[Rd] stats::poly() stopped working for Date input -- intentional?

[Martin Maechler]

>>>>> Martin Maechler 
>>>>>     on Fri, 8 Jul 2022 16:34:43 +0200 writes:

>>>>> Michael Chirico via R-devel 
>>>>>     on Thu, 7 Jul 2022 22:17:12 -0700 writes:

    >> SVN#80126 added rep.difftime, which means
    >> rep(as.difftime(0, units="secs")) retains the "datetime"
    >> class.

    > (yes, by me, March 2021), this was fixing PR#18066 ==>
    > https://bugs.r-project.org/show_bug.cgi?id=18066 )

    > Thank you, Michael, for the report!

    >> A consequence of this is that stats::poly() no longer
    >> accepts Date/POSIXct input (because poly calls outer() on
    >> the de-meaned input, which rep()):

    >> # works on R 3.6.3 (and probably everything < 4.1.0) #
    >> but on R 4.1.3 (and probably everything >= 4.1.0):
    >> stats::poly(Sys.Date() - 0:4, 3) # Error in
    >> Ops.difftime(X, Y, ...) : '^' not defined for "difftime"
    >> objects

    >> Is this intentional?

    > Well, actually I think it was not intentional that poly()
    > worked at all with Date/POSIXct input, ..  OTOH you *did*
    > encounter it.

    > Note that

    >> poly(as.Date("2020-2-2") - 0:3, 2, raw = TRUE)
    >   Error in Ops.Date(X, Y, ...) : ^ not defined for "Date"
    > objects
    >> 

    > happens (I think) in all versions of R, i.e., even before
    > the rep() extension.

    >> If not, a simple patch is to call 'x <- as.double(x)'
    >> before de-meaning.

    > well, yes, in that branch of the source code.  ... and a
    > similar call for the raw = TRUE case.

    > At first, this seems to make sense to me, but actually it
    > will break when someone uses

    >    poly(<complex>, ..)
   
    > [ Also: what about the "prediction" case (coef =
    > <specified>) ?  could you use prediction of an lm() for
    > your real use case ? ]

    > ---

    > Maybe it makes most sense if you open an R bugzilla entry
    > for this (including part of our current dialogue).

Even though there hasn't been any such formal bug report,
I've now committed a change (to R-devel only for the time being,
svn revision 82681) 
which re-enables the working of poly() in such cases and even 
for the  raw=TRUE  case where it had never worked.
Also, this is now documented.

The only change to the source was the insertion of

    if(is.object(x) && mode(x) == "numeric") x <- as.numeric(x) 

into the body of poly().


Thank you once more, Michael, for raising the issue.

Martin