How to use natural history and real-world data to augment clinical trial control groups (Part 3)

Recruiting enough patients for a conventional randomised trial can be difficult in rare disease. To say the least.

One option to address this challenge head-on is to use information from patients outside the trial. These patients might come from a natural history study, disease registry, electronic health record, insurance claims database or previous clinical trial.

But adding external patients to a control group is not a data collection exercise. It is a methodological and comparative evidence opportunity.

The external patients need to represent what would have happened to the trial patients under the control treatment. Any important differences need to be addressed. The analysis then needs to decide how much external information can safely contribute.

This is the third part of our series on borrowing external evidence. Part 1 explains why and when external information may help.[1] Part 2 explains how a meta-analytic predictive prior can borrow study-level results from historical trials.[2]

Part 3 focuses on individual patient data and a different approach: using external patients to augment the control group in a new trial.

Patient-level data can add information to a control group

A published paper may report that 60% of control patients responded. Patient-level data show who those patients were, when their follow-up began, what treatment they received and what happened to each of them.

That detail matters.

It allows the trial eligibility criteria to be applied to the external data. It also allows measured differences in age, disease severity, previous treatment and other prognostic factors to be adjusted.[3-7]

The aim is not to find as many external patients as possible. It is to identify patients who can credibly represent the control outcome in the new trial.

Fully external and hybrid controls are different designs

A fully external control has no randomised control group inside the trial. Everyone in the trial receives the experimental medicine. Their outcomes are compared with patients from an external source.

A hybrid control retains a randomised control group and supplements it with external patients. This may allow more trial participants to receive the experimental medicine while preserving a current randomised comparison.[5,8]

That retained control group matters. It shows what is happening under the control treatment in the trial now. It also provides a reference against which the external outcomes can be checked.

FDA generally prefers a randomised concurrent control when it is ethical and practical. A fully external control is most persuasive in narrower situations, such as a predictable disease course, an objective outcome and an effect large enough to be distinguished from residual bias.[3-5]

A brief explanation of the terminology

External control: patients outside the new trial who represent what would happen under placebo, standard care or another control treatment.
Natural history data: information on how a disease develops over time, with or without treatment.
Real-world data: information collected during routine care, such as registries, electronic health records or insurance claims.
Prognostic factor: a patient characteristic associated with the outcome, regardless of treatment.
Target trial: the ideal trial that the external analysis is trying to reproduce.
Patient adjustment: matching, weighting or modelling used to address measured differences between external and randomised patients.
Dynamic borrowing: using more external information when the external and randomised control outcomes agree, and less when they differ.
Commensurate prior: one method of dynamic borrowing. It links the external and randomised control outcomes, with a stronger link when the results are more compatible.
Effective sample size: the approximate amount of information the external data finally contribute, expressed as a number of trial patients.[7-11]

Dynamic borrowing cannot repair a poor external control

There are two separate jobs.

First, construct a credible external cohort and adjust measured differences between patients.

Second, decide how much outcome information to borrow from that cohort.

Matching, weighting and regression do the first job. Dynamic borrowing does the second.[9-11]

Neither replaces the other.

Dynamic borrowing may reduce the influence of an external cohort when its outcomes differ from the randomised controls. It cannot fix the wrong time zero, an incompatible outcome, poor follow-up or an important prognostic factor that was never recorded.

It also cannot detect every hidden bias. An external cohort may have a similar outcome for the wrong reasons.

The external patients must answer the same question as the trial

Before choosing a database, define the trial you are trying to reproduce.

The external cohort should align with the new trial on:

patient eligibility;
the control treatment;
the date follow-up begins;
outcome definitions;
length and frequency of follow-up;
treatment changes and rescue medication;
handling of treatment stopping, missing data and death; and
the analysis population and effect measure.[3-7]

This is often called target trial emulation. In plain English, it means writing down the trial you wish had been run and then checking whether the external data can reproduce it.[7]

Time zero deserves particular attention. Eligibility, treatment assignment and the start of follow-up should refer to the same point in the patient’s pathway. Misalignment can introduce immortal time and other biases that no later model can reliably remove.[5,7]

Plan the external cohort before the trial starts

The external evidence plan should be developed before recruitment wherever possible.

