explicit strategies to address potential biases arising from confounding, selection mechanisms, missing data, and measurement error.

Study design choices—including cohort studies, case-control studies, self-controlled designs, target trial emulation, and pragmatic clinical trials—should be driven by the scientific question rather than convenience. The strongest observational studies increasingly seek to emulate hypothetical randomized trials, thereby making assumptions more transparent and reducing opportunities for bias.

5. The Central Role of Estimands

One of the most important advances in modern biopharmaceutical statistics has been the emphasis on estimands. Estimands provide a precise description of the treatment effect being estimated, linking scientific objectives with study design, data collection, and statistical analysis. (ICH E9(R1), 2019; Chen et al., 2024)

In real-world settings, ambiguity regarding treatment effects can easily arise. Questions such as whether interest lies in treatment initiation, treatment persistence, treatment switching, or real-world use patterns can substantially alter the interpretation of results.

Clearly defining estimands helps ensure alignment across stakeholders and minimizes confusion regarding what the evidence actually represents. Moreover, estimand thinking encourages investigators to explicitly address intercurrent events, treatment discontinuation, adherence patterns, and competing risks—factors that are often highly relevant in routine clinical practice.

6. Causal Inference: The Bridge Between Data and Decisions

As RWE becomes increasingly influential, causal inference has emerged as one of the most important methodological developments in biopharmaceutical statistics. (Hernán & Robins, 2020; Ho et al., 2021)

Traditional statistical methods often focus on identifying associations. However, healthcare decision-makers are rarely interested in association alone. Instead, they seek answers to causal questions:

• What would happen if a patient received treatment A instead of treatment B?

• How much benefit can be attributed to treatment?

• What is the impact of earlier intervention?

• What outcomes would have occurred under alternative treatment strategies?

Causal inference provides a principled framework for addressing these questions. Techniques such as propensity score methods, inverse probability weighting, marginal structural models, instrumental variable approaches, g-methods (Hernán & Robins, 2020), targeted learning (Van der Laan et al., 2011), and target trial emulation (Hernán et al., 2016, Wang et al., 2023) are increasingly used to estimate causal effects from observational data.

Importantly, causal inference is not simply a collection of statistical techniques. It begins with explicit assumptions regarding treatment assignment mechanisms, confounding structures, and causal pathways. Careful articulation of these assumptions allows stakeholders to evaluate the credibility of resulting conclusions.

Sensitivity analyses play a particularly important role. Because causal conclusions often depend on assumptions that cannot be fully verified, investigators should routinely assess the robustness of findings to alternative assumptions, unmeasured confounding, model specification choices, and potential sources of bias.

7. The Emergence of AI in Evidence Generation

Artificial intelligence is rapidly transforming the evidence-generation landscape. Machine learning, natural language processing (NLP), large language models, and other AI-enabled methods offer unprecedented opportunities to leverage complex and previously inaccessible data sources. (FDA AI/ML Discussion Paper, 2023/2025)

AI technologies can improve efficiency across multiple stages of the evidence-generation process:

• Identifying eligible patients

• Extracting information from unstructured clinical notes

• Classifying outcomes and adverse events

• Predicting disease progression

• Detecting data quality issues

• Automating literature review and evidence synthesis

• Supporting protocol development and study execution

These capabilities are particularly valuable because much of healthcare information resides in unstructured formats. NLP techniques can convert clinical narratives, pathology reports, radiology reports, and physician notes into analyzable data, substantially expanding the usable evidence base.

Yet the excitement surrounding AI must be balanced with caution.

8. Why AI Cannot Replace Statistical Thinking

While AI methods can identify highly complex patterns within large datasets, they do not inherently solve fundamental challenges related to causal inference, study design, bias, or decision-making. A highly accurate prediction model may still generate misleading conclusions regarding treatment effects because prediction and causal inference address fundamentally different scientific questions (Hernán & Robins, 2020; Ho et al., 2021). In many healthcare decisions, the key question is not what is likely to happen, but rather what would have happened under an alternative treatment strategy. Answering such counterfactual questions requires explicit assumptions, careful study design, and causal reasoning that extend beyond algorithmic performance alone.

AI models are also highly dependent on the quality and representativeness of the data used for training and validation. Large datasets do not automatically translate into reliable evidence. Fit-for-purpose data sources, missing information, measurement error, selection bias, and inadequate capture of important confounders can lead AI systems to learn spurious relationships that do not reflect true causal effects. As emphasized in recent work on fit-for-purpose RWD, rigorous assessment of data relevance, completeness, provenance, and suitability remains essential regardless of the sophistication of the analytic methods applied (Levenson et al., 2023; He et al., 2023).

Another limitation is that AI systems generally optimize prediction accuracy rather than alignment with a clearly defined scientific estimand. Without careful specification of the research objective, target population, treatment strategies, outcomes, and handling of intercurrent events, it may be unclear what quantity is actually being estimated. Defining estimands and aligning analyses with decision-relevant questions remain fundamentally statistical activities that require scientific judgment and stakeholder engagement (ICH E9(R1), 2019; Chen et al., 2024).

Several concerns warrant particular attention:

• Lack of transparency and explainability

• Limited interpretability of complex models

• Reproducibility and model governance challenges

• Algorithmic bias and fairness concerns

• Data drift and model degradation over time

• Generalizability across healthcare settings and populations

• Regulatory acceptability and validation requirements

• Difficulty distinguishing association from causation

These concerns become particularly important in regulatory and HTA decision-making, where stakeholders must understand not only the results, but also the assumptions, biases, uncertainties, and limitations underlying those results. Regulatory decisions require evidence that is scientifically credible, reproducible, and explainable. Black-box predictions alone are rarely sufficient for high-stakes decisions involving patient outcomes, product approval, reimbursement, or clinical practice (FDA AI/ML Discussion Paper, 2023/2025).

For these reasons, AI should be viewed as a powerful complement to—not a replacement for—rigorous statistical methodology. AI can greatly enhance evidence generation by extracting information from unstructured data, automating routine analyses, identifying complex patterns, and improving operational efficiency. However, statistical thinking remains essential for defining research questions, establishing causal frameworks, evaluating data fitness, selecting appropriate study designs, assessing uncertainty, conducting sensitivity analyses, and interpreting findings in a decision-making context. The future will likely belong not to statisticians versus AI, but to statisticians who effectively leverage AI while maintaining scientific rigor, transparency, and causal reasoning (Pearl & Mackenzie, 2018).

9. The Future of Biopharmaceutical Statistics

The future of biopharmaceutical statistics will be defined by integration. Rather than viewing randomized trials, RWE, causal inference, digital health technologies, innovative study designs, and AI as separate disciplines, the field increasingly requires a coherent framework that brings these approaches together to support evidence generation and decision-making across the entire product lifecycle.

One emerging trend is the gradual convergence of clinical research and clinical practice. Historically, evidence generation occurred primarily within the boundaries of controlled clinical trials, while healthcare delivery operated separately. Increasing availability of electronic health records, registries, digital biomarkers, wearable devices, and patient-generated data is blurring this distinction. Future evidence ecosystems may support continuous learning, where data generated during routine care contribute to ongoing assessment of treatment effectiveness, safety, and value. In such learning health systems, statisticians will play a critical role in ensuring that evidence generated from real-world settings remains scientifically credible and decision-relevant.

Another important development is the growing emphasis on decision-focused evidence generation. Future stakeholders will increasingly ask not whether a treatment works under ideal conditions, but whether it delivers meaningful value for specific patients, healthcare systems, and populations. As a result, statisticians will need to move beyond hypothesis testing alone and become experts in defining decision problems, articulating estimands, quantifying uncertainty, and evaluating trade-offs among benefits, risks, costs, and societal outcomes. The integration of regulatory and HTA evidence requirements may further accelerate the need for unified evidence-generation strategies that support multiple decision-makers simultaneously.

AI will continue to transform the evidence-generation process, but its greatest impact may not be replacing statistical analyses. Instead, AI is likely to augment nearly every stage of study design, data acquisition, variable construction, evidence synthesis, and communication of results. Large language models and other emerging technologies may help automate routine analytical tasks, facilitate protocol development, identify evidence gaps, and accelerate evidence synthesis. However, as AI becomes more pervasive, the value of statistical expertise may increase rather than diminish. Human judgment will remain essential for determining whether the right question is being asked, whether the available data are fit-for-purpose, whether assumptions are plausible, and whether conclusions are supported by the evidence.

The future may also witness the emergence of digital twins, synthetic control populations, continuously updated external control arms, federated data networks, and adaptive evidence-generation platforms. These innovations have the potential to substantially improve efficiency and accelerate access to evidence. However, they will also raise important methodological and regulatory questions regarding data quality, robustness, transportability, privacy, fairness, and reproducibility. Statisticians will be expected not only to develop novel methodologies but also to establish principles and standards that ensure these innovations are implemented responsibly.

Perhaps most importantly, the future statistician will increasingly serve as an evidence architect rather than solely an analyst. Success will require expertise spanning statistics, epidemiology, causal inference, data science, machine learning, health economics, and decision science. Equally important will be the ability to collaborate across disciplines and communicate complex methodological concepts to regulators, clinicians, payers, patients, and policymakers.

As healthcare decisions become more complex and time-sensitive, the demand for trustworthy evidence will continue to grow. The challenge is not merely generating more evidence, but generating evidence that is relevant, credible, transparent, explainable, and actionable. The statisticians who thrive in the future will be those who can combine scientific rigor with innovation, leveraging new data sources and technologies while preserving the fundamental principles that underpin reliable inference.

Ultimately, the future of biopharmaceutical statistics is not about choosing between randomized trials and RWE, between causal inference and machine learning, or between statisticians and AI. It is about integrating these complementary approaches within a unified framework that supports timely, robust, and trustworthy decisions—helping ensure that patients, clinicians, regulators, HTA bodies, and payers can make better decisions based on the best available evidence.

