Overview of Regulatory Guidelines Related to Methods and Models for Small Populations Authors: Laurens Sluijterman, Letao Li, Kristin E. Karlsson, Vincent Damotte, Sarah Zohar, Kit Roes, Joanna IntHout Published: 2026 ABSTRACT This paper provides an overview of approaches and related guidelines from the EMA, FDA, and ICH that are relevant for obtaining clinical evidence for small populations to support regulatory decision making. The approaches are grouped into three categories: (1) maximizing the attainable sample size, (2) adapting the trial design to deal with a small sample size, and (3) including external information, while also examining the role of modeling. Guidelines for clinical trials, guidelines specifically for small populations, as well as guidelines that include the use of modeling to support clinical evidence were included in the review. A recurring theme throughout the guidelines is the critical importance of the clinical context, affecting both the evidentiary standard and which approaches and models are acceptable. Two key gaps were identified: (1) guidance on Bayesian approaches is limited, and (2) there is no clear guidance on how to synthesize evidence across different modalities, such as trial results and Real‐World Data, from both a practical and regulatory perspective. Study Highlights WHAT IS THE CURRENT KNOWLEDGE ON THE TOPIC? Numerous approaches have been proposed for confirmatory trials in small populations, including innovative trial designs, the use of external information, and advanced modeling. However, translating these wide‐ranging approaches into practice requires a clear overview of the relevant regulatory guidance. WHAT QUESTION DID THIS STUDY ADDRESS? This paper examined which methodological approaches are available for obtaining confirmatory evidence in small populations, which EMA, FDA, and ICH guidelines are relevant to their use, and where important gaps in regulatory guidance remain. WHAT DOES THIS STUDY ADD TO OUR KNOWLEDGE? This paper provides a concrete overview of methodological approaches and related regulatory guidance, structured around maximizing the attainable sample size, adapting the trial design and analysis to a limited sample size, and incorporating external information, while also considering modeling and evidence synthesis. It identifies the importance of clinical context and prespecification as recurring regulatory principles and highlights limited regulatory guidance on Bayesian approaches and on synthesizing evidence across different modalities. HOW MIGHT THIS CHANGE CLINICAL PHARMACOLOGY OR TRANSLATIONAL SCIENCE? This overview may support more informed choices that align with regulatory expectations when designing and analyzing confirmatory studies in small populations. Study Highlights WHAT IS THE CURRENT KNOWLEDGE ON THE TOPIC? WHAT IS THE CURRENT KNOWLEDGE ON THE TOPIC? WHAT IS THE CURRENT KNOWLEDGE ON THE TOPIC? Numerous approaches have been proposed for confirmatory trials in small populations, including innovative trial designs, the use of external information, and advanced modeling. However, translating these wide‐ranging approaches into practice requires a clear overview of the relevant regulatory guidance. WHAT QUESTION DID THIS STUDY ADDRESS? WHAT QUESTION DID THIS STUDY ADDRESS? WHAT QUESTION DID THIS STUDY ADDRESS? This paper examined which methodological approaches are available for obtaining confirmatory evidence in small populations, which EMA, FDA, and ICH guidelines are relevant to their use, and where important gaps in regulatory guidance remain. WHAT DOES THIS STUDY ADD TO OUR KNOWLEDGE? WHAT DOES THIS STUDY ADD TO OUR KNOWLEDGE? WHAT DOES THIS STUDY ADD TO OUR KNOWLEDGE? This paper provides a concrete overview of methodological approaches and related regulatory guidance, structured around maximizing the attainable sample size, adapting the trial design and analysis to a limited sample size, and incorporating external information, while also considering modeling and evidence synthesis. It identifies the importance of clinical context and prespecification as recurring regulatory principles and highlights limited regulatory guidance on Bayesian approaches and on synthesizing evidence across different modalities. HOW MIGHT THIS CHANGE CLINICAL PHARMACOLOGY OR TRANSLATIONAL SCIENCE? HOW MIGHT THIS CHANGE CLINICAL PHARMACOLOGY OR TRANSLATIONAL SCIENCE? HOW MIGHT THIS CHANGE CLINICAL PHARMACOLOGY OR TRANSLATIONAL SCIENCE? This overview may support more informed choices that align with regulatory expectations when designing and analyzing confirmatory studies in small populations. Clinical trials to evaluate medicines for rare or pediatric diseases face unique challenges. They are typically smaller 1 and it may be infeasible to conduct a full randomized controlled trial (RCT). 2 Providing confirmatory clinical evidence with such a small and possibly non‐standard trial is difficult. Proposed solutions range from innovative trial designs and leveraging external data sources to the use of complex computational models 3 , 4 , 5 , 6 to provide confirmatory evidence. For challenging scenarios—such as studying new medicines in very rare pediatric diseases—a combination of all these approaches may be necessary, but it is unclear how to navigate the multitude of approaches and regulatory guidelines. Thus, as part of the research project INVENTS, 7 the goal of this paper is to capture (1) methodological approaches for obtaining confirmatory evidence in small populations, (2) related regulatory guidelines, and (3) potential gaps in guidance. The scope is restricted to guidelines from the EMA, FDA, and harmonized guidance from the ICH. An existing framework for small‐sample clinical trials, first presented by Parmar et al., 8 was modified to organize the methodological approaches and related guidance considered in this paper. The original framework starts from the premise that RCTs are the gold standard for evidence‐based decision making for (new) medicines, and deviations from this standard may come at the cost of increased uncertainty about treatment effects (positive and negative) and hence less informed decision making. The Parmar et al. framework suggests new design alternatives, progressively departing from the standard RCT, if certain aspects are not feasible in a given context. Implementation of these adaptations is to be continued until a design solution has been reached that still attains an acceptable