At that stage, the team can:

Define the target trial and estimand, meaning the exact outcome or treatment effect being estimated.
Search all plausible natural history, registry, clinical trial and real-world sources.
Assess whether each source contains the required patients, treatments, variables, outcomes and follow-up.
Prespecify the external-cohort rules, patient adjustment, borrowing model and maximum contribution.
Simulate how the design performs when the external and randomised controls agree and when they do not.[3-6,8]

A candidate external cohort can be assembled at this point. The final adjustment to the trial population usually cannot be completed yet because the characteristics of the randomised patients are not known.

Patient adjustment happens once trial patients are known

During recruitment, the trial begins to reveal the baseline characteristics of its patients.

Once enough baseline data are available, the external cohort can be matched or weighted towards the randomised control population. This may happen at a prespecified interim point or at the final analysis. It does not always require waiting for the entire trial to finish.[8-10]

The timing should be fixed in advance. Outcome data should not be used to choose whichever patients or adjustment model produce the most favourable result.

Only measured differences can be adjusted. If an important prognostic factor is missing from the external source, the analysis cannot make the groups comparable on that factor.

Outcome agreement determines how much is borrowed

After patient adjustment, the analysis compares outcomes in the external and randomised control groups.

If their outcomes are similar, the model can use more of the external information. If they differ, it uses less.

This is dynamic borrowing.

A commensurate prior is one way to do it. The method estimates the control outcome in each source and links the results. Strong agreement creates a tighter link. Disagreement weakens it.[11]

The final contribution is reported as an effective sample size. One hundred external records might contribute information equivalent to 40 trial controls, five trial controls or almost none.

External sample size is therefore not the same as the amount borrowed.

A smaller randomised control can leave the trial underpowered

A hybrid design may use 2:1 randomisation so that twice as many participants receive the experimental medicine as the control.

That can work when the external patients contribute enough control information. It creates a risk when they do not.

Tan and colleagues simulated a modified oncology trial using 2:1 randomisation. When the external and randomised controls agreed, the borrowing analyses achieved the target power of 88%. The trial-only analysis had 74% power. As disagreement reduced borrowing, power fell.[8]

The problem is not 2:1 randomisation itself. The problem arises when the number of randomised controls has been reduced on the assumption that external information will be available.

The design must therefore be tested under little or no borrowing before the randomisation ratio and sample size are fixed.

FDA has scrutinised patient matching and dynamic borrowing together

An FDA Complex Innovative Trial Design case study considered a proposed first-line lymphoma trial using a hybrid control.[12]

Patients would still be randomised 2:1. For overall survival, external patients would make up about half of the combined control group.

The sponsor proposed propensity-score matching to address measured patient differences. A Bayesian commensurate prior would then determine how much external survival information to borrow.

FDA asked how the design would behave if the populations differed, an important prognostic factor was unmeasured, the survival model was wrong or the patient-matching model was misspecified.[12]

This is not an approval precedent. It is a useful example of the questions a regulator may ask before accepting the design.

REBYOTA shows that dynamic historical borrowing can support approval

REBYOTA provides a completed FDA example, with an important distinction.

