as.numeric(<factor>) [Difference R/S]

From  R-core;  this should interest most R-devel'ers (to some extent):

Since 0.60,  the semantics of   as.numeric(<factor>)  has changed,
e.g.

R> as.integer(factor(c("A","BB")))
[1] NA NA
R> as.integer(factor(c(100,40,100)))
[1] 100  40 100

whereas older R and S:

S> as.integer(factor(c("A","BB")))
[1] 1 2
S> as.integer(factor(c(100,40,100)))
[1] 2 1 2

Hmm,  I first had advocated your view above, myself.

Later, I started to discover in how much S-code
	as.numeric(ff) 
is just used to extract the factor codes (in {1:M})  from a factor.

This lead me (and Peter Dalgaard, I think) to the conclusion that
- yes, the present R behavior maybe ``cleaner'' than S's
- no, it is a pain to keep it, because it breaks S code too often.

However, as you see, we haven't agreed yet on the topic.
I think we should agree ASAP, since it involves code in several places
(outside R base).

(1:5)[factor(1:5,labels=5:1)]

as.numeric(factor(1:5,labels=5:1))

Peter Dalgaard BSA writes:

Martin Maechler <maechler@stat.math.ethz.ch> writes:

From  R-core;  this should interest most R-devel'ers (to some extent):

Since 0.60,  the semantics of   as.numeric(<factor>)  has changed,
e.g.

[1] NA NA

[1] 100  40 100

whereas older R and S:

[1] 1 2

[1] 2 1 2

...

Hmm,  I first had advocated your view above, myself.

Later, I started to discover in how much S-code
as.numeric(ff) 
is just used to extract the factor codes (in {1:M})  from a factor.

This lead me (and Peter Dalgaard, I think) to the conclusion that
- yes, the present R behavior maybe ``cleaner'' than S's
- no, it is a pain to keep it, because it breaks S code too often.

However, as you see, we haven't agreed yet on the topic.
I think we should agree ASAP, since it involves code in several places
(outside R base).

Actually, I'm even stronger in favour of the S semantics. In addition
to the above

	- you can always get current behaviour with
        as.numeric(as.character(f)) or as.numeric(levels(f))[f]

	- one should avoid generating NA's unless absolutely necessary

	- when a factor is used for subscripting, you mean the codes,
        not the levels. Currently, we have

(1:5)[factor(1:5,labels=5:1)]

[1] 1 2 3 4 5

but

as.numeric(factor(1:5,labels=5:1))

[1] 5 4 3 2 1

	I.e. *sometimes* when a factor is coerced to numeric you get
	something different. (And if you change the index semantics,
	code for trend tests and the like is likely to break!).

But that is really a matter of how subscripting treats factors, and not
necessarily what coercion does.

As much as I am in favor of compatibility (remember I do a lot of
porting):

* Suppose f is a factor with numeric levels other than 1 to n.  Then
as.numeric(f) returning the codes rather than the levels is strange.

* You also cannot coerce a character vector to numeric without getting
NA's.

Btw:

	x <- factor(c(10, 5, 6, 7))

Then levels(x) gives the CHARACTER vector c("5", "6", "7", "10") [in
both R and S+], why that?

And:

R> codes(x)
[1] 4 1 2 3

S> codes(x)
[1] 1 2 3 4

???