Should the grid functions retain `NA` values in a factor?

[DavisVaughan]

It currently drops them because of the way sorted_unique() works. (This also would affect forcats::fct_unique() since the implementation came from there). We get them back in complete() because that does a full_join() and the NA was in the original data.

library(tidyr)

df <- tibble(a = factor(c("x", NA), levels = c("x", "y")))

# dropped `NA`
expand(df, a)
#> # A tibble: 2 × 1
#>   a    
#>   <fct>
#> 1 x    
#> 2 y

# we get it back from the full-join
complete(df, a)
#> # A tibble: 3 × 1
#>   a    
#>   <fct>
#> 1 x    
#> 2 y    
#> 3 <NA>

tidyr:::sorted_unique
#> function (x) 
#> {
#>     if (is.factor(x)) {
#>         factor(levels(x), levels(x), exclude = NULL, ordered = is.ordered(x))
#>     }
#>     else if (is_bare_list(x)) {
#>         vec_unique(x)
#>     }
#>     else {
#>         vec_sort(vec_unique(x))
#>     }
#> }
#> <bytecode: 0x7fa27f1ac8b0>
#> <environment: namespace:tidyr>

I am fairly certain we want to keep them, since it keeps NAs with non-factor types?

library(tidyr)

df <- tibble(a = c(NA, 1))

# keeps NA at the end
expand(df, a)
#> # A tibble: 2 × 1
#>       a
#>   <dbl>
#> 1     1
#> 2    NA

[DavisVaughan]

This affects crossing() and nesting() too

library(tidyr)

a <- factor(c("x", NA), levels = c("x", "y"))

crossing(a)
#> # A tibble: 2 × 1
#>   a    
#>   <fct>
#> 1 x    
#> 2 y

^{Created on 2021-12-15 by the reprex package (v2.0.1)}

[DavisVaughan]

If we kept the NA, we'd add it to the end of the levels() set if:

• The original data contained an NA AND
• The factor levels() don't already contain NA as a level

This is a little tricky, but I think that is correct

[DavisVaughan]

TIL base R has addNA() for this