There’s one worked example that teaches Bayesian updating better than any abstract formula, and you’ve already met it twice — once in the last course, once in this course’s pretest. It’s time to take it apart completely, because once the medical test stops feeling like a trick, every other update in your life gets easier. The setup is deliberately simple: a disease, a test, a positive result. The answer is famously counter-intuitive, and the cure for the bad intuition is a picture, not an equation.
Here’s the scenario we’ll use for the whole lesson. A disease affects 1 in 1,000 people. A test for it is “99% accurate”, which we’ll pin down to mean two things: if you have the disease it correctly says positive 99% of the time (a 99% true-positive rate), and if you don’t have it, it wrongly says positive 1% of the time (a 1% false-positive rate). You take the test. It comes back positive. What’s the chance you actually have the disease?
Before you read — take a guess
Commit to a gut answer before we work it out. A disease hits 1 in 1,000. The test has a 99% true-positive rate and a 1% false-positive rate. You test positive. The chance you're actually sick is closest to…
Don’t use percentages — use people
The reason this puzzle feels impossible in percentages is that percentages hide the base rate. So throw them out and count actual people instead. This is the single most powerful trick in all of Bayesian reasoning, and it has a name: natural frequencies. Instead of juggling “1%” and “0.1%” and “99%”, we drop a concrete crowd of people into the problem and just tally who ends up where.
Take 100,000 people and walk them through the scenario step by step.
- The disease affects 1 in 1,000, so of our 100,000 people, 100 have the disease and 99,900 do not. That split — 100 versus 99,900 — is the base rate, and it’s doing almost all the work.
- Of the 100 who have it, the test correctly flags 99% → 99 true positives (and 1 sick person wrongly cleared).
- Of the 99,900 who don’t have it, the test wrongly flags 1% → 999 false positives.
Now just look at everyone who tested positive. There are 99 + 999 = 1,098 of them. Of those, only 99 actually have the disease. So the chance that a positive-testing person is truly sick is:
99 ÷ 1,098 ≈ 0.090, about 9%.
The whole “paradox” evaporates. The test is genuinely good — but the healthy crowd is so enormous that even 1% of it (999 people) swamps the tiny sick crowd (99 people). Most positives are false alarms not because the test is bad, but because there’s so much more “healthy” to wrongly flag than “sick” to correctly flag.
| Group | How many | Test says positive | Of those, who’s truly sick? |
|---|---|---|---|
| Have the disease | 100 | 99 (true positives) | 99 |
| Don’t have it | 99,900 | 999 (false positives) | 0 |
| Everyone who tests positive | — | 1,098 | 99 → about 9% |
The natural-frequency habit
Whenever a probability puzzle has you stuck, stop computing percentages and imagine a fixed crowd — 1,000 or 100,000 people — then walk them through the branches and count who lands where. Bayes’ theorem is exactly this tally; the formula just hides the people. Research on doctors and judges shows that the same problem they get wrong in percentages, they get right in natural frequencies.
Watch the answer move: the posterior calculator
Reading the tally once convinces your head; driving it convinces your gut. Below is the same problem as a live calculator. It starts on our scenario — a 0.1% base rate (that’s 1 in 1,000), a 99% true-positive rate, and a 1% false-positive rate — and shows the posterior (your real chance of disease after a positive test) as both a percentage and a crowd of people.
Play with all three sliders and watch the posterior lurch:
Bayesian updating
A rare disease and a 99%-accurate test
Set the three numbers. The posterior — your chance of the hypothesis after the evidence — updates live.
- 0.1%
- 99%
- 1%
After a positive test, the chance you actually have the disease is 9%.
Out of 100,000 people
Of 1,098 who test positive, only 99 truly have the disease — the other 999 are false alarms.
■ true positives · ■ false positives · 100 of 100,000 truly have the disease
Three experiments worth running before you move on, because each one teaches a different lever:
- Make the disease ten times more common — drag the prior from 0.1% up to 1%. The posterior jumps from ~9% to roughly 50%. Same test, same accuracy; the only thing that changed is how much “sick” there was to find. The base rate is the dominant dial.
