I recently read the excellent Judgment under uncertainty edited by Kahneman, Tversky and Slovic. I was particularly struck by a simple observation that's mentioned in a couple of the chapters.
Suppose you're a parent or manager or teacher. When one of your charges does something particularly good, you're likely to respond positively (praise, reward, promotion, ...). When they do something particularly bad, you're likely to respond negatively.
Now, suppose that your positive and negative responses are equally effective (or ineffective) in producing improved performance. Because of regression to the mean, they'll tend to do worse after a really good performance and better after a really bad one. Therefore, they'll tend to do worse after you respond positively and better after you respond negatively.
Therefore, we are all repeatedly exposed to misleading evidence that tends to make us think that negative responses are more effective than they are, and positive responses less effective.
(The Wikipedia article on regression to the mean quotes Kahneman making the same point. He calls the moment when he noticed it "the most satisfying Eureka experience of my career".)
There's a lot of other really good stuff in it, on the general theme of cognitive heuristics and biases. Opening it at random five times and looking for something interesting within a few pages I find:
If you ask people to estimate a bound for some quantity with 99.9% confidence, they'll be wrong something like 10% of the time.
If you tell people about something like the Milgram experiment and ask them what they think they'd do and why they predict that, it makes essentially no difference to either part of their answer whether or not you tell them that so-and-so-many-percent of people behaved in such-and-such a way.
Estimated answers to a question that amounts to "what is 10 choose r?" for various values of r are monotone decreasing in r. (For combinatorically naïve subjects.)
Two very plausible-sounding ways of estimating the amount of theft carried out by drug addicts in New York differ by about a factor of 10.
Describe a tragic traffic accident and ask people to complete the sentence "If only ..." as it might be uttered by people concerned; they absolutely consistently don't mention the single most likely change that would have made it not happen, apparently because it involves changing a continuous rather than a discrete variable.
Estimated answers to a question that amounts to "what is 10 choose r?" for various values of r are monotone decreasing in r.
For 5 <= r <= 10, perhaps?
Ho ho. No, the estimates are decreasing across the full range of values of r. And, since I see you tried both ways, I think the Markdown thing leaves bare less-than signs as they are, which is suboptimal but usually harmless, so thanks for escaping them :-).
Post a comment:
This one very beautiful observation has sold me on the book.