#Hi
I have two factors (treatment and site) for permanova/adonis2 analysis. Admittedly, I do only have 18 samples for community assessment in a pilot scale study. All diversity indices are good, r.a abundances of targeted genera are all well. When I run adonis2, r2 for interaction is higher and interaction effect is also highly significant. But the model drops the two individual factors. I run, interaction with "margin", it become non-significant. I run the two factors individually, they both significant but their r2 was not as high as when they combined. Pcoa on weighted uni-frac matrix clearly showed vivid and meaningful separation for treatment but not for site/spatial. SO...What do I do? Thank you everyone,
1 Like
Hi @Wossata,
This does sound like a problem. On the face of it your model is telling us that the treatment effect depends on the site, which... isn't great news if those sites are merely independent replicates.
This is expected I believe, margin isn't generally what you want as it will only test your interaction.
This is also expected, since either can pull some of the residuals towards themselves, and since you have a significant interaction, it's almost certainly that exact overlap that either can absorb, giving a better r^2 then when they have to share.
Ok let's see if we can figure out this bit:
One thing that is maybe missing is PCoA is really a large N-dimensional object that we happen to look at the best 3 dimensions of. It is possible that treatement is the larger effect and takes up a main axis. What happens when we exclude that one (aka in our eye-ball-the-coordinates-test, condition on it) and look at the other axes. Does it begin to separate by site at that point?
Plausibly we might expect the PCoA to choose something like the community substructure of treatement + treatment:site as Axis 1. And then some unrelated thing as Axis 2 and perhaps site as Axis 3. Your scree plot (the barplot of percent variation explained) would also give us some perspective on how many axes are meaningful.
2 Likes
Oh one last thought:
Compare this formula: treatment + site + treatment:site to treatement:site + treatment + site. There should be a difference because adonis isn't a Type III test, so we have an accumulation of residuals as you go through the terms. Does the second form of this formula result in no-longer significant main-effects?
Thank you Evan, as always. I have tried all combination as some people suggested exchanging the position in the model might have some impact. I will try again. When I said, meaningful clustering, I might have overstated 
but, its a soil fumigation small trial- the hypothesis appeared to be holding alright until adonis2 showed these results. There is no way f choosing one from the other right? running individually or considering the interaction? happy to share the r version of pcoa
I mean you can leave the interaction out, but I don't think that really helps the situation, since the natural question becomes, well why is site significant?.
It could be that beta diversity just isn't the tool you need. Could you elaborate a bit more on the study design? Do you have longitudinal data?
You might also have a peek at this plugin: longitudinal - Microbiome marker gene analysis with QIIME 2