# [Rd] rank(*) with NAs -- new option "keep" desired

Martin Maechler maechler at stat.math.ethz.ch
Thu Sep 11 17:57:22 MEST 2003

```Thank you, Gabor!

>>>>> "Gabor" == Gabor Grothendieck <ggrothendieck at volcanomail.com>
>>>>>     on Thu, 11 Sep 2003 07:09:42 -0700 (PDT) writes:

Gabor> I could have used this also.  Currently I do this:
Gabor> z <- ifelse(is.na(y),NA,rank(y))
Gabor> names(z) <- names(y)

Gabor> The following also works but is less transparent:

Gabor> z <- y
Gabor> z[z==z] <- rank(y)

the next version of R should definitely have a builtin version,
as a matter of fact, today's (or tomorrow's) R-alpha version
My point was rather enquiring on how the "API" should be setup
(argument name, semantic, ..).

Gabor> Another extension of rank that could be useful would be to have
Gabor> an argument that causes it NOT to do tie averaging.
Gabor> This is useful when you are using rank(x) in the
Gabor> sense of an inverse permutation.  Currently I do
Gabor> this:

Gabor> z <- order(order(y))
Gabor> names(z) <- names(y)

Good point!  However,  help(rank) has been mentioning this and recommends
(and does forget about the names() !)

[ For efficiency reasons, sort.list() is preferable to order for
the simple order()ing on only one argument. ]

Gabor> --- Martin Maechler <maechler at stat.math.ethz.ch> wrote:
>> In some contexts, I find the current behavior of rank() very
>> `suboptimal'.
>>
>> We have the argument na.last = {TRUE | FALSE | NA }
>> where the first two cases treating NAs (almost) as if they were
>> == +Inf or == -Inf  whereas the 3rd case just drops NAs.
>> For the typical ``Rank Transformation'' that is recommended in
>> EDA in several contexts, I would however want something else,
>> namely keep the NAs !
>>
>> An example -- including the new option as I'm proposing it ---
>> makes things more clear :
>>
>>> y <- c(2:1,NA,0)
>>> rank(y, na.last = TRUE)## ==== rank(y)
>> [1] 3 2 4 1
>>> rank(y, na.last = FALSE)
>> [1] 4 3 1 2
>>> rank(y, na.last = NA)
>> [1] 3 2 1
>>> rank(y, na.last = "keep") ### <<<<< NEW >>
>> [1]  3  2 NA  1
>>>
>> ---
>>
>> Alternatively to extending the possible values of `na.last' I
>> first thought of a new (boolean) argument, but found the current
>> solution less ugly.
>>
>> Feedback welcome!
>>
>> Martin Maechler <maechler at stat.math.ethz.ch>	http://stat.ethz.ch/~maechler/
>> Seminar fuer Statistik, ETH-Zentrum  LEO C16	Leonhardstr. 27
>> ETH (Federal Inst. Technology)	8092 Zurich	SWITZERLAND
>> phone: x-41-1-632-3408		fax: ...-1228			<><
>>
>> PS: Stumbled over this while implementing  cor.test()s
>> method = c("pearson", "spearman", "kendall")  for  cor()
>> itself.

```