case_when() doesn't return NA for NA by default

[Pozdniakov]

The problem of case_when()'s .default =.
I think that current logic of case_when() is flawed in terms of how it handles NA. case_when() is positioned as a "A general vectorised if-else" that "allows you to vectorise multiple if_else() statements". So, I expect case_when() to behave the same way as if_else()/ifelse() if i put minimum options:

vec <- c(1L, 2L, NA_integer_)
> dplyr::if_else(vec == 1, "one", "not one")
[1] "one"     "not one" NA       

this behaviour is the same for the base R ifelse()

ifelse(vec == 1, "one", "not one")
[1] "one"     "not one" NA       

Non-vectorised if-else returns error for NA:

> if(vec[3] == 1) {"one"} else {"not one"}
Error in if (vec[3] == 1) { : missing value where TRUE/FALSE needed

However, naive attemption to reproduce the same results as for if_else()/ifelse() using case_when() will fail:

> dplyr::case_when(
    vec == 1 ~ "one", # first condition is analog of "if"
    .default = "not one" # .default value is analog of "else"
)
[1] "one"     "not one" "not one"

...due to NA are considered FALSE by all conditions, BUT .default = value will be returned.
Thus, NA values need some special treatment. For example, to mimic if_else()/ifelse() behaviour we need to set up is.na(vec) ~ NA.

There is a note on this trick in docs for case_when(), however, it is not on the surface: inside description of .default = argument and only in one code example. So, the function works not the way many users (including me) expect it to work and not consistent with if_else()/ifelse(). Finally, it contradicts the logic of NA handling in general: for missing value we don't know in which group a value behind the NA will fall and we cannot say that it belongs (or doesn't) to the .default group.

I think of this possible solution

in addition to .default = argument add .na = argument (the value that willl be returned in case of NA in condition) with default .na = NA that can be overwritten either by setting up .na = parameter or by is.na(ver) ~ value. The last one has prority over .na = argument.

[DavisVaughan]

We considered adding a .missing argument for dplyr 1.1.0, but I ultimately found that it was too confusing and too hard to define in a way that always made sense. I have discussed this in detail here #6300 (comment), but we don't currently have any plans to add this argument back in. Ultimately explicitly handling missing values ends up being the clearest way to handle them.