"Estimate a range for the population of Nevada, with 80% confidence."
The book's definition of 'better' here is that, if asked a large number of questions:
- Right about 80% of the time: good calibration
- Right much more than 80% of the time: your upper/lower bounds are too far apart, i.e. you're being too conservative. e.g. I could estimate the population as being between 0 and 7bn, and I'd be right, but not usefully right.
- Right much less than 80% of the time: too confident, with too narrow ranges.
Anyway, if there were a free app to train your calibration, would I be the only one who would use it?
I'm imagining something which presents you with a question, after which you enter an upper and lower bound. You get asked a few questions and then shown the answers. After you've answered more than, say, 20 questions, cumulatively (not per session) the app will start to show an indicator about whether you are over-confident, under-confident, or just right. (The confidence intervals for each question need not be the same.)
[0] http://www.amazon.com/dp/0470539399