Difference between ifelse and if_else for conditionals returning NA

I can't work this out.

I posted yesterday asking how to apply which.max rowwise in dplyr. I got a great answer that solved my problem. However, as often happens, getting one question answered raised another one. I will lay the problem out with some toy data

Here is a data frame with three variables.

data.frame(v1 = c(3,2,NA,6),
           v2 = c(NA,1,NA,7),
           v3 = c(1,1,NA,1)) -> df2

df2

# output
#
#   v1 v2 v3
# 1  3 NA  1
# 2  2  1  1
# 3 NA NA NA
# 4  6  7  1

Using which.max within dplyr, I want to calculate a new variable that returns the column position of the maximum value in each row. However, for reasons explained in my previous post linked above, if the row is all NA I want it to return NA instead of integer(0). So I use ifelse to apply a conditional whereby if the row is all NA it returns NA and, if not, it performs the rowwise which.max operation.

library(dplyr)

df2 %>% 
  rowwise() %>% 
    mutate(maxCol = ifelse(test = all(is.na(c_across(everything()))),
                           yes = NA, 
                           no = which.max(c_across(everything()))))

# output
# # A tibble: 4 × 4
# # Rowwise: 
#      v1    v2    v3 maxCol
#   <dbl> <dbl> <dbl>  <int>
# 1     3    NA     1      1
# 2     2     1     1      1
# 3    NA    NA    NA     NA
# 4     6     7     1      2

However if I do the same thing using the dplyr if_else verb it throws an error.

df2 %>% 
  rowwise() %>% 
  mutate(maxCol = if_else(condition = all(is.na(c_across(everything()))), 
                          true = NA, 
                          false = which.max(c_across(everything()))))

# Error in `mutate()`:
# ℹ In argument: `maxCol = if_else(...)`.
# ℹ In row 3.
# Caused by error in `if_else()`:
# ! `false` must have size 1, not size 0.
# Run `rlang::last_trace()` to see where the error occurred.

Why is this happening? and how do I make if_else work the same way as ifelse?

My solution (to the given problem) would be to not use ifelse or if_else. To me, the easiest appears to be to define the following:

which.max2 <- function(x) {
  x <- which.max(x)
  if (length(x) == 0) {
    NA
  } else {
    x
  }
}

And then run:

df2 %>% 
  mutate(
    maxCol = apply(across(everything()), 1, which.max2)
  )

Alternatively, we could put it in the the if_else statement:

df2 %>% 
  mutate(
    maxCol = if_else(if_all(everything(), is.na),
      NA,
      apply(across(everything()), 1, which.max2)
    )
  )

But that is just more verbose.

To illustrate the insights from the comments (and answer the "why" question), consider the output of

if_else(all(is.na(c(NA,NA,NA))),
        NA,
        which.max(c(NA,0,NA)))
# [1] NA

compared to

if_else(all(is.na(c(NA,NA,NA))),
        NA,
        which.max(c(NA,NA,NA)))
# Error in `if_else()`:
# ! `false` must have size 1, not size 0.
# Run `rlang::last_trace()` to see where the error occurred.

where the latter is what is happening in your code. if_else is trying to evaluate the false statement, even though the condition is true. The output of which.max(c(NA,NA,NA)) is integer(0), which if_else really does not like.

Because of this, the second part of your question (how do I make if_else work the same way as ifelse) could only be obtained through tryCatch(), or through some verbose workaround, as the other answer suggests.