R devel - Jan 2013 - na.omit option in prcomp: formula interface only

Dear r-devel list, dear Ben

I came across a post of Ben Bolker from Feb 2012 (see below) on handling NA
values in prcomp(). As I faced the same issue and found Ben's suggestions
interesting, I was wondering whether this led to further discussions I
might have missed? I understand handling NA values is far from trivial, but
would it be possible to add a warning in the documentation, and/or whenever
na.action is used with  prcomp() on a data frame (suggesting to use the
formula instead?)?

Thanks!

Matthieu Stigler


This is a wishlist/request for discussion about the behaviour of the
na.action option in prcomp, specifically the fact that it only applies to
the formula interface.

   I had a question from a friend (who is smart and careful and generally
R's TFM, although like all of us he misses things sometimes) asking why the
na.action= argument didn't seem to be doing anything in prcomp (i.e. one
gets an "Error in svd(x, nu=0): infinite or missing values in
'x'"). Some
poking later, I realized that na.action only applied to the formula
interface (so I told him to try prcomp(~.,data=x,...) instead).
Sufficiently careful reading of the help page, with hindsight, revealed
that na.action only appears in the arguments for the formula method, not
the default (on the other hand,

'scale.' only appears in the default formula, but it *does* work with
prcomp.formula as well, because prcomp.formula passes ... through to
prcomp.default ...)


   Would it be reasonable to (at least) add a sentence to the documentation
saying that na.action applies only to the formula interface or (possibly)
to add some NA-processing machinery to prcomp.default to allow it to handle
na.action as well? (I can appreciate from looking at stats:::prcomp.formula
that the NA-processing is not completely trivial ...)

  Ben Bolker

	[[alternative HTML version deleted]]