[BioC] differential expression, model selection and F-tests of nested models

[Ramon Diaz-Uriarte]

Dear All,

When we use limma, we apply the same linear model to all genes. However, at 
least some times, we can think of different genes requiring slightly 
different models (e.g., geneA ~ class + age, geneC ~ class).

When having a set of prespecified, and nested, set of models, and only one 
dependent var., I'd use F-tests to compare among models. Sure, I can still do 
the same for thousands of genes, but I'd be working gene-by-gene (i.e., not 
taking advantage of the ebayes approaches available in limma) and incurring 
in massive multiple testing and sequential testing problems. Has anybody 
thought about a more principled and rigorous framework? I haven't seen this 
issue dealt with in lists or the literature.

Best,

R.

P.S. I am not looking for any stepwise-like solution. In the particular case I 
am working on now, we are considering three pre-specified models (age*trt, 
age + trt, trt). And, also in this case, my collaborators do not like at all 
the idea of fitting an age term when it is obviously not needed for at least 
some genes that differ strongly in non-age-adjusted intercept.



-- 
Ramón Díaz-Uriarte
Bioinformatics Unit
Centro Nacional de Investigaciones Oncológicas (CNIO)
(Spanish National Cancer Center)
Melchor Fernández Almagro, 3
28029 Madrid (Spain)
Fax: +-34-91-224-6972
Phone: +-34-91-224-6900

http://bioinfo.cnio.es/~rdiaz
PGP KeyID: 0xE89B3462
(http://bioinfo.cnio.es/~rdiaz/0xE89B3462.asc)