10. Conclusion

Biopharmaceutical statistics is entering a new era in which evidence generation extends far beyond the confines of traditional clinical trials. The convergence of real-world evidence, causal inference, digital health technologies, and artificial intelligence offers unprecedented opportunities to improve healthcare decision-making across the product lifecycle.

Yet the future of the discipline will not be defined by data volume, computational power, or increasingly sophisticated algorithms alone. It will be defined by our ability to transform those capabilities into trustworthy evidence that can guide meaningful decisions. As healthcare systems become more interconnected and data-rich, the fundamental challenge is no longer access to information—it is determining which evidence is credible, relevant, and actionable.

In this evolving landscape, statisticians will play a more important role than ever before. The profession is transitioning from its traditional focus on study analysis toward a broader responsibility for designing evidence ecosystems that connect scientific questions, data sources, methodologies, technologies, and decision-makers. Future statisticians will increasingly serve as evidence architects—integrating randomized and real-world evidence, combining causal inference and machine learning, and ensuring that innovation remains grounded in scientific rigor.

Artificial intelligence will undoubtedly reshape how evidence is generated, analyzed, and communicated. However, AI will be most powerful when guided by sound statistical principles, thoughtful study design, and a clear understanding of the decisions that evidence is intended to inform. The goal is not to replace human judgment, but to augment our ability to learn from increasingly complex data and to answer increasingly complex questions.

Looking ahead, the greatest opportunities may emerge from the integration of historically separate domains: clinical research and clinical practice, efficacy and effectiveness, regulatory and HTA evidence, prediction and causal inference, human expertise and artificial intelligence. Statisticians are uniquely positioned to help bridge these domains and create a more continuous, efficient, and reliable evidence-generation paradigm. Ultimately, the future of biopharmaceutical statistics is not about choosing between randomized trials and real-world evidence, between causal inference and machine learning, or between statisticians and AI. It is about bringing these complementary approaches together within a unified framework that supports timely, robust, transparent, and trustworthy decisions.

In the coming decade, the most valuable statisticians may not be those who build the most sophisticated models, but those who can most effectively connect scientific questions, diverse data sources, methodological rigor, and emerging technologies to generate evidence that improves decisions and ultimately advances patient care. If statistics has always been the science of learning from data, its future may be even more ambitious: enabling a continuously learning healthcare ecosystem that delivers better outcomes for patients, providers, regulators, payers, and society as a whole.

Subscribe now

References

· Berger ML, Sox H, Willke RJ, et al. Good practices for real-world data studies of treatment and/or comparative effectiveness. Value Health. 2017;20(8):1003-1008.

· Chen J, Scharfstein D, Wang W, Yu B, Song Y, He W, et al. Estimands in real-world evidence studies. Stat Biopharm Res. 2024.

· Fang Y, Wang H, He W. A statistical roadmap for the journey from real-world data to real-world evidence. Ther Innov Regul Sci. 2020.

· Fang Y, He W. Key considerations in forming research questions and conducting research in real-world settings. In: He W, Fang Y, Wang H, eds. Real-World Evidence in Medical Product Development. Springer; 2023.

· U.S. Food and Drug Administration. Framework for FDA’s Real-World Evidence Program. FDA; 2018.

· U.S. Food and Drug Administration. Using Artificial Intelligence and Machine Learning in the Development of Drug and Biological Products: Discussion Paper and Request for Feedback. FDA; 2023. Revised 2025.

· He W, Zhang Z, Dharmarajan S. Assessment of fit-for-use real-world data sources and applications. In: He W, Fang Y, Wang H, eds. Real-World Evidence in Medical Product Development. Springer; 2023.

· He W, Fang Y, Wang H, Chan I. Applying quantitative approaches in the use of RWE in clinical development and life-cycle management. Stat Biopharm Res. 2021.

· He W, Fang Y, Wang H, Lee C. The need for real-world evidence in medical product development and future directions. In: He W, Fang Y, Wang H, eds. Real-World Evidence in Medical Product Development. Springer; 2023.

· Hernán MA, Robins JM. Causal Inference: What If. Chapman & Hall/CRC; 2020.

· Hernán MA, Robins JM. Using big data to emulate a target trial when a randomized trial is not available. Am J Epidemiol. 2016;183(8):758-764.

· Ho M, van der Laan M, Lee H, Chen J, Lee K, Fang Y, He W, et al. The current landscape in biostatistics of real-world data and evidence: causal inference frameworks for study design and analysis. Stat Biopharm Res. 2021.

· International Council for Harmonisation. ICH E9(R1): Addendum on Estimands and Sensitivity Analysis in Clinical Trials. ICH; 2019.

· Levenson M, He W, Dharmarajan S, Izem R, Meng Z, Pang H, Rockhold F. Statistical considerations for fit-for-use real-world data to support regulatory decision making in drug development. Stat Biopharm Res. 2023.

· Pearl J, Mackenzie D. The Book of Why: The New Science of Cause and Effect. Basic Books; 2018.

· van der Laan MJ, Rose S. Targeted Learning: Causal Inference for Observational and Experimental Data. Springer; 2011.

· Wang S, Schneeweiss S; RCT-DUPLICATE Initiative. Emulation of randomized clinical trials with nonrandomized database analyses: results of 32 clinical trials. JAMA. 2023;329(16):1376-1385.

Lessons from More Than a Decade of Recruiting Biostatistics Leaders

Over the past twenty-plus years, we have recruited and advised hundreds of biostatistics professionals across pharmaceutical companies, biotechnology organizations, and CROs. Watching careers unfold from Associate Director to Vice President has given us a unique perspective on what separates professionals who continue to advance from those whose careers plateau.

One observation has remained remarkably consistent.

Technical excellence gets professionals to mid-career. It rarely takes them beyond it.

By the time someone reaches the Associate Director level, most professionals are already highly capable statisticians. Leadership is no longer asking, Can this person perform the analysis? Instead, they begin asking different questions.

Can this person influence decisions?

Can they communicate through uncertainty?

Can they build trust across functions?

Can they develop future leaders?

The professionals who continue ascending are not always the smartest people in the room. They are the ones who consistently expand their influence, increase the scope of their ownership, and position themselves where opportunity continues to grow.

The observations that follow are not rules. Rather, they are recurring patterns we have watched shape leadership careers across the biostatistics profession.

1. Choose Environments That Compound Your Growth - Where should I grow?

One of the most common mistakes we see mid-career professionals make is evaluating opportunities primarily by title.

Titles matter, but what matters far more is what that role allows you to become over the next three to five years. Every career decision should be evaluated not only by the position you are accepting today, but by the opportunities that position is likely to create tomorrow.

The best environments continually stretch you. They increase the complexity of your decisions, broaden your cross-functional relationships, and expand the number of people who rely on your judgment. Hiring managers understand that biotech is volatile; what they look for is a pattern of increasing responsibility, influence, and organizational impact.

The better question is not, ‘Is this a better title?’ It is, ‘Will this environment make me a stronger leader three years from now?’

2. Build Your Career Around Your Strengths - What Should I Grow Into?

One of the biggest advantages of mid-career is that you finally have enough experience to understand what you do exceptionally well.

Early in a career, breadth builds capability. Mid-career is when professionals should become intentional. The leaders who continue to advance rarely try to excel at everything. Instead, they identify where they create disproportionate value and deliberately build their careers around those strengths.

Organizations promote people because of the value they consistently create, not because they possess the longest list of competencies. The more your career reflects your strengths, the more likely you are to produce exceptional results, enjoy your work, and build a reputation that compounds over decades.

3. From Communication to Leadership - Can I influence people and decisions beyond my technical expertise?

One of the biggest misconceptions in career progression is that leadership begins when someone receives their first direct report. In reality, leadership starts much earlier.

Strong communication is the foundation of every promotion after the Associate Director level. Before an organization trusts you to manage people, it must first trust you to influence meetings, align cross-functional teams, resolve conflict, and communicate complex statistical concepts in ways that drive better decisions. Managing people is the next evolution of those same skills.

The statisticians who ultimately become Directors, Senior Directors, Executive Directors, and Vice Presidents are rarely promoted solely because they are exceptional statisticians. They are promoted because they consistently develop people, build high-performing teams, and create environments where others can succeed.

As careers progress, expectations continue to expand.

Most executive-level positions in biostatistics require years of successful people management. Organizations want evidence that candidates have hired exceptional talent, coached developing leaders, managed multiple direct reports, and led complex development programs simultaneously.

This does not diminish the importance of technical expertise. Quite the opposite. Technical excellence earns credibility, communication creates influence, management builds trust, and leadership multiplies impact across an organization.

4. Shift from Execution to Ownership - Can I lead outcomes, not just analyses?

Titles may open doors, but the programs you lead build your reputation.

As careers progress, the depth of your involvement and the scope of what you own become stronger indicators of your market value than the titles on your resume. Quality is largely inferred from the type and stage of the program itself. A pivotal Phase III trial, a first-in-class asset, or a filing under active health authority review carries its own signal. What separates candidates is how much of that program they actually own. The strongest signal is meaningful regulatory experience, especially direct interaction with health authorities. There is a real difference between contributing analyses to a submission package and leading the statistical discussions that shape regulatory strategy. A statistician who has defended a position in a Type B or Type C meeting carries a credibility that someone who only supported analyses behind the scenes does not.

Working on programs that reach approval and improve patient outcomes builds credibility that lasts for years. The most influential senior biometrics leaders we have placed built their careers around a handful of strategically important programs where they held deep ownership and broad scope, not by collecting larger titles across unrelated assets.

This is the shift that defines the next stage of growth: moving from supporting programs to owning them. Early in a career, success is measured by execution, delivering quality analyses, meeting timelines, and solving technical problems. At the Associate Director level and above, leadership evaluates something different: can this person lead through uncertainty and drive organizational decisions when the answer is not obvious?

Strong contributors often plateau when leadership cannot yet picture them operating one level up. The question worth asking is whether your current leadership can picture you there. The gap is rarely technical. It is the visible ability to own outcomes that extend beyond the analysis itself.

