combn unclasses factor variables

Question

UPDATE: FIXED

This is fixed in the upcoming release of R 3.1.0. From the CHANGELOG:

combn(x, simplify = TRUE) now gives a factor result for factor input x (previously user error).
Related to PR#15442

---

I just noticed a curious thing. Why does combn appear to unclass factor variables to their underlying numeric values for all except the first combination?

x <- as.factor( letters[1:3] )

combn( x , 2 )
#     [,1] [,2] [,3]
#[1,] "a"  "1"  "2" 
#[2,] "b"  "3"  "3" 

This doesn't occur when x is a character:

x <- as.character( letters[1:3] )

combn( x , 2 )
#     [,1] [,2] [,3]
#[1,] "a"  "a"  "b" 
#[2,] "b"  "c"  "c"

Reproducible on R64 on OS X 10.7.5 and Windows 7.

Answer 1

I think it is due to the conversion to matrix done by the simplify parameter. If you don't use it you get:

combn( x , 2 , simplify=FALSE)
[[1]]
[1] a b
Levels: a b c

[[2]]
[1] a c
Levels: a b c

[[3]]
[1] b c
Levels: a b c

The fact that the first column is OK is due to the way combn works: the first column is specified separately and the other columns are then changed from the existing matrix using [<-. Consider:

m <- matrix(x,3,3)
m[,2] <- sample(x)
m
     [,1] [,2] [,3]
[1,] "a"  "1"  "a" 
[2,] "b"  "3"  "b" 
[3,] "c"  "2"  "c" 

I think the offending function is therefore [<-.

Answer 2

As Konrad said, the treatment of factors is often odd, or at least inconsistent. In this case I think the behaviour is weird enough to constitute a bug. Try submitting it, and see what the response is.

Since the result is a matrix, and there is no factor matrix type, I think that the correct behaviour would be to convert factor inputs to character somewhere near the start of the function.