[FEAT] Equivalent to group_by(.drop = FALSE) in group_by() and group_by_vars()

[phgrosjean]

Currently, group_by() and group_by_vars() do not allow to compute groups, keeping levels that are not present in the dataset, something like what dplyr::group_by(.drop = FALSE) does.
Yet, using groups from group_by() into fsummarise() or converting them to a GRP object shows that {collapse} honors .drop = FALSE. So, I guess it could be added as an argument to group_by() and group_by_vars() (and GRP.default()) quite easily?

Reproducible example:

library(dplyr)
library(collapse)

iris2 <- iris[iris$Species != "versicolor", ]

# dplyr, do not drop unused factor levels
iris2 |>
  group_by(Species, .drop = FALSE) |>
  summarise(mean = mean(Sepal.Length))

# collapse... how to do the same with fgroup_by() or group_by_vars()?
iris2 |>
  fgroup_by(Species) |>
  fsummarise(mean = fmean(Sepal.Length))

# Yet, collapse is handling it correctly when using group_by():
iris2 |>
  group_by(Species, .drop = FALSE) |>
  fsummarise(mean = fmean(Sepal.Length))

# Indeed, conversion from grouped_df to GRP is considering .drop = FALSE
iris2 |>
  group_by(Species, .drop = TRUE) |>
  GRP()

iris2 |>
  group_by(Species, .drop = FALSE) |>
  GRP()

[SebKrantz]

Thanks. I can consider it, although enabling this may have unintended side effects. For now, fgroup_by() always calls GRP.default() to give consistent behavior across univariate and multivariate groupings. There exists, however, GRP.factor(), and by default it has drop = FALSE. So a workaround could be:

library(fastverse)
iris2 <- iris[iris$Species != "versicolor", ]
  
iris2 %>%
  structure(groups = GRP(.$Species), 
            class = c("grouped_df", class(.))) %>%
  fsummarise(mean = fmean(Sepal.Length))
#>      Species  mean
#> 1     setosa 5.006
#> 2 versicolor    NA
#> 3  virginica 6.588

^{Created on 2025-07-26 with reprex v2.0.2}

[SebKrantz]

I will push a small fix that allows you to do fgroup_by(iris2, Species = GRP(Species)) shortly. Is that an acceptable solution within the scope you had envisioned? Note that after aggregation in this case Species is not a factor anymore (as in the structure() solution above).

[SebKrantz]

Please try master branch.

[phgrosjean]

Just group_by(iris2, GRP(Species)) produces the same result. Is there no way to include the GRP(<factor>) inside the logic of fgroup_by() with a new argument drop = FALSE to mimic a little bit better dplyr::group_by()? If not, I'll figure out how I could manage this case. I will also have to convert Species in a factor with same levels, so:

iris2 |>
  fgroup_by(GRP(Species)) |>
  fsummarise(mean = fmean(Sepal.Length)) |>
  fmutate(Species = factor(Species, levels = levels(iris2$Species)))

This is not easy to remember...
Thanks for your answer and your great work on 'collapse'.

[SebKrantz]

Ok, then I'm afraid I cannot help you further. The algorithms inside GRP.default() are not designed to handle factors in a special way (to them this is an integer vector with some attributes). So a general drop argument to fgroup_by() is not feasible.

[phgrosjean]

Oh I see. Thanks anyway.