[R-sig-ME] Questions - the stepwise selection issue.

[Ben Bolker]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

On 10-11-07 08:22 PM, Andrew Kosydar wrote:
> Hello All,
> 
> Instead of using step-wise selection, I would suggest instead using
> multimodel inference (Burnham & Anderson).  The technique avoids
> having to choose one "right" model and, in my opinion, is a more
> accurate method than traditional step-wise procedures.

  OK, but (in my opinion) only if: (1) you actually do multi-model
averaging (and don't just pick the best-AIC model to work with), and (2)
you actually want to maximize prediction accuracy rather than testing
statistical hypotheses. I am leery of people using information-theoretic
measures as a shortcut when the language in which they express their
questions strongly suggests that they are really interested in testing
hypotheses rather than making predictions.
One can be led to some marginally sensible statements when trying to
make inferences about the multi-model averaged values of parameters that
average across models containing different sets of correlated
predictors, or containing models with different interaction terms.

  Ben Bolker

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.10 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/

iEYEARECAAYFAkzYAdkACgkQc5UpGjwzenNP5wCggEa5yYIA/Nr4ebfpUkLB3Y9d
mxUAn1MpWojZ39nmvMCJQiV2fUEFA1XY
=P4am
-----END PGP SIGNATURE-----