level of evidence. The first step, also proposed by Day et al., 9 is to assess whether a “conventional” RCT is feasible. This is defined here as an adequately powered study with a clinically relevant end point and type‐I error control at the two‐sided 0.05 level. If this is feasible, a “conventional” RCT is recommended. If necessary to make a “conventional” RCT possible, Parmar et al. suggest broadening patient eligibility criteria if possible, increasing accrual and follow‐up time, or collaborating, for instance in a multi‐center trial. These approaches all aim to maximize the attainable trial size, but they may lead to answering a somewhat different research question. However, the maximum feasible sample size may still be very small, in which case the trial design and analysis must be adapted to deal with this limitation. A more efficient design could be considered, for example by measuring patients more frequently and making full use of longitudinal data. Though more expensive and more demanding for the patients, this likely results in a smaller required sample size by extracting more information from the limited sample size. Other options include not dichotomizing a continuous end point or switching to another end point altogether, although this comes at the cost of altering the research question. If this is still insufficient, exploring less common approaches such as innovative designs, including adaptive trials or master protocols, or leveraging external information may be necessary. 8 The modified version of this framework used here to structure the overview of methods and guidelines is illustrated in Figure 1 . All approaches are grouped into three main categories: maximizing the attainable sample size, dealing with a limited sample size, and incorporating external information. The bounding box emphasizes that the clinical context is always relevant, which is defined here as a combination of the disease, experimental treatment, and standard of care. The acceptability of innovative trial designs, modeling, and leveraging external information for regulatory decision making depends on the clinical context. An overview of the different approaches to obtain confirmatory clinical evidence for small populations. The methods are grouped around the three strategies of (1) maximizing the attainable trial size, (2) dealing with the small sample size, and (3) leveraging external information, potentially using modeling. Evidence synthesis combines the different sources of information. The bounding box visualizes that everything takes place in an all‐encompassing clinical context. The multiple modeling icons illustrate the potential applications of computational models. Their deliberate placement at the arrows, rather than as a separate block, underscores that models do not provide evidence in isolation. Instead, they utilize data, whether from the trial or external sources, as input to generate new insights. Finally, multiple sources of information can be combined during evidence synthesis, possibly using modeling. A meta‐analysis is a common example, but with the use of increasingly complex models and real‐world data, new challenges arise in combining these different modalities of evidence. The remainder of this paper is structured as follows. The Methods section outlines the specific scope and inclusion and exclusion criteria of our review. Subsequently, in the Results section, an overview of relevant methods and related guidance is provided, organized according to the elements shown in Figure 1 . First, the evidentiary standard for confirmatory clinical evidence for licensure is described, followed by the three strategies to reach this evidence. Subsequently, the use of modeling and evidence synthesis are addressed. Each subsection is divided into two parts: the first part summarizes the available methods and guidelines, and the second part highlights key regulatory principles. Finally, overarching aspects, such as the central role of the clinical context and identified gaps in guidance, are discussed. METHODS For the review, guidance from the ICH, EMA, and FDA was considered from the following sources: The ICH guidance database 10 The scientific EMA guidance under Biostatistics, General, and Real‐World Evidence 11 The FDA clinical trial guidelines 12 These sources were small enough for a fully exhaustive search. An additional keyword search through the entire FDA guidelines database 13 was performed using the following keywords: Adaptive, Artificial, Bayesian, External, In silico, Innovative, Kinetics, Master, Rare, Real World, Sequential, Small Sample. Included guidelines have relevant content related to: Increasing the trial sizeMaking the trial more efficientIncluding external informationUsing modeling techniquesThe following guidance documents were excluded: Guidelines on trial conduct, monitoring, laboratory standardsConcept papersHighly specific subgroup guidelines (e.g., impaired kidney function)Pre‐ and nonclinical guidelines The ICH guidance database 10 The ICH guidance database 10 The scientific EMA guidance under Biostatistics, General, and Real‐World Evidence 11 The scientific EMA guidance under Biostatistics, General, and Real‐World Evidence 11 The FDA clinical trial guidelines 12 The FDA clinical trial guidelines 12 Increasing the trial size Increasing the trial size Making the trial more efficient Making the trial more efficient Including external information Including external information Using modeling techniques Using modeling techniques Guidelines on trial conduct, monitoring, laboratory standards Guidelines on trial conduct, monitoring, laboratory standards Concept papers Concept papers Highly specific subgroup guidelines (e.g., impaired kidney function) Highly specific subgroup guidelines (e.g., impaired kidney function) Pre‐ and nonclinical guidelines Pre‐ and nonclinical guidelines RESULTS Evidentiary standard Before presenting approaches to provide confirmatory clinical evidence, the concept of evidentiary standard must be understood. Multiple guidelines and regulations 14 , 15 , 16 , 17 , 18 address this topic. Crucially, the scientific data needed to achieve the required evidentiary standard depend on the context. Historically, multiple independent and well‐controlled trials have been upheld as the norm 14 , 16 , but it is recognized that there are many contexts in which one confirmatory (pivotal) trial has sufficed. 