select not working for grouped_df

[dickoa]

Hi,

I think that select verb is not working as expected for grouped_df.
Here's a small reproductible example :

Using select

require(dplyr)
group_by(mtcars, vs) %.% select(mpg)

Error: index out of bounds

Using subset

str(group_by(mtcars, vs) %.% subset(select = mpg))

Classes ‘grouped_df’, ‘tbl_df’, ‘tbl’ and 'data.frame':    32 obs. of  1 variable:
 $ mpg: num  21 21 22.8 21.4 18.7 18.1 14.3 24.4 22.8 19.2 ...

[romainfrancois]

Here is the implementation of select for grouped data frames:

select.grouped_df <- function(.data, ...) {
  grouped_df(select.data.frame(.data, ...), groups(.data))
}

So the error comes from the grouped_df call that uses variables no longer in the data. The data that goes into grouped_df only has the mpg variable, so cannot be grouped by vs.

Maybe we should lose the groupings or force selection of grouping variables. @hadley ?

[hadley]

Ooh, I didn't think of that possibility. I think it would be best to give an error in this situation: e.g. "selection doesn't include grouping variables: vs".

[piccolbo]

It seems to me Romain's proposal to implicitly select grouping variables was superior from a DRY or "war on boilerplate" point of view. Now we have to write

group_by(data, group.vars) %.% select(select.vars, group.vars)

Mandated repetition of group.vars or error. Can't be The Right Way.

[hadley]

Yes, that's an excellent point. No reason to let people do something that they shouldn't want to do.

[romainfrancois]

Perhaps this should come with some sort of message letting the user know we've added variables to the selection. In any case, this can be handled in the select.grouped_df method.

[hadley]

I think it's reasonable to say that it's implicit in the grouping - i.e. a grouped select includes the grouping variables in the same way that a grouped arrange orders by the grouped variables.

[piccolbo]

An alternative could be to remove the constraint and allow to drop grouping variables. This makes more sense than it seems: imagine you are doing resampling and you are only interested in the variation of an estimator, not exactly which group it comes from. But there are implications to this: I don't think it's natural to preserve the grouping when the grouping variables are gone. Hence it would become an implicit ungroup. plyrmr goes with implicit selection, but I am not sure it's the last word.