In terms of a general solution to the problem, I am most interested that
predict()
give correct answers, even if the user re-arranges the order of levels of a
factor
predictor. Otherwise, some serious nonsense could find its way into
publications.
So I hope that MathSoft will seriously consider having predict look
at the factor levels that were used when the fit object was stored, and then
coerce predictors to this coding. If they won't coerce, a warning should be
given
that missing predictions result. This can be done through model.matrix if an
additional argument could be implemented that essentially provides information
from a previous model.matrix construction.
Note that all predictions using the Design library are 'safe' as these checks
are
always made and levels are re-ordered if needed. A couple of months ago
Brian Ripley posted some general code related to this, on s-news.
-Frank Harrell
-----------------------------------------------------------------------
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
|