Base rate neglect: why a vivid detail beats a background number

The number that was true before you looked

Every judgement about a particular case sits on top of a background fact: how common that kind of case is to begin with. Statisticians call it the base rate, and it is the prior probability of the thing you are trying to identify, before any specific evidence about this instance arrives. If one house in a thousand on a given street has been burgled this year, that thousand is doing real work in any estimate about your house.

The reliable finding, reproduced many times and in many formats, is that people discard this number almost the moment a description arrives. Give someone a short character sketch that sounds like a librarian, and they will judge the person likely to be a librarian, largely regardless of how many librarians there are compared with, say, salespeople. The sketch feels like evidence. The population feels like trivia.

Why the story wins

The usual explanation is that we judge by resemblance. A case that matches our mental image of a category gets assigned to that category, and the strength of the match substitutes for the probability. This is efficient in a world where categories are roughly equal in size, and badly wrong in a world where they are not.

It is worth being careful about how much weight that explanation carries. The phenomenon itself is robust and has been demonstrated in many different presentations, but the size of the failure moves a great deal depending on how the problem is worded, whether the base rate is described as causally relevant, and whether the person has any stake in getting it right. Some framings nearly eliminate it. That variability is not a footnote — it tells you the effect is about how information is presented rather than a fixed defect in the machinery.

Where it does the most damage

The damage concentrates wherever the thing being detected is rare. Screening for an uncommon condition, flagging fraudulent transactions, identifying the small number of job applicants who will turn out excellent, spotting a rare failure in a manufacturing line — in all of these, the population of true cases is tiny and the population of false alarms is enormous, because the false-alarm rate is applied to a much larger group.

The consequence is uncomfortable and mathematically unavoidable. A test that is right most of the time, applied to a condition that is rare, will still produce more false positives than true ones. Nothing is wrong with the test. The arithmetic simply reflects that a small percentage of a very large number can exceed a large percentage of a very small one.

This is why sensible institutions do not treat a single flag as a finding. They treat it as a reason to look again with a second, independent method, which changes the arithmetic completely.

Presentation changes how badly people do

One of the more durable results in this area is that the format of the question matters enormously. Ask people to reason with percentages and conditional probabilities and most of them, including many trained professionals, get it wrong. Restate the identical problem in natural frequencies — out of a thousand people, this many have the condition, of those this many test positive, of the rest this many also test positive — and performance improves substantially.

Nothing has changed about the mathematics. What changed is that the frequency version keeps the base rate visible as a count of actual cases rather than burying it in a ratio. That finding has held up well across replications, and it is one of the few debiasing results with a genuinely practical implication: if you want people to reason correctly, hand them counts rather than rates.

The question that does most of the work

Out of how many? That is the whole habit, and it is unglamorous enough that people skip it. When someone reports that most of the people who did a certain thing went on to succeed, the question is how many people did it in total, and how many of them you never heard about. When a description sounds compelling, the question is how many people in the relevant population would fit that description equally well.

It helps to say the reference class out loud, because a base rate borrowed from the wrong population is worse than no base rate at all. The failure rate of restaurants in general is not the base rate for a specific restaurant with an experienced operator in a location with proven demand. Choosing the reference class is an act of judgement, and pretending otherwise gives the number a false authority.

What base rates cannot do

A base rate is not destiny and it does not settle an individual case. It sets the starting point that specific evidence then has to move, and strong specific evidence moves it a long way. The mistake runs in both directions: ignoring the base rate entirely, or treating it as though no evidence about the particular case could ever overcome it.

The honest version is that good judgement is arithmetic between the two. Start with how common the thing is, then ask how much more likely your evidence would be if the thing were true than if it were not. When those two questions are both asked, most of the classic errors do not appear. They appear when the second question is asked alone.