Bayes' Theorem

See why a positive result from an accurate test can still be unlikely to be right

Intermediate
Presets
Test the same person again
Each button treats the current posterior as the new prior, assuming the repeat test errs independently of the first.

1,000 people tested

  • Has the disease, tests positive: 9
  • Has the disease, tests negative: 1
  • Healthy, tests positive: 89
  • Healthy, tests negative: 901

Natural frequency tree(counts rounded to whole people)

1.0%99.0%90.0%10.0%9.0%91.0%1,000 people10 have it990 do not9 test +1 test −89 test +901 test −

After a positive test

P(disease | +)
9.2%
From the counts
9 / 98
P(+) overall
9.8%
Likelihood ratio LR+
10.00

After a negative test

P(no disease | −)
99.89%
From the counts
901 / 902
P(disease | −)
0.11%

Bayes' theorem with these numbers

P(D | +) = P(+ | D) P(D) / P(+)

= (0.9000 × 0.0100) / (0.9000 × 0.0100 + 0.0900 × 0.9900)

= 0.00900 / 0.0981

= 0.0917

Most positives here are false alarms: 89 healthy people test positive against 9 sick ones. Ignoring how rare the disease is makes the test look far more conclusive than it is: the base-rate fallacy.

Turning P(evidence | cause) into P(cause | evidence)

Tests are described by how they behave when you already know the truth. Sensitivity is P(+ | disease), the share of sick people who test positive. Specificity is P(− | healthy), the share of healthy people who test negative. But a patient holding a positive result wants the reverse: P(disease | +). Bayes' theorem is the rule that converts one into the other.

The conversion depends on the base rate. Take a test with 90% sensitivity and 91% specificity, used on a population where 1% have the disease. Out of 1,000 people, 10 are sick and 9 of them test positive. Of the 990 healthy people, 9% (about 89) also test positive. So 98 people test positive and only 9 of them are sick: P(disease | +) = 9/98 ≈ 9.2%, even though the test is "90% accurate".

Thinking in counts of people like this, called natural frequencies, makes Bayes' theorem almost obvious. The formula does the same calculation with probabilities instead of people.

Prior

What you believed before the evidence. For a screening test, the prior is the prevalence: the share of tested people who have the disease.

Likelihood

How likely the evidence is under each explanation: P(+ | disease) is the sensitivity and P(+ | healthy) is 1 − specificity.

Posterior

What you should believe after the evidence: P(disease | +). It weighs the likelihoods by how common each explanation was to begin with.

Updating

Today’s posterior is tomorrow’s prior. Each new independent piece of evidence updates the belief again.

Key formulas

Bayes' theorem

P(D | +) = P(+ | D) P(D) / [P(+ | D) P(D) + P(+ | D′) P(D′)]

P(D)
prior: prevalence of the disease
P(+ | D)
sensitivity
P(+ | D′)
1 − specificity, the false positive rate
P(D | +)
posterior, also called the positive predictive value

Negative result

P(D′ | −) = P(− | D′) P(D′) / [P(− | D′) P(D′) + P(− | D) P(D)]

Odds form

posterior odds = prior odds × LR+

LR+
positive likelihood ratio = sensitivity / (1 − specificity)

Same test, different base rates

A test with 90% sensitivity and 91% specificity. Only the prevalence changes.
PrevalenceP(disease | +)P(healthy | −)What it means
0.1%1.0%99.99%About 99 of every 100 positives are false
1%9.2%99.89%About 9 of every 10 positives are false
10%52.6%98.79%A positive is roughly a coin flip
50%90.9%90.10%Now a positive is usually right

How to use this visualizer

  1. 1
    Set the prevalence, sensitivity and specificity, or pick a preset. The grid sorts 1,000 people: circles have the disease, squares do not, and filled shapes tested positive.
  2. 2
    Tick Show only positive tests. Compare the filled circles (true positives) with the filled squares (false positives): P(disease | +) is the share of circles.
  3. 3
    Follow the natural frequency tree from 1,000 people down to the four groups, and check that the panel on the right gives the same answer from the formula.
  4. 4
    Lower the prevalence and watch false positives swamp the true ones, even with sensitivity and specificity at 99%.
  5. 5
    Press Positive again to retest the same person. The posterior becomes the new prior, and a second positive usually raises the probability a lot.

The base-rate fallacy

It is tempting to read "90% sensitive" as "a positive means a 90% chance of disease". That ignores the base rate: when the disease is rare, the healthy group is so large that even a small false positive rate produces more false alarms than there are sick people. Always ask how common the condition is among the people being tested.

Practice problems

Problem 1: A screening test

A disease affects 2% of the people screened. The test has 95% sensitivity and 90% specificity. What is the probability that someone who tests positive has the disease? What is the probability that someone who tests negative has it?
Show solution

Out of 1,000 people: 20 sick, 980 healthy.

True positives: 95% of 20 = 19. False negatives: 1.

False positives: 10% of 980 = 98. True negatives: 882.

P(D | +) = 19 / (19 + 98) = 19 / 117 ≈ 0.162

With the formula: 0.95 × 0.02 / (0.95 × 0.02 + 0.10 × 0.98) = 0.019 / 0.117 ≈ 0.162

P(D | −) = 1 / (1 + 882) = 1 / 883 ≈ 0.00113

P(disease | +) ≈ 16.2%; P(disease | −) ≈ 0.11%

Problem 2: Testing again

The person from Problem 1 who tested positive takes the same test again, and it is positive again. Assuming the two results are independent given the person's true status, what is the new probability that they have the disease?
Show solution

New prior: P(D) = 19/117 ≈ 0.1624

Numerator: 0.95 × 0.1624 ≈ 0.1543

Denominator: 0.1543 + 0.10 × 0.8376 ≈ 0.1543 + 0.0838 = 0.2381

P(D | +, +) ≈ 0.1543 / 0.2381 ≈ 0.648

Check with odds: prior odds 19/98 ≈ 0.194, LR+ = 0.95 / 0.10 = 9.5, posterior odds ≈ 1.842, probability = 1.842 / 2.842 ≈ 0.648

About 64.8%, up from 16.2% after one positive and 2% before testing

Related topics

Discussion

Questions, ideas or a different approach? Join the conversation.