Hey,
I am using Deseq2 in conjuncture with Phyloseq to check for deferentially abundant OTUs along a natural gradient. My sample come from different habitats and I know that the community composition differs between the habitats. I am interested in which OTUs co-vary with the continuous predictor in each habitat (4 levels). I tried the following model:
~ Habitat + continuous_predictor + Habitat:continuous_predictor
followed by results ( ... name = Habitat_1.continuous_predictor)
. Yet as is stated in Eaxmple 3 in the ?Results documentation:
"this tests if the condition effect is different in III compared to I". I don't think that this is what I want. I am not interested if the relationship (slope) of, say, OTU_1 with the continuous predictor differs in Habitat_2 compared to the reference habitat 1. I am interested which OTUs co-vary with the continuous predictor - conditional on habitat.
Therefor I think it would be better to split the data by habitat and run 4 models independently. Would you agree? I was thinking that turning of independent filtering, recording all the p-values, and then perform p-value adjustment with independent filtering on the joint pvalues. Would that make sense? Thanks a lot for your support!
Fabian
thanks for coming back to me!
would your model not only account for the main effect of habitat (i.e. different mean abundance?). I hypothesise that any OTU that covaries with the continuous predictor will only do so in some habitats but not in others (where it might not even be present). Sorry if I am missing the obvious.
Fabian
Oh I didn't understand the question at first. You should use a design of ~Habitat + Habitat:continuous_predictor, which will give a different slope for each habitat.