19 In such cases, the evidence from the single trial is expected to be stronger (e.g., lower significance threshold) and additional evidence (on disease and/or medicine mechanisms) may be required. Moreover, it is recognized that in certain situations such as a serious disease with an unmet medical need or during emergency situations other approaches to generate evidence could be considered. 15 However, this is not to say that less conventional approaches are always acceptable for rare diseases. In fact, the EMA states that “No methods exist that are relevant to small studies that are not also applicable to large studies”. 20 This may seem discouraging, but this statement is immediately followed by one of a different sentiment: “However, it may be that in conditions with small and very small populations, less conventional and/or less commonly seen methodological approaches may be acceptable if they help to improve the interpretability of the study results.” This highlights that guidelines like the ICH E8, E9, and E10 21 , 22 , 23 are equally relevant for small‐sample trials, for which the EMA and FDA also have dedicated guidance. 20 , 24 Table 1 provides a chronological overview of all regulatory documents mentioned in this subsection. It also lists the documents referenced in the following subsections on the various approaches to providing clinical evidence for small populations. Table 1Overview of regulatory documents discussed in the Results section. Documents are grouped by the same subsections, with chronological ordering within each groupYearTitleSourceReferencesGeneral trial design and evidentiary standards1998Providing Clinical Evidence of Effectiveness for Human Drug and Biological ProductsFDA 16 1998/2019Statistical Principles for Clinical Trials E9 and the corresponding addendumICH 21, 25 2000Choice of Control Group in Clinical Trials E10ICH 22 2001Points to consider on application with 1.meta‐analyses, 2.one pivotal studyEMA 19 2001Directive 2001/83/ECEU parliament and council 14 2004EU regulation 726/2004EU parliament and council 15 2006Guideline on clinical trials in small populationsEMA 20 2019Demonstrating Substantial Evidence of Effectiveness for Human Drug and Biological Products (Draft)FDA 17 2021General Considerations for Clinical Studies E8ICH 23 2023Demonstrating Substantial Evidence of Effectiveness With One Adequate and Well‐Controlled Clinical Investigation and Confirmatory Evidence Guidance for Industry (Draft)FDA 18 2023Rare Diseases: Considerations for the Development of Drugs and Biological Products – Guidance for IndustryFDA 24 Maximizing the attainable sample size2017General Principles for Planning and Design of Multi‐Regional Clinical Trials E17ICH 26 2022Recommendation paper on decentralized elements in clinical trialsEC/HMA/EMA 27 2024Conducting Clinical Trials With Decentralized Elements – Guidance for Industry, Investigators, and Other Interested PartiesFDA 28 Adapting to a small‐sample size2007Reflection paper on methodological issues in confirmatory clinical trials planned with an adaptive designEMA 29 2012Guideline on the use of pharmacogenetic methodologies in the pharmacokinetic evaluation of medicinal productsEMA 30 2015Guideline on adjustment for baseline covariatesEMA 31 2019Adaptive Designs for Clinical Trials of Drugs and Biologics – Guidance for IndustryFDA 32 2022Master Protocols: Efficient Clinical Trial Design Strategies to Expedite Development of Oncology Drugs and Biologics – Guidance for IndustryFDA 33 2023Master Protocols for Drug and Biological Product Development (Draft)FDA 34 2023Adjusting for Covariates in Randomized Clinical Trials for Drugs and Biological ProductsFDA 35 2023Pharmacogenomic Data Submissions – Guidance for Industry (Draft)FDA 36 2024Reflection paper on establishing efficacy based on single‐arm trials submitted as pivotal evidence in a marketing authorization applicationEMA 37 2025Adaptive Designs For Clinical Trials E20 (Draft)ICH 38 2026Use of Bayesian Methodology in Clinical Trials of Drug and Biological Products (Draft)FDA 39 Including external information2016Scientific guidance on post‐authorization efficacy studiesEMA 40 2018Reflection paper on the use of extrapolation in the development of medicines for pediatricsEMA 41 2018Use of Electronic Health Record Data in Clinical Investigations – Guidance for IndustryFDA 42 2019Rare Diseases: Natural History Studies for Drug Development – Guidance for Industry (Draft)FDA 43 2021Guideline on registry‐based studiesEMA 44 2023Real‐World Data: Assessing Registries to Support Regulatory Decision Making for Drug and Biological Products – Guidance for IndustryFDA 45 2023Data Quality Framework for EU medicines regulationEMA 46 2023Considerations for the Design and Conduct of Externally Controlled Trials for Drug and Biological Products – Guidance for Industry (Draft)FDA 47 2023Considerations for the Use of Real‐World Data and Real‐World Evidence to Support Regulatory Decision Making for Drug and Biological Products – Guidance for IndustryFDA 48 2024Real‐World Evidence: Considerations Regarding Non‐Interventional Studies for Drug and Biological Products – Guidance for Industry (Draft)FDA 49 2024Real‐World Data: Assessing Electronic Health Records and Medical Claims Data to Support Regulatory Decision Making for Drug and Biological Products – Guidance for IndustryFDA 50 2024Pediatric Extrapolation E11AICH 51 2025General Principles on Planning, Designing, Analyzing, and Reporting of Non‐interventional Studies That Utilize Real‐World Data for Safety Assessment of Medicines M14ICH 52 2025Reflection paper on use of real‐world data in non‐interventional studies to generate real‐world evidenceEMA 53 Modeling2008Guideline on reporting the results of population pharmacokinetic analysesEMA 54 2018Physiologically‐Based Pharmacokinetic Analyses — Format and Content Guidance for IndustryFDA 55 2019Guideline on the reporting of physiologically‐based pharmacokinetic (PBPK) modeling and simulationEMA 56 2022Population Pharmacokinetics – Guidance for IndustryFDA 57 2023Assessing the Credibility of Computational Modeling and Simulation in Medical Device Submissions – Guidance for Industry and Food and Drug Administration StaffFDA 58 2024Reflection paper on the use of Artificial Intelligence (AI) in the medicinal product lifecycleEMA 59 2025Considerations for the Use of Artificial Intelligence to Support Regulatory Decision Making for Drug and Biological Products – Guidance for Industry and Other Interested Parties (Draft)FDA 60 2026General Principles for Model‐Informed Drug Development M15ICH 61 Evidence synthesis2001Points to consider on application with 1. meta‐analyses, 2. one pivotal studyEMA 19 2018Meta‐Analyses of Randomized Controlled Clinical Trials to Evaluate the Safety of Human Drugs or Biological Products Guidance for IndustryFDA 62 Overview of regulatory documents discussed in the Results section. Documents are grouped by the same subsections, with chronological ordering within each group Maximizing the attainable sample size There are various strategies to maximize the feasible sample size, including broadening eligibility criteria, increasing accrual or follow‐up time, and extending collaboration. Increasing accrual and follow‐up time is particularly helpful for time‐to‐event data as increasing the number of observed events greatly improves power. 