When to use missing versus NULL values for passing undefined function arguments in R, and why?

Question

To date when writing R functions I've passed undefined arguments as NULL values and then tested whether they are NULL i.e.

f1 <- function (x = NULL) {
   if(is.null(x))
      ...
}

However I recently discovered the possibility of passing undefined arguments as missing i.e.

f2 <- function (x) {
   if(missing(x))
      ...
}

The R documentation states that

Currently missing can only be used in the immediate body of the function that defines the argument, not in the body of a nested function or a local call. This may change in the future.

Clearly this is one disadvantage of using missing to determine undefined values are there any others people or aware of? Or to phrase the question in a more useful form "When do you use missing versus NULL values for passing undefined function arguments in R and why?"

Answer 1

NULL is just another value you can assign to a variable. It's no different than any other default value you'd assign in your function's declaration.

missing on the other hand checks if the user supplied that argument, which you can do before the default assignment - which thanks to R's lazy evaluation only happens when that variable is used.

A couple of examples of what you can achieve with this are: arguments with no default value that you can still omit - e.g. file and text in read.table, or arguments with default values where you can only specify one - e.g. n and nmax in scan.

You'll find many other use cases by browsing through R code.

Answer 2

missing(x) seems to be a bit faster than using default arg to x equal to NULL.

> require('microbenchmark')
> f1 <- function(x=NULL) is.null(x)
> f2 <- function(x) missing(x)

> microbenchmark(f1(1), f2(1))
Unit: nanoseconds
  expr min  lq median    uq  max neval
 f1(1) 615 631  647.5 800.5 3024   100
 f2(1) 497 511  567.0 755.5 7916   100

> microbenchmark(f1(), f2())
Unit: nanoseconds
 expr min  lq median    uq  max neval
 f1() 589 619    627 745.5 3561   100
 f2() 437 448    463 479.0 2869   100

Note that in the f1 case x is still reported as missing if you make a call f1(), but it has a value that may be read within f1.

The second case is more general than the first one. missing() just means that the user did not pass any value. is.null() (with NULL default arg) states that the user either did not pass anything or he/she passed NULL.

By the way, plot.default() and chisq.test() use NULL for their second arguments. On the other hand, getS3method('t.test', 'default') uses NULL for y argument and missing() for mu (in order to be prepared for many usage scenarios).

I think that some R users will prefer f1-type functions, especially when working with the *apply family:

sapply(list(1, NULL, 2, NULL), f1)

Achieving that in the f2 case is not so straightforward.