Question about mvabund results

Dne 03. 10. 24 v 23:00 Jon Benstead napsal(a):

Dear list ? My graduate student and I are using mvabund to analyze community data from an ecosystem-level manipulation experiment. The interaction term of interest (stream x period) is highly significant in the global multivariate test, but few taxa appear to be driving the difference, as shown by the univariate tests. I realize that this sort of result is to be expected at times, as the multivariate test is more powerful than each univariate test, but does anyone have general advice about follow-up analyses to shed more light on the nature of the significant result? Any suggestions would be greatly appreciated.
Yours, Jon.

	[[alternative HTML version deleted]]

_______________________________________________
R-sig-ecology mailing list
R-sig-ecology at r-project.org
https://stat.ethz.ch/mailman/listinfo/r-sig-ecology

Dear Jon,

as usual, the correct answer depends on what you would like to show :-)

The multivariate test result itself basically says: "There is (quite 
probably) something going on." Nothing more, but that is enough for you 
to plot the differences and to be not accused of hallucinating the 
differences. You may stop there and let the readers interpret the 
(probably real) changes/differences themselves.

(If you think that the result of the overall test is driven by the 
species with "positive" univariate tests - you may try to leave the 
species out or to replace their values with some random/permuted stuff, 
but be careful: see e.g. David Zeleny's, Stephane Dray and Pedro 
Peres-Neto's papers about that)

Depending on the research question, you might follow up with plotting 
(showing, analyzing) various things:
1. Species with the highest relative change, or just showing the change 
in each species.
2. Change in some structural parameter, such as evenness
3. Change in the species abundance grouped by guilds/trophic 
levels/endangerment levels,...
4. ...

Hypotheses resulting from the point 3 are testable in the multivariate 
framework: you just add them to the model. (Beware of statistical 
fishing - pun intended).

HTH (and I hope someone corrects my naive views),

Martin

Please expect long response time (3d+) - this is because of family reasons