The Test Result That Fooled the Doctors

3 min read ▶ Listen to this post

In the late 1990s, the psychologist Gerd Gigerenzer posed a question to a room of experienced gynecologists in Germany. A woman in a routine screening program gets a positive mammogram. Breast cancer prevalence in this population is 1%. The test correctly flags 80% of women who actually have cancer. It also incorrectly flags 9.6% of women who don't. Given a positive result, what's the probability this particular woman actually has breast cancer?

The correct answer, worked out with Bayes' theorem, is about 8%. Most of the doctors didn't say anything close to 8%. Common answers clustered around 70% to 80% — many of them simply repeated the test's 80% sensitivity figure back as if it were the answer to a different question. Gigerenzer wasn't the first to find this: a 1978 study by Casscells, Schoenberger, and Grayboys asked Harvard Medical School staff and students a structurally identical question about a rare disease, and most gave an answer roughly 50 times too high. These weren't careless people. They were physicians, doing their best, getting a well-defined probability question wrong by an order of magnitude — because the question was never really about the test's accuracy. It was about how rare the disease was to begin with, and that number is the one everyone's eye slides past.

The mechanism: accuracy and rarity aren't the same input

This is base-rate neglect. A test's accuracy tells you how well it distinguishes sick from healthy people within each group. It says nothing on its own about what a positive result means once you factor in how rare the condition is in the population being tested. When a condition is rare, even a test with a low false-positive rate will, in absolute terms, generate more false positives than true positives — simply because there are so many more healthy people to draw false alarms from than sick people to correctly flag. The 80%-sensitivity number is real and it's doing real work, but it's answering "if you're sick, will the test catch it?" — not "if the test caught something, are you sick?" Those are different questions with different answers, and mixing them up is the single most common way smart people misread a positive test result.

Gigerenzer's more useful finding was about the fix. When he gave the same physicians the identical problem restated as natural frequencies — "out of 1,000 women, 10 have cancer; of those, 8 will test positive; of the 990 without cancer, about 95 will also test positive" — accuracy on the same underlying math jumped from roughly 1 in 5 correct to more than 4 in 5. The doctors weren't bad at reasoning. They were being handed the numbers in a format built to hide the base rate, and the moment it was made visible, most of them could do the arithmetic just fine.

Where else this shows up

• Academic and grant-funded research. A diagnostic or screening instrument reported as "92% accurate" can still produce a majority-false-positive result set once it's applied to a population where the target condition is uncommon — and "92% accurate" is often the only number that makes it into the abstract.
• Non-profit program evaluation. Risk-screening tools used to triage which families, students, or clients get priority services are usually validated on accuracy or sensitivity alone. If the risk condition they're screening for affects a small share of the population you actually serve, most of the people flagged "high risk" can be false positives, even with a well-validated tool.
• AI and LLM evaluation. A moderation or safety classifier reporting 99% accuracy sounds close to perfect, but if the behavior it's flagging shows up in well under 1% of real traffic, the majority of its flags in production can still be false alarms — a fact the headline accuracy number was never built to reveal.

What to ask for instead

Don't stop at a test's or a model's accuracy, sensitivity, or precision on its validation set. Ask the question those numbers can't answer by themselves: of everything this actually flags in the real population, how many flags turn out to be right? That number depends on the base rate as much as it depends on the tool, and it's rarely reported next to the headline accuracy figure — usually because nobody asked for it.

If you're evaluating a screening tool, a risk score, or a classifier and the accuracy claims don't account for how common the thing being detected actually is, talk to us before you trust what the positive results are telling you.

One email whenever we publish something new. No spam, unsubscribe anytime.