The medicine was developed to prevent recurrent Clostridioides difficile infection after antibiotic treatment. Recruitment challenges prevented the sponsor from conducting two placebo-controlled phase 3 trials.[13]

The pivotal study randomised 289 participants 2:1 to REBYOTA or placebo. It retained a current placebo group but used a Bayesian model to borrow information from an earlier randomised, placebo-controlled phase 2 trial.[13]

In the phase 3 trial, treatment success occurred in 71.2% of participants receiving REBYOTA and 62.4% receiving placebo. After borrowing from the earlier trial, the model estimated success rates of 71% and 57%. The estimated difference was 13 percentage points, with a 95% credible interval from 2 to 24 percentage points.[13]

The probability that REBYOTA was superior to placebo was 0.991. This passed one prespecified success threshold of 0.975, but not the more stringent threshold of 0.9993.[13]

FDA approved REBYOTA in November 2022. It considered the Bayesian analysis alongside the seriousness of the condition, unmet need, recruitment difficulty, supportive evidence and the advisory committee’s positive vote.[13]

This is a real precedent for approval supported by dynamic borrowing and a concurrent randomised control. But the borrowed evidence came from an earlier randomised trial, including its active and placebo groups. It did not come from natural history or real-world patients.

The example shows that acceptance is possible. It does not show that every source or model will be accepted.

Simulate the design before choosing the control-group size

The trial should not be sized around one optimistic borrowing estimate.

Simulations should test:

Strong agreement between external and randomised controls.
Moderate and severe disagreement.
Different levels of patient imbalance and overlap.
Missing or misspecified prognostic factors.
No borrowing, full pooling and the proposed dynamic approach.

The results should show bias, power, false-positive conclusions, interval coverage, precision and effective external sample size.[8-10,14]

Full pooling means treating the external and randomised controls as one group. It assumes they are fully comparable. It may look precise while producing a biased result.[9,14]

No borrowing and full pooling are useful sensitivity analyses because they show the two extremes. The prespecified dynamic analysis should sit between them and respond to the observed agreement.

Prespecify the full evidence chain

The analysis plan should cover more than the final statistical model.

Prespecify:

how external data sources will be found and assessed;
why each source is included or excluded;
eligibility and cohort rules;
time zero, follow-up and outcome definitions;
prognostic factors and the patient-adjustment method;
the dynamic-borrowing model and its prior settings;
any cap on external contribution;
the trial success rule;
simulations and sensitivity analyses;
reporting of trial-only, full-pooling and dynamic results; and
how patient-level data and source records can be made available for regulatory scrutiny.[5,6,15]

Keep an audit trail from the search to the final effective sample size. An assessor should be able to see where every external patient came from, why they were eligible and how much information they contributed.

This remains a weakness in practice. A 2026 review found that reporting of robust Bayesian borrowing often lacked enough detail to understand why particular data or prior settings had been chosen.[15]

Discuss the design with regulators before fixing the sample size

FDA recognises that natural history information and Bayesian methods may support rare disease development. It still prefers randomised concurrent controls when they are feasible.[3,4,16]

Its 2023 external-control guidance describes how patient-level data from previous trials and real-world sources should be assessed. However, the document explicitly excludes the use of external data to augment a control group within a traditional randomised trial.[5]

The lymphoma case and REBYOTA show regulatory engagement with dynamic borrowing. Neither creates automatic acceptance for a new programme.[12,13]

Discuss the proposed data sources, patient adjustment, borrowing model, simulations and role of the analysis with regulators before the sample size is fixed. HTA requirements should be considered at the same time. A design that supports regulatory efficacy may still leave uncertainty about relative effectiveness, patient-relevant outcomes or cost effectiveness.[3,5,6,16]

Evidax uses external evidence to build regulatory-grade data packages

Evidax has a mapped network of real-world, natural history and published data sources that can be used to identify and build credible external control arms. We combine this data reach with the methodological expertise to design and execute hybrid-control approaches that can reduce sample size requirements in randomised trials and make fuller use of existing evidence.

