Peter Flom <[EMAIL PROTECTED]> wrote: > > One sort of crossvalidation I have done (to demonstrate to others that > stepwise is, in fact, a bad idea) is to divide the data set into 2 equal > parts on somethig like id number, then run the same stepwise twice. > THis is, of course, not full cross validation, but it's easy to do in > any software, and easy to explain.
There's a paper by Steyerberg and Harrell demostrating via simulation that the split-half method is probably too conservative (this is for conventional regression with a fixed set of predictors), and that bootstrap is probably the better route. I would imagine that the same might apply to stepwise. Interestingly, in another set of simulations, Steyerberg (I believe it's the paper with the title something like, "In search of a sensible strategy for small sample sizes") showed that even models with way too many predictors per observation via forced entry were less biased than those derived by stepwise selection. Mike Babyak . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
