Hi,

I am using limma to get differentially methylated probes from MethylEPIC data, using M values. I have 9 groups of samples (from groupA to group I) within 3 arrays, containing 2 to 5 samples each.

I get very different results depending on the number of comparison I perform ie the number of contrasts in my contrast matrix.

For example I have a number of significant differential probes much lower in the groupD-groupC comparison when using :

Group2 <- factor(pheno$Sample_Group2) design2 <- model.matrix(~0+Group2, data=targets) colnames(design2) <- levels(Group2) fit2a <- lmFit(M_flt, design2) x2<-c("groupA-groupB", "groupA-groupC", "groupD-groupC", "groupE-groupC") contMatrix2 <- makeContrasts(contrasts=x2, levels=design2) fit2b <- contrasts.fit(fit2a, contMatrix2) fit2b <- eBayes(fit2b) summ_group2 <- summary(decideTests(fit2b, adjust.method="BH", p.value=0.05))

than when I add several other comparisons:

Group2 <- factor(pheno$Sample_Group2) design2 <- model.matrix(~0+Group2, data=targets) colnames(design2) <- levels(Group2) fit2a <- lmFit(M_flt, design2) x2<-c("groupA-groupB", "groupA-groupC", "groupD-groupC", "groupE-groupC", "groupD-groupE", "groupF-groupG", "groupH-groupI") contMatrix2 <- makeContrasts(contrasts=x2, levels=design2) fit2b <- contrasts.fit(fit2a, contMatrix2) fit2b <- eBayes(fit2b) summ_group2 <- summary(decideTests(fit2b, adjust.method="BH", p.value=0.05))

I'm moving from 100000 differentially methylated probes in case 1 to 200000 in case 2! With different logFC values as well. Moreover some probes in the 100000 are not in the 200000 and the top ten most significant are not the same..

Is it correct to use all my data in the design matrix and the linear model if I perform comparisons on a subset of group only as in case 1? Would it be more powerful to remove the unused data before performing the linear model?

If I keep all data in my design matrix and linear model, which method is the most accurate? I don't know if i'm underestimating significance in case 1 or overestimating it in case 2..

Hope you can help, sorry if the question has already been asked by newbies like me!

That's good to hear. For future reference, reply to answers with "add comment" rather than "add answer".