[BioC] Limma with multiple samples but only 2 reps

Matthew Hannah Hannah at mpimp-golm.mpg.de
Fri Oct 1 18:58:12 CEST 2004


Lots of posts today....

Actually this may also be of interest to others with similar designs.

I remember reading on a previous thread that even with low (<3 reps) you
could still use a linear model to estimate significant changes. Rather
than just ask if this is really possible I decided to do it myself
before seeing what others have to say. I know limma shrinks sample
variance towards a pooled mean so in principle it should work.

I used 7 genotypes, 2 treatments first with 3 replicates and then with
just 2 replicates. For contrasts I compared each genotype untreated vs.
treated (ebfit p<0.01). On average 3 replica detected 2500 changes, 2
reps detected 60% of these + 15% false +ves. I haven't looked at the
effect of fdr correction or including fold change yet, but looking for
other opinions. (actual #'s for the 7 comparisons are below).

Obviously more reps are better, but the planned analysis isn't my data,
just a suggestion for a collegue. Hopefully with a view to lowering the
arbitary cutoff from 2x, to include some consistent changes that just
miss the cutoff at present. Their design is slightly different - 10-15
treatments + 1 control x 2 reps.

Thanks,
Matt

3rep	Both	2rep
1146	1386	161
1004	1509	248
1103	2006	404
899	1013	185
772	1258	197
958	1887	402
1165	1795	298

Average
1007	1551	271



More information about the Bioconductor mailing list