Hello, 

I am using limma to compute the differently express genes between two conditions from microarray.

The dataset, a GSE from GEO contains 38 arrays of senescent / quiescent and growing samples, I am only interested in senescent vs growing DE genes. 

So far I have been removing the extra array with the code below, but I also recently came across some post about keeping all arrays in regards to the "level of freedom". While keeping all arrays I am getting completely different results, wondering which is the of the methods is the most appropriate.

idx <- which(cell_status == "Growing" | cell_status == "Senescent")
gse <- gse[,idx]

pheno <- droplevels(pData(gse))

mod <- model.matrix(~cell_status+ 0, data=pheno) 
colnames(mod) <- levels(pheno$cell_status)
contrast_mat <- makeContrasts(Senescent-Growing, levels=mod)

fit = lmFit(gse, mod)
fit2 <- contrasts.fit(fit, contrast_mat)
fit2 <- eBayes(fit2)

topTable(fit2, adjust="fdr", sort.by="B", number=Inf)

Generally, if you include more samples, they'll provide more residual degrees of freedom in the model to estimate the variance, even if they're not related to the contrast of interest. This leads to more precise estimates and greater power to detect significant differences. It would help if you were more specific about how your results are "completely different" with and without these extra samples, but I'd naturally expect an increase in the number of putative DE genes when you leave them in.

P.S. You don't mention what the accession number is, but I'm guessing it's GSE19864. It would seem unwise to use only the growing/senescent status of the cells to formulate your design matrix, given that there's a whole bunch of shRNA treatments happening at the same time.