I think this question applies both to DEXSeq and DRIMSeq.
I have RNA-Seq data across many tissues, and would like to get DEU and/or DTU between these tissues. Most genes are not expressed in all these tissues, so when I run the standard analysis, I get statistical significance due to small changes in read counts of non-expressed genes.
I have already generated a table of genes that are confidently expressed in each tissue, so I'm tempted to use this table to subset the genes, so that each subset is expressed in the same tissues, subset the design matrix accordingly, then run the LRT for each subset, and finally reassemble these subsets and run the p value adjustment.
Would this method be statistically sound? Is there a better approach?