Adam Feuerstein opens his follow-up on Vykat XR, the Neurocrine Biosciences drug for Prader-Willi syndrome now under safety scrutiny, with a scenario stated in the abstract. A rare-disease drug is tested in a small trial. It proves modestly effective with manageable side effects, and against the devastating nature of the disease, regulators judge that it helps more than it harms and approve it. Then, months later, with the drug widely prescribed, physicians begin noting deaths that never occurred in the trial, and serious side effects appearing far more often, some severe enough to require hospitalization. The real-world safety profile looks worse than the trial suggested.
I wrote earlier this week about the attribution problem this raises, the difficulty of separating harm caused by a drug from harm caused by the disease it treats, and I will not repeat it here beyond noting that in this case nothing has been definitively linked to the drug. The question worth taking up now is a different one, and it is the one Feuerstein's framing puts squarely on the table: why does this pattern recur so reliably, and what should anyone actually do when it does?
The trial population is not the real-world population
The most important reason has nothing to do with the drug changing. It is that the people taking it changed.
Clinical trials, and small rare-disease trials especially, enroll a selected group. Participants tend to be healthier than the average patient with the condition, with fewer complicating illnesses, a narrower age range, fewer other medications that might interact, and enough stability and support to meet the trial's demands. They are also watched with an intensity no ordinary clinic can match: scheduled visits, regular labs, clinicians alert for the first hint of trouble and empowered to intervene early.
Approval changes all of that at once. The drug becomes available to everyone with the diagnosis, including the sickest, the oldest, the most medically complex, those taking many other medicines, those with limited access to specialist care, and those monitored far less closely. So the molecule that regulators approved and the molecule now circulating are the same, but the population receiving it is not. The risk-benefit ratio that was measured is therefore not the risk-benefit ratio that now exists in the world, and the direction of the shift is predictable: a more fragile, less closely watched population will experience more serious events, from the drug and from everything else.
Small trials cannot see rare events at all
The second reason is statistical, and it is worth stating precisely, because it is routinely misread. A small trial does not have the power to detect a rare adverse event. If a drug causes a serious event in, say, one patient in five hundred, a trial of a hundred patients will very probably record none at all. The absence of such events in the trial is not evidence that the drug does not cause them. It is evidence that the trial was too small to find out.
There is a rough rule statisticians use: if an event never occurs among a given number of patients, the data are still consistent with a true rate as high as roughly three divided by that number. Observe zero deaths in a hundred patients, and a true death rate as high as about three percent remains compatible with what you saw. That is a wide range, and it is the range regulators knowingly accept when approving a drug for a disease so rare that a larger trial is not possible.
Which means "no deaths in the trial" and "several deaths once tens of thousands of doses have been given" are not in contradiction. They are exactly what you would expect to observe if the drug's true risk never changed at all, and the only thing that changed was how many patients were exposed and how carefully anyone was looking. Rare events require large numbers to become visible, and large numbers only arrive after approval.
So a worse-looking profile is expected, but not meaningless
Put those two together and a hard truth follows: for rare-disease drugs, the post-approval safety profile looking worse than the trial's is the normal case, not the aberration. It reflects a broader population and a bigger sample, and it would happen even for a drug whose risks are exactly what the trial implied.
This matters because the reflexive interpretation, that a company or a regulator missed something or hid something, is often simply wrong. The trial did not lie. It answered a smaller question than the one now being asked. But the reverse conclusion is equally wrong, and worth guarding against just as firmly: none of this means the new signals should be discounted. Post-approval data is where a drug's rare risks become knowable for the first time, and treating expected as unimportant would be its own failure. The correct posture is neither alarm nor dismissal. It is to recognize that the picture is only now coming into focus, and to update carefully as it does.
What new safety information should actually do
Here is where the practical question lies, and where the common framing goes wrong. When troubling safety data emerges, the debate is usually posed as a single binary for everyone at once: keep the drug available, or pull it. But that framing misdescribes the decision, because the risk-benefit balance is not one number that applies to all patients equally. It differs, sometimes enormously, from person to person.
For a patient whose hyperphagia is severe and dominates every hour of their life and their family's, a given level of risk may be clearly worth accepting. For a patient with milder symptoms, or with heart or kidney conditions that make a particular side effect more dangerous, the same risk may not be. New safety information, properly used, does not flip a switch. It narrows the population for whom the balance still favors treatment, and sharpens the monitoring that should accompany it, who should be watched more closely, for what, and how often.
That is why the most useful response to a signal like this is rarely a verdict and usually a refinement: identifying which patients face elevated risk, what warning signs precede serious events, and what monitoring catches them early. Regulators and clinicians are making a population-level decision about a whole class of patients. An individual family is making a decision about one person, with their own particular severity, their own other conditions, their own alternatives. Those are different questions, and they can honestly have different answers.
The harm that gets counted and the harm that doesn't
There is one more asymmetry worth naming, because it quietly shapes how these episodes unfold. If a drug stays available and a patient is harmed by it, that harm is visible, countable, and attributable. There is a name, a date, a report. If a drug is withdrawn or restricted and patients deteriorate from the untreated disease, that harm is diffuse and largely invisible. Nobody files a report on the hunger that returned, and no database records the family that lost the only treatment that had helped.
Both harms are real. Only one is easy to see, which creates a structural pull toward avoiding the visible error, sometimes at the cost of the invisible one. That pull deserves acknowledgment particularly in rare disease, where the untreated condition is often severe and the alternatives few. It is worth being clear that this cuts in both directions, though. There is also commercial pressure running the other way, toward keeping a drug on the market, and a company's financial interest in continued sales is exactly why independent scrutiny of safety signals matters. The point is not that caution is wrong. It is that both columns should be counted, and one of them is much harder to count.
For the families in the middle of this, none of the statistical structure changes the difficulty of what they face, and it would be wrong to pretend otherwise. What it can do is clarify what the new information means and what it does not. A drug looking riskier after approval than during its trial is the expected consequence of a broader population and a larger sample, not automatic proof that something was missed or that the drug should go. And a signal that is expected is still a signal, still worth investigating hard, and still capable of turning out to be real. Both of those things are true at once. What the emerging data should produce is not a single verdict handed down to everyone, but a sharper map: which patients are most at risk, which are most likely to benefit, and what to watch for along the way. For any individual family weighing whether the benefit still outweighs the risk for their own child, that map, drawn with their own physician and their own circumstances in view, is the thing worth waiting for and asking about. It is a harder answer than a headline provides, and it is the only honest one.
Primary sources
- STAT, Adam's Biotech Scorecard by Adam Feuerstein, for the framing of the risk-benefit recalculation facing patients and caregivers, and for the described pattern in which a rare-disease drug is approved on a small trial showing modest efficacy and manageable side effects, and then, months into wide prescribing, physicians note a small number of deaths that did not occur in the trial along with a marked rise in serious side effects including hospitalizations.
- STAT's earlier reporting, also by Feuerstein, that Prader-Willi syndrome experts warned clinicians of a safety signal for Vykat XR following reports of deaths and serious adverse events, while noting that none has been definitively linked to the drug.
- General, well-established background on clinical-trial methodology and pharmacoepidemiology, including the selection effects that make trial populations healthier and more closely monitored than real-world populations, the limited statistical power of small trials to detect rare events, including the "rule of three" upper bound when zero events are observed, and the role of post-marketing surveillance in identifying rare risks.