`na.rm` Argument to Deal with Missing Values

[my-R-help]

Something like this:

> library(zoo)
> ( x <- c(2, 4, 6, NA, 8) )
[1]  2  4  6 NA  8
> rollapplyr(x, 2, mean, na.rm = F)   # same as `roll_mean(x, 2)`
[1]  3  5 NA NA
> rollapplyr(x, 2, mean, na.rm = T)   # `roll_mean(x, 2, na.rm = T)`
[1] 3 5 6 8

[my-R-help]

Problem with sum and na.rm=T

If such a na.rm argument would be implemented, it might also be useful to add a special rolling sum version that, for a window with all-NAs, returns NA and not zero. The code below illustrates:

library(zoo)
x <- c(2, 4, NA, NA, 8)
rollapplyr(x, 2, sum, na.rm = T)
[1] 6 4 0 8

For many applications, I think it would be more natural to have the following output in this case:

[1]  6  4 NA  8

Background Information

This is because (as explained in ?sum) the sum of an empty set is zero.

sum(NA, na.rm = T)
[1] 0

However, for many applications, I think it would be more natural for sum to return NA in this case. Here is an amended sum function that does what I have in mind:

s <- function(x, na.rm = FALSE) {
  if (!na.rm) return(sum(x))
  if (all(is.na(x))) {
    o <- NA
    class(o) <- class(x)
    return(o)}
  sum(x, na.rm = TRUE)}

Using this with zoo::rollapplyr returns the desired result:

rollapplyr(x, 2, s, na.rm = T)
[1]  6  4 NA  8

[kevinushey]

I've added na.rm, but haven't yet made a decision about the behaviour of sum. I could imagine someone arguing for this behavior for all rolling functions, really.

[my-R-help]

Thanks, that was quick!

My reasoning for rolling sum is that if it gives NA, you can later convert the resulting NAs back to zero if needed. However, if it returns zero (as base::sum does), you don't know whether this zero is because the sum is really zero (e.g. 0+0), or because it was all-NA. In that sense, we are losing information that would be potentially valuable.

[my-R-help]

I think it might be a good idea to keep both options, i.e. the one that works as base::sum and the one suggested by me above. Maybe add an argument to roll_sum that switches between both behaviors, with the default being set consistent with base::sum.

[hsintone]

Now that the number of values used to calculate the return value for a given window can vary (if na.rm=T) it would be helpful to have an access to the n of a given window. This would also help solve the problem @my-R-help talked about. If the rolling function returns 0 when the input was all-NA and in addition to that 0 for n, you could change the return value to NA yourself if you wish to do so by looking at the n. Maybe this could be optional with return.n=T or similar, especially if it affects performance.