8 However, this increases the price and, more importantly, the time it takes to reach the patient. Broadening eligibility criteria, also recommended for rare diseases by the FDA, 24 is a balancing act. Relaxing them increases the recruitment base 21 but simultaneously decreases homogeneity, which complicates estimating the treatment effect and potentially decreases power to detect the relevant effect size. Making the trial more attractive for patients can also increase participation. Not using a standard 1:1 randomization ratio increases the odds of receiving treatment, though generally at the cost of an increased required sample size. A decentralized trial, where certain parts are conducted outside the central trial site, reduces travel times. Context is again relevant: depending on the nature of the disease, patients may be less willing, or able, to travel or less willing to take the risk of being randomized to the control arm. A guideline on decentralized trials has been issued by the FDA, 28 and a recommendation paper has been published by a collaboration of the EMA, Heads of Medicines Agencies, and European Commission. 27 Finally, collaborating via a multi‐center or multi‐regional trial may be essential. ICH E9 has a dedicated section on multi‐center trials. Besides the increased sample size, generalization may be improved because subjects are recruited from a wider population and investigated in a broader range of clinical settings. 21 Additionally, as mentioned in the dedicated ICH guideline, 26 a multi‐regional clinical trial (MRCT) can facilitate simultaneous marketing authorization applications in different regions, providing earlier world‐wide access. Main regulatory considerations for maximizing the attainable sample size The guidelines offer multiple considerations on eligibility criteria, decentralized trials, and multi‐center trials. Eligibility criteria must naturally be chosen with the proposed indication of use in mind and cannot haphazardly be relaxed to increase sample size, particularly when broader inclusion would introduce clinically meaningful heterogeneity and undermine the interpretation of efficacy or safety findings. 21 , 23 , 24 For decentralized trials, both European and FDA guidance emphasize that decentralized elements must not increase risks to patients, and the resulting data must meet the same expectations as for on‐site procedures. For multi‐center or regional trials, an important consideration is regional variability 26 that may arise, for example, due to differential principles for trial recruitment, differential concomitant or supportive care, as well as physiological differences. It is therefore recommended to stratify per center in the statistical analysis, or per region if the sample sizes per center are too small. 31 Adapting design and analysis to a small sample size Even the maximum attainable sample size may be a rather small one. The trial design must be adapted to provide reliable evidence despite this limitation. Although efficient trial design is always important, it becomes essential in a small population setting. Relevant guidelines are the ICH E8, E9, and E10. Basic elements include selecting an informative end point and a treatment arm with a reasonable expected difference to the control, although making changes here effectively changes the research question. It may be possible to carry out more extensive measurements of baseline covariates or make more frequent measurements to facilitate a longitudinal analysis. Adjusting for baseline covariates can improve efficiency. Baseline covariates, especially those with a modest or strong association with the primary outcome, can explain part of the variance, thereby improving power. The ICH E9 briefly touches upon the topic but does not go into detail. These details can be found in the dedicated guidelines from the EMA and FDA. 31 , 35 Additionally, the EMA and FDA have specific guidelines on the use of genetics. 30 , 36 A different strategy is to consider alternative trial designs, as visualized in Figure 2 . For specific scenarios, it may be possible to conduct a single‐arm trial (SAT) where all patients are enrolled in the treatment arm. The EMA has a dedicated reflection paper on SATs, 37 describing specific conditions where they could provide a causal interpretation of the observed effects. In crossover and N‐of‐1 trials, patients receive both the experimental treatment and the control. This typically reduces the required sample size and makes the trial more attractive by guaranteeing receiving treatment. The latter designs are more suitable for symptomatic treatments where treatment provides a rapid and measurable short‐term response. Carryover and period effects are typical concerns, again exemplifying how the nature of the disease and treatment dictate what is possible. Various examples of alternative trial designs. (a) Illustrates a single‐arm trial, where all participants receive treatment, and crossover and N‐of‐1 trials, where individuals receive both the treatment and control, potentially repeated over multiple periods. (b) Demonstrates the three main types of master protocols. An umbrella trial studies multiple therapies for a single disease, a basket trial studies a single therapy in the context of multiple diseases or disease subtypes, and a platform trial studies multiple therapies for a single disease in a perpetual manner where therapies are continuously added or dropped. Figure 2 b visualizes the three main types of master protocols—trials that combine multiple sub‐trials into an overarching structure. A basket trial groups multiple diseases or subtypes together that share a common biological feature such as a gene mutation. 