5. Influence Beyond Statistics - Can I influence decisions beyond my function?

By the time someone reaches the Associate Director level, most professionals under consideration for advancement are already technically strong. Promotions become less about analytical capability and more about whether the broader organization trusts someone in strategic environments.

The professionals who keep advancing consistently do the following. Ask yourself honestly whether each describes you:

• Translate statistical uncertainty into language that informs business decisions

• Remain composed when results are ambiguous or unfavorable

• Act as collaborative problem-solvers rather than gatekeepers

• Communicate clearly in leadership meetings without relying on technical jargon

Clinical development is inherently cross-functional. The statisticians who become highly influential are rarely viewed as isolated technical experts. They become trusted business partners who help organizations make better decisions.

Visibility and sponsorship matter

You may be producing exceptional work that senior leadership never fully sees. The common misconception is that outstanding work naturally creates its own visibility. It does not. Visibility is not self-promotion. It is ensuring your judgment and decision-making are consistently experienced by the people responsible for advancement.

It is equally important to distinguish mentorship from sponsorship. Mentors provide guidance. Sponsors advocate for your advancement when opportunities arise and you are not in the room. Both matter, but sponsorship is often what converts a strong reputation into the next opportunity.

Biometrics is a much smaller community than many professionals appreciate. Reputation compounds over decades. People remember how you operated under pressure, how you communicated difficult results, and whether you were someone they wanted to work with again.

6. Think Like a Business Leader, not Just a Statistician - Do I understand where the organization and industry are going?

Leaders don’t just understand statistics. They understand where the business is going. The strongest mid-career professionals understand not only the science but the business environment around it. Hiring across biotech and pharma is driven by funding cycles, clinical readouts, mergers and acquisitions, regulatory milestones, and therapeutic area momentum. If you track these dynamics, you will see where investment is flowing and which organizations are expanding before the broader market catches on, and you can position yourself accordingly.

Inside larger organizations, this awareness creates opportunity directly for you. Statisticians are often pulled in to evaluate acquisition targets, assess development programs, and contribute to portfolio strategy. Those experiences build visibility with senior leadership and demonstrate business judgment beyond statistical expertise.

The same judgment applies to your own moves. Joining the right organization at the right inflection point can create far more long-term value than accepting a higher title at a company with limited room for growth.

Final thoughts

Mid-career progression is rarely linear. There will be reorganizations, failed programs, changing markets, and moments of uncertainty. The professionals who continue to advance are not the ones who avoid these challenges, but the ones who make thoughtful decisions about where they invest their expertise and leadership.

Mid-career is also the ideal time to assess your strengths, identify where you create the greatest value, and align your career path accordingly. Sustainable growth comes less from chasing every opportunity and more from building on what you do best. Doing so often leads not only to greater success, but also to deeper satisfaction and long-term fulfillment.

After recruiting and advising biostatistics professionals for more than 20 years, we’ve observed that career-defining decisions are rarely the most obvious ones. The people who keep ascending are not necessarily those who pursue the highest title or the largest organization. They are the ones who consistently place themselves in environments where responsibility, influence, and opportunity compound over time.

Technical excellence builds credibility. Leadership multiplies impact.

The Lotus Group is a minority women-owned executive search firm specializing in Biometrics, Clinical Development, Regulatory Affairs, Medical Affairs, Clinical Pharmacology, HEOR, and related life sciences functions. We partner with professionals at every stage of their careers, helping them make decisions that pay off over the long term.

Subscribe now

Selection as a Fellow of the American Statistical Association (ASA) is a prestigious honor that many ASA members aspire to achieve. Each year, new Fellows are chosen for their outstanding achievements and significant contributions to the field, as highlighted in nomination packages submitted to the ASA Committee on Fellows. Fellows are announced in the spring and formally recognized during the Joint Statistical Meeting (JSM). The Biopharmaceutical Section (BIOP) has a strong tradition of representation among these distinguished Fellows, with 7 current BIOP members receiving the honor in 2026.

To support BIOP members interested in pursuing ASA Fellowship—whether as nominators or nominees—the BIOP operates a Fellows Committee comprised of ASA Fellows experienced in the nomination process. This year’s committee is chaired by Yongming Qu and includes Weili He, Paul Gallo, Inna Perevozskaya, Amarjot Kaur, and Bill Wang. The committee’s primary role is to assist BIOP members considering or preparing a Fellowship application. Guidance includes evaluating readiness, identifying areas for improvement, advising on obtaining letters of support, recommending resources, reviewing application materials, and providing constructive feedback for enhancement.

Nominators or nominees wishing to utilize this service should submit their draft nomination packages to committee chair Yongming Qu (qu_yongming@lilly.com) by the end of January 2027 to ensure ample time for review and revisions ahead of the ASA deadline of end of February 2027. Note there is a change for candidate eligibility. All nominees are required continuous membership in the American Statistical Association for at least five of the six years between March 1, 2021, and February 28, 2027. Candidate eligibility can be verified by emailing Rachel Mills at rachel@amstat.org. More information can be found at https://www.amstat.org/your-career/awards/asa-fellows.

There are already several good sources of information readily available to prospective candidates and nominators. Certainly, those planning a nomination should familiarize themselves with the process and deadlines, along with suggestions for an effective nomination, described on the ASA website: ASA Fellows (amstat.org). In addition, a helpful article with available resources can be found under the Fellow Nomination Committee of the Biopharmaceutical section website: https://community.amstat.org/biop/aboutus/sub-committees/fellows141. Further, an ASA-sponsored webinar was presented in September 2020, “Biopharmaceutical Section Offers Advice on Strategic Planning for ASA Fellow Nomination”, containing presentations and panel discussions featuring a large group of BIOP members with experience in the Fellows process, and has been saved for viewing at: www.youtube.com/watch?v=YLkXund_p7I .

Best of luck to all nominators and nominees aspiring to join the ASA Fellows Class of 2027! The BIOP Fellows Committee is here to support your journey.

Note from the Editor:

Welcome to the Spring 2026 Issue! This year’s theme is “Statistics Reimagined for Discovery and Decision Making”, and the articles in this issue showcase how our community is pushing the boundaries of statistical thinking to shape the future of medicine development.

We open with the BIOP 2025 Chair Transition Report from Steven Novick and Erik Bloomquist, capturing a year of remarkable achievements and previewing plans for 2026.

A central thread in this issue is the evolving relationship between AI and statistical reasoning. Yanxun Xu explores how AI is transforming the statistical workflow in healthcare — serving as a force multiplier for rigorous thinking, not a replacement. Jingyun Jia and Ben Lengerich highlight the complementary roles of machine learning and causal inference, while Antonio Remiro-Azócar features machine learning for comparative effectiveness research. Jinfeng Zhang argues that generative AI’s true power depends on structured knowledge and quantified uncertainty.

Regulatory perspectives are well represented. Susan Mayo draws on her dual industry and FDA experience to discuss statistical frameworks — from estimands to benefit-risk planning — that improve drug development clarity. We also continue the Project SignifiCanT series with two new regulatory discussion summaries: one on interpreting duration of response in cancer trials, and another on the challenges when the standard of care changes during ongoing randomized trials.

Scott Evans offers a thought-provoking reflection on the opportunities and threats facing clinical trials. In operational innovation, Fei Chen and the Efficiency+ Scientific Working Group show how statisticians are bringing statistical modeling to trial operations — from drug supply to site selection. The openstatsware group shares updates on open-source software, including the relaunched CRAN Task View for Clinical Trials and the new openstatsguide.

Career development is a highlight of this issue. Aloka Chakravarty reflects on three decades spanning FDA, academia, and industry with guidance on navigating sector transitions. The team at The Lotus Group shares a candid letter — Cracking the Industry Code — with practical advice on what hiring managers really evaluate beyond technical skills.

We proudly announce the 2026 Student Paper Award winners, whose work spans IPD reconstruction, treatment effects with competing intercurrent events, and nonconcurrent data integration in platform trials. Congratulations to all the winners.

We round out with the recap insights by Wanjie Sun and colleagues from FDA, EMA, and industry leaders on statistical innovation from the 2025 RISW plenary panel, the STATBOLIC 2026 Conference Report, documenting the growth of this cardiometabolic forum to nearly 190 attendees, and a comprehensive Upcoming Conferences column highlighting MBSW, JSM in Boston, the Regulatory-Industry Statistics Workshop, and more.

Our sincere gratitude to all contributors for sharing their expertise, and to our ASA colleagues for their continued support in production. We hope this issue inspires fresh thinking as our community reimagines what statistics can achieve.

2025 ASA Biopharmaceutical Report Editorial Board:

Di Zhang (Eli Lilly, Editor), Charlotte Baidoo (BMS, Clinical Associate Editor), Yi Pan (BMS, Clinical Associate Editor), Junjing “Jane” Lin (Takeda, Clinical Associate Editor), Francis Rogan (Merck, Non-Clinical Associate Editor), Andrew Gehman (GSK, Non-Clinical Associate Editor)

ASA Biopharmaceutical Section Chairs:

Erik Bloomquist (2025), Steven Novick (2026), Judy Li (2027)

Transition Report

BIOP 2025 Chair Transition Report

Steven Novick (Takeda), Erik Bloomquist (Neurocrine)

Featured Articles on Theme

Machine learning for comparative effectiveness research: What is the catch?

Antonio Remiro-Azócar (Novo Nordisk)

From Data to Evidence: How AI Is Transforming the Statistical Workflow in Healthcare

Yanxun Xu (Johns Hopkins University)

Structured Knowledge with Quantified Uncertainty: Unlocking the True Power of GenAI

Jinfeng Zhang (Insilicom)

AI Finds Patterns. When Do They Become Evidence?

