[BioC] loess and limma

[john seers (IFR)]

Sean

Yes, that does help. Thank you.

This worries me as people routinely slap in a loess normalisation. But
from what you are saying this is not likely to be valid unless you have
an independent reference. i.e. pooled or from another source.

This means if you are using your reference channel as in a "Direct
Two-Color Design" (limma user guide Chapter 7) there is likely to be a
correlation so you should not use loess?

Does that make sense to you?


John



 
---

John Seers
Institute of Food Research
Norwich Research Park
Colney
Norwich
NR4 7UA
 

tel +44 (0)1603 251497
fax +44 (0)1603 507723
e-mail john.seers at bbsrc.ac.uk                         
e-disclaimer at http://www.ifr.ac.uk/edisclaimer/ 
 
Web sites:

www.ifr.ac.uk   
www.foodandhealthnetwork.com


-----Original Message-----
From: Sean Davis [mailto:sdavis2 at mail.nih.gov] 
Sent: 14 September 2006 13:57
To: john seers (IFR); bioconductor at stat.math.ethz.ch
Subject: Re: [BioC] loess and limma


On Thursday 14 September 2006 08:46, you wrote:
> Hi Sean
>
> Thanks for the reply. I guess what you are confirming that it is
because
> there are too few points. I guess what surprised me was that there was
> absolutely no trace of a result. Have you any idea how many points are
> needed for the data to become "real"?

John,

Loess assumes that there is supposed to be no correlation between
average 
intensity and ratio--that is why it works.  What you did with your
experiment 
was to create a situation where there was a strong correlation between
ratio 
and average intensity, thus violating that assumption.  If that
assumption is 
wrong, as it is for your data, loess will still do the calculations, but
it 
will remove that correlation, which in your case represents signal.  

So, "real" data doesn't relate to the number of points, it means that
the 
assumption of no correlation between ratio and average intensity holds,
which 
I have generally found to be the case for gene expression data, at least
to a 
working approximation.  If that assumption DOES NOT hold, as in your
data and 
array CGH, loess is not the correct method for normalization.

Hope that helps.
Sean