6 Umbrella and platform trials may be relevant to rare diseases where it may be impossible to investigate multiple therapies in separate competing trials. Setting up the necessary infrastructure, the added planning requirements, and the greater coordination between stakeholders all come at a cost. 63 , 64 The FDA has two guidance documents on master protocols: one general (draft), 34 and one aimed at oncology. 33 Additionally, efficiency could be increased by incorporating design elements that are adapted based on observed interim results. The EMA defines this as follows in a reflection paper 29 : “A study design is called ‘adaptive’ if statistical methodology allows the modification of a design element (e.g. sample‐size, randomisation ratio, number of treatment arms) at an interim analysis with full control of the type I error.” The FDA guideline 32 defines it as “[A] clinical trial design that allows for prospectively planned modifications to one or more aspects of the design based on accumulating data from subjects in the trial.” The ICH recently issued a draft guideline. 38 Many different types of adaptations exist. The FDA lists six types of adaptations: (1) group sequential designs, where a trial is planned with prospectively planned interim analyses with prespecified stopping rules; adaptations related to (2) sample size; (3) the patient population; (4) treatment arm selection, for example adding or removing treatment arms; (5) patient allocation; (6) end‐point selection, changing the choice of the primary end point based on interim results. For a comprehensive overview of adaptive designs, we refer to Bauer et al. and Kaizer et al. 65 , 66 Potential gains of adaptive elements may be accompanied by losses in other areas. For example, a group sequential trial may stop earlier but may also need a larger maximum sample size. An adaptive design also comes with additional logistical difficulties, and efficiency gains could be small if the required follow‐up is long. Main regulatory considerations for adapting to a small sample size On baseline covariates, the EMA and FDA guidance is relatively aligned. Any baseline covariates used in the analysis must be (1) prespecified, (2) likely to be at least moderately associated with the primary outcome, and (3) measured prior to randomization whenever they are potentially affected by the trial. For master protocols, the only guidance is provided by the FDA, which advises using only concurrent controls in the analysis and excluding patients who could not have been randomized to the arms that are being analyzed. The FDA also advises allocating more patients to the control than to the individual treatment arms, as this can increase the power of the individual comparisons given a fixed sample size. Multiplicity adjustments to strongly control type‐I error may not always be required, especially in settings where the comparisons would otherwise have been done in separate trials. If SATs are to be used as pivotal clinical evidence, it is essential that results can be causally interpreted. Therefore, design, choice of end point and other aspects need to be such that the treatment effect can be reliably isolated without direct comparison to a control. For instance, improvement on the defined end‐point needs to be highly unlikely to occur spontaneously, in the absence of treatment, for example, tumor shrinkage. 37 This may limit the type of end point that can be used in comparison to an RCT. For example, although biomarkers can be objectively measured, they may often not be able to isolate the treatment effect in a clinically meaningful way in an SAT. For adaptive designs, the EMA, FDA, and ICH all state that it is crucial that type‐I error is controlled (this is even in the EMA definition of a confirmatory adaptive trial) and that trial integrity is maintained. Integrity is a concern since adaptations may give indications of ongoing trial results. Additionally, adaptations should always be prespecified and limited in number. Adaptive designs and master protocols may potentially have Bayesian elements. As discussed later, Bayesian approaches are also used to incorporate external information. However, while there are multiple best practice papers 67 , 68 and unofficial guidelines (e.g., BayesWatch, 69 ROBUST, 70 and BARG 71 ), official regulatory guidance on Bayesian approaches is limited. The ICH E9 only states that “the use of Bayesian […] and other approaches may be considered when the reasons for their use are clear and when the resulting conclusions are sufficiently robust”, a statement also found in the draft ICH E20. In other words, a clear scientific justification for the use of Bayesian approaches is needed, but more detailed guidance, besides the FDA guidance, which is still in draft status, on how to apply them in the context of regulatory decision making is lacking. Including external information It may still be impossible to meet evidentiary standards solely by maximizing the attainable size and adapting the trial design to deal with a small sample. External information could potentially bridge the remaining gap. Figure 1 lists three broad groups of external information. A first source comprises preclinical and clinical data. Preclinical data, such as in vitro, animal, toxicological, or pharmacokinetic data, may be incorporated through PK/PD or other mechanistic models, as discussed later. Clinical data may come either from an earlier phase, such as in a seamless phase I–II trial, or from an entirely different randomized trial, for example, when control‐arm data are used as an external source of information. A second source is real‐world data (RWD), defined by the FDA as “data regarding the usage, or potential benefits or risks, of a drug derived from sources other than traditional clinical trials” 72 and by the EMA 73 as “routinely collected data relating to a patient's health status or the delivery of healthcare from a variety of sources other than traditional clinical trials.” Examples include electronic health records, patient registries, and medical claims data. The third category consists of opinions and expertise. These can come in the form of expert opinions, regulatory opinions, and patient preferences. 