In article <[EMAIL PROTECTED]>, >> Why should I care about data more discrepant than what I've >> observed? I haven't seen them. Why should they affect the way I >> judge what I did observe? :-)
Rich Ulrich <[EMAIL PROTECTED]> wrote: >The above sounds has a point, but seems more than >a little naive. The big problem with statistical education is that, when it doesn't fail to teach anything at all, it ends up convincing people that certain pieces of nonsense are actually sophisticated. Being more than a little naive is a big plus when discarding this junk. >Likelihood, the height of the curve, is not "probability" >in the way we do define it. > >A.W.F. Edwards wrote a book about "Likelihood" (1972) >which is still a classic. Maybe I should look at my copy again? Yes, you should definitely look at it again. The "likelihood principle" is exactly the idea that all that matters is the likelihood function, which is based on the data actually observed. Any statistical method that looks at the probability of getting data that you didn't actually observe violates the likelihood principle. Since there are various good arguments in favour of the likelihood principle, any method (such as traditional hypothesis testing) that is incompatible with it is rather suspect. Radford Neal . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
