How to use historical trials to reduce control-group size in rare disease (Part 2)

You are developing a medicine for a rare disease.

Only a small number of patients can join the trial. A conventional randomised trial would assign many of them to placebo or standard care.

But earlier trials have already recorded what happened to similar patients receiving the same control treatment.

Can those earlier results reduce the number of new patients needed in the control group?

Sometimes.

One option is a Bayesian meta-analytic predictive prior, usually shortened to a MAP prior. The method combines historical studies, predicts what should happen in the new trial and then updates that prediction using the new data.[1-3]

Historical evidence can reduce the size of a new control group

MAP priors are particularly relevant in rare disease and paediatric development. Patient numbers may be small, recruitment may be slow and allocating many patients to control can make a trial difficult to complete.[4-6]

A MAP design can retain a randomised control group while using suitable historical evidence to reduce its size. The new control data still provide a direct, contemporary check on whether the past applies to the current trial.[2-6]

This does not lower the evidence standard. It changes how the evidence is generated.

Regulators accept MAP designs when the assumptions are justified

MAP is not an unaccepted experimental idea.

FDA’s draft Bayesian guidance identifies paediatrics and rare disease as areas where informative priors have often been proposed. ICH E11A describes Bayesian borrowing, robust mixture priors, simulations and tipping point analyses as possible parts of paediatric extrapolation.[4,5]

EMA has also accepted Bayesian borrowing within specific development programmes. Its assessment of paediatric nintedanib considered a MAP analysis alongside the randomised trial, adult evidence, pharmacology, safety and expert input.[17]

That does not mean a regulator will accept every MAP design. It means acceptance is possible when the historical evidence is relevant, the assumptions are justified, the trial still collects enough new data and the operating characteristics are acceptable.[4,5]

Discuss the proposal with regulators early. Ideally, agree the evidence sources, model, amount of borrowing, success rule and sensitivity analyses before the new trial results are known.[4,5]

A brief explanation of terminology

• Prior: the evidence used before the new trial result is analysed.
• Posterior: the updated evidence after the prior and the new trial data are combined.
• MAP prior: a prediction for the new trial built from a meta-analysis of historical studies.
• Prediction interval: the range of results that could plausibly occur in the new trial.
• Tau: a measure of how much the underlying results vary between studies. A larger tau means the studies are less consistent.
• Effective sample size: an estimate of how much statistical information the prior contributes, expressed as an approximate number of current-trial participants.
• Prior-data conflict: the new trial produces a result that disagrees with the historical prediction.
• Robustification: adding a weak alternative to the historical prior so that its influence can fall when conflict occurs.
• Dynamic borrowing: the amount borrowed changes after the new data are observed. More is retained under agreement and less under conflict.
• Operating characteristics: how the proposed design behaves in different simulated situations, including when the treatment works and when it does not.
• Tipping point: the assumption or prior weight at which the study conclusion changes.[1-5,7,8]

A MAP prior predicts the result in the new trial

A standard meta-analysis estimates the average result across the historical studies.

A MAP analysis goes one step further. It predicts the result in a new trial.[1,3]

The calculation uses a Bayesian random-effects model. This allows the underlying result to differ between studies rather than assuming that every trial has one identical true value.[1,3]

The model estimates two important quantities:

1. The average result across the historical studies.
2. Tau, which describes how much the underlying study results vary around that average.

Tau squared is the between-study variance used inside the model.

The model estimates the historical average and tau together. For every plausible combination of the two, it predicts what result could occur in the new trial. The complete distribution of those predictions becomes the MAP prior. Its central 95% range is the 95% prediction interval.[1,3,10,11]

A conventional meta-analysis may calculate a prediction interval by inserting one estimate of tau squared into a formula. A Bayesian MAP analysis does not treat tau as one known number. It carries uncertainty in both the historical average and tau into the prediction for the new trial.[1,3,10,11]

This distinction matters when only a few historical studies exist. Tau may then be very uncertain, even when the observed study results look similar.[10,11]

Historical sample size is not the amount borrowed

Five hundred patients in the historical studies do not automatically provide 500 patients’ worth of information to the new trial.

An effective sample size, or ESS, estimates how much information the prior contributes for the parameter being analysed.[4,7,8]

An RBesT example combined eight historical placebo groups containing 513 patients. The resulting MAP prior had an ESS of about 35. Adding a 20% weak component reduced it to about 25.[9]

In that example, 513 historical patients provided roughly 25 current patients’ worth of information.

ESS is useful, but it is not one fixed property of a prior. Different calculation methods can produce different answers. The method must be named and the result interpreted alongside the planned size of the new trial.[4,7,8]

ESS is also not simply subtracted from the original control-group sample size. The new sample size should be selected through simulations of the complete design. These should test what happens when the historical evidence is right, partly wrong or seriously wrong.[4,5,12,13]

Tau shows how confidently the past predicts the new trial

Tau describes real variation in the underlying study results. Tau squared is the corresponding variance.[1,10,11]

A small tau means the historical studies give similar underlying results. The prediction for the new trial will usually be narrower, so the historical evidence may contribute more information.

