Tversky and Kahneman on heuristics
📚SOURCE
💎PERMANENT NOTES
🧵SYNTHESIS
Paraphrase, not quotation: the whole paper, 🔬 Judgment under Uncertainty - Heuristics and Biases.
The paper says that people judge probability with a few mental shortcuts, which work well enough most of the time and fail in predictable directions. It names three.
Representativeness: we judge how likely something is by how much it resembles our picture of the category. This makes us ignore how common the category is, treat small samples as if they were large, and expect chance to look tidy.
Availability: we judge frequency by how easily examples come to mind, so the recent, the vivid and the personal count for too much.
Anchoring: we start from whatever number is present, relevant or not, and adjust too little.
The sting is at the end. Trained people show the same biases when they reason intuitively, and nobody notices, because no one keeps score of their own probability judgements.
✨KEY IDEAS
Told that a group is 70 percent lawyers or 70 percent engineers, people use that number correctly, until they are given a personality sketch. Even a worthless sketch makes them drop it.
A small hospital will have far more days with over 60 percent boys than a large one, and most people say the two are about equal.
People estimated the share of African countries in the United Nations higher or lower depending on a number from a wheel they had just watched spin.
A long chain of likely steps is judged more probable than it is, and a risk made of many small independent chances is judged less probable than it is.
Statistically trained researchers made the same errors on intuitive problems.
📝READING NOTES
Why I am reading a 1974 paper on a Sunday
It is the set reading for the probability lectures, and the response is due on the 25th: 📌 Statistics I - Reading response - Heuristics and biases. It is also the week after I recommended the slip, and I read it with rather more attention than Hale intended.
The three heuristics, with what each one costs
Representativeness
- Ignores prior probability (the lawyers and engineers).
- Ignores sample size (the two hospitals).
- Misreads chance: after a run of reds, black feels due.
- Ignores predictability: a glowing description produces a glowing forecast, however unreliable the description.
- Misses regression: they use the flight instructors here, years before the book.
Availability
- What is easy to retrieve seems frequent.
- What is easy to imagine seems likely.
- Things that go together in memory seem to go together in the world.
Anchoring
- Adjustment from a starting number stops too soon.
- The same people who multiplied from the large end of a sequence guessed far higher than those who started from the small end.
- Conjunctions are overrated, disjunctions underrated.
The one that describes February
The plan for 18 February needed the load test to pass and the migration to run and support to be trained and the release notes signed off and several other things. I would have given each of them nine chances in ten or better. I never multiplied. Ten steps at nine in ten is about one in three. The paper says people anchor on the probability of a single step and do not adjust enough for the chain. That is a description of my slide.
Where I would push back
- Every experiment is a puzzle designed to trip the heuristic. The paper says the shortcuts are usually useful and then shows only the failures. A reader could leave thinking intuition is worthless, which the authors do not claim.
- The subjects are students answering questions with nothing at stake. I would like to know how much the effects shrink when it is someone's own money or their own shipment.
For work
How often does an exception of a given type resolve itself if nobody touches it? That is a base rate, we have the data, and nobody has ever looked. Until we do, every triage rule is a story.
What it produced
💎 A number without its base rate is a story, this afternoon.
Added 24 February: response handed in a day early. 600 words, and the hospital problem is in it.
Reading record
- Read for: the course
🗣️IN SHORT
- In : Judgment under Uncertainty: Heuristics and Biases (1974), the whole paper
- Tversky and Kahneman argue that : people estimate probabilities with a few simple heuristics that are economical and usually adequate, and that produce systematic, predictable errors
- by stating that : representativeness leads us to neglect base rates and sample size, availability ties frequency to ease of recall, and anchoring biases estimates toward arbitrary starting points, in experts as well as laypeople