Hi,

I read some previous posts and learned that it is statistically not really valid to filter results based on fold change (padj<0.05), for example only DEGs above/below 1.5 as this "flavours" lowly-expressed genes. Instead we should use lfc argument if we want to filter for fold change. I also learned that lfcShrink can be used to try and correct unreliable fold changes. Now my question: When I use lfcShrink is it then justified to filter by fold change, for example padj < 0.05 and abs(FC)>1.5, with no lfc threshold? I wonder about this because I have many datasets, and find it difficult to decide for lfc threshold, because some sets have many replicates, and so high power, others only have n=2, so low power. It would be much easier, and give me more flexibilty to just filter by fold change, is this somehow ok with lfcShrink and no lfc argument?

The idea about using lfcThreshold is that it incorporates the LFC into the testing procedure (whether the p-values in results or the s-values in lfcShrink).

I would say, if you want to find sets of genes that reject the point null (LFC=0), _and_, separately, have evidence of large effect (abs. posterior LFC > 1.5), then that is fine. I personally prefer the former procedure because the inference is connected to the relevant effect size, but in the end the important thing is if you know how to interpret the set of genes you are looking at.