Add function that creates factor in order of case_when matches

[dchiu911]

A common workflow I do is map one vector to another using some (possibly complex) conditions, then coerce to a factor with the level order the same as parsed in dplyr::case_when(). It would be helpful if there was a wrapper that created the factor without having to manually specify the levels.
Currently, I'd do something like this:

library(dplyr)
#> 
#> Attaching package: 'dplyr'
#> The following objects are masked from 'package:stats':
#> 
#>     filter, lag
#> The following objects are masked from 'package:base':
#> 
#>     intersect, setdiff, setequal, union

set.seed(2022)
x <- sample(
  c("low", "intermediate", "high"),
  prob = c(0.5, 0.2, 0.3),
  size = 100,
  replace = TRUE
)
z <- rbinom(
  n = 100,
  size = 100,
  prob = 0.3
)
y <- case_when(
  x == "intermediate" | (x == "low" & z < 30) ~ "B",
  x == "low" ~ "A",
  x == "high" ~ "C",
  TRUE ~ NA_character_
) %>%
  factor(levels = c("B", "A", "C"))
str(y)
#>  Factor w/ 3 levels "B","A","C": 1 3 2 3 2 3 2 1 1 3 ...

^{Created on 2022-02-01 by the reprex package (v2.0.1)}

Can we add a function that makes y into a factor with the level order the same as specified in the case_when()? For example,

y <- fct_case(
  x == "intermediate" | (x == "low" & z < 30) ~ "B",
  x == "low" ~ "A",
  x == "high" ~ "C",
  TRUE ~ NA_character_
)

[hadley]

I think we'd need to make the syntax more limiting than case_when() because the RHS of a case_when() can itself use data values, and reasoning through how those values should interact between conditions seems hard.

Since we'd want to restrict each expression to a single character level, we could put it in the LHS of =, something like:

something(
  "B" = x == "intermediate" | (x == "low" & z < 30),
  "A" = x == "low",
  "C" = x == "high",
)

But I don't know if any existing tidyverse function uses similar syntax.

[dchiu911]

I do think removing the usage of ~ would make it more consistent as case_when() syntax is quite unique

[DavisVaughan]

But I don't know if any existing tidyverse function uses similar syntax.

FWIW this is basically how fct_recode() works (name represents new level, value was the old level), so it wouldn't be unheard to let the name represent the new level, and the value be the logical condition

[hadley]

Will wait until lower level functions are exposed by vctrs.

[brianmsm]

I would think it would be convenient to solve this from the case_when() itself:

Something like this:

set.seed(2022)
x <- sample(
  c("low", "intermediate", "high"),
  prob = c(0.5, 0.2, 0.3),
  size = 100,
  replace = TRUE
)
z <- rbinom(
  n = 100,
  size = 100,
  prob = 0.3
)
y <- case_when(
  x == "intermediate" | (x == "low" & z < 30) ~ "B",
  x == "low" ~ "A",
  x == "high" ~ "C",
  TRUE ~ NA_character_,
  .ptype = "factor"
)

[DavisVaughan]

Also read tidyverse/tidyups#29 (comment) for some inspiration and use cases

I think providing forcats specific versions of dplyr functions that create new factors would be the best way to tackle this:

• A dplyr::recode_values() specification for creating new from lookup tables or 1:1 value matching
• A dplyr::case_when() specification for creating new factors from arbitrary logical conditions

I don't think we need forcats versions of the replace_*() helpers because forcats already has many tools for manipulating factors, but few tools for creating new ones.

As mentioned elsewhere it will be interesting to consider the fact that the RHS of these is typically just a single string, which could affect how we write these functions, but it is worth considering how people would provide NA as the value you map to - that's not easy to provide as a name so fct_case(NA = is.na(x)) doesn't really work, that might be enough to keep the case_when() formula API in forcats too.

Also, if we provide a default fallthrough value, does the level we get from that go at the start or end of the factor level list? Particularly important for ordered factors if we allow direct creation of them.