Juliana Kasmanas
Take the test

False Positive

Data visualization of Bayes' theorem.

False Positive is an interactive simulator that illustrates Bayes' theorem through the mammography problem, a 2007 study in which most doctors misjudged what a positive test result means.

  • Personal project
  • 2026
  • Web
  • Live

The studies

The test comes from a 2007 study in which psychologist Gerd Gigerenzer asked 160 gynecologists: A woman tests positive on a screening mammogram. What is the chance she has breast cancer?1 Gigerenzer gave them three numbers: 1% of women have breast cancer, 90% of the women who have it test positive, and 9% of the women who do not have it test positive anyway. Most of the doctors got the answer wrong by a wide margin:1

About 9% the best answer

60% of the doctors said 81% or 90%

Gigerenzer then repeated the experiment with the same facts expressed as head counts: Of 1,000 women, 10 have breast cancer, and 9 of them test positive. Of the 990 women without cancer, about 89 test positive too. A woman tests positive. What is the chance she has breast cancer?1 This time, most of the doctors found the best answer:1

87% of the doctors now found the best answer

60% the first time had said 81% or 90%

Gigerenzer concluded that the problem is not the doctors, but how the numbers are presented.1 Percentages have to be combined mentally. Head counts put everyone into one group, so the ratio is there to see: 9 of the 98 who test positive.

Formula vs image

Below, the same information is presented in two ways: how many of the people who test positive actually have the disease. Presenting the information as a formula requires calculation and math knowledge. Presenting it as an image turns the ratio into something visible.

9% as a formula

0.90 × 0.01 0.90 × 0.01+0.09 × 0.99 = 0.009 ÷ 0.0981

Requires knowing Bayes' theorem.

9% as an image

Only requires looking at the group.

Design

Both studies show that how information is presented shapes how well people understand a problem, even when they are specialists in the field. That finding led to six design choices in the simulator:

1%10 of 1,000 people

Showing people instead of percentages

Presenting the problem in percentages requires mental calculation and is easy to lose track of, which is what fooled the doctors. Turning the numbers into 1,000 dots gives each person a visible place in the group.

Your guess 50%

Asking for the user's guess first

Committing to a number first turns the final answer into feedback. If the page simply showed the 9% up front, it would be a fact to read and forget. Measured against an initial guess, the gap is clear.

Explaining one step at a time

Each step adds a single fact to the problem. Step 1 presents the problem and asks for a guess. Step 2 turns the numbers into 1,000 dots. Step 3 marks the people who are sick. Step 4 runs the test across everyone, and Step 5 isolates the positive results: 9 of 98.

Keeping the same picture throughout

The 1,000 dots never leave the screen; each step only re-sorts them. Nothing is swapped behind the scenes, so the 9 out of 98 at the end is visibly drawn from the same crowd shown at the start.

  • sick, tested positive
  • sick, tested negative
  • healthy, tested positive
  • healthy, tested negative

Displaying more than one meaning

Every dot answers two questions at once. Color shows who really has the disease, and fill shows the result of the test. Where they disagree, the test's mistakes stand out immediately. The palette uses high-contrast colors so the states remain distinct for users with color-vision deficiencies.

prevalencehow common is it?sensitivityhow many sick test positive?false-positive ratehow many healthy test positive?

Using common words

Standard probability tools label the three sliders prevalence, sensitivity, and false-positive rate, terms that keep people out of this subject. The simulator asks the same three questions in everyday words.

Process

Version 1

The interactive test lived on one page and the case study on another, forcing readers to jump between them.

Version 2

Bringing them onto one page put the icon, title, pills, and facts beside the question, crowding the frame.

Now

The test has its own page again, and the frame opens straight with the test question.

AI workflow

Prototyping

I built every version of the page as a working prototype with Claude so I could test the flow directly in the browser. Trying each layout live is how the project went from two separate pages to one, and then back to two.

Pressure testing

I used Claude to test the simulator across extreme slider values and screen sizes to make sure the dots and labels held up. When extreme inputs pushed a label off the edge, I updated it to flip inward and read the other way.

Facts

Rather than relying on AI summaries of the research, I checked every claim, citation, and number directly against the original 1978, 2007, and 2014 papers to make sure the math and study results were exact.

Sources

  1. Gerd Gigerenzer, "Helping doctors and patients make sense of health statistics: towards an evidence-based society", ICOTS8, 2010. Read the paper (opens in a new tab) The study was first reported in Gigerenzer, Gaissmaier, Kurz-Milcke, Schwartz and Woloshin, Psychological Science in the Public Interest, 2007. Journal page (opens in a new tab)

  2. Ward Casscells, Arno Schoenberger and Thomas Graboys, "Interpretation by physicians of clinical laboratory results", New England Journal of Medicine, 1978 (the "Harvard" example in the Simulator). Journal page (opens in a new tab)

  3. Arjun Manrai, Gaurav Bhatia, Judith Strymish, Isaac Kohane and Sachin Jain, "Medicine's Uncomfortable Relationship With Math: Calculating Positive Predictive Value", JAMA Internal Medicine, 2014. It reports the results of the 1978 study and of its 2013 repeat. Read the paper (opens in a new tab)