[BioC] limma linear model [was: (no subject)]

[Gordon K Smyth]

> Date: Tue, 12 Oct 2004 12:49:38 -0400
> From: karen <kschlauc at vt.edu>
> Subject: [BioC] (no subject)
> To: bioconductor at stat.math.ethz.ch
> Message-ID: <41E2F73D at zathras>
> Content-Type: text/plain; charset="ISO-8859-1"
>
> Hi Folks,
>
> Thanks, Fangxin, for your replies.
>
> I still have a few questions on my linear model for the following experiment.
> If someone could help, I'd be grateful.
>
> This is an an Affymetrix time series experiment, 7 time points,
> 2 genotypes, 2 replicates of each (28 arrays).
>
> Of interest are genes that are differentially expressed
> between genotypes across all but the first time state.
>
> The design I set up with 14 treatments and 6 contrasts:
>
> genotype1_time1, genotype1_time2....
> genotype2_time1, genotype2_time2....
>
> genotype1_time2 - genotype2_time2
> genotype1_time3 - gentotype2_time3 ...
>
> The code:
>
> treatment.vector<-c(rep(1,2),rep(2,2),rep(3,2), rep(4,2), rep(5,2), rep(6,2),
> rep(7,2),
>
> rep(8,2),rep(9,2),rep(10,2),rep(11,2),rep(12,2),rep(13,2),rep(14,2))
> treatments<-factor(treatment.vector,labels=exp.labels)
> design<-model.matrix(~-1+treatments)
> fit <- lmFit(Mrma, design)
> contrast.matrix<-makeContrasts(KO.15min-WT.15min,
>                                             KO.30min-WT.30min,
>                                             KO.90min-WT.90min,
>                                             KO.3hr-WT.3hr,
>                                             KO.6hr-WT.6hr,
>                                             KO.24hr-WT.24hr,levels=design)
>
> fit2 <- contrasts.fit(fit, contrast.matrix)
> fit3 <- eBayes(fit2)
> clas <- classifyTestsF(fit3,fstat.only=FALSE)
> FStats<-FStat(fit3)
>
> The Questions:
> 1) Is this an acceptable model to use?

You seem to be fitting a separate coefficient for each factor combination, and then extracting
contrasts of interest between the coefficients.  This is a basic approach which is generally
applicable.  So, yes.

> 2) How would I report the model via an equation?
> Even using contrasts, (2 main effects) will the model be written as Y=xij + e

You're just fitting a linear regression model, and then testing hypotheses about the coefficients,
so it could in principle be written as a regression equation.  But do you really need to?

> 3) Should significant FStats be significant in 1 or more contrasts, but not
> necessarily in all six?

Here, as elsewhere in statistics, there is no apriori rule about what must be significant.  Any
combination is presumably possible.

Gordon

> Thank you for any help,
> Karen