Probability & Uncertainty

Update Your Belief

A test says positive. Should you panic? Count the dots to find out, then test again and watch your belief move.

Slide the sliders and watch the dots. Red are real, orange are false alarms. A positive is only as good as that ratio!

1000 pretend people

sick and flagged healthy but flagged sick, missed healthy, quiet

Challenges

Trust a positive: change the sliders until a positive result means 90% or more.
The rare surprise: make it very rare (1% or less) with a test that catches 90%+. Positive now means under 25%!
Second opinions: start at 5% or less, then press Another positive until you believe 90% or more.
Good news: start at 10% or more, then use negatives to push the belief under 1%.

What just happened?

For grown-ups: what's really going on

This is Bayes' rule: P(H | +) = P(+ | H) P(H) / [P(+ | H) P(H) + P(+ | ¬H) P(¬H)]. Here P(H) is the prior (prevalence), P(+ | H) is the sensitivity (true-positive rate), P(+ | ¬H) is the false-positive rate, and the result is the posterior.

  • The dot grid uses natural frequencies (Gigerenzer): counting people instead of multiplying probabilities makes the answer much easier to see. Posterior = red / (red + orange).
  • Base-rate neglect is the common mistake of ignoring the prior. With 1% prevalence, 90% sensitivity and 10% false positives, only about 8% of positives are real.
  • Repeated tests are applied sequentially: yesterday's posterior is today's prior. This assumes the tests are conditionally independent given the truth, which real repeated tests often are not (the same flaw can fool both).
  • In odds form it is simpler: posterior odds = prior odds × likelihood ratio, where the positive likelihood ratio is sens / fpr.
  • Spam filters (the earlier lesson) and many classifiers use exactly this update, word after word.

Further reading: Gigerenzer and Hoffrage (1995), How to Improve Bayesian Reasoning Without Instruction · 3Blue1Brown, Bayes' theorem.