74 This includes domain knowledge such as mechanistic understanding of the disease and treatment. External information is always used in trial design, for instance to define eligibility criteria, motivate end‐point selection, or estimate the sample size. Besides this standard use, it might also be used to provide evidence, for instance by (partially) replacing the control arm with external information, as is illustrated in Figure 3 . The FDA has a dedicated draft guideline 47 and ICH E10 discusses several advantages and disadvantages of external control arms. The disadvantages are substantial: In the absence of randomization, the absence of bias cannot be ascertained. Moreover, the study cannot be blinded. 75 On the other hand, everyone in the current trial receives the active treatment, which may be more appealing to patients. The spectrum of leveraging external controls. This ranges from an RCT and SAT that rely on trial results only, to a fully externally controlled trial, and everything in between. In this graphic, we visualize only the use of external control arms, but it is also possible to leverage external information on the treatment effect directly. Instead of a full external control, it is also possible to combine a (smaller) concurrent control arm with an external control arm 76 or to use external information to inform treatment‐effect estimation directly. Both frequentist and, more frequently, Bayesian techniques can be used to incorporate external information. 77 , 78 , 79 , 80 A known issue with the so‐called Bayesian borrowing approaches is that the type‐I error can generally not be strictly controlled. 81 , 82 As mentioned earlier, definitive guidance for these Bayesian methodologies is lacking. There are situations, especially in pediatrics, where the sample size of newly collected data – whether in an RCT, SAT, or externally controlled trial—may inevitably be too limited to provide independent evidence. Reliance is often placed, at least in part, on the extrapolation of other trial results, for example from an adult trial of the same drug and disease. In the context of pediatrics, where it is frequently used, both the EMA and ICH have guidance on extrapolation. 41 , 51 In some cases, development programs rely not only on prospectively collected clinical trial data but also on RWD. A review by Chen et al. 83 presents many different applications, one example being natural history studies for which the FDA has a dedicated guideline. 43 As part of the recent real‐world evidence (RWE) program, 72 the FDA has general guidance on RWD. 48 Additionally, guidance for specific data sources like electronic health records, medical claims data, and registries has also been provided by both the FDA and the EMA. 42 , 44 , 45 , 50 The importance of data quality is emphasized, for which the Data Quality Framework published by the EMA 46 is considered highly relevant. RWD can also be used to assess efficacy and safety in non‐interventional studies, 84 defined as “a type of study in which patients received the marketed drug of interest during routine medical practice and are not assigned to an intervention according to a protocol”. 49 The EMA, FDA, and ICH have (draft) guidance on this type of study. 40 , 49 , 52 , 53 , 85 Main regulatory considerations for including external information The EMA acknowledges that what is known about a (related) product and disease mechanisms influences the required amount of confirmatory phase III data. 19 The FDA similarly recognizes that various sources of external information—such as scientific knowledge about similar drugs, natural history studies, or mechanistic understanding—can contribute to evidence in combination with an adequate and well‐controlled clinical investigation. 17 More specifically, the FDA writes that sufficient evidence could be obtained by: “One adequate and well‐controlled clinical investigation on a new indication for an approved drug, supported by existing adequate and well‐controlled clinical investigation(s) that demonstrated the effectiveness of the drug for its other, closely related approved indication(s)”“One adequate and well‐controlled clinical investigation supported by data that provide strong mechanistic support”“One adequate and well‐controlled clinical investigation with compelling results, supported by additional data from the natural history of the disease”“One adequate and well‐controlled clinical investigation of the new drug, supported by scientific knowledge about the effectiveness of other drugs in the same pharmacological class” “One adequate and well‐controlled clinical investigation on a new indication for an approved drug, supported by existing adequate and well‐controlled clinical investigation(s) that demonstrated the effectiveness of the drug for its other, closely related approved indication(s)” “One adequate and well‐controlled clinical investigation on a new indication for an approved drug, supported by existing adequate and well‐controlled clinical investigation(s) that demonstrated the effectiveness of the drug for its other, closely related approved indication(s)” “One adequate and well‐controlled clinical investigation supported by data that provide strong mechanistic support” “One adequate and well‐controlled clinical investigation supported by data that provide strong mechanistic support” “One adequate and well‐controlled clinical investigation with compelling results, supported by additional data from the natural history of the disease” “One adequate and well‐controlled clinical investigation with compelling results, supported by additional data from the natural history of the disease” “One adequate and well‐controlled clinical investigation of the new drug, supported by scientific knowledge about the effectiveness of other drugs in the same pharmacological class” “One adequate and well‐controlled clinical investigation of the new drug, supported by scientific knowledge about the effectiveness of other drugs in the same pharmacological class” A fundamental principle shared across regulatory agencies is that the reliance on external information and the acceptability of its sources are highly context‐dependent. For instance, external controls may be deemed acceptable only in specific settings due to their inherent drawbacks. The ICH E8 guideline states that external controls should be considered only when other design options are unacceptable, such as when there is a strong prior belief in superiority. Additionally, the treatment effect must be substantial, and disease progression must be highly predictable, as also emphasized by the FDA, 24 although such judgments can often be made only retrospectively. Similarly, a data source is never universally acceptable but must be fit‐for‐purpose. This concept is fundamental and serves as the basis for the EMA's definition of data quality 46 : “fitness for purpose for users' needs in relation to health research, policy making, and regulation and that the data reflect the reality, which they aim to represent.” In line with this, fit‐for‐purpose itself is defined as: “Possessing all required data quality characteristic needed to address a specific goal.” Modeling The use of models plays a central part in many drug development programs to include and integrate sources of data, prior information, and knowledge to generate evidence. 