FDR in RNA-Seq with Multiple Pairwise Tests

[NateP]

Okay, I guess this is more of a statistics question than a program question, but I just want to clarify something dealing with the correct usage of FDR with multiple tests.

General Experiment Set:
3 Biological Replicates of samples collected from 4 time points (t1, t2, t3, t4).

The question is how is t1 different from each of t2, t3, and t4 individually.

So using DESeq, I basically did 3 pairwise tests: t1 vs t2, t1 vs t3, t1 vs t4. And of course, DESeq gives FDR values for each of those test pairs.

However, since I am technically doing 3 tests on the same data set, is the appopriate statistical approach to carry out an FDR correction on all of the tests as a whole?

In other works, take the raw p-values from all 3 tests, and combine them in a single list and then FDR correct (like using p.adjust, method="BH"), and afterwords break the whole set back into the original 3 tests again.

Or is just using the FDR values generated from the 3 pairwise tests separately statistically valid?

[mbblack]

The three time point comparisons are done independently, not simultaneously. You are comparing three hypotheses - that each time point, independently, is different from the controls. So the only multiplicity you need correct for is the simultaneous pairwise tests you do each time you test one of those three hyptotheses.

It would be analogous to performing orthogonal linear contrasts in an ANOVA, where you would correct each linear contrast for multiple tests independently of each other.