[Rd] [R] Why does R replace all row values with NAs

Hervé Pagès hpages at fredhutch.org
Tue Mar 3 22:26:55 CET 2015



On 03/03/2015 02:28 AM, Martin Maechler wrote:
> Diverted from R-help :
> .... as it gets into musing about new R language "primitives"
>
>>>>>> William Dunlap <wdunlap at tibco.com>
>>>>>>      on Fri, 27 Feb 2015 08:04:36 -0800 writes:
>
>      > You could define functions like
>
>      > is.true <- function(x) !is.na(x) & x
>      > is.false <- function(x) !is.na(x) & !x
>
>      > and use them in your selections.  E.g.,
>      >> x <- data.frame(a=1:10,b=2:11,c=c(1,NA,3,NA,5,NA,7,NA,NA,10))
>      >> x[is.true(x$c >= 6), ]
>      > a  b  c
>      > 7   7  8  7
>      > 10 10 11 10
>
>      > Bill Dunlap
>      > TIBCO Software
>      > wdunlap tibco.com
>
> Yes; the Matrix package has had these
>
> is0  <- function(x) !is.na(x) & x == 0
> isN0 <- function(x)  is.na(x) | x != 0
> is1  <- function(x) !is.na(x) & x   # also == "isTRUE componentwise"

Note that using %in% to block propagation of NAs is about 2x faster:

 > x <- sample(c(NA_integer_, 1:10000), 500000, replace=TRUE)
 > microbenchmark(as.logical(x) %in% TRUE, !is.na(x) & x)
Unit: milliseconds
                     expr       min        lq      mean   median        uq
  as.logical(x) %in% TRUE  6.034744  6.264382  6.999083  6.29488  6.346028
            !is.na(x) & x 11.202808 11.402437 11.469101 11.44848 11.517576
       max neval
  40.36472   100
  11.90916   100



>
> namespace hidden for a while  [note the comment of the last one!]
> and using them for readibility in its own code.
>
> Maybe we should (again) consider providing some versions of
> these with R ?
>
> The Matrix package also has had fast
>
> allFalse <- all0 <- function(x) .Call(R_all0, x)
> anyFalse <- any0 <- function(x) .Call(R_any0, x)
> ##
> ## anyFalse <- function(x) isTRUE(any(!x))		 ## ~= any0
> ## any0 <- function(x) isTRUE(any(x == 0))	      ## ~= anyFalse
>
> namespace hidden as well, already, which probably could also be
> brought to base R.
>
> One big reason to *not* go there (to internal C code) at all with R is that
> S3 and S4 dispatch for '==' ('!=', etc, the 'Compare' group generics)
> and 'is.na() have been known and package writers have
> programmed methods for these.
> To ensure that S3 and S4 dispatch works "correctly" also inside
> such new internals is much less easily achieved, and so
> such a C-based internal function  is0() would no longer be
> equivalent with    !is.na(x) & x == 0
> as soon as 'x' is an "object" with a '==', 'Compare' and/or an is.na() method.

Excellent point. Thank you! It really makes a big difference for 
developers who maintain a complex hierarchy of S4 classes and methods,
when functions like is.true, anyFalse, etc..., which can be expressed in
terms of more basic operations like ==, !=, !, is.na, etc..., just work
out-of-the-box on objects for which these basic operations are defined.

There is conceptually a small set of "building blocks", at least for
objects with a vector-like or list-like semantic, that can be used
to formally describe the semantic of many functions in base R. This
is what the man page for anyNA does by saying:

   anyNA implements any(is.na(x))

even though the actual implementation differs, but that's ok, as long
as anyNA is equivalent to doing any(is.na(x)) on any object for which
building block is.na() is implemented.

Unfortunately there is no clearly identified set of building blocks
in base R. For example, if I want the comparison operations to work
on my object, I need to implement ==, >, <, !=, <=, and >= (the
'Compare' group generics) even though it should be enough to implement
== and >=, because all the others can be described in terms of these
2 building blocks. unique/duplicated is another example (unique(x) is
conceptually x[!duplicated(x)]). And so on...

Cheers,
H.

