Hi,

I am using Limma to analyze my data, which is a simple two-group comparison (treated vs. control). The analysis is very straigtforward, but it becomes not as easy when the gene list is to be generated. I noticed that some genes with small P values (<0.05 adjusted with FDR) have negative B logodd scores, which mean the chance of DE is <50%. For situation like this, should I use log odd for cutoff or should I use P value or combined?

Xiwei Wu

Dear Xiwei,

I wouldn't use the B statistic for cutoff because it depends on the prior probability that genes are DE, and this prior probability is by default just set to an arbitrary values (proportion=0.01).

So we usually use the FDR values to choose a cutoff.

In principle, you could use the propTrueHyp() function to estimate the prior probability that genes are DE, input this to eBayes() and then the B-statistic would be a meaningful cutoff. Usually though we prefer the FDR because it has a clear interpretation and doesn't require estimation of the prior probability.