Jingyun Jia (UW-Madison), Ben Lengerich (UW-Madison)

Bridging Industry and Regulatory Perspectives: Statistical Frameworks for Clearer Drug Development and Review

Susan Mayo (FDA)

Leadership and Career Development

Seizing the Opportunities in and Protecting the Integrity of Clinical Trials

Scott Evans (George Washington University)

Navigating the Statistical Career Landscape: Lessons from Across Sectors of Industry, Government, and Academia

Aloka Chakravarty (Eli Lilly and Company)

Cracking the Industry Code: A Letter to Future Biostatisticians Entering the Pharmaceutical Industry

Bill Emker, Lei Wang, Helena Fan & Dean Grimm (The Lotus Group)

Working Group and Conference Update

ASA Biopharmaceutical Section Student Paper Awards

Francis Rogan (Merck)

Upcoming Conferences

Andrew Gehman (GSK)

Statistical Innovation in the AI/ML Era: Insights from the 2025 Regulatory Industry Statistics Workshop Plenary Panel Discussion

Wanjie Sun (FDA), Mark Levenson (FDA), John Scott (FDA), Gregory Alexander (FDA), Florian Lasch (EMA), Li Wang (AbbVie), Devan Mehrotra (Merck), Jared Lunceford (Merck)

STATBOLIC 2026 Conference Report

Hiya Banerjee (Eli Lilly), Henrik Ravn (Novo Nordisk), Yingwen Dong (Roche) on behalf of STATBOLIC Working Group

How Many Vials for Brazil? Efficiency+ Brings Statistical Innovation to Trial Operations

Fei Chen (Johnson & Johnson) on behalf of the Efficiency+ Scientific Working Group

Summary of ASA BIOP Section’s Virtual Discussion with Regulators on Statistical Considerations for Changes in Standard of Care During Ongoing Randomized Cancer Clinical Trials

Rajeshwari Sridhara (FDA), Olga Marchenko (Bayer), Qi Jiang (Pfizer), Brittany McKelvey (LUNGevity), Yiyi Chen (Pfizer), Gautam Mehta (FDA), Richard Pazdur (FDA)

Summary of ASA BIOP Section’s Virtual Discussion with Regulators on Statistical Considerations for Interpreting Duration of Response in Cancer Clinical Trials

Rajeshwari Sridhara (FDA), Gautam Mehta (FDA), Olga Marchenko (Bayer), Brittany McKelvey (LUNGevity Foundation), Qi Jiang (Pfizer), Yiyi Chen (Pfizer), Richard Pazdur (FDA)

openstatsware: Recent Highlights and Ongoing Work

Daniel Sabanés Bové (RCONIS), Alessandro Gasparini (Red Door Analytics), Ya Wang (Gilead Sciences)

Subscribe now

Susan draws her perspective from 32 years serving as a biostatistician in the biopharmaceutical industry (biotech, small and large companies, attaining expertise in safety and benefit-risk planning and assessment) and 8 years as a primary statistical reviewer in FDA's Office of Biostatistics, Division III, in the Center for Drug Evaluation and Research. Her BS training in biology/ecology and second MS in marine ecology shape the systems thinking in this article.

Disclaimer

This publication reflects the views of the author and should not be construed to represent the FDA’s views or policies.

Highlights:

• The Estimand Framework Improves Drug Development Clarity

• Additional Statistical Frameworks Support Comprehensive Drug Development

• Early Sponsor-Regulator Engagement Reduces Regulatory Risk

• Statisticians are Well-Positioned to Serve as Drug Development Data Stewards in the AI Era

  ---

Introduction

As a pharmaceutical industry statistician, I wrote statistical content for many protocols, statistical analysis plans (SAPs), study reports and marketing applications. As an FDA statistical drug reviewer, I reviewed hundreds of sponsor submissions relating to Investigational New Drugs (INDs), Biologics License Applications (BLAs), and New Drug Applications (NDAs) across multiple sponsors and therapeutic areas.

The estimand framework has demonstrably improved drug development clarity and efficiency. When the framework is practiced in the spirit of gaining understanding and clarity, we have seen this lead to clearer marketing applications with fewer review issues. Early sponsor-regulator engagement resolves questions prior to Phase 3 initiation, resulting in analyses well-aligned with trial objectives. Two additional frameworks—Aggregate Safety Assessment Plans (ASAP) and benefit-risk planning—have similar potential to uncover unforeseen challenges earlier, reducing risks in study design and data collection. These frameworks reduce regulatory risk while enabling potential artificial intelligence (AI) automation in commonly approved indications, provided we preserve the critical human judgment essential to protecting public health.

Dual Perspectives: Industry and Regulatory Environments

Industry statisticians manage multiple concurrent protocols, analysis plans, study reports, and regulatory submissions under tight timelines, where the strategic through-line connecting individual studies to the final marketing application can become obscured by immediate operational demands.

Regulatory statisticians face different constraints. While reviewing individual drug projects over shorter timeframes, we evaluate significantly more drug products than industry counterparts, gaining comprehensive therapeutic area knowledge across many drugs under development in a given therapeutic area. This broader exposure enables comparative assessment of endpoints, trial designs, estimands, and potential safety concerns while maintaining focus on patient benefit through multidisciplinary collaboration. The multi-layered review process—where team leaders review primary reviews and disciplines review each other’s responses—ensures comprehensive evaluation, facilitates knowledge transfer, and creates accountability for thorough, defensible evaluations that sponsors will scrutinize.

This perspective reveals how implementing robust statistical frameworks during the IND phase directly translates into clearer BLAs and NDAs.

Statistical and Assessment Frameworks for Clarity

Estimands framework

The estimand framework¹ addresses ambiguity in statistical planning and analysis for addressing the primary clinical question. The role of intercurrent events (ICEs) that occur post-randomization are a prominent feature. (Planning to address missing data,² while not attributes of the estimand framework itself, are included in my definition of the framework for purposes of this article.) When primary analyses are prespecified in ways that don’t fully address the primary clinical question, ICEs can lead to erroneous interpretation. For example, treating deaths as missing data in long-term trials of fatal diseases could be misleading, especially if the primary endpoint trends differently than mortality.

The five estimand attributes serve as a checklist for key statistical considerations that, when considered comprehensively, are rigorously defined before protocol finalization and detailed in the SAP. Reflecting and refining these attributes back to the primary clinical question may require several iterations between statisticians and clinicians, and between sponsors and regulators, to achieve maximum clarity. Figure 1 demonstrates how clinical questions are translated into precise estimand attributes.

Since sponsors began implementing estimand guidance in their protocols and SAPs, my observation from reviewing applications during their IND and NDA stages is that marketing applications utilizing this framework are more straightforward to review and require fewer clarifications. Issues that previously triggered information requests during application review are getting resolved earlier in development, potentially reducing uncertainty and increasing approval likelihood for safe and effective drugs.

Figure 1 Steps used to identify a study’s primary estimand

Clinical knowledge about the indication and associated clinical practices is essential for establishing rigorous linkage between a trial’s primary clinical question and its estimand, making statistician-clinician communication critical. The estimand framework introduces terminology initially unfamiliar to clinicians, while statisticians must understand clinical terminology and disease management practices to identify expected ICEs.

Effective communication helps clinicians recognize the value of these considerations while enabling statisticians to gain essential clinical background. This collaboration becomes particularly evident when choosing the most appropriate strategy for each ICE. Clinicians are ideally positioned to weigh in on strategy selection based on the study question but need translation of technical terms into familiar concepts. When I’ve asked clinical colleagues what has been most useful in understanding estimand terminology, they report that statisticians describing strategies in layperson terms has been most helpful. Examples are shown in Figure 2.

Figure 2 Examples of estimand strategies applied to three clinical questions (modified from Regulatory Industry Statistics Workshop short course on estimands, 2022; Instructors A. Ionan, J. Scott, M, Paterniti, S. Mayo)

Based on my review experience and conversations over the years with review staff and decision makers across clinical and statistical divisions covering guidance development, reviews, manuscripts, and training development, I created a short list of points to consider about estimands, ICEs, and their strategies:

• ICE Strategy Declaration: Each ICE needs a declared strategy in the protocol and SAP for how it will be handled in analysis. Multiple ICEs may occur in the same study and may have different strategies applied, depending on how each ICE affects interpretability of results for a given endpoint.

• Strategy Selection Considerations: ICH E9(R1), Section III.B provides guidance for selecting the appropriate strategy for an intercurrent event, with training slides offering further detail for appropriately assigning strategies.³ These concepts are not always practiced in IND briefing documents and marketing applications.

• Prespecification of Several Estimand Strategies: Applying various estimand strategies to the primary endpoint results in answers to different questions; typically, only one best fits the clinical question for that study. Using other strategies in supplementary analyses may be reasonable, but declaring many additional prespecified analyses without clarifying the alignment of these alternative strategies to a clinical question creates ambiguity.

• Data Handling to Implement a Strategy May Not Be Mathematically Unique: E9 Addendum Figure 1 illustrates that the appropriate approach to creating an estimand is to begin with the clinical question, with the goal of identifying the strategy and its corresponding data handling method that best align with this question. It is important to focus on whether the proposed data handling method makes sense given the clinical question of interest even if the proposed data handling strategy ends up being mathematically the same as another estimand strategy. What matters is focusing on the clinical question.

• Clinical Question for Public Health: A concern warranting further discussion relates to handling study medication discontinuation. Patients randomized to placebo may subsequently receive alternative medications improving their outcomes, while patients on the active arm may discontinue due to drug toxicity. From a public health perspective, toxicity events should not be disregarded, yet applying a hypothetical strategy to these discontinuations effectively ignores them. A composite strategy addressing treatment failure—encompassing both study drug discontinuation and use of certain alternative medications—would more appropriately cover placebo arm failures with active arm drug toxicities (treatment policy strategy) considered as a separate ICE.

Aggregate Safety Assessment Planning

While statisticians have typically focused on clinical trial efficacy, statistical expertise in drug safety is increasingly important. In 2010, a Final Rule for IND expedited reporting of serious adverse events (SAEs) was issued,⁴ with guidance finalized in 2025.⁵ Reportable events are those meeting the criteria of serious, unexpected, and with evidence of a causal relationship to the investigational drug. Sponsors should perform an aggregate analysis comparing the rate of serious suspected adverse reactions in ongoing trials to an expected rate, while protecting trial integrity.

