[BioC] limma p.values, tech replication and dup cor
Matthew Hannah
Hannah at mpimp-golm.mpg.de
Wed Nov 10 15:51:07 CET 2004
Hi,
I've got a question about df.prior, df.residual and p.value from eBayes.
df.prior seems to be number of arrays minus no of coefficients.
ebayes$p.value is on (df.prior + df.residual). This leads to a strange?
situation when I compare between a technically replicated (single
channel) spotted oligo array and non-tech replicated affy experiment.
I'd like to keep the information in the tech replicates so used
duplicateCorrelation. However, I'm now concerned that the tech
replication may be treated in some way like biological variation by
lmFit().
Basically if you mean the tech replicates and do the fit below without
the dupCor then the p.values appear to be generally higher, particularly
at the low end (see graph). Is this due to the t-statistic being based
on a higher df? If so is this valid as the biological information hasn't
changed?
If I now compare to my affy data then I get a similar situation (ie:
less significant p.values), and wonder if this is for a similar reason
based of the way the p.values are calculated.
I realise this is probably 'statistical' rather than anything
specifically wrong with limma but wonder if I should worry about it.
My logical reasoning is-
1. Spotted oligo >> Affy in tech variability
2. This technical variability should be included in the model.
3. It's 'unfair' on the affy data as IF I had tech replicates then the
consensus cor would be higher than for the oligo, and more genes would
be significant?
4. You can increase the significant changes by increasing the tech
replication....
Thoughts or answers appreciated.
Cheers,
Matt
#2 treatments, 3 bio reps, 2 tech reps. mut1a, mut1b, con1a, con1b,
mut2a, mut2b...etc
Exp <- c(1,1,2,2,3,3,4,4,5,5,6,6)
design <- cbind(Con=1,MutvsCon=c(1,1,0,0,1,1,0,0,1,1,0,0))
corfit <- duplicateCorrelation(signal, design, block =Exp)
fit <- lmFit(signal, design, block = Exp, correlation=corfit$consensus)
fit <- eBayes(fit)
-------------- next part --------------
A non-text attachment was scrubbed...
Name: dupCor.png
Type: image/png
Size: 17921 bytes
Desc: dupCor.png
Url : https://stat.ethz.ch/pipermail/bioconductor/attachments/20041110/28b29b95/dupCor.png
More information about the Bioconductor
mailing list