Fewer than 300 patients may be globally eligible for a rare disease trial. At that scale, conventional randomized trial assumptions can fail in ways standard sample size formulas do not show. Type I and Type II error tradeoffs at N=20 or N=40 per arm differ materially from those at N=150. Protocol amendments also carry greater operational cost: replacing 10% of enrolled patients may delay enrollment by six months, whereas the same replacement in a large oncology study is negligible. Decisions made before screening begins therefore have disproportionate consequences.
Why Power Calculations Break in Rare Disease
A conventional power calculation asks how many patients are required to detect an assumed treatment effect, given assumed outcome variability, a chosen significance level, and a specified probability of detection. In rare disease, the effect estimate may come from a small Phase 1/2 study or natural history cohort, while treatment and control variability remain uncertain at small sample sizes. Both principal inputs therefore contain substantial estimation error, which is transmitted to the resulting sample size.
The result is frequent underpowering despite stated 80% or 90% targets. An assumed 30% reduction in a functional outcome, estimated from a 15-patient natural history study, may have confidence bounds compatible with both a 10% and 55% true effect. A design based on the 30% point estimate will be underpowered when the actual effect lies near the lower bound.
Adding patients is not necessarily an option because global population may be exhausted. Instead, uncertainty in design inputs should be reduced before protocol finalization, while the design should remain efficient across the uncertainty that persists. Both objectives call for prospective modeling of disease progression and compound effects.
Natural History Data for Virtual Cohorts
Rare disease virtual cohorts typically start with natural history evidence: longitudinal disease observations from untreated or minimally treated patients in registries, academic cohorts, and prospective publications. These data characterize the control-arm distribution. Rather than enrolling a concurrent placebo group, which may be ethically difficult and statistically inefficient at small N, investigators can derive a virtual control arm for a single-arm interventional study.
In 2023, FDA published draft guidance addressing external control arms and recognizing their scientific use under specified conditions. Natural history participants should be comparable to trial participants on important prognostic factors, outcomes should be measured by the same or an equivalent method, and comparisons between treated patients and the external control require an accompanying bias analysis.
Building a virtual cohort involves more than resampling the observed data. Registry participants may receive more systematic follow-up than nonparticipants, creating selection effects, immortal time bias, and healthy user effects that require explicit adjustment. Registries may also underrepresent disease severity or particular phenotypes, forcing extrapolation to the intended trial population. Every adjustment adds assumptions that should undergo sensitivity analysis and be reported transparently.
Bayesian Designs and External Data Borrowing
A natural progression from virtual cohorts is a Bayesian adaptive design that dynamically borrows external information. The natural history control becomes a prior for the control response rate or outcome trajectory. As interventional data accrue, the treatment-effect posterior is updated, with greater borrowing when trial data are sparse and less borrowing when observed results conflict with the historical prior.
Power priors and hierarchical Bayesian models are common implementations. They discount historical observations according to similarity between the registry population and the trial population. Agreement with the historical control can yield meaningful effective sample size and a decision threshold with fewer enrolled patients. Divergence reduces the external contribution, limiting results driven by prior-data conflict.
For sufficiently large rare disease studies, group sequential methods using O'Brien-Fleming or Lan-DeMets alpha-spending remain applicable. In very small studies, N=15 per arm, interim stopping is usually impractical because conditional power is too unstable for dependable rules. Bayesian posterior probability thresholds for efficacy and futility are often more useful and can be calibrated to approximate frequentist type I error when the prior is not overly informative.
Building a Virtual Cohort in Practice
QuantHealth's rare disease workflow has four stages. First, available natural history evidence is assembled and curated for follow-up completeness, outcome ascertainment, and covariate coverage. Second, a hierarchical Bayesian progression model separates between-patient variation, which sets virtual-cohort dispersion, from within-patient trajectory variation, which governs uncertainty in individual outcomes. Third, patients are sampled from the fitted between-patient distribution subject to the planned trial criteria. Fourth, alternative borrowing assumptions are evaluated to measure effects on power and operating characteristics.
In a QuantHealth pilot, a team planning a Phase 2 single-arm study for a pediatric rare neurological condition had 87 longitudinal observations spanning two-year follow-up. The primary endpoint was a functional composite score at month 12. After adjustment for enrollment-criterion differences, the registry-based virtual cohort generated a more precise month 12 control-arm posterior predictive distribution than the registry mean alone. Adaptive borrowing reduced planned enrollment from 48 patients, based on a conventional frequentist calculation using the registry mean and variance, to 32 patients without increasing expected type I error, as trial data updated the posterior in real time.
Limits of the Approach
Virtual cohorts perform best when natural history data are well characterized and their outcome measure matches the trial endpoint. They are less suitable when important phenotypic heterogeneity is absent from the dataset, the registry population differs materially in breadth from the planned trial population, or the trial uses an endpoint that is newer or more precise than historical assessments.
Regulatory acceptability is distinct from statistical performance. FDA consideration of a virtual control varies with the indication, the depth and breadth of natural history evidence, and the clinical question. Well documented conditions, including certain lysosomal storage disorders, may support external controls. Poorly characterized diseases or underrepresented trial subgroups may still require a concurrent control despite the enrollment burden. Type B or pre-IND discussions with FDA should occur before selecting this design.
We state the boundary plainly: virtual cohort quality cannot exceed the natural history evidence beneath it. Sparse, variable, or demographically mismatched data produce wide posterior credible intervals and strong sensitivity to prior assumptions. That warning informs protocol design, but it is not validation for a design unsupported by its underlying data.
When Virtual Cohorts Work Well
A carefully constructed virtual cohort does not remove the core constraints of rare disease research: small populations, variable outcomes, and limited replication across independent studies. It can, however, reduce concurrent placebo enrollment while preserving the intended evidence threshold, an ethical and operational benefit when the global patient pool is limited.
The principal methodological benefit is that virtual cohort construction requires the control-arm assumptions to be stated before enrollment. Those assumptions can then be documented, stress-tested through sensitivity analyses, and presented for regulatory review. A conventional power calculation may simply import a point estimate from a precedent trial without making the same assumptions explicit. Virtual cohorts therefore turn control-arm uncertainty into an inspectable quantity, supporting a more transparent basis for design decisions than treating an uncertain value as known.