×
About Test Positive?
You test positive for something. How worried should you be? It depends on three numbers: how common the thing is among people like you, how often the test catches it (its sensitivity ), and how often it correctly clears people who don't have it (its specificity ). The answer is often far lower than people expect, doctors included, and this page shows why.
TEST : pick an example or set the three sliders. The rack holds 1,000 (or 10,000) sample tubes, one per person, sorted with everyone who has it first, then the false alarms. Dark caps have it, light caps don't; red caps with a + tested positive. Click a group under the rack to pick it out. The sentence underneath says the same thing in whole numbers. TEST AGAIN retests everyone who was positive with a second, independent test of the same accuracy, and then a third.
TREE draws the same crowd as a tree, and works the answer out in odds. COURT is the same mistake in a courtroom. SIMULATE tests made-up people one at a time and shows the share of positives who really have it settling on the Expected value. COPY LINK copies the current settings.
THE STORY
The rule is named after Thomas Bayes, an English minister and mathematician. His essay on reasoning backwards from what you see to its likely cause was found among his papers after his death and published by his friend Richard Price in 1763. Pierre-Simon Laplace found the rule for himself in 1774 and turned it into a working tool.
In 1978 Ward Casscells and colleagues asked 60 doctors and medical students at Harvard Medical School's teaching hospitals about a test with a 5% false-positive rate for a disease 1 in 1,000 people have: if someone tests positive, how likely are they to have it? The most common answer was 95%; only a minority said about 2%, the right answer. The psychologist Gerd Gigerenzer has spent years on why. His answer is that chances like "90% sensitive" are hard for people to combine, but the same facts as natural frequencies ("of 1,000 women, 10 have it, 9 of them test positive...") are easy, and in his studies doctors given the figures that way got the answer right far more often. The Doctors' quiz example uses his mammography figures: 1% prevalence, 90% sensitivity, 9% false positives. The answer is about 1 in 10.
The same slip in a courtroom is the prosecutor's fallacy (named by William Thompson and Edward Schumann in 1987): mixing up "the chance an innocent person would match" with "the chance this matching person is innocent". The COURT tab tells the case of Sally Clark, where a misused statistic played a part in a wrongful conviction.
HOW IT WORKS
Bayes' theorem: P(has it | positive) = P(positive | has it) × P(has it) ÷ P(positive), where P(positive) counts both the true and the false positives. With prevalence p, sensitivity se and specificity sp, that is p·se ÷ (p·se + (1 − p)(1 − sp)). The odds form says the same more simply: the odds after a positive are the odds before it times the likelihood ratio , se ÷ (1 − sp). A second positive multiplies by the likelihood ratio again.
That retest rule assumes the tests are independent : each one's mistakes have nothing to do with the last one's. Real retests often aren't (the same thing that fooled the first test can fool the second), so in real life a second positive is usually worth less than the page shows. The examples' figures are illustrative, except the Doctors' quiz.
The rack and the sentence use whole numbers, rounded down the tree the way you'd write them by hand, so they can differ a little from the exact percentage. When a group would be too small to count in 1,000 people the sentence uses a bigger crowd and says so. The simulation draws a random number for each person and each test (a seeded generator), so its counts wobble around the Expected ones; the shaded band on its chart is a rough 95% range.
FURTHER WATCHING AND READING