- Cut the false-positive rate to 0.1% (a far more specific test). The posterior leaps to about 50% even at the original rare base rate. False alarms, not missed cases, are what poison a screening test for a rare condition.
- Crank the true-positive rate up and down between, say, 90% and 100%. The posterior barely flinches. Catching more of the rare sick people can’t help much when they were rare to begin with — there just aren’t many of them to catch.
Using the calculator's logic: which single change would raise the chance-you're-actually-sick (the posterior) the MOST, starting from the rare-disease scenario?
The second test: updating again
So you tested positive and the honest read is “about 9%, probably a false alarm — but worth checking.” What does a careful person do? Run a second, independent test. And here’s the beautiful part: Bayesian updating tells you exactly what to do with the result, because your posterior from the first test becomes your prior for the second.
You no longer walk into the second test believing “1 in 1,000.” After the first positive, your honest belief is “about 9%.” So plug a 9% prior into the same machinery. If the second independent test also comes back positive (99% true-positive, 1% false-positive again):
- Out of 100,000 people who are each at a 9% prior: 9,000 have the disease, 91,000 don’t.
- True positives:
9,000 × 0.99 = 8,910. False positives:91,000 × 0.01 = 910. - Posterior:
8,910 ÷ (8,910 + 910) ≈ 0.907, about 91%.
One positive test took you from 0.1% to 9%. A second took you from 9% to 91%. That’s updating by degrees — each piece of evidence moves you the right amount, and the moves compound. The single test that felt useless (“only 9%?”) was actually a huge update; it multiplied your odds nearly a hundredfold. It just started from such a tiny base that even a hundredfold jump landed at 9%.
“Independent” is doing real work
The second-test trick only works if the second test is independent of the first — a different method, not the same test repeated (which would just re-confirm the same false positive). If two tests share the same flaw, a second positive tells you almost nothing new. We’ll meet this trap again as double-counting evidence: correlated clues feel like two confirmations but are closer to one. When sources aren’t independent, updating twice over-counts.
After one positive test your chance of disease is about 9%. You take a second, independent test and it ALSO comes back positive. Roughly what should you now believe?
Where this worked example lies to you
The medical test is the perfect teacher, but lift two warnings out of it before you generalize.
First: the base rate has to be the right one. Our 1-in-1,000 figure was the base rate for the general population. If you’re tested because you already have symptoms, or because the disease runs in your family, your prior is no longer 1 in 1,000 — it might be 1 in 10, and then a positive test is genuinely alarming. The whole calculation is only as good as the prior you feed it, which is why choosing the right reference class (from the base rates lesson) is half the battle. Screening a low-risk crowd and testing a high-risk patient are different problems with different priors and different posteriors.
Second: “9%” is not “ignore it.” A 9% chance of a serious disease is not nothing — it’s 90 times your starting risk and absolutely worth a follow-up test. Bayesian updating doesn’t tell you to dismiss the evidence; it tells you to size it correctly and act accordingly. The mistake the panicking patient makes isn’t taking the test seriously — it’s jumping straight to “I have it” and skipping the second test that would actually settle the question.
When to reach for it
Reach for the natural-frequency tree any time a test, signal, or filter flags something rare — a medical screen, a fraud alert, a spam filter, a security scanner, a “this candidate fits the profile” hunch. The trigger is the phrase “the test is X% accurate” (or “the model is X% accurate”) attached to a rare target. The instant you hear it, refuse to read accuracy as the answer. Instead: name the base rate, drop 100,000 imaginary cases into the branches, and count the true positives against the false positives. It takes thirty seconds and it routinely turns a panic into a sensible “let’s confirm that.”
Check yourself: the medical test
Why is the chance of disease after a positive test (about 9%) so much lower than the test's 99% accuracy?
Check your answer to continue.
Where this goes next
You’ve now solved the puzzle that breaks most people’s intuition — and you did it by counting people, not by memorizing a formula. But the counting is a formula in disguise, and naming it lets you update without drawing a tree every time. Next, in “Bayes’ Theorem,” we lift the equation straight out of the natural-frequency tree you just built, then meet its friendlier twin — the odds form — that lets you do these updates in your head: prior odds, times a likelihood ratio, equals posterior odds.