Which kills more: Jev vs 1978's people

How many Americans a year die of botulism, tornadoes, diabetes or stroke? Does Jev show the famous 1978 pattern of overestimating rare, dramatic deaths and underestimating common, quiet ones?

result77 questions

In 1978, people asked how many Americans die of each cause squashed their answers toward the middle: rare, dramatic deaths overestimated, common quiet ones underestimated. Jev squashes far less (slope 0.65 on a log scale, where 1 is unbiased; people 0.45), orders the causes just as well (rank correlation 0.94), and barely overrates dramatic causes (1.2 times, against 3.3 for people). Its biggest misses: measles (1,732 a year vs 5) and fireworks (547 vs 6).

Jevpeople

How to read this: The classic 1978 chart, redrawn: true yearly deaths across, estimates up, both on log scales. Ink diamonds are the 1978 public, magenta dots are Jev; the diagonal is a perfect estimate. A cloud flatter than the diagonal means rare causes are overestimated and common ones underestimated.

37 causes with the study's reference; Jev's median bin is right for 46%. Without the study's reference number, Jev's median estimate does not move (people in the study were always given one). Causes with zero 1970s deaths are left out of the log-scale fits.

In short

Asked yearly US deaths from 37 causes in the mid-1970s, Jev ordered them as well as the 1978 public did (rank correlation 0.94 for both).
Unlike people, Jev barely inflates dramatic deaths like tornadoes and homicide (1.2 times, against 3.3) and squashes its estimates less (slope 0.65 vs 0.45).
Jev may simply know this famous study; its own big misses, such as measles at 1,732 deaths a year against 5, look like outdated history.

What the data shows

Grim Reaper Knocking Door meme: measles in the 1970s

Same order: Jev and the 1978 public both order the 37 causes very well against the truth (0.94 each).
Much less squash: Jev's line has a slope of 0.65; the public's was 0.45. Jev's median bin is the right one for 46% of causes.
Almost no drama bias: the 1978 public overestimated dramatic causes (homicide, tornadoes, floods) 3.3 times as much as quiet ones (diabetes, stroke); for Jev the factor is 1.2.
Its misses are its own: it wildly overestimates measles (1,732 a year vs 5) and fireworks (547 vs 6), and underestimates all disease (547,722 vs 1,740,450) and stomach cancer (54,772 vs 95,530).
The reference doesn't matter to it: without the 50,000 motor-vehicle anchor, its median estimate doesn't move (the 1978 participants always had one).

What it means, and what it doesn't

Jev doesn't show the famous availability bias: it doesn't inflate headline deaths over quiet ones the way people did. Its errors look more like misremembered history (measles was a big killer before the vaccine) than like fear.

It doesn't prove Jev reasons from statistics: it may simply know this classic study. A test on causes and years nobody has published estimates for would separate the two.

Caveats

People's side is an average. The study published one geometric-mean estimate per cause, not each person's answer, so the human dots are averages and can't show how spread out people were.
Jev knows later statistics. The questions ask about the mid-1970s and are scored against the 1970s counts, but Jev has read decades of later statistics and the 1978 paper itself, one of the most cited in the psychology of risk. Knowing the famous result could help it avoid the famous bias.
Numbers are a known weak spot. TypeSafe lists raw numeric values as a known weakness of Jev. It was given ordered answer bins (1 to 9, 10 to 29, ... 1 million or more) instead of asking for a number, and each bin spans about a factor of three.
Which causes count. Four of the 41 causes had no deaths at all in the 1970s counts and can't go on a log scale, so the fits use 37. Which causes count as "dramatic" follows Pachur's 2024 compilation, not a judgment made for this project.

Jev on this experiment

Would a person find it interesting to read?: Yes78%
Does it describe you?: Yes61%
Would you have predicted it?: No51%
How fair is the comparison?: The comparison is reasonable
How much should a reader rely on it?: Moderately
Which caveat matters most?: Jev knows later statistics60%

Why ask this

In 1978 Lichtenstein and colleagues asked Americans how many people die each year of 41 causes, from botulism and tornadoes to diabetes and stroke. The result became the textbook picture of the availability bias: people's estimates were squashed toward the middle. Rare, vivid causes that make the news were overestimated; common, quiet killers were underestimated.

A model trained on news-heavy text might inherit that squash, or it might have read the statistics instead. The answer says something about where its sense of risk comes from.

How this was done

The people and the data

The 1978 study's 41 causes, each with the yearly US death count from the vital statistics of the time and the participants' average estimate (a geometric mean, so a few wild guesses don't dominate), as compiled by Pachur in 2024 (open data on OSF). The participants were given one reference point, as Jev was: about 50,000 people a year died in motor-vehicle accidents. Four causes had no deaths at all in the 1970s counts and can't go on a log scale, so the comparison uses 37.

What Jev was asked

One question per cause, with the study's reference, in ordered bins:

For reference, about 50,000 people a year died in motor vehicle accidents. In the United States in the mid-1970s, about how many people died each year from measles?
None · 1 to 9 · 10 to 29 · 30 to 99 · 100 to 299 · 300 to 999 · 1,000 to 2,999 · 3,000 to 9,999 · 10,000 to 29,999 · 30,000 to 99,999 · 100,000 to 299,999 · 300,000 to 999,999 · 1 million or more

Jev's most likely answers were 3,000 to 9,999 and 1,000 to 2,999; the real count was in single digits. Each cause was also asked without the reference. That's 81 new questions, each asked with the bins in three shuffled orders and averaged.

How it was measured

Jev's estimate is the middle of its median bin (on a log scale). On log scales, the analysis compares how well estimates order the causes, how steep the estimate-versus-truth line is (1 means no squash, lower means rare causes are pushed up and common ones down), and how much more dramatic causes are overestimated than quiet ones, for Jev and for the 1978 public.

Where these questions live

77 questions across 1 topic of the map. Each opens on the map with every question in it.

Health & Body

Causes of Death77

Every question

All 77 questions behind this result, the telling ones first: the examples the analysis points to, then the ones where Jev misses, biggest gap first.

Jev’s own answer

Showing 0 of 77