How to Change Your Mind

In 2009, a radiologist in San Diego looked at a mammogram and said the word no patient wants to hear: positive. The woman was 40 years old, no family history, no symptoms. She went home and Googled "mammogram accuracy" and found a number — 90% — and spent two weeks believing she was almost certainly dying. She wasn't. The probability she actually had cancer was about 9%. Bayes' theorem could have told her that in thirty seconds. Nobody taught her Bayes' theorem.

The Reverend Who Solved Everything

Thomas Bayes was not a mathematician. Well, not primarily. He was a Presbyterian minister in Tunbridge Wells, England, who spent his days tending to souls and his evenings tending to a problem that had nagged natural philosophers since the dawn of probability theory: how do you reason backwards from evidence to causes?

Forward reasoning is easy. If I tell you a bag is 90% red balls, you can predict what you'll pull out. But life never works that way. You pull out a red ball and need to figure out which bag you're dealing with. You see a symptom and need to figure out the disease. You observe data and need to figure out the theory. The world gives you effects and demands you deduce causes.

Bayes figured out the math. He wrote it up in a manuscript called "An Essay towards solving a Problem in the Doctrine of Chances" and then, as mathematicians sometimes do, he died before publishing it. His friend Richard Price found the manuscript, polished it up, and sent it to the Royal Society in 1763.1

The core insight is simple enough to state in English: what you should believe after seeing the evidence depends on what you believed before. This sounds obvious. It is obvious. And yet it is the single most frequently violated principle in all of human reasoning.

The Formula on the Napkin

Here it is. The whole thing. The theorem that launched a thousand arguments in statistics departments and quietly runs half the algorithms on your phone:

P(H|E)

Posterior — your updated belief after seeing the evidence

P(H)

Prior — your belief before seeing the evidence

P(E|H)

Likelihood — probability of seeing this evidence if the hypothesis is true

P(E)

Evidence — total probability of seeing this evidence under all scenarios

Let me translate this into something you can hold in your head. Imagine you're at a party and someone mentions they went to Harvard. You want to know: are they actually brilliant, or did they just have rich parents?

The prior is what you believed before they opened their mouth — maybe a 20% chance any given person at this party is genuinely brilliant. The likelihood is: how probable is it that a brilliant person would mention Harvard? (Pretty high — they're proud of it.) But here's the crucial part: the evidence term P(E) accounts for the fact that people of all intelligence levels mention Harvard at parties, possibly with even greater frequency if they have nothing else to brag about.

The denominator is where all the magic lives. It forces you to ask: what else could explain this evidence? Skip that question, and you get the wrong answer. Every time.

There's a beautiful alternative form that makes the machinery even clearer. Instead of thinking about raw probabilities, think about odds and likelihood ratios:

This form is what practicing Bayesians actually use, because it makes the update step trivially simple. If you started at 1:99 odds (1% prior) and the evidence has a likelihood ratio of 20 — meaning it's 20 times more likely if the hypothesis is true — your new odds are 20:99, or about 17%. One multiplication. No need to compute P(E) at all.

The Doctor's Dilemma

This is the example that makes Bayes' theorem famous, and it is the example that makes doctors squirm.

A rare disease affects 1 in 1,000 people. There's a screening test: if you have the disease, it correctly returns positive 99% of the time (sensitivity = 99%). If you don't have the disease, it correctly returns negative 95% of the time (specificity = 95%, meaning a 5% false positive rate).

You take the test. It comes back positive.

Quick: what's the probability you actually have the disease?

Here's why. Picture a stadium with 100,000 people — a useful number because it makes the arithmetic visible:

100 have the disease. The test catches 99 of them. (One unlucky person is missed.)

99,900 are healthy. But the 5% false positive rate tags 4,995 of them anyway.

Total positive results: 99 + 4,995 = 5,094. Of those, only 99 are truly sick. So if your test is positive, the probability you actually have the disease is 99 / 5,094 ≈ 1.9%.

The intuition-killer is the base rate. The disease is so rare that even a good test generates a flood of false alarms from the vast healthy population. The 4,995 false positives swamp the 99 true positives. It's not that the test is bad — it's that the prior is doing almost all the work.

Gerd Gigerenzer, the German psychologist who has spent decades studying this problem, discovered something hopeful. When you rephrase the problem using natural frequencies — "Out of every 10,000 people, 10 have the disease; of those 10, about 9 test positive; meanwhile, about 500 healthy people also test positive" — suddenly people get it right.3 Our brains aren't bad at Bayesian reasoning. They're bad at probabilities. Give them counts, and they do fine. Evolution trained us on frequencies, not fractions.

See It for Yourself

Enough reading. Play with the numbers. Each dot below represents a person. Watch what happens to the crowd of positive results as you change the base rate — that's where your intuition fails hardest.

Now try this: crank the base rate up to 10% and watch the posterior leap. The same test, the same accuracy — but because the disease is more common, a positive result means something completely different. The test didn't change. The context did. That's Bayes' theorem in one sentence.

The Second Test (Where the Magic Happens)

Your first test came back positive and the probability you're sick is only 2%. So your doctor says: "Let's run it again." You take the same test, and it's positive again.

Now what? This is where Bayes' theorem reveals its most beautiful property: today's posterior becomes tomorrow's prior. The 2% from your first test becomes the starting belief for interpreting the second test. And 2% is a much higher prior than 0.1%.

With a 2% prior, 99% sensitivity, and 95% specificity, the second positive test pushes you to about 29%. A third positive? About 89%. Each test ratchets the probability upward — slowly at first, when the prior is resisting, then faster as the evidence accumulates.