Aggregate safety monitoring may demonstrate (or rule out) causality by comparing the aggregate rate in the test drug to control or to a known rate from the investigator brochure or labeling. The simultaneous effort to identify increased aggregate event rates indicative of causality while protecting ongoing trials’ blinded data integrity benefits from careful program planning (e.g., using consistent definitions and data structures) and clear communication on processes and roles.

An early recommendation for safety planning was the Program Safety Analysis Plan (PSAP). ⁶ The ASA Biopharmaceutical Section’s Safety Scientific Working Group⁷—composed of statisticians, clinicians, data scientists, and epidemiologists—has led efforts to expand the PSAP into the Aggregate Safety Assessment Plan (ASAP) ^8,9 to address the Final Rule. Members have also contributed Bayesian approaches for monitoring safety in ongoing, blinded trials. ^10,11

This careful planning and automation—which involves centralizing a drug program’s safety data in a single repository, predefining MedDRA coding terms and safety issue handling procedures across studies, and enabling on-demand safety signal detection rather than reactive data assembly—has already proven valuable beyond aggregate safety monitoring, including use in due diligence for in-licensing. ¹² The Bradford Hill Criteria can further benefit ASAP implementation with scientific rigor and operational consistency in organizing safety data from multiple sources. ¹³ One sponsor company has proposed ASAP implementation using an R markdown application.¹⁴

As an advocate for safety statistics planning and analysis, note this is one of several areas of safety and benefit-risk planning and assessment being progressed by several working groups. In addition to ASA BIOP’s Safety group, PhUSE¹⁵ and PSI¹⁶ are progressing efforts as well.

Benefit-Risk Planning

Benefit-risk planning involves less data complexity than IND aggregate safety signal detection, but, like ASAP, requires multidisciplinary input and is less commonly used than SAPs for efficacy planning. In many drug programs, benefit-risk decisions are straightforward—either clearly favorable or unfavorable—making formal planning less critical. However, planning during the IND phase for how key benefits and risks will be assessed is being implemented in sponsor companies¹⁷ and considered valuable by stakeholders. ¹⁸

As a reviewer, this planning becomes especially valuable when anticipated benefits are modest relative to potential risks. In these marginal cases, structured planning creates opportunities for sponsors to engage with regulators (or vice versa) to establish a shared understanding of how benefits and risks might be assessed together, positively influencing study design, outcome selection, incorporation of patient preferences, and data collection strategies. Without this planning step, both sponsors and regulatory reviewers are left with what may be suboptimal design, outcomes, and/or data collected to craft the application and review it from the standpoint of understanding benefits compared to risks.

Extensive literature exists on this topic, including an ICH guideline,¹⁹ FDA guidance,²⁰ a 2025 CIOMS report,²¹ a European Medicines Agency website describing a benefit-risk methodology project,²² and the Innovative Medicines Initiative’s PROTECT project benefit-risk recommendations.²³ These describe the benefit-risk landscape, promote structured frameworks, and give an overview of methods across the drug development life cycle.

Particularly in drug programs anticipating a potential higher risk and only modest benefit, the advice offered in ICH, CIOMS, and regulatory agency documents can serve as a helpful checklist in ensuring a program is robust as it can be regarding its potential benefit-risk challenges. Such an approach has potential for sponsors to reduce risk with their application and make the regulatory review a smoother process because these issues were already discussed and addressed during the IND phase.

Sponsors have published their experiences with benefit-risk planning and assessment, e.g., documenting their approaches, with visualizations for several benefit-risk scenarios²⁴ and descriptions of organizational and procedural restructuring to support benefit-risk evaluation.²⁵ Some of these authors, along with other clinicians and statisticians offered presentations on their insights at ASA Biopharmaceutical Section’s Safety Scientific Working Group Q2 2024 webinars as well.²⁶

In my opinion, for marginal cases—such as non-life-threatening diseases with concerning preclinical or clinical safety signals—sponsors may benefit from structured pre-market qualitative or quantitative benefit-risk planning and discussion with the Agency during drug development. Appropriate engagement during end of Phase 2 or Phase 3 trial planning allows sponsor and regulatory teams to discuss and align on which efficacy and safety outcomes to collect and how to assess these outcomes together. Clarity on regulatory priorities for public health can reduce program risks by identifying potential hurdles for marketing applications with marginal benefit-risk outcomes.

The Role of AI and Human Judgement

Recent perspectives from thought leaders in biostatistics, including Margaret Gamalo²⁷ and Feiming Chen, ²⁸ have prompted reflection on how artificial intelligence is reshaping our profession. Dr. Gamalo emphasized a critical principle: we must automate processes while preserving human judgment. This principle has important implications for implementing statistical frameworks in drug development.

Automating Processes, Preserving Judgement

Dr. Gamalo’s insight, “Automate the plumbing but never the judgement,” resonates for both our AI-enabled future and current work. Statisticians are uniquely equipped to discern which steps can be automated from those requiring human critical thinking in the clinical trial data flow. While AI can efficiently process vast data and identify patterns, it cannot replace the nuanced judgment required to determine whether patterns are clinically meaningful, reflect true treatment effects or should inform regulatory decisions.

Statisticians are uniquely positioned to co-lead development of automated data infrastructure, particularly in identifying critical human decision points where judgment is essential. Our expertise spans from theoretical inference to practical implementation, from data structure design to IT system requirements —areas where data scientists, clinicians, and other disciplines may lack depth. This dual capability positions us as natural architects of AI-enabled systems, envisioning how data should flow, where quality checks are needed, which processes can be safely automated, and where human oversight must be preserved.

The frameworks discussed—estimands, ASAP, and benefit-risk planning—represent structured junctures where human judgment must be applied, regardless of automation level.

The Imperative to Speak Up

As we move toward greater automation, statisticians must actively engage when AI-based approaches are developed or implemented. If you have concerns—for example, whether a touchpoint for human decision-making should be included at a specific juncture—raise them. The statistician may be the only person in the room thinking about this issue, given our inferential perspective, clinical data expertise, and practice of beginning with the end in mind.

Statisticians should proactively participate in designing AI systems, ensuring automated processes include appropriate checkpoints for human review. Critical questions include: Does this automation preserve our ability to assess data quality? Does it allow for the clinical and statistical judgment needed to interpret results appropriately? Will it enable us to identify unexpected safety signals or efficacy patterns warranting further investigation?

Our unique vantage point—understanding both statistical requirements and clinical context—make us essential contributors to these discussions. Failing to engage risks systems that optimize efficiency at the expense of scientific rigor or patient safety.

AI and the Future of Estimands

As the AI era unfolds, establishing standard estimands for common indications could integrate consensus solutions directly into AI infrastructure. For prevalent diseases with approved medications, there is already agreement on key estimand attributes: target population, endpoint, and summary measure. Treatment specifications will always depend on the specific drug under investigation, making intercurrent events and their handling strategies the primary domains where standardization would yield greatest benefit.

A designated entity could develop and publish these standards, analogous to CDISC’s therapeutic area-specific data standards, streamlining protocol development, reducing ambiguity in regulatory submissions, and enabling more efficient AI-assisted review processes.

However, this vision of standardization must be balanced against the fundamental need for clinical judgment in estimand development.

Statisticians as Data Stewards in the AI Era

Statisticians are uniquely positioned to serve as data stewards in drug development, drawing on our knowledge of inference, methods, simulations, data structures, and requirements for data flow through IT systems. We are trained not only to generate and interpret technical results but also to translate their meaning for decision-makers in ways that drive informed action.

In the AI era, this stewardship role becomes even more critical. As routine analytical tasks become automated, our value increasingly lies in the judgment we apply to complex questions, the context we provide for interpreting results, and the safeguards we build into automated systems to ensure AI tools enhance rather than replace the careful reasoning that protects public health.

This stewardship extends to the frameworks discussed here. These frameworks represent structured junctures where human judgment must be applied, regardless of automation level. They embody the principle of automating the plumbing while preserving the judgment, creating clear decision points where statistical and clinical expertise must guide the path forward.

Conclusions and Recommendations

The frameworks discussed—estimands, ASAP, and benefit-risk planning—represent essential tools for navigating the complexities of modern drug development. As artificial intelligence increasingly automates routine analytical tasks, the human judgment required to properly implement these frameworks becomes even more critical.

Three key recommendations emerge:

1. Early engagement: Sponsors should engage with regulators during the IND phase to clarify estimands, safety and benefit-risk planning approaches, reducing downstream regulatory risk.

2. Multidisciplinary collaboration: Effective implementation requires ongoing dialogue between statisticians and clinicians, with statisticians serving as translators of statistically technical concepts.

3. Standardization with flexibility: While standardized estimands for common indications may streamline development, these must evolve as clinical practice advances and new therapies emerge.

Statisticians are uniquely positioned to serve as stewards of data integrity and analytical clarity in this evolving landscape. By beginning with the end in mind and applying these structured frameworks, we can better serve both the organizations we work for and, ultimately, the patients who depend on safe and effective therapies.

Subscribe now

Reference

[1] ICH E9(R1) https://www.fda.gov/regulatory-information/search-fda-guidance-documents/e9r1-statistical-principles-clinical-trials-addendum-estimands-and-sensitivity-analysis-clinical

[2] National Research Council. 2010. The Prevention and Treatment of Missing Data in Clinical Trials. Washington, DC: The National Academies Press.

[3] ICH E9(R1) training slides https://database.ich.org/sites/default/files/E9-R1_EWG_Step2_TrainingMaterial.pdf

[4] U.S. Department of Health and Human Services and Food and Drug Administration. (2010), Code of Federal Regulations Title 21 Food and Drugs Chapter I Food and Drug Administration Department of Health and Human Services Subchapter D Drugs for Human Use Part 312 Investigational New Drug Application. 75 FR 59935 - Investigational New Drug Safety Reporting Requirements for Human Drug and Biological Products and Safety Reporting Requirements for Bioavailability and Bioequivalence Studies in Humans.