>
> OTOH, simple R versions such as your  'is.true',  called 'is1'
> inside Matrix maybe optimizable a bit by the byte compiler (and
> jit and other such tricks) and still keep the full
> semantic including correct method dispatch.
>
> Martin Maechler, ETH Zurich
>
>
>      > On Fri, Feb 27, 2015 at 7:27 AM, Dimitri Liakhovitski <
>      > dimitri.liakhovitski at gmail.com> wrote:
>
>      >> Thank you very much, Duncan.
>      >> All this being said:
>      >>
>      >> What would you say is the most elegant and most safe way to solve such
>      >> a seemingly simple task?
>      >>
>      >> Thank you!
>      >>
>      >> On Fri, Feb 27, 2015 at 10:02 AM, Duncan Murdoch
>      >> <murdoch.duncan at gmail.com> wrote:
>      >> > On 27/02/2015 9:49 AM, Dimitri Liakhovitski wrote:
>      >> >> So, Duncan, do I understand you correctly:
>      >> >>
>      >> >> When I use x$x<6, R doesn't know if it's TRUE or FALSE, so it returns
>      >> >> a logical value of NA.
>      >> >
>      >> > Yes, when x$x is NA.  (Though I think you meant x$c.)
>      >> >
>      >> >> When this logical value is applied to a row, the R says: hell, I don't
>      >> >> know if I should keep it or not, so, just in case, I am going to keep
>      >> >> it, but I'll replace all the values in this row with NAs?
>      >> >
>      >> > Yes.  Indexing with a logical NA is probably a mistake, and this is one
>      >> > way to signal it without actually triggering a warning or error.
>      >> >
>      >> > BTW, I should have mentioned that the example where you indexed using
>      >> > -which(x$c>=6) is a bad idea:  if none of the entries were 6 or more,
>      >> > this would be indexing with an empty vector, and you'd get nothing, not
>      >> > everything.
>      >> >
>      >> > Duncan Murdoch
>      >> >
>      >> >
>      >> >>
>      >> >> On Fri, Feb 27, 2015 at 9:13 AM, Duncan Murdoch
>      >> >> <murdoch.duncan at gmail.com> wrote:
>      >> >>> On 27/02/2015 9:04 AM, Dimitri Liakhovitski wrote:
>      >> >>>> I know how to get the output I need, but I would benefit from an
>      >> >>>> explanation why R behaves the way it does.
>      >> >>>>
>      >> >>>> # I have a data frame x:
>      >> >>>> x = data.frame(a=1:10,b=2:11,c=c(1,NA,3,NA,5,NA,7,NA,NA,10))
>      >> >>>> x
>      >> >>>> # I want to toss rows in x that contain values >=6. But I don't want
>      >> >>>> to toss my NAs there.
>      >> >>>>
>      >> >>>> subset(x,c<6) # Works correctly, but removes NAs in c, understand why
>      >> >>>> x[which(x$c<6),] # Works correctly, but removes NAs in c, understand
>      >> why
>      >> >>>> x[-which(x$c>=6),] # output I need
>      >> >>>>
>      >> >>>> # Here is my question: why does the following line replace the values
>      >> >>>> of all rows that contain an NA # in x$c with NAs?
>      >> >>>>
>      >> >>>> x[x$c<6,]  # Leaves rows with c=NA, but makes the whole row an NA.
>      >> Why???
>      >> >>>> x[(x$c<6) | is.na(x$c),] # output I need - I have to be
>      >> super-explicit
>      >> >>>>
>      >> >>>> Thank you very much!
>      >> >>>
>      >> >>> Most of your examples (except the ones using which()) are doing logical
>      >> >>> indexing.  In logical indexing, TRUE keeps a line, FALSE drops the
>      >> line,
>      >> >>> and NA returns NA.  Since "x$c < 6" is NA if x$c is NA, you get the
>      >> >>> third kind of indexing.
>      >> >>>
>      >> >>> Your last example works because in the cases where x$c is NA, it
>      >> >>> evaluates NA | TRUE, and that evaluates to TRUE.  In the cases where
>      >> x$c
>      >> >>> is not NA, you get x$c < 6 | FALSE, and that's the same as x$c < 6,
>      >> >>> which will be either TRUE or FALSE.
>      >> >>>
>      >> >>> Duncan Murdoch
>      >> >>>
>      >> >>
>      >> >>
>      >> >>
>      >> >
>      >>
>      >>
>      >>
>      >> --
>      >> Dimitri Liakhovitski
>      >>
>      >> ______________________________________________
>      >> R-help at r-project.org mailing list -- To UNSUBSCRIBE and more, see
>      >> https://stat.ethz.ch/mailman/listinfo/r-help
>      >> PLEASE do read the posting guide
>      >> http://www.R-project.org/posting-guide.html
>      >> and provide commented, minimal, self-contained, reproducible code.
>      >>
>
>      > [[alternative HTML version deleted]]
>
>      > ______________________________________________
>      > R-help at r-project.org mailing list -- To UNSUBSCRIBE and more, see
>      > https://stat.ethz.ch/mailman/listinfo/r-help
>      > PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
>      > and provide commented, minimal, self-contained, reproducible code.
>
> ______________________________________________
> R-devel at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-devel
>

-- 
Hervé Pagès

Program in Computational Biology
Division of Public Health Sciences
Fred Hutchinson Cancer Research Center
1100 Fairview Ave. N, M1-B514
P.O. Box 19024
Seattle, WA 98109-1024

E-mail: hpages at fredhutch.org
Phone:  (206) 667-5791
Fax:    (206) 667-1319



More information about the R-devel mailing list