The War That Wouldn't End

For most of the 20th century, statistics had a civil war. On one side: the frequentists, led by the intellectual descendants of Ronald Fisher and Jerzy Neyman. On the other: the Bayesians, a ragtag band who thought the frequentists had made a catastrophic wrong turn somewhere around 1925.

The disagreement sounds technical, but it's philosophical, and it cuts deep. Frequentists say probability is about long-run frequencies. A fair coin means: flip it infinitely many times, half come up heads. Probability is a property of the physical world.

Bayesians say that's lovely but useless. When a climate scientist says there's a 90% chance humans are causing global warming, she doesn't mean that in 90% of parallel Earths we'd see warming. That's gibberish. She means: given her evidence and expertise, she is 90% confident. Probability is a degree of belief.

Fisher despised Bayesian methods. He called inverse probability "a mare's nest of unsupported claims."4 The Bayesians fired back that Fisher's p-values were a statistical abomination — a number that doesn't answer the question anyone actually wants to ask.

The war has softened in recent decades. Most working scientists use whatever works — frequentist methods for simple problems, Bayesian methods for complex ones. But the philosophical tension never went away, because it can't. It's a question about what probability means, and that's not the kind of question data can settle.

Spam, Suspects, and the Replication Crisis

If Bayes' theorem were only useful for terrifying people about medical tests, it would be a curiosity. But it quietly runs the modern world.

Your Inbox

Every time Gmail decides whether an email is spam, it's running Bayes' theorem. The prior is the base rate of spam (~45% of all email). The evidence is the words in the message. The likelihood ratio: how much more likely is the phrase "Nigerian prince" in spam versus legitimate mail? Paul Graham's 2002 essay "A Plan for Spam" kickstarted modern Bayesian spam filtering, and the core idea remains unchanged: each word in the email is evidence, each evidence has a likelihood ratio, and the ratios multiply together to produce a verdict.6

The Courtroom

DNA evidence is a Bayesian problem in disguise. A forensic analyst tells the jury: "The probability of a random person matching this DNA profile is 1 in 10 million." The jury hears: "There's a 1 in 10 million chance the defendant is innocent." Those are completely different statements.

To get from the first to the second, you need Bayes' theorem, and you need a prior: how many plausible suspects exist? If the city has 5 million people and no other evidence narrows the field, the expected number of random matches is 0.5. The DNA evidence is strong but not overwhelming. The prosecutor's fallacy — collapsing the likelihood into a posterior by ignoring the prior — is not an abstraction. It is a machine for producing wrongful convictions.7

The Replication Crisis

The crisis in psychology, medicine, and other fields is, at its heart, a Bayesian crisis. It's the medical test problem again: most hypotheses tested are wrong (that's why we test them), sample sizes are small (the test has low specificity), and the significance threshold is lenient (p < 0.05 means a 5% false positive rate). Run this through Bayes' theorem and you get exactly what Ioannidis found in his landmark 2005 paper: most published research findings are false.8

That bottom row is the medical test problem wearing a lab coat. The "disease" is "true finding." The "test" is "p < 0.05." And the base rate is not in your favor.

How to Actually Change Your Mind

Bayes' theorem isn't just a formula. It's a discipline of thought — maybe the best one we have for navigating a world that throws evidence of varying quality at you all day. Here are the lessons, stripped to bone:

1. Your prior matters. Admit you have one. Everyone walks into every question with existing beliefs. The Bayesian says: write them down. Make them visible. Then update rationally. The alternative — pretending to start from a blank slate — means your priors are hidden, unexamined, and probably wrong.

2. Extraordinary claims require extraordinary evidence. Carl Sagan was being Bayesian. If your prior for "aliens visited Earth" is 0.0001%, you need a likelihood ratio in the millions to budge it. A blurry photo — which is roughly equally likely whether aliens exist or teenagers exist — has a likelihood ratio near 1. It tells you nothing. This isn't closed-mindedness; it's arithmetic.

3. Update incrementally, not dramatically. You don't go from "I believe X" to "I believe not-X" in one step. Each piece of evidence nudges you. This is how rational belief change actually works — not as a dramatic conversion, but as a slow, relentless accumulation of signal. (This is what the sequential calculator above was showing you.)

4. Always ask: "how common is the thing I'm testing for?" Before asking "how good is this test?" ask "how prevalent is the condition?" Before evaluating evidence, establish the base rate. The denominator is not optional.

5. Beware the likelihood ratio of 1. Evidence that's equally likely whether the hypothesis is true or false tells you nothing. Your posterior equals your prior. This eliminates a shocking amount of "evidence" people treat as meaningful. ("He looks like a criminal" — well, what does a criminal look like? If the feature is equally common in criminals and non-criminals, it's noise, not signal.)

6. Seek disconfirming evidence. Confirmation bias is the anti-Bayesian sin. If you only look for evidence that confirms your hypothesis, you're only computing the numerator. You need the denominator — the probability of the evidence under alternative hypotheses — to update correctly. The most informative evidence is the kind that differentiates between hypotheses, not the kind that's consistent with your favorite one.

Reverend Bayes didn't set out to start a revolution. He was a minister with a math hobby, curious about a puzzle involving chances. But the thing he found — that there is a precise, mechanical, inarguable way to update beliefs in the light of evidence — turns out to be one of the most important ideas humans have ever had. Not because it's complicated. Because it's right. And because we so reliably get it wrong without it.