FR: New keep_empty = FALSE argument to list_c() and list_rbind()

[krlmlr]

I'm looking for the following behavior so that I can use pivot_wider() %>% mutate(across(..., ~ list_c(.x, keep_empty = TRUE))) as a more robust replacement for tidyr::spread(convert = TRUE) .

I can achieve it today with the lower-level list_unchop(), in a very clumsy way.

library(purrr)

# Existing behavior:
list_c(list("a", NULL, "b"))
#> [1] "a" "b"
vctrs::list_unchop(list("a", NULL, "b"))
#> [1] "a" "b"
vctrs::list_unchop(list("a", NULL, "b"), indices = as.list(1:3))
#> [1] "a" NA  "b"

# I'd like to write this:
list_c(list("a", NULL, "b"), keep_empty = TRUE)
#> [1] "a" NA  "b"

^{Created on 2023-08-20 with reprex v2.0.2}

[hadley]

Hmmmm, this doesn't feel quite right for purrr, but maybe it's the least worst place?

[hadley]

I think the way to implement this for both list_c() and list_rbind() is to replace any list elements that are identical to NULL with NA. Then the vctrs machinery can take over and ensure that the output gets a missing value of the correct type.

[krlmlr]

The original use case was a data set with a column with varying-type data.

library(tidyverse)

variant_data <- tibble(
  key = c(1, 1, 2, 2),
  name = letters[c(1:2, 1:2)],
  value = list("a", 1, NULL, NULL)
)

variant_data |>
  pivot_wider()
#> # A tibble: 2 × 3
#>     key a         b        
#>   <dbl> <list>    <list>   
#> 1     1 <chr [1]> <dbl [1]>
#> 2     2 <NULL>    <NULL>

# How to achieve this output:
tibble(
  key = c(1, 2),
  a = c("a", NA),
  b = c(1, NA),
)
#> # A tibble: 2 × 3
#>     key a         b
#>   <dbl> <chr> <dbl>
#> 1     1 a         1
#> 2     2 <NA>     NA

^{Created on 2024-07-16 with reprex v2.1.0}

[DavisVaughan]

I've come back to look at @krlmlr's original example. It's just unchop(keep_empty = TRUE)

library(tidyverse)

variant_data <- tibble(
  key = c(1, 1, 2, 2),
  name = letters[c(1:2, 1:2)],
  value = list("a", 1, NULL, NULL)
)

variant_data |>
  pivot_wider() |>
  unchop(c(a, b), keep_empty = TRUE)
#> # A tibble: 2 × 3
#>     key a         b
#>   <dbl> <chr> <dbl>
#> 1     1 a         1
#> 2     2 <NA>     NA

So now I'm questioning if list_c() and friends really should get this argument or not.

The other thing that is nice about unchop() over forcing list_c() here is that list_c(keep_empty = TRUE) would only work on this:

variant_data |>
  pivot_wider()
#> # A tibble: 2 × 3
#>     key a         b        
#>   <dbl> <list>    <list>   
#> 1     1 <chr [1]> <dbl [1]>
#> 2     2 <NULL>    <NULL>

If those list elements are size 1. Otherwise it won't recycle to the right size. unchop() is a better tool for the job here because it gets the whole data frame and can handle any recycling that's required (like if it was <chr [2]> for example).

[DavisVaughan]

See #1144 (comment)