A larger tau means the historical results vary more. The prediction becomes wider and the prior usually contributes less.[1,10]

I-squared and the Q test may describe the variation observed in a conventional meta-analysis. They do not directly determine MAP borrowing. The MAP model uses tau and the uncertainty around tau.[1,10,11]

The nintedanib analysis provides a useful example. The conventional adult meta-analysis reported I-squared of 0% and tau squared of 0. But the Bayesian analysis still allowed substantial uncertainty around tau. No observed heterogeneity did not prove that the true variation was exactly zero.[15]

Historical studies must answer the same clinical question

A sophisticated model cannot make unsuitable studies relevant.

The historical and new trials should align on the question being asked. This includes:

• the patients;
• the treatment and control conditions;
• the outcome and how it is measured;
• when follow-up begins;
• how long patients are followed;
• changes in supportive care or clinical practice; and
• how treatment changes, rescue medication, stopping treatment and missing data are handled.[4-6]

These components define the estimand, meaning the exact treatment effect or outcome the trial is trying to estimate.

Matching a few baseline percentages is not enough. The historical evidence must be suitable for predicting the parameter in the new trial.[4-6]

Robustification reduces borrowing when the new trial disagrees

A standard MAP prior may place substantial weight on the historical prediction.

A robust MAP prior adds a second, weakly informative component. This component contributes little evidence and allows a wider range of results.[2,3,5]

When the new trial agrees with the historical prediction, more of the informative component is retained. When the results conflict, the balance can move towards the weak component and reduce borrowing.

This is dynamic borrowing.[2,3]

Robustification is protection, not a repair tool. It cannot correct poor historical data, an irrelevant outcome or important clinical differences that are not visible in the observed result.[2,5,6]

It can also create a practical problem. If the new control group is extremely small and historical borrowing falls sharply, too little information may remain for a useful conclusion. The design needs to anticipate that possibility before recruitment begins.[2,4,5]

Find every relevant source and preserve the audit trail

FDA’s guidance is direct:

“The process of determining a prior should begin with the identification and review of all the available relevant external information.”[4]

This means searching systematically before deciding which studies will inform the prior. The eligibility rules should be fixed in advance. Evidence should not be excluded because it weakens the expected result or makes the prediction less precise.[4,5]

The full audit trail should preserve:

• where and how the evidence was searched;
• why each study was included or excluded;
• the outcome data and study characteristics extracted;
• the source of every extracted value;
• differences between the historical and new trials;
• judgements about whether each study can be borrowed from; and
• how those judgements affect the model and sensitivity analyses.[4,5]

Evidax TRACE can map and structure the complete published and real-world evidence landscape while preserving the link between each source, extracted value and design decision.

The technology supports coverage, speed and traceability. Clinical and statistical experts still decide whether an evidence source is suitable for borrowing and how its uncertainty should enter the design.

Build the MAP prior in ten transparent steps

The following workflow brings together the methodological papers and the expectations described by FDA and ICH. It is a practical synthesis, not a checklist copied from one guidance document.[1-5]

Step 1. Define the new trial question

Specify the patients, treatment, control, outcome, measurement time and handling of events after treatment starts. This defines the estimand that the new trial and historical evidence need to address.[4,5]

Step 2. Speak to regulators early

Discuss whether borrowing is acceptable, what new evidence is still required and how the prior may affect the control-group size and success rule. Allow time for agreement before the trial is analysed.[4,5]

Step 3. Define what the prior will predict

The prior might predict a control response rate, mean outcome, event rate or treatment effect transferred from another population. Prespecify the success criterion that will be applied after the prior and new data are combined.[4,5]

Step 4. Find all relevant historical evidence

Search trials, systematic reviews, registries, natural history studies and suitable real-world data. Prespecify the eligibility rules before knowing which studies give the most favourable results.[4-6]

Step 5. Check whether the evidence is suitable

Assess data quality and alignment with the new trial’s estimand. Document differences in patients, care, outcomes, time zero, follow-up and events after treatment begins.[4-6]

Step 6. Fit the Bayesian random-effects meta-analysis

Estimate the historical average and tau. Justify the prior distribution used for tau, particularly when only a few studies are available.[1,3,10,11]

Step 7. Predict the result in the new trial

Carry uncertainty in the historical average and tau into the posterior prediction for the new trial. This distribution becomes the MAP prior. Report its prediction interval.[1,3]

Step 8. Add protection against conflict

Combine the informative MAP prior with a weak component. Justify the starting weight and show how borrowing changes when the new trial agrees or disagrees with the historical prediction.[2,3,5]

Step 9. Quantify and test the borrowing

Report ESS using a named method. Simulate the design under agreement, moderate conflict, strong conflict and no treatment effect. Examine sample size, precision, bias, false-positive conclusions and the probability of meeting the success rule. Use tipping point analyses to show when the conclusion changes.[4,5,7,8,12,13]

Step 10. Prespecify and report the full analysis

Document the evidence sources, model, priors, weights, success rule, simulations and sensitivity analyses. Report an analysis using the new trial data alone. Provide the software and code needed to reproduce the work.[4,5]

Nintedanib shows why the amount borrowed matters

Nintedanib provides a real example of MAP borrowing in rare disease development.

