External Control Arms From Real-World Data: Methods Guide
An external control arm is a comparison group built from patients outside the trial, usually drawn from real-world data such as electronic health records, registries, or claims, and used in place of or alongside a randomized control group. It needs patient-level data that matches the trial’s eligibility criteria, index date, and outcome definitions closely enough to support statistical adjustment for confounding, most often through propensity score methods within a target trial emulation framework. Regulators accept external controls in specific settings, mainly rare diseases and serious conditions with large expected effects, and they judge them on data quality and pre-specification more than on method choice.
Why this matters now
Randomization is the gold standard, but it is not always feasible or ethical. In rare diseases there may not be enough patients to fill two arms. In oncology subgroups defined by a single biomarker, placebo control can be unacceptable when the natural history is poor. Sponsors increasingly run single-arm trials in these settings and need a credible comparator.
Regulatory guidance has caught up with that demand. The International Council for Harmonisation (ICH) E10 guideline on the choice of control group has recognized external controls since 2000, while warning about their bias. The US Food and Drug Administration (FDA) published its Framework for Real-World Evidence in 2018 and a draft guidance, Considerations for the Design and Conduct of Externally Controlled Trials, in 2023. The European Medicines Agency (EMA) has also issued a reflection paper on single-arm trials submitted as pivotal evidence. The common message across these documents is consistent: the external data must be fit for purpose, the analysis must be pre-specified, and the sponsor must show why randomization was not feasible.
The practical bottleneck is access. The data that makes a good external control, detailed longitudinal records with lab values, biomarkers, and treatment lines, sits with hospitals, cancer registries, and national programs that will not ship patient-level extracts to a sponsor’s servers.
The Sovereign AI angle: bring the analysis to the comparator data
Building an external control arm is a data access problem before it is a statistics problem. A sponsor has to find eligible patients across many custodians, apply the trial’s criteria identically at each, and fit adjustment models on combined data, all while each custodian keeps legal control of its records.
A federated Trusted Research Environment (TRE) handles that by moving the analysis rather than the data. The sponsor’s cohort definition and analysis code run inside each custodian’s environment, and only aggregate results, model coefficients, or balance diagnostics come back after an automated airlock checks them. The data never leaves the source, so custodians can take part without negotiating a new data transfer agreement for each study. Federated regression and federated propensity score estimation are established methods; custodians share sufficient statistics or model updates, not rows.
This is the same architecture Lifebit uses for multi-site observational studies without moving data, applied to a regulatory question.
Methods: how external control arms are built
Target trial emulation
Target trial emulation, formalized by Hernán and Robins in 2016, asks the analyst to write down the randomized trial they would have run, then emulate each element with observational data: eligibility, treatment strategies, assignment, start of follow-up, outcomes, and the causal contrast. For external controls the most important element is time zero. The index date for external patients must correspond to the point at which trial patients would have started treatment, or the comparison inherits immortal time bias.
Confounding adjustment
Propensity score methods are the workhorse. The propensity score is the probability of being in the trial arm given baseline covariates. Analysts then use matching, inverse probability of treatment weighting, or stratification to balance the groups, and check balance with standardized mean differences. Outcome regression and doubly robust estimators, which combine a propensity model with an outcome model, add protection if one model is misspecified.
Estimands and outcome choice
Define the estimand before choosing a method. The ICH E9(R1) addendum on estimands asks sponsors to state the population, treatment, variable, handling of intercurrent events, and summary measure. For external controls this discipline exposes problems early. Overall survival is usually the most robust outcome because death is recorded reliably across sources. Endpoints such as tumor response or progression depend on imaging schedules and assessment criteria that differ between a trial protocol and routine care, so they need careful justification, and sometimes blinded central re-review of external records, before they can support a comparison.
Bayesian borrowing
When a small randomized control arm exists, Bayesian dynamic borrowing, such as power priors or commensurate priors, can add external patients while down-weighting them if they look different from the concurrent controls. This hybrid design is often more acceptable to regulators than a fully external comparator.
Sensitivity and bias analysis
No adjustment removes unmeasured confounding. Quantitative bias analysis and the E-value, introduced by VanderWeele and Ding in 2017, express how strong an unmeasured confounder would have to be to explain away the observed effect. Regulators expect to see these analyses planned in advance.
