Whatever Comes to Mind First

Ask someone whether more English words begin with the letter K or have K as their third letter, and most people say the first. It is not close: there are about three times as many words with K in third position. But you cannot search your memory by third letters. You can search it by first letters, effortlessly, and so the K-initial words arrive in a rush while the others barely arrive at all, and that difference in ease gets read as a difference in number.

This is the availability heuristic, described by Amos Tversky and Daniel Kahneman in the early 1970s, and the first thing to understand about it is that it usually works.

That point gets lost, because the examples that make a good story are always the failures. But ease of recall is, on the whole, a decent proxy for frequency. Common things happen more, so you encounter them more, so they are easier to bring to mind. A shortcut that reads frequency off fluency is getting something genuinely informative for almost no effort, which is what a good heuristic is. The trouble is only that fluency has other causes.

Consider what the news is. An event's newsworthiness is very nearly its rarity: the reason a shark attack is reported and a heart attack is not is that one is unusual and the other is not. Coverage is therefore roughly inversely proportional to frequency, which makes the news a systematically inverted sample of the world. Anyone estimating risk from how easily examples come to mind is estimating from that sample, and will reliably overweight terrorism, kidnapping, plane crashes and shark attacks while underweighting the things that kill almost everyone: heart disease, stroke, the road.

The companion error is worse and less discussed. Suppose a test for a condition is ninety-five per cent accurate, and you test positive. How worried should you be? Most people answer something near ninety-five per cent. But if the condition affects one person in a thousand, then out of ten thousand people tested, ten have it, and of the nine thousand nine hundred and ninety who do not, around five hundred will test positive anyway. Your odds are ten in five hundred and ten: about two per cent. The base rate is doing nearly all the work in that calculation and it feels, subjectively, like it is doing none, because it is not a vivid fact about you. It is a boring fact about a population.

So what do you do about it? The intuitive answer (learn about the bias, and then be careful) has a poor track record. Knowing the Müller-Lyer lines are the same length does not make them look the same length, and knowing about availability does not stop the vivid example arriving first or feeling like evidence when it does. Debiasing by explanation is roughly as effective as telling someone to be taller.

What does work is procedural, and it is much less satisfying. Write your estimate down before you look anything up, so you can see the size of your own error later. Go and find the base rate, since it is the one number intuition cannot supply. Force a comparison against something countable: not "is this common?" but "out of a hundred thousand people a year, how many?" These are not insights. They are habits, and they work because they route around the judgement rather than trying to improve it.

One last subtlety, because it explains a great deal about how surveys and conversations go wrong. Availability is sensitive to the question. Asking "how likely is a recession?" and asking "how likely is a recession caused by a credit crunch in commercial property?" are, logically, a question and a narrower version of it, so the second must be less probable. Reliably, people rate the second higher. The detail makes a story retrievable, and retrievability is what is being read as probability. Which means that adding specifics to a claim can make it feel more likely while actually making it less so, and that anyone who wants to be believed has an obvious incentive to add them.