[BioC] limma and blocks

[Fangxin Hong]

Françoise;
I am glad that helps. Generally speaking, block effect is random from a
distribution. But if you don't have enough blocks to estimate this
distribution ( in the case of two levels within block), I do think that
a fixed effect model is better, especially when you have many units within
each block.

Bests;
Fangxin


> Fangxin,
>
> Thank you very much for your answer. I derived my models from the
> technical replicates examples in the Limma manual. I am glad you can
> confirm that model 1 corresponds to a random effect model and model 2 to a
> fixed effect model. You seem to hint that with two levels for block a
> fixed-effects model would be more appropriate. Is this correct? For a
> given factor, is a fixed-effects model better if you have fewer levels?
>
> Thanks,
>
> Françoise
>
> -----Original Message-----
> From: Fangxin Hong [mailto:fhong at salk.edu]
> Sent: Friday, November 05, 2004 6:17 PM
> To: Thibaud-Nissen, Francoise
> Cc: bioconductor at stat.math.ethz.ch
> Subject: Re: [BioC] limma and blocks
>
> I think you can read the new limma User's Guide, section 9.4. should
> answer your questions. I am sorry that I did notice that you have only two
> blocks (growth condition), tht way maybe model 2 is better.
>
> Limma User's Guide, available from
> http://bioinf.wehi.edu.au/limma/usersguide.pdf
>
> Fx
>
>> Hi,
>>
>>
>>
>> I am analyzing an experiment using 32 Arabidopsis Affymetrix chips. It
> is
>> basically a 2x2x4 design repeated twice using different biological
> replicates (the third replicate will be provided later). The plants
> within
>> each biological replicate were grown at the same time, and in that sense
> are related and form a block. There are no technical replicates within
> each block.
>>
>>
>>
>> I am using limma. In the model I calculate separate coefficients for
> each
>> of the 16 conditions. I then use contrasts matrices to evaluate
> contrasts
>> of interest.
>>
>>
>>
>> I now would like to incorporate the block effect in my model in order to
> account for random variation in the growth conditions between the two
> biological replicates.
>>
>>
>>
>> I tried two models that give different results, but I am not sure any of
> them is correct:
>>
>>
>>
>> If the first biological replicate appears first in my design, and
> "design"
>> is my design matrix for the 16 coefficients:
>>
>>
>>
>> Model 1:
>>
>> biorep <- c(rep(1,16),rep(2,16))
>>
>> fit <- lmFit(mydata, design, block= biorep)
>>
>> fit <- eBayes(fit)
>>
>> ...
>
>> Model 2:
>>
>> blockdiff <- c(rep(1,16),rep(-1,16))
>>
>> blockdesign <- cbind(design, Block=blockdiff)
>>
>> fitblock <-lmFit(mydata, blockdesign)
>>
>> fitblock <- eBayes(fitblock)
>>
>> ...
>>
>> I would appreciate any tip that could put me in the right track!
>>
>>
>>
>> Thanks,
>>
>>
>>
>> Françoise
>>
>
>
>


-- 
Fangxin Hong, Ph.D.
Plant Biology Laboratory
The Salk Institute
10010 N. Torrey Pines Rd.
La Jolla, CA 92037
E-mail: fhong at salk.edu