Skip to content

Wilcoxon Mann-Whitney Rank Sum Test in R

6 messages · Bob Green, Peter Dalgaard, Torsten Hothorn +1 more

#
An earlier post had posed the question: "Does anybody know what is relation 
between 'T' value calculated by 'wilcox_test' function (coin package) and 
more common 'W' value?"

I found the question interesting and ran the commands in R and SPSS. The W 
reported by R did not seem to correspond to either   Mann-Whitney U, 
Wilcoxon W or the Z which I have more commonly used. Correction for ties 
may have affected my results.

Can anyone else explain what the reported W is and the relation to the 
reported T?

regards

bob
#
Bob Green <bgreen at dyson.brisnet.org.au> writes:
Well, it's open source... You could just go check.

W is the sum of the ranks in the first group, minus the minimum value
it can attain, namely sum(1:n1) == n1*(n1+1)/2. In the tied cases, the
actual minimum could be larger.

The T would seem to be asymptotically normal
Asymptotic Wilcoxon Mann-Whitney Rank Sum Test

data:  pd by groups 12-26 Weeks, At term
T = -1.2247, p-value = 0.2207
alternative hypothesis: true mu is not equal to 0
[1] 0.2206883

so a good guess at its definition is that it is obtained from W or one
of the others by subtracting the mean and dividing with the SD.
#
Peter Dalgaard wrote:
With the SD adjusted for ties, of course. (See, e.g., Conover's book.)

Peter Ehlers
University of Calgary
#
P Ehlers <ehlers at math.ucalgary.ca> writes:
...which is actually the exact SD, conditional on the set of tied
ranks, not just a correction term. See my discussion with Torsten a
month or so ago.
#
On Wed, 21 Dec 2005, Peter Dalgaard wrote:

            
yes, exactly. Thanks, Peter!

The `T' values reported by functions in the `coin' package are
_standardized_ statistics. Standardization is done utilizing the
conditional expectation and conditional variance of the underlying linear
statistics as given by Strasser & Weber (1999). Note that _no_
`continuity correction' whatsoever is applied. The limit distribution is
normal (or chisq, when the test statistic is a quadratic form).

The vignette explains the theoretical framework `coin' maps into
software in more detail. It _definitively_ is worse the effort to
have a look at it. At first glance it might seem a little bit
abstract but after this you'll see how general and powerful the
tools are.

We are currently working on a manuscript showing more applications, so
watch out for the new `coin' version in a few days.

Best,

Torsten
#
Peter,

You're right, of course, as usual. Sorry about that.

Peter E.
Peter Dalgaard wrote: