PERMANOVA - choice of distance measure

[Trond]

Hi,
I want to run a (PRIMER based) PERMANOVA analysis on my univariate data (output var.= weight of organisms, and 3 factors = 2 fixed, 1 random) and I'm wondering which distance measure (Bray-Curtis or Euclidean distance) to use to create the resemblance matrix. In most literature that I found Euclidean distance is used for univariate analysis (thereby getting an F-value similar to the F obtained by a normal ANOVA) and Bray-Curtis for multivariate analysis. Now I'm wondering if this is how it is supposed to be done, or if there is a more general rule how to determine which distance measure to use, especially as there are additional possibilities in the PERMANOVA menu...

Thanks a lot for your help!

 

[bugman]

The million dollar question.

In Euclidean space, univariate permanovas are equiviant to your family of linear models. The choicce of distant measure is not easy and in the end quite subjective. But given your data are univariate, I am a wee bit unsure why you are using PERMANOVA and not a different model. At a very general scale though, euclidean distances are usually used for environmental data - when zeros actually are included (i.e. temperature at site 1 and 2 are zero and are therefore simialr) whearas BC ignores zeros, which is why it is used for abundance data (i.e. zeros at site 1 and 2 do not neccessarily mean simailrity because they might be absent for different environmental reasons, or lower sampling effort...).

 

[Trond]

Hi Bugman,
thanks for your reply, this definitely helps! and it would mean that I probably should go with Bray-Curtis since I'm analysing information about the abundance of species under certain factors. However, just to get it right: there is no general problem with using Bray-Curtis in a univariate analysis?

 

[bugman]

No, not as such but your F-ratio will be based on simulations (psuedo -F). If it were my data I'd be looking at a mixed model ANOVA.

Your p-values will be based on distances between sites or treatments not the actual weights, but I guess it depends on your question.

 

[Trond]

Thanks again, that's good to know
I will have a look at the mixed-model ANOVA again. We had been considering this option as well...
Thanks for your help!