Skip to content

Power calculations for Wilcox.test

2 messages · Collin Lynch, David Winsemius

Original
Collin Lynch
# 
Greetings, I'm working on some analyses where I need to calculate wilcox
tests for paired samples.  In my current literature search I've found a
few papers on sample size determination for the wilcox test notably:

Sample Size Determination for Some Common Nonparametric Tests
Gottfried E. Noether
Journal of the American Statistical Association

http://www.jstor.org.pitt.idm.oclc.org/stable/2289477

My question is: are there any implementations of power calculations for
the wilcox test in R based either on Noether's methods for sample size or
another method?

	Thanks,
	Collin.
David Winsemius
#
On Dec 16, 2013, at 8:27 AM, Collin Lynch wrote:

            

      
      
      
    

        
You've offered a citation that is perhaps only accessible from computers within your own institution: at any rate it's not accessible to this non-academic viewer. Doing a tiny bit of searching shows it to be found on an alternate website:

www.stat.purdue.edu/~jennings/stat582/notes/docs3-9-10/noether.pdf?

Paired tests are really one-sample tests against a location parameter of 0.

The equations presented in that article are fairly simple. It should be easy to implement with basic R programming methods.

These days statisticians would approach the problem by setting up a simulation. It allows investiagation of more complex analysis strategies.

To look for pre-canned approaches.
To search R functions, the sos package is useful:

library(sos)

 findFn("power nonparametric")

findFn("sample size power wilcox")
found 25 matches;  retrieving 2 pages
2 
Downloaded 12 links in 10 packages.

I'm not sure why ciNparN {EnvStats} didn't show up in that set of links since the title of the help page is "Sample Size for Nonparametric Confidence Interval for a Quantile". That should be able to estimate a sample size for a specified width of CI for the median. That's a precision oriented determination which I suspect should be the same as a method that were based on sampling properties of a distribution, since you are working with ranks.