A sponsor with a disease too rare to power a conventional trial has two ways to go. Run a single-arm study against an external control, measure a biomarker, and hold the conventional significance threshold. Or run a randomized, double-blind trial against a clinically meaningful endpoint, and ask the agency to accept a looser threshold in exchange.
The first option looks more rigorous. It keeps the familiar one-sided 0.025. At a workshop convened on September 29 at the National Press Club in Washington by the Duke-Margolis Institute for Health Policy with the FDA’s Rare Disease Innovation Hub, an agency statistician said the second is generally the better trial.
That is a more consequential statement than it first sounds. It tells sponsors the agency would rather see a randomized comparison with a 5 or 10 percent false positive rate than a single-arm study that preserves the traditional threshold by giving up the control arm.
Imprecision Is Visible. Bias Is Not.
The agency framed the whole day around a single asymmetry. A small trial produces an imprecise estimate, and imprecision can be measured, reported and reasoned about. A biased trial produces a tight estimate centered on the wrong answer, and nothing in the data announces it. More patients will not fix it.
From that asymmetry the FDA statistician drew a principle, offered explicitly for debate rather than as settled policy: efficiency gains should not come at the cost of the reliability or interpretability of the evidence. In rare disease, where trials are small and rarely repeated, a biased result is not a temporary error. It is the answer the field lives with.
“We are favoring accepting some imprecision over accepting bias, and we are insisting that the question being asked must matter to patients.”
FDA statistician, CDER Office of Biostatistics
The second half of that sentence carries as much weight as the first. A trial can be feasible and unbiased and still fail, if the thing it measures is not the thing patients care about. The agency’s endpoint flowchart starts with how patients feel, function or survive, then applies practical filters in order: can the outcome be measured reliably, is it sensitive to the mechanism of action, over what time frame would an effect appear, and is any of it achievable at the sample size the disease allows. In rare disease those criteria are almost never all satisfied at once.
What the Agency Says Randomization Is Not
The FDA presenter was direct about the objection the rare disease community raises most often, and answered it in terms of care rather than statistics. Randomization is not withholding effective treatment. Control patients are expected to receive high quality care, enhanced monitoring and closer clinical oversight than they would outside a trial. Where an approved therapy exists, controls receive it, as background standard of care or in an active control design.
In rare disease specifically, the agency argued, specialist care and systematic follow-up are often thin on the ground, so enrolling can improve the quality of care a patient receives regardless of which arm they land in.
The same presentation then conceded the concerns are legitimate, and listed four things that would make randomized designs more workable in practice rather than merely more defensible: engage patients and families about randomization from the outset, build in early stopping rules so no patient stays on control longer than the evidence requires, write protocol procedures that protect the integrity of the randomization, and consider alternatives to parallel-group designs such as randomized crossover.
Five Principles, and a Number That Makes Them Affordable
A biostatistician and professor emeritus took the argument further, setting out five design principles for registrational trials in rare disease.
On the second principle he used an example from oncology that generalizes. Suppose patients with greater tumor shrinkage live longer. It is tempting to treat shrinkage as the endpoint. But the patient-level correlation is not causal evidence, and the causality can run the other way: the patients who responded may have had more intact immune systems, which is also why they survived longer.
“All I did with my treatment was put a label on who would have done better.”
Biostatistician, on mistaking a prognostic marker for a treatment effect
The principles cost precision, which is the objection sponsors raise immediately. His answer was to move the cost somewhere it can be paid. Hold the discussion with regulators before the trial, and ask whether a higher false positive rate is acceptable for this programme: 5 percent, or 10, instead of the traditional 2.5.
Then he quantified it. Relaxing the threshold that way can take a trial that would have needed 75 to 100 primary endpoint events down to roughly 30. In a disease with a few hundred diagnosed patients worldwide, that is the difference between a trial that can be run and one that cannot.
The Community and the Family Are Not the Same Respondent
A child psychiatrist and pharmacologist on the panel, who has worked in drug development and has lived with a rare cancer for thirteen years, drew a distinction that rarely survives into trial design documents. The patient community and the individual family are two different respondents, and they are routinely consulted as though they were one.
