Skip to content

xyTable(x,y) versus table(x,y) with NAs

6 messages · Wolfgang Viechtbauer, Serguei Sokol, Bill Dunlap

Original
Wolfgang Viechtbauer
# 
Hi all,

Posted this many years ago (https://stat.ethz.ch/pipermail/r-devel/2017-December/075224.html), but either this slipped under the radar or my feeble mind is unable to understand what xyTable() is doing here and nobody bothered to correct me. I now stumbled again across this issue.

x <- c(1, 1, 2, 2,  2, 3)
y <- c(1, 2, 1, 3, NA, 3)
table(x, y, useNA="always")
xyTable(x, y)

Why does xyTable() report that there are NA instances of (2,3)? I could understand the logic that the NA could be anything, including a 3, so the $number value for (2,3) is therefore unknown, but then the same should apply so (2,1), but here $number is 1, so the logic is then inconsistent.

I stared at the xyTable code for a while and I suspect this is coming from order() using na.last=TRUE by default, but in any case, to me the behavior above is surprising.

Best,
Wolfgang
Serguei Sokol
# 
Le 25/04/2023 ? 10:24, Viechtbauer, Wolfgang (NP) a ?crit?:
Not really. The variable 'first' in xyTable() is supposed to detect 
positions of first values in repeated pair sequences. Then it is used to 
retained only their indexes in a vector of type 1:n. Finally, by taking 
diff(), a number of repeated pairs is obtained. However, as 'first' will 
contain one NA? for your example, the diff() call will produce two NAs 
by taking the difference with precedent and following number. Hence, the 
result.

Here is a slightly modified code ox xyTable to handle NA too.

xyTableNA <- function (x, y = NULL, digits)
{
 ??? x <- xy.coords(x, y, setLab = FALSE)
 ??? y <- signif(x$y, digits = digits)
 ??? x <- signif(x$x, digits = digits)
 ??? n <- length(x)
 ??? number <- if (n > 0) {
 ??????? orderxy <- order(x, y)
 ??????? x <- x[orderxy]
 ??????? y <- y[orderxy]
 ??????? first <- c(TRUE, (x[-1L] != x[-n]) | (y[-1L] != y[-n]))
 ??????? firstNA <- c(TRUE, xor(is.na(x[-1L]), is.na(x[-n])) | 
xor(is.na(y[-1L]), is.na(y[-n])))
 ??????? first[firstNA] <- TRUE
 ??????? first[is.na(first) | isFALSE(first)] <- FALSE
 ??????? x <- x[first]
 ??????? y <- y[first]
 ??????? diff(c((1L:n)[first], n + 1L))
 ??? }
 ??? else integer()
 ??? list(x = x, y = y, number = number)
}

Best,
Serguei.
Serguei Sokol
# 
I correct myself. Obviously, the line

first[is.na(first) | isFALSE(first)] <- FALSE

should read

first[is.na(first)] <- FALSE

Serguei.

Le 25/04/2023 ? 11:30, Serguei Sokol a ?crit?:

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Wolfgang Viechtbauer
    

    

      
      #
    

  


  

      
Nice! Would this be something to consider as either a permanent fix to xyTable() (to me, the function is right now behaving in a rather unexpected manner, if not to say, buggy) or via an argument (for backwards compatability)?

Best,
Wolfgang

            

      
      
      
    

  
  


  


      

    

    

      
      

        


  

        

      Bill Dunlap
    

    

      
      #
    

  


  

      
x <- c(1, 1, 2, 2,  2, 3)
y <- c(1, 2, 1, 3, NA, 3)

            

      
      
      
    

        
List of 3
 $ x     : num [1:6] 1 1 2 2 NA 3
 $ y     : num [1:6] 1 2 1 3 NA 3
 $ number: int [1:6] 1 1 1 NA NA 1


How many (2,3)s do we have?  At least one, the third entry, but the fourth
entry, (2,NA), is possibly a (2,3) so we don't know and make the count NA.
I suspect this is not the intended logic, but a byproduct of finding value
changes in a sorted vector with the idiom x[-1]!=x[-length(x).  Also the
following does follow that logic:

            

      
      
      
    

        
List of 3
 $ x     : num [1:5] 1 2 2 5 6
 $ y     : num [1:5] 2 2 4 NA 3
 $ number: int [1:5] 2 1 1 1 1



table() does not use this logic, as one NA in a vector would make all the
counts NA.  Should xyTable have a way to handle NAs the way table() does?

-Bill

On Tue, Apr 25, 2023 at 1:26?AM Viechtbauer, Wolfgang (NP) <

        
wolfgang.viechtbauer at maastrichtuniversity.nl> wrote:

        

            

      
      
      
    

        

      
      
    

        

  
  


  


      

    

    

      
      

        


  

        

      Serguei Sokol
    

    

      
      #
    

  


  

      
Le 25/04/2023 ? 17:39, Bill Dunlap a ?crit?:

            

      
      
      
    

        
Not really. If we take

   x <- c(1, 1, 2, 2,  5, 6, 5, 5)
   y <- c(2, 2, 2, 4, NA, 3, 3, 4)

we get

   str(xyTable(x,y))

List of 3
  $ x     : num [1:7] 1 2 2 5 5 NA 6
  $ y     : num [1:7] 2 2 4 3 4 NA 3
  $ number: int [1:7] 2 1 1 1 NA NA 1

How many (5, 3) we have? At least 1 but (5, NA) is possibly (5,3) so we 
should have NA but we have 1.
How many (5, 4) we have? At least 1 but (5, NA) is possibly (5,4) and we 
do get NA. So restored logic is not consistent.
Without talking about a pair (NA, NA) appeared and not producing (5, NA) 
pair.

Best,
Serguei.