[5] U.S. Department of Health and Human Services and Food and Drug Administration (2025), Sponsor Responsibilities — Safety Reporting Requirements and Safety Assessment for IND and Bioavailability/Bioequivalence Studies https://www.fda.gov/media/150356/download

[6] Crowe BJ, Xia HA, Berlin JA, et. al. Recommendations for safety planning, data collection, evaluation and reporting during drug, biologic and vaccine development: a report of the safety planning, evaluation, and reporting team. Clin Trials. 2009 Oct;6(5):430-40. doi: 10.1177/1740774509344101. PMID: 19846894.

[7]https://community.amstat.org/biop/workinggroups/safety-home

[8] Hendrickson, B.A., Wang, W., Ball, G., et al. Aggregate Safety Assessment Planning for the Drug Development Life-Cycle. Therapeutic Innovation & Regulatory Science 55(4):717-732, 2021.

[9] Several other publications on PSAP (among others), by this Working Group are noted here: https://community.amstat.org/biop/workinggroups/safety-home/safety-publications

[10] Ball G. Continuous safety monitoring for randomized controlled clinical trials with blinded treatment information. Part 4: One method. Contemp Clin Trials. 2011 Sep;32 Suppl 1:S11-7. doi: 10.1016/j.cct.2011.05.008. Epub 2011 Jun 1. PMID: 21651993.

[11] Mukhopadhyay, S, BR Waterhouse and A Hartford. “Bayesian detection of potential risk using inference on blinded safety data.” Pharmaceutical Statistics 17 (2018): 823 - 834.

[12] Youssef, A., Hammad, T.A. Pharmacovigilance Due Diligence in Drug Development: A Practical Playbook for Risk Identification, Compliance Assessment, and Strategic Decision Making. Drug Saf (2025). https://doi.org/10.1007/s40264-025-01640-8

[13] Loos A, Khazneh E, Kübler J. Enhancing Safety Evaluations: A Comprehensive Framework for Evidence-Based Safety Assessment Using the Bradford Hill Criteria. Ther Innov Regul Sci. 2026 Jan 8. doi: 10.1007/s43441-025-00910-y. Epub ahead of print. PMID: 41507617.

[14] Ghosh, D and R Gordon, Enhancing the (ASAP) Aggregate Safety Assessment Planning Framework with an Interactive R Markdown Application: Seeing is believing!, Paper ET06 https://www.lexjansen.com/phuse-us/2025/et/PAP_ET06.pdf&ved=2ahUKEwib0PTN9uCSAxV8D1kFHc3oATYQqYcPegQIAxAG&opi=89978449&cd&psig=AOvVaw31T5DOlRYdfDcIRQQobz9T&ust=1771431471122000

[15] https://advance.hub.phuse.global/wiki/spaces/WEL/pages/26804326/Safety+Analytics

[16] https://www.psiweb.org/sigs-special-interest-groups/benefit-risk/safety-implementation

[17] Gebel M, Renz C, Rodriguez L., et al. Survey to Assess the Current Status of Structured Benefit-Risk Assessment in the Global Drug and Medical Device Industry. Ther Innov Regul Sci. 2024 Jul;58(4):756-765. doi: 10.1007/s43441-024-00650-5. Epub 2024 Apr 22. PMID: 38649524.

[18] Simonetti A, Colilla S, Edwards B, et.al. Key Opinion Leaders’ Interviews to Inform the Future of Benefit-Risk Planning in the Medical Total Product Life Cycle of Global Pharmaceutical and Medical Device Organizations. Drug Saf. 2024 Sep;47(9):853-868. doi: 10.1007/s40264-024-01442-4. Epub 2024 Jun 1. PMID: 38824267; PMCID: PMC11324710.

[19] ICH M4E(R2) Revision of M4E Guideline on Enhancing the Format and Structure of Benefit-Risk Information in ICH (2016) https://database.ich.org/sites/default/files/M4E_R2__Guideline.pdf

[20] U.S. Department of Health and Human Services and Food and Drug Administration (2023). Benefit-Risk Assessment for New Drug and Biological Products https://www.fda.gov/regulatory-information/search-fda-guidance-documents/benefit-risk-assessment-new-drug-and-biological-products

[21] Benefit-risk balance for medicinal products, 2025. Report of the CIOMS Working Group XII https://cioms.ch/publications/product/benefit-risk-balance-for-medicinal-products/

[22] https://www.ema.europa.eu/en/about-us/what-we-do/regulatory-science-research/benefit-risk-methodology

[23] https://imi-protect-eu.cc.ic.ac.uk/

[24] Colopy M, Gakava L, Chen C. Planning Benefit‑Risk Assessments Using Visualizations. Ther Innov Regul Sci. 2023 Sep; 57:1123-1135. https://doi.org/10.1007/s43441-023-00563-9

[25] Sullivan T, Zorenyi G, Feron J, Smith M, Nord M. A Structured Benefit-Risk Assessment Operating Model for Investigational Medicinal Products in the Pharmaceutical Industry. Ther Innov Regul Sci. 2023 Jul;57(4):849-864. doi: 10.1007/s43441-023-00508-2. Epub 2023 Apr 1. PMID: 37005972; PMCID: PMC10276786.

[26] https://community.amstat.org/biop/workinggroups/safety-home/safety-presentations

[27] M. Gamalo. Systems Biostatistics: Making Speed in Drug Development Safe. Dec 2025 Biopharmaceutical Report

[28] F Chen. How Regulatory Statisticians Can Adapt to New Challenges in the AI Era. Dec 2025 Biopharmaceutical Report

We’re excited to share some recent achievements and ongoing initiatives from openstatsware, also known as the ASA BIOP Software Engineering Working Group. Over the past year, the group has continued to advance high‑quality statistical software practices through curation, education, and the development of open‑source tools for the clinical and biostatistical community. Please visit our working group website, https://openstatsware.org and feel free to join us!

Recent Highlights

Relaunch of the CRAN Task View for Clinical Trials

In 2025, openstatsware successfully relaunched the CRAN Task View for Clinical Trials, which provides a curated overview of R packages used across the clinical trial lifecycle—from design and monitoring to analysis and reporting. Maintained by the working group, the Task View reflects collective expertise in biostatistics and software engineering, with a strong emphasis on practical relevance and quality. Its goal is to improve package discoverability, promote well‑maintained tools, and support reproducible clinical trial analyses within the R ecosystem.

Publication of openstatsguide

The working group published openstatsguide, a concise, community‑driven guide that defines Minimum Viable Good Practices for statistical software development. The guide focuses on practical, fit‑for‑purpose recommendations that support code quality, sustainability, and collaboration—without imposing unnecessary overhead. It serves as a lightweight reference for developers and contributors, particularly in pharmaceutical and clinical research environments, and reflects openstatsware’s commitment to robust, transparent, and reproducible software.

Workshops on Good Software Engineering Practices for R Packages

Throughout 2025, openstatsware continued its active program of workshops on Good Software Engineering Practices for R Packages, delivering hands‑on training across multiple locations worldwide. These workshops translated core principles—such as reproducibility, testing, documentation, and collaborative development—into practical guidance for real‑world applications. Highlights included four short courses delivered in Tokyo (April 2025) and Philadelphia (August 2025), as well as courses at ISCB 2025 in Basel, Switzerland (August 2025) and the 10th Statistics and Biopharmacy Conference in Paris (October 2025). Together, these activities reinforced the group’s ongoing commitment to education, community engagement, and advancing software engineering standards within the clinical and biostatistical R ecosystem.

Conference Participation

In 2025, members of the working group actively engaged with the professional community by organizing the session Statistical Software Engineering at PSI Conference 2025 in London (June 2025). The group also presented posters at the EFSPI Regulatory Statistics Workshop 2025 in Basel (September 2025) and Biopharmaceutical Section Workshop (BBSW) 2025 in Foster City, California (November 2025). These contributions highlighted openstatsware’s activities and its ongoing impact on advancing statistical software engineering practices in industry settings.

Ongoing Projects

R package mmrm

mmrm is an open‑source R package developed by openstatsware to support robust and efficient analyses using Mixed Models for Repeated Measures (MMRM), a method widely used in clinical trials. The package provides a comprehensive and flexible implementation, with careful attention to statistical rigor, usability, and alignment with regulatory and industry practices. Available on CRAN, mmrm offers modern features, clear documentation, and a well‑tested code base, supporting reproducible and transparent clinical trial analyses.

R package brms.mmrm

brms.mmrm extends MMRM to a Bayesian framework using brms, enabling flexible model specification and full Bayesian inference while retaining a workflow familiar to applied clinical trial statisticians. Developed by openstatsware and available on CRAN, the package brings modern Bayesian methodology together with strong software engineering principles.

R package maicplus

maicplus is an open‑source R package designed to facilitate Matching‑Adjusted Indirect Comparison (MAIC) analyses, commonly used to support health technology assessment (HTA) and reimbursement submissions. The package provides a streamlined, user‑friendly workflow for population adjustment, estimation, and diagnostics, with a strong emphasis on transparency and reproducibility. maicplus is available on CRAN.

Completed Projects

Julia package SafetySignalDetection.jl

SafetySignalDetection.jl is an open‑source Julia package developed by openstatsware to support Bayesian safety signal detection in clinical and post‑marketing surveillance. Built on the Turing.jl probabilistic programming framework, the package provides flexible and extensible modeling tools for detecting and evaluating potential safety signals, with an emphasis on transparency and reproducibility. It is available via the Julia package registry.

Looking Ahead

Looking to the future, openstatsware remains committed to strengthening the statistical software ecosystem through continued package development, curation, and community engagement. Plans include further expanding educational offerings, evolving existing tools based on user feedback, and exploring new methodologies and programming languages where they can add value. The working group looks forward to continued collaboration with the broader biostatistical community to promote sustainable, high‑quality, and impactful statistical software.

