summarise_each_q naming consistency

[tverbeke]

Currently the naming of output variables does not seem consistent:

summarise_each_q(mtcars, funs = "mean", vars = "disp")
#       disp
#1 230.7219

summarise_each_q(mtcars, funs = c("mean", "sd"), vars = "disp")
#       mean       sd
#1 230.7219 123.9387

summarise_each_q(mtcars, funs = c("mean", "sd"), vars = c("disp", "cyl"))
#   disp_mean cyl_mean  disp_sd   cyl_sd
#1  230.7219   6.1875 123.9387 1.785922

Maybe in each of the cases <var>_<fun> could be used, as in

summarise_each_q(mtcars, funs = "mean", vars = "disp")
#    disp_mean
#1 230.7219

summarise_each_q(mtcars, funs = c("mean", "sd"), vars = "disp")
#  disp_mean  disp_sd
#1 230.7219 123.9387

summarise_each_q(mtcars, funs = c("mean", "sd"), vars = c("disp", "cyl"))
#   disp_mean cyl_mean  disp_sd   cyl_sd
#1  230.7219   6.1875 123.9387 1.785922

[hadley]

It is consistent - it always uses the minimum needed to disambiguate between columns.

[tverbeke]

I understand the underlying logic, but if one wants to use the results of summarise_each_q downstream (i.e. programmatically in a larger piece of code), it requires quite some special casing to e.g. pick the right columns depending on the number of funs and vars that were specified.

[hadley]

I prefer to solve this conundrum by some how describing how the outputs should be named. Any ideas for interfaces?

[tverbeke]

Thank you, Hadley. Allowing to explicitly name outputs is a good solution. Below a first idea for an interface. If outputs is NULL (default) the current system could be applied in the presence of a single 'var' (vars of length one):

summarise_each_q(mtcars, funs = c("mean", "sd"), vars = c("disp", "cyl"), 
      outputs = list(vars = c("displace", "cylinder"), funs = c("avg", "stdev"), sep = "_"))
#   displace_avg cylinder_avg  displace_stdev   cylinder_stdev
# 1      230.7219          6.1875          123.9387      1.785922

summarise_each_q(mtcars, funs = c("mean", "sd"), vars = c("disp", "cyl"), 
      outputs = list(vars = vars, funs = funs, sep = "_"))
#   disp_mean cyl_mean    disp_sd        cyl_sd
# 1  230.7219      6.1875  123.9387  1.785922

[lionel-]

If summarise() gets the ability to return multi-row outputs (see #154), then it could make sense for summarise_each() to report one row by function with a new .funs column.

This way we'd get a tidy data frame.
See also this SO question.

[anhqle]

The issue here is not just naming inconsistency. More problematic is dplyr's behavior to add variables when there are two summarizing functions and replace existing variables when there is only one summarizing function (see #1259)

I would suggest that dplyr adds variable and renames them when the user gives a named function, e.g. by_species %>% mutate_each(funs(blah = min)). In this case, the columns will be Sepal.Length, Sepal.Width, ... , Sepal.Length_blah, Sepal.Width_blah, ...

This way, there is no need for an extra argument and the behavior is consistent with when there are multiple summarizing functions specified.

[hadley]

@LaDilettante I like that idea!