Data requirements for a credible external control
| Requirement | What it means in practice | Common failure |
|---|---|---|
| Eligibility alignment | Every inclusion and exclusion criterion can be computed from the source data | Performance status or prior therapy lines missing, so the cohort is broader than the trial |
| Index date definition | A clear equivalent of trial enrollment for each external patient | Follow-up starts at diagnosis, creating immortal time |
| Covariate completeness | Key prognostic factors recorded with acceptable missingness | Biomarkers only available for a subset of sites |
| Outcome comparability | Outcomes defined and ascertained the same way as in the trial | Response assessed by clinicians rather than a standardized criterion |
| Contemporaneity | External patients treated in a similar era of standard care | Historical controls from before a practice-changing therapy |
| Provenance and traceability | Each data element traced to its source with a documented transformation | Undocumented mappings that an auditor cannot reproduce |
| Harmonization | All sources mapped to one common data model and vocabulary | Local codes interpreted differently at each site |
Harmonization deserves particular attention. Mapping every contributing source to the Observational Medical Outcomes Partnership (OMOP) Common Data Model lets one cohort definition run unchanged at every site, which is the only way to claim identical eligibility logic. Our OMOP implementation checklist covers the mapping and quality steps.
A practical workflow
- Justify the design. Document why randomization is infeasible or unethical and why the expected effect size is large enough for an external comparison to be informative.
- Run feasibility first. Before writing the protocol, count eligible patients per site using the full criteria. Federated feasibility queries return counts without exposing records; see our guide to clinical trial feasibility with real-world data.
- Assess fitness for purpose. Measure completeness of every covariate and outcome per site, and drop sources that cannot support the analysis.
- Pre-specify and share the plan. Lock the statistical analysis plan, including primary method, balance thresholds, and sensitivity analyses, and discuss it with regulators before unblinding trial results.
- Execute in place. Run the analysis code inside each custodian’s environment, return only checked aggregates, and keep a full audit trail of code versions and outputs.
- Report transparently. Present balance diagnostics, per-site results, and bias analyses alongside the primary estimate.
Real-world context
External controls are not new to regulatory decisions. Several accelerated approvals in oncology and rare disease have relied on single-arm trials interpreted against historical or natural history data, and FDA reviews of these submissions repeatedly point to the same weaknesses: mismatched eligibility, different outcome assessment, and missing prognostic variables. The agencies have been clearer about what fails than about any single approved template.
On the data side, industry and public programs are moving toward governed access instead of extracts. Boehringer Ingelheim works with Lifebit on reuse of clinical data in early discovery, a setting where historical trial data can inform comparator design without creating uncontrolled copies. National programs with federated infrastructure, such as Genomics England and CanPath, both Lifebit customers, show that detailed clinical and genomic data can be analyzed at scale while custodians keep control, which is exactly the property an external control arm across several countries needs.
Common pitfalls and objections
“Real-world data is too messy”
Some of it is. The answer is a documented fitness-for-purpose assessment per source, not a blanket judgment. Structured oncology registries and curated EHR cohorts can support external controls; unmapped claims data often cannot.
“Federated analysis limits the methods we can use”
Propensity score estimation, weighting, and Cox regression all have federated implementations that give results equivalent or very close to pooled analysis. Where a method needs row-level pooling, the data can often be combined inside one governed environment with output checks rather than exported.
“Regulators will not accept it”
Acceptance depends on early engagement and pre-specification. Treat the external control plan as a regulatory document from day one, and expect it to be one line of evidence among several.
What to do next
Start with a feasibility count across candidate custodians using your draft eligibility criteria, then score each source on the requirements in the table above. That exercise usually tells you within weeks whether an external control arm is viable and which sites it needs. If you are evaluating infrastructure for this, the federated Trusted Research Environment overview explains how cohort discovery, harmonization, and output review work together.
Frequently asked questions
What is the difference between an external control arm and a synthetic control arm?
The terms are often used interchangeably. Both describe a comparator built from patients outside the trial. Some authors reserve “synthetic” for controls built by weighting or modeling external patients to resemble the trial population, but regulators generally use “externally controlled trial” for the whole category.
When do regulators accept external control arms?
Mainly when randomization is infeasible or unethical, the disease course is well characterized, and the expected treatment effect is large relative to potential bias. Rare diseases and serious conditions with poor prognosis are the typical settings.
Which statistical method is best for an external control arm?
There is no single best method. Propensity score weighting or matching within a target trial emulation framework is the most common starting point, with doubly robust estimators and quantitative bias analysis as supporting analyses. Pre-specification matters more than the specific choice.
Can external control arms be built without moving patient data?
Yes. In a federated Trusted Research Environment, cohort definitions and analysis code run inside each custodian’s environment, and only aggregate statistics or model parameters return after output review. Patient-level records stay with the custodian.
Why does data harmonization matter for external controls?
Eligibility criteria and outcomes must be applied identically across all sources. Mapping each source to a common data model such as OMOP lets one cohort definition run unchanged everywhere, which supports a defensible claim that the external cohort matches the trial.
What is immortal time bias?
It occurs when follow-up for one group starts before the point at which patients could have experienced the outcome under the comparison, giving that group guaranteed event-free time. Aligning the external index date with trial enrollment prevents it.