I'm analysing an RNA-Seq dataset with the following study design: Three locations (e.g. A, B, C) at two different times (e.g. time 1 and 2), everything in triplicates. I'm interested in finding DE genes within each time point: A1 vs B1, A1 vs C1, B1 vs C1, and similarly for the second time point.
I set limma up for doing all the pairwise combinations between location-time combinations (~0+Group and makeContrast). I can then extract both t-tests for each pairwise comparison with topTable(coef="A1-B1") or F-tests for all pairwise comparisons using topTable(coef=c("A1-B1", "A1-C1", "B1-C1")), etc.
I'm now considering if it would be appropriate to use some of the more advanced options for correcting for multiple testing in decideTests in a scenario like this where three closely related contrasts are analyzed. Would the nestedF method be appropriate and how would one handle having to different F-tests in a single setup (all comparisons within timepoint 1 and all comparisons within timepoint 2) as only one F.p.value is stored in an MArrayLM-object?
Another option would be to use the recent stageR package (https://genomebiology.biomedcentral.com/articles/10.1186/s13059-017-1277-0). In the discussion section of the paper they discuss a similar study setup:
"For example, a DGE study that compares three drugs (e.g. a new drug, the current state of the art and a placebo) would require exactly the same data analysis paradigm as the Hammer dataset: three different hypotheses of interest (mean differential expression between the drugs) and, according to Shaffer’s modified sequentially rejective Bonferroni (MSRB) procedure, no correction is needed in stage II for FWER control."
Would an appropriate implementation of this be to extract F-test p-values using topTable to use as the screening-tests and use individual t-test p-values for confirmation tests? Would a stage II correction be necessary in this case?
For the second option, you can achieve it by using
method = "global"
in thedecideTests
function, right? (Provided that he/she used the factorized contrast matrix).Thanks for the detailed comment! I don't think the the two-stage approach you suggest would be entirely appropriate here, as the two timepoints have a very different number of pairwise differences (more than twice as many difference between sites on the segment time point).
Would it still be correct to analyse the data using two stageR analyses, where the approach is: 1) Extract F-test p-values for each time point 2) adjust t-test p-values using stageWiseAdjustment with method="none"?
Yes, two stageR analyses would be appropriate in that case, and you are correct that according to Shaffer's procedure, you are allowed to not adopt a FWER correction in the confirmation stage, so indeed
method="none"
will be valid here.