[datatable-help] Follow-up on subsetting data.table with NAs

[Arunkumar Srinivasan]

Matthew, 

Regarding your recent answer here: http://stackoverflow.com/a/17008872/559784 I'd a few questions/thoughts and I thought it may be more appropriate to share here (even though I've already written 3 comments!).

1) First, you write that, DT[ColA == ColB] is simpler than DF[!is.na(ColA) & !is.na(ColB) & ColA == ColB,]
However, you can write this long expression as: DF[which(DF$ColA == DF$ColB), ]

2) Second, you mention that the motivation is not just convenience but speed. By checking:

require(data.table)
set.seed(45)
df <- as.data.frame(matrix(sample(c(1,2,3,NA), 2e6, replace=TRUE), ncol=2))
dt <- data.table(df)
system.time(dt[V1 == V2])
# 0.077 seconds
system.time(df[!is.na(df$V1) & !is.na(df$V2) & df$V1 == df$V2, ])
# 0.252 seconds
system.time(df[which(df$V1 == df$V2), ])

# 0.038 seconds

We see that using `which` (in addition to removing NA) is also faster than `DT[V1 == V2]`. In fact, `DT[which(V1 == V2)]` is faster than `DT[V1 == V2]`. I suspect this is because of the snippet below in `[.data.table`:

        if (is.logical(i)) {
            if (identical(i,NA)) i = NA_integer_  # see DT[NA] thread re recycling of NA logical
            else i[is.na(i)] = FALSE              # avoids DT[!is.na(ColA) & !is.na(ColB) & ColA==ColB], just DT[ColA==ColB]
        }


But at the end `irows <- which(i)` is being done:

            if (is.logical(i)) {
                if (length(i)==nrow(x)) irows=which(i)   # e.g. DT[colA>3,which=TRUE]


And this "irows" is what's used to index the corresponding rows. So, is the replacement of `NA` to FALSE really necessary? I may very well have overlooked the purpose of the NA replacement to FALSE for other scenarios, but just by looking at this case, it doesn't seem like it's necessary as you fetch index/row numbers later.

3) And finally, more of a philosophical point. If we agree that subsetting can be done conveniently (using "which") and with no loss of speed (again using "which"), then are there other reasons to change the default behaviour of R's philosophy of handling NAs as unknowns/missing observations? I find I can relate more to the native concept of handling NAs. For example:

x <- c(1,2,3,NA)
x != 3
# TRUE TRUE FALSE NA

makes more sense because `NA != 3` doesn't fall in either TRUE or FALSE, if NA is a missing observation/unknown data. The answer "unknown/missing" seems more appropriate, therefore.

I'd be interested in hearing, in addition to Matthew's, other's thoughts and inputs as well.

Best regards,

Arun

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.r-forge.r-project.org/pipermail/datatable-help/attachments/20130609/28fb57d7/attachment.html>