[BioC] globaltest: pairwise design?

[Goeman, J.J. (MSTAT)]

 Hi Mette,

> We have one-color ABI-data from 3 persons at 2 timepoints (a pilot
> study), and want to compare the gene expression at the two timepoints
> for a group of genes.

> We could forget that the same person gave rise to two correlated
values
> in time and run globaltest with y=0,0,0 for the samples at timepoint 1
> and and y=1,1,1 for the samples at timepoint 2. But, we really want to
> do a paired test. Is that possible with globaltest?
> We tried making differences between timepoint 2 and 1 for each person
> and aim at testing difference=0, but I guess globaltest is not
designed
> for that?

> I'm sorry for bothering you with these simple question - and
appologise
> if I have missed this information in the documentation.
> All our collaborators are now wanting to test groups of genes instead
of
> individual genes -- so your globaltest is really helping us!


To do globaltest in a pairwise design, make your phenoData object as
follows

array   timepoint   individual
1       0           1
2       1           1
3       0           2
4       1           2
5       0           3
6       1           3

Then tell globaltest that the individuals have correlated measurements
by putting an effect per individual into the model.

globaltest(X, timepoint ~ factor(individual), genesets)

Note that this is very similar to the way you can make a paired t-test
from a two-sample t-test by putting a pair effect into the model. Adding
such a pair effect could gain you much power if there is much
heterogeneity between pairs.

The permutation version of globaltest is not yet implemented when
covariates are present, so you'll have to do with the asymptotic
distribution. This is not ideal for 6 samples, but for a pilot it may be
good enough.

Best,

Jelle