[BioC] A question about paired data and how to set up the designmatrix in LIMMA

[Gordon Smyth]

>Date: Sun, 17 Jul 2005 08:36:01 -0700
>From: Robert Gentleman <rgentlem at fhcrc.org>
>Subject: Re: [BioC] A question about paired data and how to set up the
>         designmatrix in LIMMA
>To: Johan Lindberg <johanl at biotech.kth.se>
>Cc: bioconductor at stat.math.ethz.ch
>
>Hi Johan,
>    The simple answer is that your data do not quite fit the paired
>t-test model. You probably want some form of random effects model as can
>be fit by lme (nlme pacakge) or lmer (Matrix package). Where patients
>are treated as a random effect and before/after as a fixed effect. I do
>not believe that limma fits these models, although it may be extended at
>some time.

Actually it does, when there is only one random factor. It does this 
through the 'block' argument to duplicateCorrelation() and lmFit(). The 
actual fitting of the mixed models is done by a call to the 
randomizedBlock() function in the statmod package. Limma uses a particular 
approach to Bayes-type moderation of the random effects across genes. See 
my reply to Johan Lindberg.

Gordon

>  And neither of those pieces of software (lme or lmer) fits an
>empirical Bayes model to lots of arrays although they too could be
>extended. Hence, there is no easy solution. Nor is fitting and
>interpreting mixed effects models simple - if you have little
>statistical experience, the best advice is to find someone with a lot
>more who can help, as the analysis is not likely to be simple or
>straightforward.
>
>You can always take the expedient of discarding some of the data, so
>that you are reduced to one before and one after sample per patient.
>Then standard paired t-tests can be used. But you have lost data.
>
>   Best wishes,
>     Robert