[BioC] multtest problem...

[Katherine Pollard]

Jo,

The statistic is the same between multtest and t.test (for t-tests) or
wilcox.test (for wilcoxen tests), but the p-value is computed differently.

The functions in t.test and wilcox.test use tabled distributions to get a
p-value for each test. These assume a certain distribution for the test
statistics and also treat each test independently. The multtest package
provides methods that use the bootstrap (MTP function) or permutations to
estimate the joint null distribution of the test statistics and then
compute p-values from these. These are empirical null distributions which
differ from those used in t.test and wilcox.test in two important ways:
(i) they do not assume a particular parametric form for the test
statistics distribution, and (ii) they account for correlations between
test statistics by using the *joint* null distribution. For these reasons,
the p-values computed can be very different from those using tabled,
marginal distributions.

Hope this helps,
Katie


> ------------------------------
>
> Message: 18
> Date: Fri, 29 Apr 2005 11:58:26 +0200
> From: "Dipl.-Ing. Johannes Rainer" <johannes.rainer at tugraz.at>
> Subject: [BioC] multtest problem...
> To: "bioconductor at stat.math.ethz.ch" <bioconductor at stat.math.ethz.ch>
> Message-ID: <20050429115826.v1sqjl58kk80c88g at webmail.tugraz.at>
> Content-Type: text/plain;	charset=ISO-8859-1;	format="flowed"
>
> hi,
>
> i used the mt.teststat function from the package multtest and wonder
> why the p values that the function calculates differ from those i get
> using the wilcox.test or t.test function.
> i tried it with the golub data set (data(golub)) and
> wilcox.test(x=golub[,golub.cl==0],y=golub[,golub.cl==1])
>
> returns a p-value = 0.8987
>
> while teststat<-mt.teststat(golub,golub.cl,test="wilcoxon")
>
> teststat[1] = 1.75419
>
> i thought that the mt.teststat funciton uses simple t tests or wilcox
> test to calculate p values...
> perhaps i am wrong with the class labels?
>
> from the help i understood, that if i have two groups in my samples i
> submit as classlabels for examples c(0,0,0,1,1,1) and this means that
> the first 3 columns of my data correspond to group 0 and the other 3 to
> group 1. is this correct?
>
> thanks, jo