The Test Was Right and the Doctor Was Wrong

The letter arrives on a Tuesday. The screening result is positive, and the leaflet that comes with it says, in careful type, that the test correctly identifies the condition in ninety percent of cases. The person holding the letter does what any of us would do. They read ninety percent as their odds. They spend a week not sleeping.

At the follow-up appointment the specialist is oddly calm, and explains why. The condition is rare: it affects about one person in a thousand. So imagine a thousand people took the test. One of them has the condition, and the test very probably catches them. But of the nine hundred and ninety-nine who are healthy, a test that is wrong one time in ten flags about a hundred of them as positive too. A positive letter, then, went to roughly a hundred and one people, and only one of them is ill. The odds on the Tuesday were never nine in ten. They were closer to one in a hundred.

Nothing about the test changed between the letter and the appointment. What changed was that the specialist asked a question the leaflet had left out: before the result arrived, how likely was this?

The half we throw away

That question is the whole of the idea, and the arithmetic that follows from it was written down by a Presbyterian minister in the eighteenth century and has been making people uncomfortable ever since. The discomfort has a shape. We like to think that evidence should be taken on its own terms. A result is a result. Bringing in what we believed beforehand feels like cheating, like letting our prejudices vote.

But look at what happens when we refuse. The test above is identical in a hospital ward where half the patients have the condition and in a shopping centre where almost nobody does. Same machine, same accuracy, same positive reading. In the ward it is strong evidence; in the shopping centre it is nearly noise. If we insist on reading the result without asking where it was taken, we are not being objective. We are silently assuming that the ward and the shopping centre are the same place, which is a belief about the world, and a false one.

There is no view from nowhere. Every reading of evidence comes with an assumption about how likely things were before it arrived, and the only choice we have is whether to state that assumption out loud or hide it from ourselves. The rare, uncomfortable discipline is to say the prior plainly, hold it loosely, and let each new fact move it by the right amount.

Three places it shows up

Consider an argument between two friends about whether a mutual acquaintance can be trusted. One has known him for ten years and has one bad story; the other met him last month and has one good one. They trade the stories and get nowhere, because a story is a piece of evidence and neither is telling the other what it landed on. The ten-year friend's single bad story lands on a decade of good ones and barely moves the needle. The new friend's good story lands on nothing and moves it a long way. They are not disagreeing about the evidence. They are disagreeing about the prior, and neither has said so. The arguments that actually resolve are the ones where people trade histories, not verdicts.

Consider a spam filter, or any system that learns. It does not decide that an email is junk from the email alone. It carries a running estimate of how much junk arrives, and each word in the message nudges that estimate one way or the other. The estimate after the first word becomes the starting point for the second. By the end of the message it has done a few hundred small updates in a chain, and the verdict is the last link. Nobody programmed the answer. They programmed the updating.

Consider a detective, the fictional kind. The pleasure of the genre is watching someone hold a suspicion at exactly the right strength. The butler is unlikely, so the mud on his boots moves the case only a little. The heir is likely, so the same mud moves it a lot. The bad detective treats every clue as equally decisive. The good one knows that a clue is only as strong as the world it arrives in.

Where it sits on the map

This idea sits at the end of a journey about the foundations of thought, and it is placed there for a reason. It leans on metacognition, the habit of watching our own reasoning while it runs, because the prior is invisible until we look for it. It sits close to base-rate neglect, which is the name for the error in the opening story: the moment the leaflet's ninety percent is read as the patient's odds. It is a near neighbour of Occam's razor, which is really a rule about priors: the simpler story starts with more weight, so it takes more evidence to dislodge it.

And it opens onto a harder question, the one about whether two honest people who share their reasoning can go on disagreeing. If both are updating correctly on the same evidence, their starting points get washed out over time and they converge. The disagreements that never converge are the ones where somebody's prior is not being updated at all, and is not on the table to be examined.

The question it leaves open

The example with the letter is clean because the numbers are known. Most of the priors we carry are not. How likely, before this conversation, was it that the colleague meant what we think they meant? How likely, before this quarter's figures, was it that the strategy was sound? We do not know those numbers, and yet we update on evidence every day as if we did, and the updating is where the errors live.

So the real question is not how to do the sum. It is how to notice, in the moment, that there is a sum being done at all, and what we had quietly put on the scale before the evidence arrived. That is not something an essay can do for us. On Constellate, Bayesian Thinking is one guided conversation that ends when we can find our own prior in a case we actually care about. The first concept of every course is free.