Take home

Patient-level natural history and real-world data can sometimes augment a randomised control group.
A fully external design has no current randomised control. A hybrid design retains one and adds external information.
Define the external cohort before recruitment. Final patient adjustment occurs once enough randomised baseline data are available.
Patient adjustment addresses measured differences. Dynamic borrowing reduces the external contribution when control outcomes disagree.
Dynamic borrowing cannot correct poor data, the wrong time zero or an important variable that was never measured.
A trial sized around expected borrowing may be underpowered if the external and randomised controls conflict.
Simulate agreement, disagreement and no borrowing before fixing the randomisation ratio and sample size.
REBYOTA shows that dynamic historical borrowing can contribute to an FDA-approved evidence package. It does not establish automatic acceptance of natural history or real-world controls.
Prespecify the complete evidence chain, keep an audit trail and discuss the design with regulators early.

References

Wason JS. Augmenting clinical trials with external information: why, when and how? Evidax. 2026.
Evidax. How to use historical trials to reduce control-group size in rare disease. 2026.
US Food and Drug Administration. Rare Diseases: Considerations for the Development of Drugs and Biological Products: Guidance for Industry. December 2023. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/rare-diseases-considerations-development-drugs-and-biological-products
US Food and Drug Administration. Rare Diseases: Natural History Studies for Drug Development: Draft Guidance for Industry. March 2019. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/rare-diseases-natural-history-studies-drug-development
US Food and Drug Administration. Considerations for the Design and Conduct of Externally Controlled Trials for Drug and Biological Products: Draft Guidance for Industry. February 2023. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/considerations-design-and-conduct-externally-controlled-trials-drug-and-biological-products
National Institute for Health and Care Excellence. NICE real-world evidence framework. Updated 2026. https://www.nice.org.uk/corporate/ecd9
Hernán MA, Robins JM. Using big data to emulate a target trial when a randomised trial is not available. American Journal of Epidemiology. 2016;183(8):758-764. https://doi.org/10.1093/aje/kwv254
Tan WK, et al. Augmenting control arms with real-world data for cancer trials: hybrid control arm methods and considerations. Contemporary Clinical Trials Communications. 2022;30:101000. https://doi.org/10.1016/j.conctc.2022.101000
Fu C, Pang H, Zhou S, Zhu J. Covariate handling approaches in combination with dynamic borrowing for hybrid control studies. Pharmaceutical Statistics. 2023;22:619-632. https://doi.org/10.1002/pst.2297
Jesse D, et al. Bayesian methods incorporating baseline covariates for borrowing historical control data in randomised clinical trials: a neutral comparison study. Statistics in Medicine. 2026. https://doi.org/10.1002/sim.70785
Hobbs BP, Carlin BP, Mandrekar SJ, Sargent DJ. Adaptive incorporation of historical information in clinical trials. Biometrics. 2011;67(3):1047-1056. https://doi.org/10.1111/j.1541-0420.2011.01564.x
US Food and Drug Administration. CID Case Study: External Control in Diffuse B-Cell Lymphoma. Complex Innovative Trial Design Pilot Meeting Program. https://www.fda.gov/media/155405/download
US Food and Drug Administration. Summary Basis for Regulatory Action: REBYOTA (fecal microbiota, live-jslm). BLA 125739/0. 30 November 2022. https://www.fda.gov/media/163879/download
Dron L, Golchi S, Hsu G, Thorlund K. Minimizing control group allocation in randomized trials using dynamic borrowing of external control data: an application to second-line therapy for non-small-cell lung cancer. Contemporary Clinical Trials Communications. 2019;16:100446. https://doi.org/10.1016/j.conctc.2019.100446
Weru V, Calderazzo S, Held L, Wiesenfarth M, Kopp-Schneider A. Uptake of Bayesian robust information borrowing in phase II clinical trials: a scoping review. BMC Medical Research Methodology. 2026. https://doi.org/10.1186/s12874-026-03052-3
US Food and Drug Administration. Use of Bayesian Methodology in Clinical Trials of Drug and Biological Products: Draft Guidance for Industry. January 2026. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/use-bayesian-methodology-clinical-trials-drug-and-biological-products