[R] strange behavior of cor() with pairwise.complete.obs

[Peter Dalgaard]

Hilmar Berger wrote:
> Sorry, 
>
> I obviously did not state clearly what the problem is (thanks Daniel):
>
> 1. minor problem: cor() does return different types of variables for
> methods "kendall" and pearson (matrix vs. scalar) when 
> pairwise.complete.obs is selected. 
>
> 2. major problem: cor() does return with an error if both x and y 
> are matrices with method="kendall" when pairwise.complete.obs is
> selected and one column of one of the two matrices is completely NA.
> This does not happen for method "pearson".
>
> Regards,
> Hilmar
>
> Hilmar Berger <hilmar.berger <at> imise.uni-leipzig.de> writes:
>   
>> Hi all, 
>>
>> I'm not quite sure if this is a feature or a bug or if I just fail to
>> understand
>> the documentation: 
>>
>> If I use cor() with pairwise.complete.obs and method=pearson, the
>> result is a
>> scalar: 
>>
>> ->cor(c(1,2,3),c(3,4,6),use="pairwise.complete.obs",method="pearson")
>> [1] 0.9819805
>>
>> The documentation says that 
>> " '"pairwise.complete.obs"' only works with the '"pearson"' method
>>      for 'cov' and 'var'."
>>
>> Thus, I guess that cor() should work for pairwise.complete.obs and
>> method =
>> "kendall", or am I misinterpreting that statement ?
>>
>>     
I would interpret it to mean that it does NOT work for "kendall" and 
"spearman" and I don't see how you can possibly interpret otherwise.

>> -> c(1,2,3),c(3,4,6),use="pairwise.complete.obs",method="kendall")
>>      [,1]
>> [1,]    1
>>
>> Now the result is a matrix with dimensions (1,1) - strange enough.
>>
>> Note that when I use "all.obs" or "complete.obs" I get a scalar for
>> method
>> kendall, too.
>>
>> It gets worse if one tries to calculate the correlation between the
>> columns of
>> two matrices (i.e. cor(x,y) with x and y being a matrix). Then
>>
>> -> c=matrix(c(1,2,3,3,4,5),nrow=3,ncol=2)
>> -> d=matrix(c(2,3,4,NA,NA,NA),nrow=3,ncol=2)
>> -> cor(c,d,use="pairwise.complete.obs",method="pearson")
>>      [,1] [,2]
>> [1,]    1   NA
>> [2,]    1   NA
>>
>> -> cor(c,d,use="pairwise.complete.obs",method="kendall")
>> Error: 'x' is empty (*translated from german error message*)
>>
>> The behavior is reproducible in R 2.4.1 and 2.6.1 (WinXP). I noticed
>> that in
>> 2.7.0 something was fixed in cor() related to "complete.obs" handling
>> - would
>> that fix my problems ?
>>     
Apparently. There are ways to find this out yourself, you know....

The help page still claims that it doesn't work, though.

-- 
   O__  ---- Peter Dalgaard             Øster Farimagsgade 5, Entr.B
  c/ /'_ --- Dept. of Biostatistics     PO Box 2099, 1014 Cph. K
 (*) \(*) -- University of Copenhagen   Denmark          Ph:  (+45) 35327918
~~~~~~~~~~ - (p.dalgaard at biostat.ku.dk)                  FAX: (+45) 35327907