61 For small samples, their relative importance increases and the ways in which they are used may shift. For example, in the setting of pediatric extrapolation, model‐based analyses often provide pivotal evidence. 51 The focus in this paper is on more advanced (mechanistic) pharmacological models, rather than on standard statistical models used in the analyses of clinical trials, such as linear and survival models. These models are diverse and cover various domains, as illustrated in Figure 4 . Many models are either PK models, modeling the effect of the body on a drug; PD models, modeling the effect of a drug on the body; disease progression models, modeling the time evolution of a disease; or broader mechanistic models, such as QSP models, that capture interactions between a drug, the body, and biological mechanisms underlying a disease, potentially at a molecular level. Domains of pharmacological models. The top row illustrates several commonly used modeling domains, including pharmacokinetic (PK) models, which describe drug concentration over time; pharmacodynamic (PD) models, which relate the drug concentration to a biological effect; and disease progression models, which describe the natural evolution of a disease. Additional domains exist, and multiple domains can be integrated within a single framework, as is done in QSP modeling. The bottom row illustrates that, within each domain, multiple model types can be defined. As an example, three different PK model types are shown. Within each domain, models vary in basis and complexity. For example, PK models range from comparatively simple compartmental models relying on basic assumptions and solving differential equations to highly complex physiologically‐based pharmacokinetic (PBPK) models 86 that combine physiological measurements and drug properties. Models can accomplish wide‐ranging goals such as designing a trial, combining information from various sources, or analyzing trial results. For example, PBPK models have been used to estimate initial pediatric dosages based on adult data. 87 QSP, popPK, and PBPK models have also been used to estimate treatment effects or dosages. 88 , 89 , 90 Simpler models, like (semi‐mechanistic) PK/PD models and dose–response models, can be used to obtain as much information from trial results as possible. The suitability of a model depends on its intended purpose. 91 A central guideline is the ICH M15, 61 which shares many elements with previous papers on the subject, 92 , 93 , 94 which in turn are influenced by the ASME V&V 40. 95 The FDA has guidance on the credibility of modeling and simulation in a medical device setting. 58 Although medical devices are out of scope, it contains insights into the FDA's thinking on topics such as risk assessment and defining a Question of Interest and the Context of Use, all central parts of the ICH M15. For popPK and PBPK models, the FDA and EMA have dedicated guidelines. 54 , 55 , 56 , 57 More recently, both agencies have released a reflection paper 59 and draft guideline, 60 respectively, and a set of common principles on Artificial Intelligence. 96 Main regulatory considerations for modeling Figure 5 visualizes the important concepts from the ICH M15 guidance. The process begins by defining the Question of Interest that the model aims to address, followed by describing the Context of Use, which outlines the model's role and scope in addressing that question. Based on these, the model risk (its influence on decision making and potential consequences of error) and impact (in relation to regulatory standards) are assessed. Another essential component is the model evaluation, consisting of verification, validation, and an applicability assessment. Verification ensures correct implementation and error‐free code; validation assesses robustness and performance, including data relevance, structural assumptions, method‐specific concerns, and uncertainty; and the applicability assessment judges whether it is fit‐for‐purpose. A visualization of core concepts of the ICH M15 guideline on model‐informed drug development (MIDD). The key assessment elements include defining the Question of Interest, Context of Use, and a subsequent risk and impact analysis. During the evaluation, the model is verified, validated, and undergoes an applicability assessment. The broader framework also considers the specific technical criteria for evaluating the model and its outcomes and the appropriateness of the proposed MIDD approach. During planning, these criteria should be defined in relation to the Question of Interest and model risk, and it should be explained why the proposed approach is suitable. Once the analyses have been completed, the evaluation results are assessed against these criteria. Finally, a multidisciplinary team considers all elements and concludes whether the model outcomes constitute MIDD evidence that can be used to answer the Question of Interest. Evidence synthesis Evidence may arise from diverse sources ranging from clinical trials to RWD and complex mechanistic models. Evidence synthesis is the process of combining such sources to strengthen decision making. This can occur across studies (e.g., multiple RCTs) or within a single study integrating different data types (e.g., RCTs and RWD). Combining clinical studies is typically done through meta‐analysis, defined in the ICH E9 as “a formal evaluation of the quantitative evidence from two or more trials bearing on the same question”. Besides a more precise estimate of treatment effect or its heterogeneity, evidence synthesis can also be used to detect adverse events that individual RCTs, powered on an efficacy end point, may be