Functions lda and qda have a new argument CV, which if true performs
leave-one-out cross-validation using fast updating. This had been
requested, but we do not advocate its use in preference to V-fold
cross-validation.
stepAIC has been enhanced. If the fit is one that it recognises as
linear least-squares (lm, aov, glm with family=Gaussian(link=identity)),
it uses the Wald/score test equivalences to compute the change in
log-likelihood. If the scale is specified in these cases, the constant
in AIC is chosen so that it is exactly equal to Mallows' Cp. Thus
stepAIC now provides a complete replacement for step.lm, using
(Akaike's) AIC or (Mallows') Cp as appropriate. With iterative glm
fits there is now an option to start the fitting process at the linear
predictor of the current model. This usually (but not always) speeds up
the process.
We have included a replacement for predict.glm that fixes an error of
several years' standing. The standard errors given by predict.glm use
the estimated scale even when the scale is known, either explicitly or
implicitly as in a binomial or Poisson fit. This also applies to
predict.lm: our predict.glm used on lm objects will allow the scale to
be specified.
Brian Ripley
Bill Venables
-----------------------------------------------------------------------
This message was distributed by s-news@wubios.wustl.edu. To unsubscribe
send e-mail to s-news-request@wubios.wustl.edu with the BODY of the
message: unsubscribe s-news