Subscribe now

References

Ya Wang, Thomas Jaki, Laura Pascasio Harris, Orla Doyle, Elias Laurin Meyer, Wilmar Igl (2026). CRAN Task View: Clinical Trial Design, Monitoring, Analysis and Reporting. Version 2026-02-11. https://CRAN.R-project.org/view=ClinicalTrials.

openstatsware Working Group. (2024). OpenStatsGuide: Minimum viable good practices for high-quality statistical software packages. Version 0.1-0. https://www.openstatsware.org/guide.html

On September 30, 2025, the American Statistical Association (ASA) Biopharmaceutical Section (BIOP) and LUNGevity Foundation hosted a virtual forum to discuss Statistical Considerations for Interpreting Duration of Response in Cancer Clinical Trials. This forum was part of a series conducted under the guidance of the U.S. FDA Oncology Center of Excellence’s Project SignifiCanT (Statistics in Cancer Trials). The goal of Project SignifiCanT is to advance cancer drug development through collaboration and engagement among various stakeholders in the design and analysis of cancer clinical trials. The discussion was organized jointly by the ASA BIOP Statistical Methods in Oncology Scientific Working Group, the FDA Oncology Center of Excellence (OCE), and LUNGevity Foundation. 

Cancer clinical trials that evaluate patients for anti-tumor response often include duration of response (DoR) as a study endpoint. Interpretation of DoR, a time-to-event endpoint, can be difficult, particularly in single-arm trials where the therapeutic effect cannot be distinguished from the natural history of the disease. Additionally, including in randomized trials, assessment only in responders can create bias since response is a post-randomization outcome and the follow-up duration and trial sample size can greatly impact the measure. Kaplan-Meier methods to characterize durability among responders may address follow-up concerns but are still affected by sample size. Alternative methods to describe the durability of response in randomized trials have been recently described (e.g., restricted mean DoR). This open forum discussion among multi-disciplinary experts considered the utility and interpretation of DoR as an endpoint in cancer clinical trials. The panel discussed the appropriate settings for including DoR as an endpoint as well as how the endpoint should be reported.

The speakers/panelists* for the discussion included members of the BIOP Statistical Methods in Oncology Scientific Working Group representing pharmaceutical companies, representatives from international regulatory agencies (Food and Drug Administration (FDA), Medicines and Healthcare products Regulatory Agency (MHRA, UK), and Medicines Evaluation Board (MEB, Netherlands)), clinicians, academicians, patient advocates, and expert statisticians. In addition, over 100 participants attended the virtual meeting, including representatives from other international regulatory agencies (European Medicines Agency (EMA), Health Canada (HC), Paul-Ehrlich-Institut (PEI, Germany), Therapeutic Goods Administration (TGA, Australia), Brazilian Health Regulatory Agency (ANVIS, Brazil), Health Sciences Authority (HAS, Singapore), Ministry of Health (MOH, Israel), and Pharmaceuticals and Medical Devices Agency (PMDA, Japan)). The discussions were moderated by the BIOP Statistical Methods in Oncology Scientific Working Group co-chair, Dr. Olga Marchenko from Bayer, Brittany McKelvey from LUNGevity Foundation, and Dr. Rajeshwari Sridhara, consultant from OCE, FDA.

In the introductory presentation, the OCE leadership presenters discussed DoR as a commonly used endpoint in cancer clinical trials, including in single-arm studies. While DoR provides insights into how long treatment responses may last, its interpretation and estimation present notable challenges. As a time-to-event endpoint measured only among responders, DoR can be difficult to disentangle from the natural history of disease in single-arm trials and may introduce dependent censoring or imbalance in randomized settings. Estimation is also influenced by factors such as sample size and follow-up duration. Panelists and presenters were invited to discuss whether DoR is interpretable in both single-arm and randomized trials, under what circumstances it should be reported, and how it can be most reliably estimated.

The presenter from industry introduced restricted mean duration of response (RMDoR) as a novel statistical approach designed to address key limitations of traditional DoR analyses. RMDoR applies an intention-to-treat (ITT) framework as a composite of a binary response endpoint and a time-to-event DoR endpoint (restricted mean DoR), estimating the expected time a patient spends in response before progression or death while incorporating both responders and non-responders. Using a multi-state modeling approach, RMDoR computes the area between the PFS and progression/death/response-free curves (or alternatively, the area under the probability of being in response curve), providing a nonparametric, model-free estimate bounded by the study’s observation window. Compared to PFS, RMDoR is a measure of tumor size reduction and its durability, while PFS is a measure of disease stability and reflects the natural history of the disease. If a drug has a cytostatic mechanism that induces low response rate but delays progression, RMDoR may not be an appropriate statistical measure. On the other hand, if a drug has curative effect for some patients or can significantly reduce tumor size for a moderate to large percentage of the target population, and can also prolong duration while in response, then RMDoR may be more sensitive than PFS to predict OS as shown in previous publications. Advantages of RMDoR include statistical rigor, interpretability, and greater sensitivity to treatment effects, while limitations include dependence on follow-up duration and reduced interpretability in studies with low response rates.

The key points raised in the panel discussion following the presentations are as follows: 

• Estimand framework considerations for RMDoR: The panel emphasized that RMDoR implementation should follow ICH E9(R1) estimand framework principles by clearly defining the clinical question of interest before selecting analytical approaches. RMDoR’s composite nature, combining response occurrence with duration, requires careful consideration of how intercurrent events (such as treatment discontinuation or subsequent therapies) should be handled to align with the intended estimand and clinical interpretation.

• RMDoR advantages in randomized trials: RMDoR in randomized trials provides a statistically rigorous, interpretable, and ITT-based framework for comparing treatment arms, addressing key limitations of traditional DoR and enabling valid inference across the entire study population. RMDoR may help inform early-phase go/no-go decisions. RMDoR may potentially support initiatives such as a single trial to support accelerated and traditional approval (FDA’s Project FrontRunner).

• RMDoR limitations and considerations: RMDoR’s interpretability is limited in single-arm settings, as it cannot fully account for natural history or selection bias. In large, randomized trials, if used as a primary endpoint, RMDoR may offer limited added value beyond established endpoints such as PFS and OS, given its complexity, potential for bias, and increase in measurement error.

• Need for additional research: Additional empirical research across tumor types and therapeutic classes is needed to assess the robustness of RMDoR and its correlation with long-term outcomes, including PFS and OS. The panel noted that more experience is needed to understand optimal implementation and interpretation across different clinical contexts.

• Traditional DoR remains valuable: DoR remains an important descriptive measure for evaluating benefit–risk for regulatory approvals, particularly for rare cancers or small single-arm studies, but interpretation should be made cautiously given the dependency on censoring, sample size, and disease natural history.

This forum provided an opportunity to have open scientific discussion among a diverse multidisciplinary stakeholder group – clinicians, epidemiologists, and statisticians from academia and pharmaceutical companies, patient advocates, and international regulators- focused on emerging statistical issues in cancer drug development.  

References:

• Cui Y, Huang B, Mao L, Uno H, Wei LJ, Tian L. Inferences for the distribution of the duration of response in a comparative clinical study. Clin Trials. 2024 Oct;21(5):541-552. doi: 10.1177/17407745241264188. Epub 2024 Aug 8. PMID: 39114952; PMCID: PMC12113349.

• Huang B, Tian L. Utilizing restricted mean duration of response for efficacy evaluation of cancer treatments. Pharm Stat. 2022 Sep;21(5):865-878. doi: 10.1002/pst.2198. Epub 2022 Feb 21. PMID: 35191170; PMCID: PMC10676756.

Acknowledgement: Authors thank Joan Todd (FDA) and Syed Shah (FDA) for technical support.

* Speakers/ Panelists:  

Dr. Keaven Anderson (Merck), Dr. Arunava Chakravartty (Lilly), Janet Freeman-Daily (Patient Advocate), Dr. Boris Freidlin (National Cancer Institute, NIH), Prof. Tim Friede (University Medical Center Göttingen), Dr. Mithat Gönen (Memorial Sloan Kettering Cancer Center), Dr. Bo Huang (Pfizer), Dr. Alexei Ionan (FDA), Dr. Lang Li (FDA), Dr. Brittany McKelvey (LUNGevity Foundation), Dr. Olga Marchenko (Bayer), Dr. Gautam Mehta (FDA), Dr. Richard Pazdur (FDA), Dr. Khadija Rantell (MHRA, UK), Dr. Laura Rodwell (Medicines Evaluation Board, Netherlands), Dr. Rajeshwari Sridhara (FDA), Dr. Yevgen Tymofyeyev (Johnson & Johnson), Dr. Stephanie Wethington (FDA), Prof. Ying Yuan (University of Texas, MD Anderson Cancer Center).

Subscribe now

Scan to follow Efficiency+ on LinkedIn

The Importance of Statistical Innovation in Clinical Trial Operations

Most clinical trial statisticians have probably never been asked: How many vials of drug should we ship to a site in Brazil next month? Or: Should we activate five more sites in Japan to stay on track for enrollment?

These are clinical operations questions, but they are fundamentally statistical problems. For decades, operational decisions have been made largely by deterministic rules of thumb, spreadsheets, previous experience, and intuition. The result? Sites are selected based on historical familiarity rather than data; more than 30% of clinical trials globally fail to enroll on time, leading to costly extensions and opening of additional sites (Lamberti et al., 2024); and drug supply waste in clinical trials is estimated at 50% (McKinsey & Company, 2021). The cost of these inefficiencies is staggering as delays alone can cost sponsors millions of dollars per day for a late-stage program.

Statisticians are well-equipped to contribute innovative solutions to address operational inefficiencies. This motivated the formation of Efficiency+.

How We Got Here

The idea for Efficiency+ grew out of conversations among statisticians who found themselves increasingly involved in several aspects of operational decisions, including enrollment forecasting, site selection, and drug supply planning. Yet, there is no organized statistical community to share methods, compare approaches, or develop best practices.

