Approved is not the end

A new drug clears every test. Phase 1 said it was safe. Phase 2 saw a signal. Phase 3 beat the standard treatment. Three thousand people took it, the regulators read the file, and it goes on sale. Then, two years later, a warning is added: a rare heart problem, about one case in every ten thousand people. The headlines call it a failure of the system. It is not. It is the system working exactly as it has to — because of a piece of arithmetic that was true the whole time, and that no trial could ever have caught.

What approval actually proves

Approval is not a certificate of perfect safety. It is a bet, made on the best evidence available, that the drug helps more than it harms — on average, across a few thousand people.

Read those last words carefully, because each one is a limit.

A few thousand people. A phase 3 trial is large by trial standards and tiny by world standards. It might enrol three thousand. The drug, once approved, may be taken by three million. That is a thousand-fold jump in exposure, and rare things that hide in three thousand cannot hide in three million.

On average. The trial reports what happened to the group. It says the average person did better. It does not promise that you will — only that, across everyone, the benefit outweighed the harm.

So approval answers a real question well: is this worth releasing? It does not, and cannot, answer two others: what happens at scale, and what happens to one specific person. Those only reveal themselves later.

The arithmetic of a rare harm

Here is the piece of maths that explains the two-years-later warning.

Suppose a side effect strikes 1 person in every 10,000 who take the drug. Now ask how many cases you would expect to see in a trial of 3,000 people.

The sum is simple: 3,000 ÷ 10,000 = 0.3 expected cases.

You cannot have a third of a case. In real life this means: most trials of this size see zero. A few unlucky ones see one. The expected number is below one, so the most likely count is none at all. The trial ends, the safety table is clean, and everyone involved is being honest — the effect genuinely did not appear, because it was too rare to.

Now release the drug.

• 50,000 people take it → 50,000 ÷ 10,000 = 5 expected cases. A signal starts to stir.
• 1,000,000 people take it → 1,000,000 ÷ 10,000 = 100 expected cases. Now it is undeniable.
• 10,000,000 people take it → 1,000 expected cases. A clear, countable pattern.

Nothing about the drug changed. The rate was 1 in 10,000 from the first day. What changed is how many people were exposed — and a rate that is invisible at 3,000 is loud at a million. The trial did not miss it through sloppiness. It missed it because the maths made it nearly impossible to see. You will run this exact arithmetic yourself in the lab in a moment.

Why this is normal, not a scandal

A trial is a filter sized to catch common and medium harms before release. Catching a 1-in-10,000 effect before approval would mean enrolling hundreds of thousands of people, which would take many extra years and keep a working drug from everyone who needs it now. Every system makes this trade: release on strong-but-incomplete evidence, then keep watching.

That watching has a name — post-market surveillance. After launch, regulators and companies track reports from the real, much larger population. It is the only place a 1-in-10,000 effect can finally be seen. So the warning that appears two years later is not the trial having failed. It is the next stage of the same investigation, running on the only sample big enough to answer the question.

This is also why “approved” should be read as “approved so far, on what we know.” It is a strong statement. It is not a final one. None of this is a reason to fear a medicine or to stop one — that is a conversation for a person and their doctor, who can weigh a known benefit against a rare risk for that individual. The point here is only how to read the news: a post-approval warning is the machine learning more, not the machine breaking.

”Works on average” is not “works for you”

There is a second gap, separate from rarity. The trial’s headline — say, “lowers the risk by a third” — is an average over everyone in it. Underneath that average, people vary. Some got most of the benefit. Some got little. A few may have got none, or had a private reaction the group number washed out.

This is why your own response to a medicine can differ from the figure in the press release, in either direction, and why doctors adjust and monitor rather than treating the average as a personal guarantee. The headline number is true and it is about the crowd, not about you.

On the whole

Almost every powerful claim in this course has been a claim about a group — a trial group, a population, an average. The honest reader holds two facts at once: the average is real, and you are one person standing inside a distribution it summarises. A drug approval is a snapshot of strong evidence at one moment, with millions of exposures and your own particular body still ahead of it. Seen that way, the after-the-fact warning stops looking like a betrayal and starts looking like what it is: a large, slow, shared experiment, still telling us the truth about ourselves — quietly correcting the picture long after the headline has moved on.