[Rd] stringsAsFactors

[Duncan Murdoch]

On 11/02/2013 12:13 PM, William Dunlap wrote:
> Note that changing this does not just mean getting rid of "silly warnings".
> Currently, predict.lm() can give wrong answers when stringsAsFactors is FALSE.
>
>    > d <- data.frame(x=1:10, f=rep(c("A","B","C"), c(4,3,3)), y=c(1:4, 15:17, 28.1,28.8,30.1))
>    > fit_ab <- lm(y ~ x + f, data = d, subset = f!="B")
>    Warning message:
>    In model.matrix.default(mt, mf, contrasts) :
>      variable 'f' converted to a factor
>    > predict(fit_ab, newdata=d)
>     1  2  3  4  5  6  7  8  9 10
>     1  2  3  4 25 26 27  8  9 10
>    Warning messages:
>    1: In model.matrix.default(Terms, m, contrasts.arg = object$contrasts) :
>      variable 'f' converted to a factor
>    2: In predict.lm(fit_ab, newdata = d) :
>      prediction from a rank-deficient fit may be misleading
>
> fit_ab is not rank-deficient and the predict should report
>     1 2 3 4 NA NA NA 28 29 30

In R-devel, the two warnings about factor conversions are no longer 
given, but the predictions are the same and the warning about rank 
deficiency still shows up.  If f is set to be a factor, an error is 
generated:

Error in model.frame.default(Terms, newdata, na.action = na.action, xlev 
= object$xlevels) :
   factor f has new levels B

I think both the warning and error are somewhat reasonable responses.  
The fit is rank deficient relative to the model that includes f == "B",  
because the column of the design matrix corresponding to f level B would 
be completely zero.  In this particular model, we could still do 
predictions for the other levels, but it also seems reasonable to quit, 
given that clearly something has gone wrong.

I do think that it's unfortunate that we don't get the same result in 
both cases, and I'd like to have gotten the predictions you suggested, 
but I don't think that's going to happen.  The reason for the difference 
is that the subsetting is done before the conversion to a factor, but I 
think that is unavoidable without really big changes.

Duncan Murdoch


>
> Bill Dunlap
> Spotfire, TIBCO Software
> wdunlap tibco.com
>
> > -----Original Message-----
> > From: r-devel-bounces at r-project.org [mailto:r-devel-bounces at r-project.org] On Behalf
> > Of Terry Therneau
> > Sent: Monday, February 11, 2013 5:50 AM
> > To: r-devel at r-project.org; Duncan Murdoch
> > Subject: Re: [Rd] stringsAsFactors
> >
> > I think your idea to remove the warnings is excellent, and a good compromise.
> > Characters
> > already work fine in modeling functions except for the silly warning.
> >
> > It is interesting how often the defaults for a program reflect the data sets in use at the
> > time the defaults were chosen.  There are some such in my own survival package whose
> > proper value is no longer as "obvious" as it was when I chose them.  Factors are very
> > handy for variables which have only a few levels and will be used in modeling.  Every
> > character variable of every dataset in "Statistical Models in S", which introduced
> > factors, is of this type so auto-transformation made a lot of sense.  The "solder" data
> > set there is one for which Helmert contrasts are proper so guess what the default
> > contrast
> > option was?  (I think there are only a few data sets in the world for which Helmert makes
> > sense, however, and R eventually changed the default.)
> >
> > For character variables that should not be factors such as a street adress
> > stringsAsFactors can be a real PITA, and I expect that people's preference for the option
> > depends almost entirely on how often these arise in their own work.  As long as there is
> > an option that can be overridden I'm okay.  Yes, I'd prefer FALSE as the default, partly
> > because the current value is a tripwire in the hallway that eventually catches every new
> > user.
> >
> > Terry Therneau
> >
> > On 02/11/2013 05:00 AM, r-devel-request at r-project.org wrote:
> > > Both of these were discussed by R Core.  I think it's unlikely the
> > > default for stringsAsFactors will be changed (some R Core members like
> > > the current behaviour), but it's fairly likely the show.signif.stars
> > > default will change.  (That's if someone gets around to it:  I
> > > personally don't care about that one.  P-values are commonly used
> > > statistics, and the stars are just a simple graphical display of them.
> > > I find some p-values to be useful, and the display to be harmless.)
> > >
> > > I think it's really unlikely the more extreme changes (i.e. dropping
> > > show.signif.stars completely, or dropping p-values) will happen.
> > >
> > > Regarding stringsAsFactors:  I'm not going to defend keeping it as is,
> > > I'll let the people who like it defend it.  What I will likely do is
> > > make a few changes so that character vectors are automatically changed
> > > to factors in modelling functions, so that operating with
> > > stringsAsFactors=FALSE doesn't trigger silly warnings.
> >
> > ______________________________________________
> > R-devel at r-project.org mailing list
> > https://stat.ethz.ch/mailman/listinfo/r-devel