[BioC] Limma Multi-level Experiments

[P.D. Moerland]

Dear Michael,

This is a perfectly valid approach and has been described by Gordon in section 3.5 'Comparisons Both Between and Within Subjects' of the edgeR manual. Note however that baseline differences at time A between diseased and normal cannot be estimated with this design matrix. This can be done with the approach described in section 8.7 in the Limma manual that you also refer to.

best,
Perry

---
Perry Moerland, PhD
Room J1B-215
Bioinformatics Laboratory, Department of Clinical Epidemiology, Biostatistics and Bioinformatics
Academic Medical Centre, University of Amsterdam
Postbus 22660, 1100 DD Amsterdam, The Netherlands
tel: +31 20 5666945
p.d.moerland at amc.uva.nl, http://www.bioinformaticslaboratory.nl/


-----Original Message-----
From: bioconductor-bounces at r-project.org [mailto:bioconductor-bounces at r-project.org] On Behalf Of Michael Breen
Sent: Friday, July 19, 2013 11:15 PM
To: bioconductor at r-project.org; Bioconductor Mailing List
Subject: [BioC] Limma Multi-level Experiments

Hi all ,

My data and question can best be related to section 8.7 in the Limma manual "Multi-level Experiments".  However, lets replace Tissue with Time with the idea to measure expression changes overtime that are different between disease and normal.

If I pursue this outlined route 8.7 and compare it to a very similar route which I do not estimate the correlation between measurements made on the same subject and use this  as input to the linear model , I get very similar with only minor differences.

In my work-flow I create a design matrix from this information, notice  the subtle change of numbering Subjects versus section 8.7:

FileName Subject Condition Time
File01 1 Diseased A
File02 1 Diseased B
File03 2 Diseased A
File04 2 Diseased B
File05 3 Diseased A
File06 3 Diseased B
File07 1 Normal A
File08 1 Normal B
File09 2 Normal A
File10 2 Normal B
File11 3 Normal A
File12 3 Normal B

Condition <- factor(targets$Condition, levels=c("Control","Case")) Time <- factor(targets$Time, levels=c("Pre","Post")) Subject <- factor(targets$Subject) design <- model.matrix(~Condition+Condition:Subject+Condition:Time)

And fit the design like this:

fit <- lmFit(exprs, design)
fit <- eBayes(fit)

Then I form a contrast to test for genes that respond differently overtime between disease and normal. I am still able to detect changes overtime that are different between the two groups with extremely similar results.

In short, am I missing anything not taking into consideration a correlation coefficient as input to my linear model?

Any insight is appreciated.

Yours,

Michael

        [[alternative HTML version deleted]]

________________________________

AMC Disclaimer : http://www.amc.nl/disclaimer