The design was not a simple historical-control analysis used to remove a fixed number of children from the placebo group. It transferred evidence about the treatment effect from adults to children.[14-17]

InPedILD randomised 39 children and adolescents with fibrosing interstitial lung disease. Twenty-six received nintedanib and 13 received placebo. The trial was designed mainly around pharmacokinetics and safety rather than being powered to establish efficacy.[14]

The trial-alone estimate for change in forced vital capacity percentage predicted favoured nintedanib by 1.21 percentage points. Its 95% confidence interval ranged from -3.40 to 5.81.[14,15]

Six adult trials were then used to construct a MAP prior for the treatment effect. A formal expert exercise assigned a prespecified weight of 0.56 to the adult evidence.[15]

The effect of borrowing was material:

Analysis	Estimated difference	95% credible interval	Probability that the effect exceeded zero
No adult borrowing	1.17	-3.18 to 5.52	70.2%
Prespecified adult weight of 0.56	1.63	-0.69 to 3.40	95.5%

The prespecified analysis passed its success rule of at least 90% probability that the effect exceeded zero. Its 95% credible interval still included no effect.[15]

FDA calculated that the 0.56 adult weight represented an adult ESS of 370 and about 90% of the total information. Its review did not support the proposed paediatric efficacy conclusion. A central concern was whether adult and paediatric disease were similar enough to justify that degree of extrapolation.[16]

EMA reached a different conclusion. It accepted the extrapolation package and supported extending the indication to children. But it did not rely on the MAP result alone. Its benefit-risk judgement also considered adult evidence, disease biology, paediatric pharmacokinetics, safety, the direction of the randomised result, expert advice and post-authorisation evidence commitments.[17]

This is the balanced lesson. MAP borrowing can be accepted by regulators. Acceptance depends on the clinical question, transferability of the evidence, amount borrowed, operating characteristics and the wider evidence package.[4,5,16,17]

Regulatory acceptance does not settle the HTA question

The same evidence later faced a different question in Germany.

IQWiG and G-BA concluded that an additional benefit was not proven. Their concern was not simply the Bayesian method or its prior weight. The adult evidence focused on forced vital capacity. The German assessment did not consider that outcome sufficient to establish additional patient-relevant benefit in the reimbursement context.[18,19]

This was not a refusal of reimbursement. The additional-benefit rating feeds into the German reimbursement and price process.

The wider point is that regulatory borrowing and HTA comparison are connected but different. A MAP prior may help establish efficacy or support extrapolation. HTA bodies may still require evidence on patient-relevant outcomes and comparisons with the treatments used in current practice.[18-20]

Evidence synthesis and trial design belong in one plan

A MAP prior begins with the full external evidence landscape.

The search determines which studies exist. The clinical review determines whether they answer the same question. The statistical model determines how uncertainty and differences are represented. Simulations determine whether the design remains credible when the past and present disagree.

These are connected decisions.

Evidax TRACE provides the evidence coverage and audit trail. Evidax Comparison Blueprints connect that evidence to the appropriate comparative method, its assumptions, diagnostics and limits. The Integrated Evidence Plan then links the regulatory design to the later evidence needed by HTA bodies and payers.

The aim is not to maximise borrowing. It is to borrow only what the evidence can support.

Take home

• MAP priors can use historical trials to reduce control-group size in rare disease.
• The method estimates the historical average and tau, then carries uncertainty in both into a prediction for the new trial.
• The historical sample size is not the amount borrowed. ESS estimates the information contributed by the prior and may be much smaller.
• Robustification allows borrowing to fall when the new result disagrees with the historical prediction.
• Regulators accept MAP designs in suitable settings, but the evidence, assumptions, amount borrowed and operating characteristics must be justified.
• Search the full evidence landscape and preserve a reproducible audit trail before constructing the prior.
• Discuss the design with regulators early and prespecify the model, success rule, simulations and sensitivity analyses.
• Regulatory acceptance does not establish additional benefit for HTA. Plan both questions together.