In early 2025, we submitted a proposal to the ASA Biopharmaceutical Section to form a new Scientific Working Group (SWG) focused on clinical trial operational efficiency. The proposal was approved and in July 2025 we held our kickoff meeting with about 20 members from over 10 pharmaceutical and biotech companies.

From the start, we wanted Efficiency+ to be different from a typical methods-focused working group. Our scope is deliberately cross-functional, emphasizing the intersection of statistics with clinical operations, drug supply, regulatory planning, data management, and functions supporting other clinical operations. The “+” in our name reflects the statisticians plus the collaborators and domains needed to move the needle.

Our Mission

Our mission is to advance clinical trial operations by championing cross-pharma and cross-functional collaborative research to drive statistical innovations. We are dedicated to fostering interdisciplinary progress in trial design and execution, ensuring the highest standards of study conduct. Through sharing insights, identifying gaps, and learning from one another, we aim to optimize efficiency, reduce waste, and ultimately accelerate the path from lab to patient (Efficiency+ Scientific Working Group, 2025).

Workstreams

After our kickoff, we surveyed members to identify the most pressing topics. Four focus areas emerged, and the group was organized into the corresponding workstreams:

1. Patient Recruitment Monitoring and Forecasting, Site Selection Led by Clara Cali Mella (Bayer) and Vlad Anisimov (Amgen)

This topic drew the most interest. The team is tackling questions like: What are the best statistical models for enrollment forecasting? How do you combine historical data, real-world data, and accumulating trial data into a single forecast? How should site selection leverage predictive analytics rather than just investigator relationships? Vlad brings deep expertise in Poisson-gamma modeling for recruitment (Anisimov, 2020), and the team is working toward a comparative review of forecasting methodologies.

2. Dynamic Trial Monitoring and Data Quality Led by Palanikumar Ravindran (BMS)

This workstream focuses on using AI/ML and statistical methods for real-time risk detection during trial conduct, identifying underperforming sites early, predicting dropout patterns, and monitoring data quality signals. The goal is to move from reactive to proactive trial oversight.

3. Study Design and Operations Impact Led by Dooti Roy (Boehringer Ingelheim)

Innovative trial designs (adaptive, platform, seamless) are statistically exciting but can create operational nightmares. This team examines the two-way relationship between design choices and operational execution, such as how design decisions affect timelines and resources, and how operational realities should inform design.

4. Clinical Supply Chain Led by Christi Kleoudis (AstraZeneca)

Drug supply planning is one of the most impactful yet underappreciated areas where statistical innovation can make a difference. This workstream is developing simulation-based approaches to forecast drug demand, optimize vial configurations, and reduce waste, potentially saving millions of dollars per program.

Our Members

We are proud of the breadth of our membership, which exemplifies our vision for cross-industry, cross-functional collaboration. Efficiency+ currently includes about 20 active members from AbbVie, Amgen, AstraZeneca, Bayer, BeOne Medicines, Boehringer Ingelheim, Bristol-Myers Squibb, Cytel, Johnson & Johnson, Eli Lilly, Regeneron, and Sanofi. Our membership comprises statisticians, data scientists, and clinical operations professionals, exactly the kind of cross-functional mix needed to make real progress.

What We Have Done So Far

It has been less than a year since our kickoff, but Efficiency+ has been quite productive in building presence in the research community:

Conference Presence. We just wrapped up a successful invited session at ENAR 2026 (March 15-18, Indianapolis) titled “Enhancing Clinical Trial Efficiency through Statistics, AI, and Collaborative Innovation.” The session featured talks by Inna Perevozskaya (BMS) on trial monitoring methods, Vlad Anisimov (Amgen) on recruitment forecasting, Forrest Williamson (Eli Lilly) on pediatric trial challenges, and Ziqian Geng (AbbVie) on randomization in platform trials, with Kyle Wathen (Cytel) chairing. We generated great discussion with the ENAR audience, and we look forward to more conversations at these upcoming Efficiency+ sessions:

• IBC 2026 (July, Seoul) - Invited session introducing Efficiency+ and showcasing member research on enrollment forecasting, AI-driven monitoring, and statistical leadership in operations

• JSM 2026 (August, Boston) - Accepted invited session on “Advancing Clinical Trial Operations through Innovative Statistical Methodologies”

• RISW 2026 (September, Rockville) - Accepted invited session on “Advanced Analytics for Operational Excellence” covering recruitment, monitoring, design-operations interplay, and supply chain optimization

• SCOPE Europe 2026 (October, Barcelona) - Planning to submit a session proposal to this premier clinical operations summit, bringing our work directly to the operations community

Simulation Tools. We are developing a clinical trial simulation engine in R (the “CSC” project) that provides end-to-end simulation capabilities, from site activation and patient enrollment through randomization, visit scheduling, and drug supply forecasting. The engine features 31 country-specific parameter types and Bayesian MCMC forecasting. A key strength of the tool is its scenario planning capability: supply chain planners can explore the trade-off between probability of stock-out and amount of drug waste under different study designs and forecast/supply strategies, enabling more informed decision-making before and during a trial.

Importantly, the CSC platform is not just a standalone tool. It is our integration backbone. Each of our four topic teams works on different aspects of trial operations, and the simulation engine serves as the common platform that ties all their work together. The approach is “specification-first”: each team defines what their methodology improves (e.g., adaptive site activation, ML-based enrollment forecasting, kit-specific safety stock policies) as structured specifications, and AI-assisted coding tools translate those specs into working simulation modules. This allows methodology experts to contribute without the necessity of writing code, while enabling rigorous comparison of isolated versus integrated improvements. The vision is one platform, four teams, infinite scenarios, where simulation becomes the common language that proves the value of each team’s contribution and reveals synergies across them. This tool is still work in progress, with active development ongoing across the working group.

Reading Group. We run a regular reading group where members discuss recent literature on clinical trial operations, keeping the group current on methods and sparking new research ideas.

Find Us Online

We are actively building our community presence beyond conferences:

• LinkedIn: Connect with us at @efficiency-plus-trials (linkedin.com/company/efficiency-plus-trials) for updates on conference sessions, working group news, and member spotlights

• Website: efficiencyplustrials.github.io serves as a hub for our mission, focus areas, and resources

• GitHub: github.com/efficiencyplustrials - our open-source tools and materials

We promote our work through these channels and through the ASA BIOP section’s own communications, including the AMSTAT newsletter and conference announcements.

Where We Are Headed

Looking ahead, our roadmap includes:

• Special journal issue - organizing a collection of white papers from each of our topic subteams into a special issue of a peer-reviewed journal (exact venue to be determined), providing a comprehensive reference for the field. Each subteam is actively drafting papers as expected deliverables for 2026

• Webinars to share our findings with the broader community

• Expanded simulation tools - continuing development of our clinical trial simulation engine with additional modules for outcome metrics, real-time dashboards, and scenario optimization

• Cross-working group collaboration - working with other ASA BIOP SWGs (e.g., Safety, Bayesian, Software Engineering) on shared challenges

Join Us

If any of this resonates with you, whether you are a statistician who has been pulled into operational discussions, a clinical operations professional curious about what advanced analytics can do, or a data scientist looking for impactful applications, we would love your input.

We are especially looking for:

• Regulatory agency participants - perspectives from FDA, EMA, and other agencies are invaluable for ensuring our methods align with evolving guidelines like ICH E6(R3)

• Clinical trial operations professionals - people who live these challenges daily and can ground our statistical work in operational reality

• Statisticians and data scientists with experience in Bayesian modeling, AI/ML, clinical trial simulation, or drug supply chain analytics

Reach out to me at fchen6@its.jnj.com, visit our website at efficiencyplustrials.github.io, or find us at any of our upcoming conference sessions. We are building something that we believe can genuinely change how our industry runs clinical trials. The more perspectives we have at the table, the better.

Acknowledgments

The author thanks all members of the Efficiency+ Scientific Working Group for their contributions and enthusiasm in the group’s first year. Efficiency+ was co-founded by Fei Chen (Johnson & Johnson) and Bohdana Ratitch (Bayer), who jointly developed the original proposal and led the working group through its formation. Bohdana has since stepped down from her co-chair role, and her contributions to launching this initiative are gratefully acknowledged. Clara Cali Mella (Bayer) has stepped in as co-chair to continue the partnership. Special thanks to our group advisors; Jose Pinheiro, Inna Perevozskaya (BMS), and Benjamin Hofner (PEI), for their guidance; to the workstream leads; Clara Cali Mella, Vlad Anisimov, Palanikumar Ravindran, Dooti Roy, and Christi Kleoudis; to our communication leads, Crystal Shaw and Leanne Goldstein (Amgen), for building our presence on LinkedIn; and to Kyle Wathen (Cytel) for chairing our ENAR session. Thanks also to the ASA Biopharmaceutical Section for supporting this initiative. We are grateful to each of our member companies for investing their people’s time and expertise in this initiative - Efficiency+ exists because our sponsors recognize that improving trial operations is a shared challenge worth solving together. A full list of Efficiency+ members and their affiliations can be found on our GitHub repository at github.com/efficiencyplustrials.

Subscribe now

References

Anisimov, V. (2020). Modern analytic techniques for predictive modeling of clinical trial operations. In O. Marchenko & N. Katenka (Eds.), Quantitative methods in pharmaceutical research and development (pp. 361–408). Springer. https://doi.org/10.1007/978-3-030-48555-9_8

Efficiency+ Scientific Working Group. (2025). Efficiency+: Enhancing clinical trial operations through advanced statistics.

https://efficiencyplustrials.github.io/

Lamberti, M. J., Dirks, A., Kikuchi, N., et al. (2024). Benchmarking site activation and patient enrollment. Therapeutic Innovation & Regulatory Science, 58, 696–703. https://doi.org/10.1007/s43441-024-00638-1

McKinsey & Company. (2021). Clinical supply chains: How to boost excellence and innovation. https://www.mckinsey.com/industries/life-sciences/our-insights/clinical-supply-chains-how-to-boost-excellence-and-innovation