Methods package is now attached by default

Prof Brian D Ripley writes:

On Fri, 17 Jan 2003, John Chambers wrote:
There are two main known differences from having methods attached:

- the definition of class() changes, in particular no object ever has a
NULL class.  If you have code that depends on test such as
`if(is.null(class(x)))...', there may be problems.

Usually code with those sort of tests is doing a workaround of the fact
that not all objects had a class before.  The best solution is usually
to ask what the code really wants to do.  If you do have to retain
exactly the old behavior, one solution is to copy the version of class
and class<- from the base package and use those (as baseClass and
baseClass<-, e.g.) instead of class and class<-).

Here is one example, which makes the MASS scripts fail.

library(MASS)
corresp
function (x, ...)
{
    if (is.null(class(x)))
        class(x) <- data.class(x)
    UseMethod("corresp", x, ...)
}

That used to work with matrices, and dispatch to corresp.matrix.  Now
a matrix has reported class "matrix" but dispatch occurs to
corresp.default.  The temporary fix is to remove the is.null
condition.

That's quite a common construction, and I think I should expect
UseMethod to dispatch on the class class() reports.  So it looks to me
as if UseMethod needs to be altered to do so.

same happens e.g. with class "numeric":

ipredbagg <- function(y, ...) UseMethod("ipredbagg")

does NOT dispatch to ipredbagg.numeric if y IS of class "numeric". With
the "old" class(), this one worked:

ipredbagg.default <- function(y, ...) {
  # "numeric" is not an S3 class: check for regression problems and
  # the method dispatch should work for class(y) == "numeric"
  if (is.numeric(y)) {
    class(y) <- "numeric"
    return(ipredbagg(y, ...))
  } else {
    stop(paste("Do not know how to handle objects of class", class(y)))
  }
}

but now I need to call

return(ipredbagg.numeric(y, ...))

I think Brian is right with altering UseMethod() in a way taking care of
what class() tells us.
...
The daily check process on all CRAN packages shows that the packages

	StatDataML geoR ipred xtable

now fail running their examples with methods loaded by default.  The
three latter can be traced immediately to variations on the above; the
first fails in

    if (is.factor(x)) {
      attr(xtmp, "levels") <- NULL
      class(xtmp) <- class(xtmp)[!class(xtmp) %in% c("ordered","factor")]
    }

with the error message

Error in "class<-"(*tmp*,
                   value = class(xtmp)[!class(xtmp) %in% c("ordered",  :
        Invalid replacement object to be a class string

I think I can tell David how to fix this, but wasn't quite sure about
the wording in the error message ...

I assume I should wait before contacting the authors of the other 3
packages for changing their code---or do we already have an official
recommendation?
I can fix the code for ipred. The only problem is making it work for both
R-1.6.2  AND  R-devel :-)

Torsten
-k

______________________________________________
R-devel@stat.math.ethz.ch mailing list
http://www.stat.math.ethz.ch/mailman/listinfo/r-devel

Methods package is now attached by default

Thread (10 messages)