References

1. Neuenschwander B, Capkun-Niggli G, Branson M, Spiegelhalter DJ. Summarizing historical information on controls in clinical trials. Clinical Trials. 2010;7(1):5-18. https://doi.org/10.1177/1740774509356002

2. Schmidli H, Gsteiger S, Roychoudhury S, O’Hagan A, Spiegelhalter D, Neuenschwander B. Robust meta-analytic-predictive priors in clinical trials with historical control information. Biometrics. 2014;70(4):1023-1032. https://doi.org/10.1111/biom.12242

3. Weber S, Li Y, Seaman JW III, Kakizume T, Schmidli H. Applying meta-analytic-predictive priors with the R Bayesian evidence synthesis tools. Journal of Statistical Software. 2021;100(19):1-32. https://doi.org/10.18637/jss.v100.i19

4. 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

5. International Council for Harmonisation. ICH E11A: Pediatric Extrapolation. Final version adopted 21 August 2024. https://database.ich.org/sites/default/files/ICH_E11A_Guideline_Step4_2024_0821.pdf

6. Ghadessi M, Tang R, Zhou J, et al. A roadmap to using historical controls in clinical trials by Drug Information Association Adaptive Design Scientific Working Group. Orphanet Journal of Rare Diseases. 2020;15:69. https://doi.org/10.1186/s13023-020-1332-x

7. Neuenschwander B, Weber S, Schmidli H, O’Hagan A. Predictively consistent prior effective sample sizes. Biometrics. 2020;76(2):578-587. https://doi.org/10.1111/biom.13252

8. Morita S, Thall PF, Müller P. Determining the effective sample size of a parametric prior. Biometrics. 2008;64(2):595-602. https://doi.org/10.1111/j.1541-0420.2007.00888.x

9. RBesT development team. Introduction to RBesT. Package vignette, version 1.12-0. 18 September 2026. https://cran.r-project.org/package=RBesT

10. Röver C, Bender R, Dias S, et al. On weakly informative prior distributions for the heterogeneity parameter in Bayesian random-effects meta-analysis. Research Synthesis Methods. 2021;12(4):448-474. https://doi.org/10.1002/jrsm.1475

11. Friede T, Röver C, Wandel S, Neuenschwander B. Meta-analysis of few small studies in orphan diseases. Research Synthesis Methods. 2017;8(1):79-91. https://doi.org/10.1002/jrsm.1217

12. Viele K, Berry S, Neuenschwander B, et al. Use of historical control data for assessing treatment effects in clinical trials. Pharmaceutical Statistics. 2014;13(1):41-54. https://doi.org/10.1002/pst.1589

13. Kopp-Schneider A, Calderazzo S, Wiesenfarth M. Power gains by using external information in clinical trials are typically not possible when requiring strict type I error control. Biometrical Journal. 2020;62(2):361-374. https://doi.org/10.1002/bimj.201800395

14. Deterding R, Young LR, DeBoer EM, et al. Nintedanib in children and adolescents with fibrosing interstitial lung diseases. European Respiratory Journal. 2023;61:2201512. https://doi.org/10.1183/13993003.01512-2022

15. Maher TM, Brown KK, Cunningham S, et al. Estimating the effect of nintedanib on forced vital capacity in children and adolescents with fibrosing interstitial lung disease using a Bayesian dynamic borrowing approach. Pediatric Pulmonology. 2024;59:1038-1046. https://doi.org/10.1002/ppul.26882

16. US Food and Drug Administration. NDA/BLA Multi-disciplinary Review and Evaluation: NDA 205832 S023, Ofev (nintedanib). 2023. https://www.accessdata.fda.gov/scripts/cder/daf/index.cfm?ApplNo=205832&event=overview.process

17. European Medicines Agency. Ofev: EPAR assessment report, variation EMEA/H/C/003821/X/0057/G. EMA/2963/2025. 12 December 2024. https://www.ema.europa.eu/en/documents/variation-report/ofev-h-c-003821-x-0057-g-epar-assessment-report-variation_en.pdf

18. Institute for Quality and Efficiency in Health Care. Nintedanib (clinically significant, progressive fibrosing interstitial lung diseases, 6 to 17 years): Benefit assessment according to §35a SGB V; Extract. Project A25-30. 12 May 2025. https://doi.org/10.60584/A25-30_en

19. Federal Joint Committee. Justification of the Resolution on an Amendment of the Pharmaceuticals Directive: Nintedanib (new therapeutic indication: clinically significant, progressive fibrosing interstitial lung diseases, 6 to under 18 years). 7 August 2025. https://www.g-ba.de/bewertungsverfahren/nutzenbewertung/1177/

20. Directorate-General for Health and Food Safety. Methodological Guideline for Quantitative Evidence Synthesis: Direct and Indirect Comparisons. 8 March 2024. https://health.ec.europa.eu/publications/methodological-guideline-quantitative-evidence-synthesis-direct-and-indirect-comparisons_en