[R] predict.lda problem with posterior probabilities
Marc R. Feldesman
feldesmanm at pdx.edu
Fri Jan 23 14:49:52 CET 2004
At 04:23 AM 1/23/2004, Mike White wrote:
>With predict.lda the posterior probabilities only relate to the existing
>Class definitions. This is fine for Class definitions like gender but it is
>a problem when new data does not necessarily belong to an existing Class.
>Is there a classification method that gives posterior probabilities for
>Class membership and does not assume the new data must belong to one of the
>existing Classes? A new class would then be indicated by low posterior
>probabilities for all the existing Classes.
For discriminant analysis, one of of doing this would be to use "typicality
probabilities", sometimes called "inverse probabilities". This is
discussed in Huberty's Applied Discriminant Analysis, Wiley 1994 (and other
sources), and is used relatively commonly in the paleoanthropological
literature where researchers try to "classify" a fossil into a matrix of
modern forms to which it is already known that the fossil does not
belong. The question is whether the fossil has a modern analog or whether
it really is dissimilar from any modern form.
More information about the R-help
mailing list