Using the BH method for FDR control

Question

I have preformed multiple tests and now have a list of all the calculated P-Values (~30K^2 values). I have a reason to believe my tests are somehow dependent.

I would like to analyze the results and correct the p-values.

Unfortunately, I only have very basic knowledge in statistics. From the little I managed to gather, I think the Benjamini Hochberg method for FDR control should be suitable for my needs.

My questions are:

1. Am I right?

2. If so, I'd like to implement a code that preforms this calculation. Is the following code correct: http://www.augix.com/wiki/index.php/Benjamini_and_Hochberg_FDR_control and does it calculate the cutoff for my case?

Thanks!!!

Answer 1

BH is reasonable in the sense it takes p-values as input for the procedure. However, it will guarantee FDR control only for certain kinds of dependence. Namely, positive regression dependence on a subset (PRDS). It is unclear from your question if this condition holds. If you believe the variables are pairwise positively correlated, it will do. If you suspect there might be some negative correlations, BH is not the way to go. At the cost of some power, you can do the more general test you mentioned in your comment. If you are using R, then the function p.adjust is all you need. Look for the BY method.