A community wants treatments to exist, and reasons in percentages and trade-offs. A parent is deciding whether to put their child into this trial, now. He put the individual calculus plainly: a family may be weighing ten more years with a child against the possibility that the trial shortens that, and a community-level argument about evidentiary standards does not reach that decision.
He also raised a practical consequence of designing around that tension rather than for it. Families try to work out which arm they are in. Online forums exist where parents compare side effects to infer assignment, which is not an ideal blinding. And an institutional review board approving a protocol settles very little, because the board’s job is to set up the consent question, not to answer it. A trial can clear review and still fail because no family chooses it.
The Denominator Problem Underneath All of It
A clinical geneticist who has worked on more than 250 rare disease trials supplied the context that makes the statistical argument urgent rather than academic. Only a couple of hundred diseases have enough diagnosed patients to support a conventional trial design. Fewer than 200 have patients numbering in the thousands.
Meanwhile the denominator keeps growing. Since the human genome project, roughly ten new genotype-phenotype correlations are being described every week, putting the count of recognized conditions somewhere north of 16,000. He also noted that the statistical traditions sponsors import from oncology do not transfer cleanly to lifetime germline conditions, which behave differently and need different designs.
So the population of diseases needing trials is expanding faster than the population of patients available to fill them. That is the structural reason the trade-off conversation cannot be deferred.
What Sponsors Should Take From It
The workshop produced no recommendations and was explicitly not a federal advisory committee. What it produced is a reading of how the agency’s statistical reviewers are currently thinking, which is useful precisely because it arrived before a guidance document rather than after one.
The operative signal is that the alpha level is negotiable and the control arm is much less so. A sponsor arriving at a pre-submission meeting with a single-arm design and an unvalidated biomarker, defending it on the grounds that it preserves a conventional p-value, is making an argument the agency’s own statisticians were publicly arguing against. A sponsor arriving with a randomized design and a request to pre-specify a 5 or 10 percent false positive rate is having a conversation the agency invited.
The qualifier that ran through the day was that none of this is a formula. Prevalence, natural history, causal pathway and mechanism all bear on what is appropriate, and a design that is right for one disease can be wrong for the next. The agency asked for the conversation to happen early in development, which is the one piece of procedural advice that costs a sponsor nothing.
The workshop was the fourth in the FDA’s Rare Disease Innovation Science and Exploration series. Under the latest reauthorization the agency has committed to at least ten more over five years beginning in fiscal 2028, funded through user fees. A Federal Register docket accompanying this one, carrying specific questions drafted by FDA biostatisticians, is open for comment until November 13.
References
- FDA, RISE Workshop 4: RISE to the Challenge: Statistical Considerations for Rare Disease Clinical Investigations, September 29, 2026, National Press Club, Washington, DC.
- RISE Workshop #4 pre-read (PDF), co-convened by Duke-Margolis and the FDA Rare Disease Innovation Hub. The background document and discussion questions the workshop was built around.
- Federal Register, “Statistical Considerations for the Design of Rare Disease Clinical Investigations; Establishment of a Public Docket; Request for Information and Comments,” published September 15, 2026. Docket FDA-2026-N-10165.
- Docket FDA-2026-N-10165 on Regulations.gov, carrying the specific questions drafted by FDA biostatisticians. The comment period closes November 13, 2026.
- Workshop agenda (PDF), Duke-Margolis Institute for Health Policy.
- Duke-Margolis, Rare Disease Innovation, Science and Exploration (RISE) Workshop Series, with materials from all four workshops to date.
- FDA meetings and workshops listing for the September 29, 2026 session.
- Full workshop recording (video). Quotations in this article are taken from the recording and are reported by role rather than by name.
Moe Alsumidaie, MBA, MSF, is founder and Chief Editor of Vanguard Publications, which publishes Clinical Trial Vanguard, Pharma Vanguard and BullScope, and Head of Research at CliniBiz. He has two decades in clinical trial operations and data science, with earlier roles at Genentech, Abbott Vascular and Stanford University Medical Center, and is a guest lecturer in clinical trial sciences at Rutgers University.
