[BioC] help with paired-test in limma

Ariel Chernomoretz ariel.chernomoretz at crchul.ulaval.ca
Tue Jul 5 18:09:46 CEST 2005

Dear list,

This is my first time with limma (and also with linear models!).
After reading the vignettes, papers and previously posted messages I came out 
with something, but I am not sure it is the correct procedure....in particular 
I have doubts related to how to handle with limma the paired nature of my 
design. I would greatly appreciate some remarks, comments, help, etc. 

I want to study the differences between two treatments: allocationA and 
allocationB, using single color affymetrix arrays, in a paired design.
Two samples were taken from 13 patients, one before, and the other after 
treatment. There were 6 and 7 patients treated with allocationA, and 
allocationB, respectively. Each sample was technically replicated, so at the 
end we have 4 chip per patient.
The experimental setup looks like this:

Sample	BeforeAfter	Patient	allocation
1		0			1		A
2		0			1		A
3		1			1		A
4		1			1		A
5		0			2		A
6		0			2		A
7		1			2		A
8		1			2		A
9		0			3		B
10		0			3		B
11		1			3		B
12		1			3		B
.		.			.		.
.		.			.		.
52		1			13		B

After reading some posts I decided to take the average of technical 
replicates. Then I used a block in order to get the lmFit, and I calculated 
the contrasts of interest:
> design<-model.matrix(~allocation*AntesDespues,data=pData(eset))
> bblock<-rep(1:13,each=2)
> fit2<-lmFit(eset,design,block=bblock)
> cont.matrix<-cbind(A.AfterVSBefore=c(0,0,1,0),
>                                    B.AfterVSBefore=c(0,0,1,1),
>                                    Interac                =c(0,0,0,1))
> fit2<-contrasts.fit(fit2,cont.matrix)
> fit2eb<-eBayes(fit2)

I think that in this way I am implicitly assuming an intrablock correlation 
level of 0.75. Is this somehow appropriate in order to consider this 
a paired t-test like analysis? 

With duplicateCorrelation, I get a value of 0.21 (!)
Is this value the one I should consider for lmFit ?
Using this rather low value isn't like disregarding
the before-after pairing for each patient?  Shouldn't I tweak the correlation 
level to an artificially higher value (for instance 0.9) instead?

Any comments would be highly appreciated

Ariel Chernomoretz, Ph.D.
Centre de recherche du CHUL
2705 Blv Laurier, bloc T-367
Sainte-Foy, Qc
G1V 4G2
(418)-525-4444 ext 46339

More information about the Bioconductor mailing list