I am trying to understand some results from DESeq2. The experiment is as follows: Before and after treatment for 11 individuals. The most significant gene has the following counts: (Before treatment, after treatment) (0,0) (1298,0) (0,0) (0,0) (0,0) (0,0) (0,0) (0,0) (0,0) (0,0)(0,0) (so all individuals have zero counts only, except from the second individual, who has 1298 counts before treatment). The p-value from DESeq2 is: 1.49e-21 (using the default Wald test). Here I have used the DESeq vignette instructions for doing paired comparisons, putting individual (a factor variable) and treatment (before/after) (factor variable) in the colData and specified design = ~ Individual+Treatment. The estimated dispersion is 22. With such a large dispersion estimate it is counterintuitive to me that the p-value becomes so small. When I use the likelihood ratio test (LRT) the gene does not become significant (p=0.41), which makes much more sense to me.
