[BioC] Limma Coefficients using lmscFit

Sean Davis sdavis2 at mail.nih.gov
Thu Dec 22 12:57:06 CET 2005




On 12/22/05 2:27 AM, "Gordon K Smyth" <smyth at wehi.EDU.AU> wrote:

>> Date: Tue, 20 Dec 2005 21:12:24 -0500
>> From: Naomi Altman <naomi at stat.psu.edu>
>> Subject: [BioC] Limma Coefficients using lmscFit
>> To: bioconductor at stat.math.ethz.ch
>> Cc: QING ZHANG <qxz5 at psu.edu>
>> 
>> Being a big believer in single-channel analysis of loop designs, I
>> used lmscFit to analyze my loop design data. All went well until the
>> investigator requested the normalized single channel data.
>> Since lmscFit actually operates on M and A, we took the same MAlist,
>> and created R and G using RG.MA.
>> 
>> To check the computation, we then computed the treatment means by
>> hand for a few genes.  These did not work out to the same treatment
>> means obtained from lmscFit (using model ~-1+Trt).
>> 
>> I hope that someone can explain why these are not equal.  (And I
>> really hope that this is not another case where I did not read the
>> documentation sufficiently carefully.)
>> 
>> 
>> 
>> Naomi S. Altman                                814-865-3791 (voice)
>> Associate Professor
>> Dept. of Statistics                              814-863-7114 (fax)
>> Penn State University                         814-865-1348 (Statistics)
>> University Park, PA 16802-2111
> 
> I assume you're taking means of log-intensities.  lmscFit() uses generalised
> least squares with
> block weights (as for a mixed model analysis), so the values from
> lmscFit()$coef will be simple
> means only for balanced designs.  The coefficients should not in a different
> ball park to the
> means however.

Spot weights could also make the simple means different from the estimates,
even for balanced designs, couldn't they?

Sean



More information about the Bioconductor mailing list