Date Format

On Tue, 2006-07-11 at 20:05 +0200, Peter Dalgaard wrote:

Martyn Plummer <plummer at iarc.fr> writes:

I moved this to R-devel because I am wondering why the base package does
not allow you to convert from numeric to Date. Could we not have
something like this?

as.Date.numeric <- function(x, epoch="1970-01-01", ...) {
   if (!is.character(epoch) || length(epoch) != 1)
      stop("invalid epoch")
   as.Date(epoch, ...) + x
}

We could, but you might as well do it explicitly. There's something to
be said for not confusing the concept of dates with a particular
implementation, which is effectively what happens if you can convert
them to and from numeric too seamlessly.

Currently you can easily convert one way, but not the other. I just find
that a bit odd.

Pierre's problem was that his Date objects were converted internally by
some function.  His first instinct, to use as.Date to convert them back,
was, I think, correct. But that doesn't work. So now we say "You have to
understand how Date objects are implemented to get your dates back"? I
don't know about that.

I'm more perplexed by the failure of adding difftimes to dates:

as.Date("2006-1-1") + (as.Date("2006-1-1") - as.Date("2006-1-2"))

[1] "2005-12-31"
Warning message:
Incompatible methods ("+.Date", "Ops.difftime") for "+"

and if you have a difftime in non-days units, you'll  actually get a
wrong result:

D1 <- as.Date("2006-1-1")
D2 <- as.Date("2006-1-2")
difftime(D2,D1,units="hours")

Time difference of 24 hours

dd <- difftime(D2,D1,units="hours")
D1+dd

[1] "2006-01-25"
Warning message:
Incompatible methods ("+.Date", "Ops.difftime") for "+"

[I raised this problem earlier in private discussions with Peter]

It certainly is perplexing. There is code in "+.Date" that correctly
handles the case where the second argument is a difftime. But it will
never get called! I wonder if it ever worked.

The warning is coming from DispatchGroup (in eval.c). When it finds
different methods for two arguments of a binary group generic, it gives
up and the default method is called - in this case R_binary in
arithmetic.c - which is why the results depends on the implementation of
the difftime object.

I guessed that this was a limitation of S3 generics, and I suppose I was
right.  To allow mixing arguments of two classes, you would need code in
Ops.foo to handle objects of class bar *and* vice versa.  It's a bad
idea to have two separate bits of code to do the same job, so I can't
fault the logic of forbidding this, but it does leave us with some
usability problems.

While we are on the topic, is there no function to convert a difftime
object from one time scale to another? I found a couple of private
functions, but nothing public.