Mental Health

Why Our Brains Misread Percentages: The Psychology of Health Statistics

Waqar Ahmad Waqar Ahmad November 20, 2026
Why Our Brains Misread Percentages: The Psychology of Health Statistics

A headline says the risk of a rare side effect has risen by 50%. A second headline, about the same study, says cases went from 4 in 100,000 people to 6 in 100,000 people. Both headlines are accurate. They do not feel like the same news, and the distance between them is where most misunderstanding of health statistics happens.

The arithmetic takes seconds. Six minus four is two. Two divided by four is 0.5, so the relative rise really is 50%. In absolute terms, the risk grew by two people in a hundred thousand, from 0.004% to 0.006%. A city of a million people would expect roughly 40 cases before and roughly 60 after, which is 20 extra cases. Neither headline lied. The first headline leaves out the starting point, making it harder for readers to judge the actual size of the risk. Researchers who study this call it statistical illiteracy, and they find it in patients, journalists and doctors alike (Gigerenzer et al., 2007).

Relative and Absolute Risk Answer Different Questions

Relative risk describes the proportional change in risk compared with the starting level. Absolute risk describes the actual difference in the probability of an event. An absolute risk change (absolute risk reduction or risk difference) tells you how much the baseline event probability changed for an actual group of people. Trouble starts when a report answers the first question and the reader hears the second one.

Take a treatment that cuts the ten-year risk of a heart attack from 2 in 100 people to 1 in 100. Described in relative terms, that is a 50% reduction, and it sounds like a drug everyone should take. Described in absolute terms at the group level over ten years, the absolute risk reduction is one percentage point, meaning one fewer heart attack per 100 people treated over ten years, on average. without the drug, 98 out of 100 would not have had a heart attack anyway, and with the drug, 1 out of 100 still has one within that ten-year window. Clinicians have a term for the absolute view, the number needed to treat. Here it is 100, because the risk fell by 1 in 100, so 100 people take the drug for one event to be prevented on average.

Both descriptions belong in any honest report. In practice, benefits tend to be advertised in relative terms, where they look largest, while harms are often given in absolute terms, where they look smallest. Gigerenzer and his co-authors call this mismatched framing, and it works because few readers stop to convert one form into the other.

Percentage Points are Not Percent

When a rate moves from 4% to 5%, it has risen by one percentage point. It has also risen by 25%, because one divided by four is 0.25. A report that says a hospital infection rate "rose by 1%" may confuse percentage points with percent. If the rate increased from 4% to 5%, the correct change is one percentage point, or a 25% relative increase, and most readers will walk away thinking the change was small. A report that says infections "jumped 25%" about the same wards produces the opposite reaction. The wards did not change. The unit did.

The Base Rate is the Part Nobody Prints

Here is a puzzle that has been put to doctors and medical students for decades. A condition affects 1 person in 100. A test for it catches 90% of the people who have the condition, and it wrongly flags 9% of the people who do not. You test positive. What is the chance you actually have the condition?

Most people guess high, often somewhere near 90%. Now run the same problem as counts instead of percentages. Take 1,000 people. Ten of them have the condition, and the test catches nine of those ten. The other 990 people are healthy, and in natural frequency terms the test wrongly flags 89 of them. Add the two groups and the test hands out roughly 98 positive results, of which nine are genuine. A positive result therefore means about a 9% chance of having the condition, roughly one in eleven.

The ingredient most readers mentally drop in the first version is the base rate—the 1 in 100 starting figure—and ignoring it has a name: base-rate neglect. Gigerenzer and Hoffrage (1995) showed that the neglect is partly a fault of format rather than of intelligence. When problems like this one were stated in natural frequencies, as counts of people, far more participants reasoned their way to the right answer. In some of their studies close to half succeeded, where only a small minority had managed the same problem stated in probabilities.

The reader matters too. Schwartz and colleagues (1997) gave women statements about screening mammography and asked them to judge how much it reduced the risk of dying of breast cancer. Women who scored zero on a short three-question numeracy test judged the benefit accurately about 6% of the time. Women who answered all three questions correctly managed it 40% of the time. Both groups faced the same set of risk statements. Their facility with basic arithmetic was not the same.

The Wording Chooses for You

Percentages are one way to steer a reader. Plain words do it too. In 1981, Tversky and Kahneman described a hypothetical epidemic expected to kill 600 people and asked students to choose between two response programmes. The first group chose between saving 200 people for certain, or a gamble with a one-in-three chance of saving all 600 and a two-in-three chance of saving nobody. Seventy-two percent picked the certain option. The second group was offered the same outcomes described as deaths: 400 people dying for certain, or a gamble with a one-in-three chance that nobody would die and a two-in-three chance that 600 would die. This time only 22% picked the certain option. The programmes were identical, person for person. Only the verbs had changed, from "saved" to "die", and the majority preference flipped with them.

Medicine runs on this effect daily. A treatment described by a 90% survival rate sounds better than the same treatment described by a 10% mortality rate, even though the two figures add up to the same hundred patients. It is one reason Gigerenzer and colleagues recommend that health statistics be reported as mortality rates and as counts of people affected, formats that give the framing less room to work in.

Four Questions that Catch Most of it

None of the errors in this article needed advanced mathematics to avoid. Every sum above is a subtraction, a division, or a count. What they needed was the habit of asking for the numbers a percentage was built from. Four questions do most of the work:

  1. Out of how many? A percentage with no denominator tells you almost nothing.
  2. Compared with what? Ask for the risk before the change, not only the size of the change.
  3. Over how long? A risk across ten years and a risk per year are different numbers wearing the same clothes.
  4. Can I see it as people? Convert the claim into counts out of 1,000 or 100,000 and read it again.

Percentages compress information, and compression always throws something away. In health reporting, what gets thrown away is usually the size of the risk itself. Two extra cases in a hundred thousand is worth knowing about. It is also worth knowing that it is two.

References

  • Gigerenzer, G., & Hoffrage, U. (1995). How to improve Bayesian reasoning without instruction: Frequency formats. Psychological Review, 102(4), 684-704.
  • Gigerenzer, G., Gaissmaier, W., Kurz-Milcke, E., Schwartz, L. M., & Woloshin, S. (2007). Helping doctors and patients make sense of health statistics. Psychological Science in the Public Interest, 8(2), 53-96.
  • Schwartz, L. M., Woloshin, S., Black, W. C., & Welch, H. G. (1997). The role of numeracy in understanding the benefit of screening mammography. Annals of Internal Medicine, 127(11), 966-972.
  • Tversky, A., & Kahneman, D. (1981). The framing of decisions and the psychology of choice. Science, 211(4481), 453-458.
Waqar Ahmad
Waqar Ahmad Founder at CalcMega