underpowered to detect, as mentioned in the regulatory documents on the topic by the EMA and FDA 19 , 62 . Main regulatory considerations for evidence synthesis The EMA notes limitations of using meta‐analyses of individually inconclusive trials as pivotal evidence, favoring a single robust trial over multiple smaller ones. 20 Both the FDA and EMA emphasize the importance of a prespecified meta‐analysis protocol which describes: the objective; inclusion and exclusion criteria; the strategy for the identification of studies; the hypotheses and end points; statistical methods, including approaches to maximize data quality; and a plan to evaluate consistency and robustness. When the meta‐analysis provides pivotal evidence, a P‐value lower than the conventional 0.05 would generally be required, where the exact level is determined on a case‐by‐case basis. Although this protocol is to be written before the results of the studies are known, it is not uncommon that the desirability of a meta‐analysis only becomes apparent retrospectively, in which case more stringent criteria may apply. 19 While there is guidance on combining clinical studies, it is less clear how to quantitatively combine evidence of different modalities, such as results from an RCT and data from electronic health records. This type of formal integration is currently rarely done, but, especially for very small populations, it may be relevant to develop a structured quantitative approach to do so. DISCUSSION A recurring theme regarding methodological approaches and related guidance is the importance of context, which we defined as a combination of the experimental treatment, disease, and standard of care. First of all, the evidentiary standard is context‐dependent. Different factors are emphasized when assessing a serious rare disease without treatment as opposed to a common disease with ample effective treatments. For example, while type‐I error control at the standard level and statistical power are properties of the trial design, their importance is ultimately context dependent and should be considered at the decision‐making level both when designing and assessing the trial. Which designs are feasible depends on factors such as disease characteristics. The same context‐dependence holds for modeling approaches. Models serve to answer a specific question within a specific context and all aspects from validation metrics, to verification steps, to the data sources must align with that question within that context. Furthermore, in the RWD guidance, it is emphasized that the data must be fit‐for‐purpose. In essence, no design, model, or data source is universally applicable; it all depends on the research question and context. The second common theme is prespecification, especially for confirmatory evidence: modifications of an adaptive design must be prospectively planned, the use of RWD must be described in a prespecified protocol that addresses the suitability of the sources and approaches to minimize biases, and SATs require a prespecified success criterion. This consistent emphasis on prespecification serves to ensure that studies have the best opportunity to provide reliable estimates of treatment effects. It was noted that methodological guidance, like the ICH M15, is by design general and not tailored to specific applications. To apply the guidance, application‐specific requirements may need to be developed. For example, using a popPK model to extrapolate to pediatric populations may require a different model validation than when using it for sample‐size calculations. In the former case, ontogeny needs to be captured accurately; in the latter case, the population‐level variance must be realistic. We identified two main gaps in regulatory guidance. First, there is limited official guidance on Bayesian approaches, besides the FDA draft guideline. 39 While the FDA also has effective guidance for medical devices, 97 this is a different domain and the guidance is from 2010, prior to considerable advances in the field. Although some other guidelines, for example the ICH E11A and E9 and draft E20, briefly mention Bayesian methods, more detailed guidance is required. Secondly, evidence synthesis is becoming increasingly complicated. For very small populations, we may need to go beyond combining multiple trials, for example by incorporating RWD and complex modeling approaches. How to rigorously integrate these different information modalities remains an open question. The more complex models, such as PBPK and QSP models, arguably are a form of evidence synthesis. They combine knowledge on physiology, drug properties, and disease progression within a single system. However, this comes with new challenges. Modeling platforms are built by different research groups over multiple years, which makes it increasingly difficult to verify all the individual components and to assess whether the individual information sources are fit‐for‐purpose. CONCLUSION We have presented the available methods and related guidelines from the FDA, EMA, and ICH relevant to obtaining confirmatory clinical evidence for small populations, structured around three categories: maximizing the attainable sample size, dealing with a small sample size, and including external information. Throughout the regulatory landscape, we encountered two common themes: everything (1) is context‐ and research question dependent, and (2) should be, as much as possible, prespecified. We encountered two gaps: (1) regulatory Bayesian guidance is limited, and (2) evidence synthesis across different modalities is challenging and requires the development of more structured approaches and related guidance. FUNDING INVENTS has received funding from the European Union's Horizon Europe Research and Innovation program under grant agreement 101,136,365, the Swiss State Secretariat for Education, Research and Innovation (SERI), and the UKRI Innovate UK under their Horizon Europe Guarantee scheme. Views and opinions expressed are, however, those of the author(s) only and do not necessarily reflect those of the European Union or the European Health and Digital Executive Agency. Neither the European Union nor the granting authority can be held responsible for them. CONFLICT OF INTEREST The authors declared no competing interests for this work. AUTHOR CONTRIBUTIONS L.S., L.L., V.D., K.K., S.Z., K.R., and J.I. designed and performed the research. L.S. analyzed the data and wrote the manuscript.