WISH: rbind on data.frames with 0 columns would preserve rows

[karoliskoncevicius]

With matrix objects rbind works as expected:

mat1 <- matrix(nrow=2, ncol=0)
mat2 <- matrix(nrow=2, ncol=0)

> dim(rbind(mat1, mat2))
[1] 4 0

Once turned into data.frames the rows are dropped:

> dim(rbind(as.data.frame(mat1), as.data.frame(mat2)))
[1] 0 0

However cbind works as expected:

> dim(cbind(as.data.frame(mat1), as.data.frame(mat2)))
[1] 2 0

[jangorecki]

you might find following discussion related: Rdatatable/data.table#2422

[karoliskoncevicius]

@jangorecki thank you for the comment and the link.

You seem to argue that zero-row/column matrices should be a thing and nonzero-row zero-column data.frames should not be a thing. I would tend to agree, however below is an example of scenario I encountered where such data.frames are a good default edge case.

For my personal uses I am developing an "annotated matrix" object. That is - a matrix with persistent annotations of rows and columns. This object holds annotation about rows and columns in two separate data.frames. Once you select the row of such a matrix - row annotations are automatically subsetted and vice-versa.

Under such conditions a data.frame with non zero rows and zero columns is quite useful - it's a state of there not being any meta-data present. So for a matrix with n rows and m columns - the column-annotation data.frame would have n rows and 0 columns.

I found that having this behaviour works well with various cbind, rbind, etc operations by default without resorting to hacks. And I would have to add some ad-hoc solutions in order to make dim(0, 0) work.

This is precisely the issue I encountered and reported here. When I cbind my special matrices - I can simply also do cbind on their annotations. However this didn't work when I did an rbind, which was confusing (but documented) behaviour in base R.

[jangorecki]

I would rather prefer to change cbind.data.frame to match rbind.data.frame behavior.