[Rd] [bug] droplevels() also drop object attributes (comment…)

[Martin Maechler]

>>>>> Suharto Anggono Suharto Anggono via R-devel <r-devel at r-project.org>
>>>>>     on Thu, 8 Jun 2017 16:43:48 +0000 writes:

    > * Be careful with "contrasts" attribute. If the number of levels is reduced, the original contrasts matrix is no longer valid.
    > Example case:
    > x <- factor(c("a", "a", "b", "b", "b"), levels = c("a", "b", "c"))
    > contrasts(x) <- contr.treatment(levels(x), contrasts=FALSE)[, -2, drop=FALSE]
    > droplevels(x)

    > * If function 'factor' is changed, make sure that as.factor(x) and factor(x) is the same for 'x' where is.integer(x) is TRUE. Currently, as.factor(<integer>) is treated specially.

    > * It is possible that names(x) is not attr(x, "names"). For example, 'x' is a "POSIXlt" object.
    > Look at this example, which works in R 3.3.2.
    > x <- as.POSIXlt("2017-01-01", tz="UTC")
    > factor(x, levels=x)


    > By the way, in NEWS, in "CHANGES IN R 3.4.0", in "SIGNIFICANT USER-VISIBLE CHANGES", there is "factor() now uses order() to sort its levels". It is false. Code of function 'factor' in R 3.4.0 (https://svn.r-project.org/R/tags/R-3-4-0/src/library/base/R/factor.R) still uses 'sort.list', not 'order'.

    > --------------------------------
>>>>> Martin Maechler <maechler at stat.math.ethz.ch>
>>>>>     on Tue, 16 May 2017 11:01:23 +0200 writes:

>>>>> Serge Bibauw <sbibauw at gmail.com>
>>>>>     on Mon, 15 May 2017 11:59:32 -0400 writes:

    >>> Hi,

    >>> Just reporting a small bug… not really a big deal, but I
    >>> don’t think that is intended: droplevels() also drops all
    >>> object’s attributes.

    >> Yes.  The help page for droplevels (or the simple
    >> definition of 'droplevels.factor') clearly indicate that
    >> the method for factors is really just a call to factor(x,
    >> exclude = *)

    >> and that _is_ quite an important base function whose
    >> semantic should not be changed lightly. Still, let's
    >> continue :

    >> Looking a bit, I see that the current behavior of factor()
    >> {and hence droplevels} has been unchanged in this respect
    >> for the whole history of R, well, at least for more than
    >> 17 years (R 1.0.1, April 2000).

    >> I'd agree there _is_ a bug, at least in the documentation
    >> which does *not* mention that currently, all attributes
    >> are dropped but "names", "levels" (and "class").

    >> OTOH, factor() would only need a small change to make it
    >> preserve all attributes (but "class" and "levels" which
    >> are set explicitly).

    >> I'm sure this will break some checks in some packages.  Is
    >> it worth it?

    >> e.g., our own R  QC checks currently check (the printing of) the
    >> following (in tests/reg-tests-2.R ):

    >> > ## some tests of factor matrices
    >> > A <- factor(7:12)
    >> > dim(A) <- c(2, 3)
    >> > A
    >> [,1] [,2] [,3]
    >> [1,] 7    9    11  
    >> [2,] 8    10   12  
    >> Levels: 7 8 9 10 11 12
    >> > str(A)
    >> factor [1:2, 1:3] 7 8 9 10 ...
    >> - attr(*, "levels")= chr [1:6] "7" "8" "9" "10" ...
    >> > A[, 1:2]
    >> [,1] [,2]
    >> [1,] 7    9   
    >> [2,] 8    10  
    >> Levels: 7 8 9 10 11 12
    >> > A[, 1:2, drop=TRUE]
    >> [1] 7  8  9  10
    >> Levels: 7 8 9 10
    >> 
    >> with the proposed change to factor(),
    >> the last call would change its result:
    >> 
    >> > A[, 1:2, drop=TRUE]
    >> [,1] [,2]
    >> [1,] 7    9   
    >> [2,] 8    10  
    >> Levels: 7 8 9 10

    >> because 'drop=TRUE' calls factor(..) and that would also
    >> preserve the "dim" attribute.  I would think that the
    >> changed behavior _is_ better, and is also according to
    >> documentation, because the help page for [.factor explains
    >> that 'drop = TRUE' drops levels, but _not_ that it
    >> transforms a factor matrix into a factor (vector).

    >> Martin

    > I'm finally coming back to this.
    > It still seems to make sense to change factor() and hence
    > droplevels() behavior here, and plan to commit this change
    > within a day.

I had committed these (including regression checks and *.Rd
changes) as  svn rev 72771  to R-devel..

As Suharto also warned, this change had more severe effects in
package space than I expected.

Notably the fact that the "dim" attribute was kept there, lead
to several errors in CRAN packages / their checks.

Of course, I could have patched the change by explicitely *not*
keeping certain attributes such as "dim" and "contrasts".

But I've decided that in this case it seems more rational,
notably for back compatibility reasons, to keep factor()'s and hence
droplevel()'s behaviour  and rather document it.
So the "bug" is solved differently, by documenting the behavior
that has been in place "forever".

With svn's 72773, we are back to the previous state with the
addition of mentioning in help(factor) that attributes are "lost".

Martin Maechler
ETH Zurich