[BioC] edgeR: GLM & residuals and model fitting & hypothesis testing

[Gordon K Smyth]

Dear Susanne,

1) You may know that there are many different definitions of "residuals" 
from a generalized linear models, and none of them have the properties of 
residuals from normal linear models.  I have not found residuals very 
useful myself in a genomic differential expression context.

What sort of residuals did you want, and what were you planning to do with 
them?

2) You must always fit the model with all relevant factors: 
individual*treatment.  You cannot do meaningful inference from a reduced 
model until you have determined that the other factors are not important. 
Hence you have to deal with the interaction first.

Best wishes
Godon

> Date: Tue, 14 Feb 2012 17:44:31 +0100
> From: Susanne Franssen <s.franssen at uni-muenster.de>
> To: bioconductor at r-project.org
> Subject: [BioC] edgeR: GLM & residuals and model fitting & hypothesis
> 	testing
>
> Dear all,
>
>
> 1) GLM & residuals:
>
> I have a question concerning the use of GLMs in edgeR and the analysis
> of the residuals after model fitting.
>
> I have followed all the steps until model fitting, e.g.:
> glmfit.D <- glmFit(D, design, dispersion = D$tagwise.dispersion)
>
> The results I obtain from the fitting are the following catgories:
>> names(glmfit.D)
> [1] "coefficients"  "fitted.values" "fail"          "not.converged"
> [5] "deviance"      "df.residual"   "abundance"     "design"
> [9] "offset"        "dispersion"    "method"        "counts"
> [13] "samples"
>
> What would be the best way to obtain the residuals for the "genewise" GLMs?
>
>
>
> 2) model fitting & hypothesis testing:
>
> I have a fully crossed design with 2 factors and 2 factor levels each:
> individual <- as.factor(c("indA","indA","indB","indB"))
> treatment <- as.factor(c("treat1","treat2","treat1","treat2"))
>
> in general I would be interested in 3 different aspects:
> a) effect of individual
> b) effect of treatment
> c) interaction between individual and treatment
>
> What would be the best way to test for those effects, would I rather
> test for all three aspects individually, i.e.:
> a) design <- model.matrix(~individual)
> b) design <- model.matrix(~treatment)
> c) design <- model.matrix(~individual*treatment)
>
> or doesn't it also make sense to model
> design <- model.matrix(~individual+treatment)
> and test for
> a) lrt.cd_ind <- glmLRT(D, glmfit.D, coef=2)
> b) lrt.cd_treat <- glmLRT(D, glmfit.D, coef=3)
> ... this way the effect of both factors could be accounted for in the model?!
>
>
> Thanks a lot!
> Susanne
>

______________________________________________________________________
The information in this email is confidential and intend...{{dropped:4}}