[R-sig-eco] relative compositional dissimilarity among sites against a continuous variable

[Pekin, Burak K]

Hello Jari,

My objectives are two fold. First, I want to determine the mean dissimiliarity among sites to assess which sites are most 'unique' in terms of their composition, i.e., contain a high number of species that are not found in many other sites. In a sense, I am trying to assess the relative rarity of species found at the sites. Working with a presence/absence data set, doing a bray-curtis or jaccard dissimiliarity should tell me this? 

Second, I want to assess the relationship between the relative 'uniqueness' of sites and other continuous site variables. So, I could create an index out of the mean compositional dissimilarity of the sites and regress it against other variables, e.g., variable X increases as compositional 'uniqueness' decreases. If I understand correctly, while Permanova/adonis tell me which continuous site variables are significantly correlated with the variation in compositional dissimiliarity among sites, they do not tell me the direction of the relationship, right? 

As for your example below, yes all sites within the matrix; 
0011
0011
1100
1100 
are equally 'unique'. So Adonis is probaly not very usefull for my purposes. 

betadisper ( ) is probably more approriate since it will tell me if the 'uniqueness' of a site is significantly greater than the 'uniqueness', or in other words, beta-diversity, of other sites? Then I can then regress the uniqueness/beta-diveristy of the sites against any other environmental variable. I assume this is the best approach for the objectives I listed above. Please let me know if you agree.

Thank you for your insights,
Burak

________________________________________
From: Jari Oksanen [jari.oksanen at oulu.fi]
Sent: Thursday, November 25, 2010 2:35 AM
To: Pekin, Burak K
Cc: 'r-sig-ecology at r-project.org'
Subject: Re: [R-sig-eco] relative compositional dissimilarity among sites against a continuous variable

Burak,

I just wonder how this turned out to being an adonis() case. Your original
idea was quite different: you asked about mean dissimilarities of
observations. Some messages suggested you how to get it. Then you asked
would analysing these mean dissimilarities be the same as NPMANOVA, and got
an answer that you cannot use mean dissimilarities for this, but you could
use adonis(). So what do you actually want to achieve?

Here a simple and extreme example of symmetric dissimilarity matrix of four
observations:
d =
0011
0011
1100
1100

That is, there are two blocks of observations: first two are an identical
set, last two are another identical set (dissimilarities = 0), and these
sets have nothing in common (dissimilarities = 1), and diagonal is = 0. All
these observations have the same row means, but adonis(d ~ c(0,0,1,1)) finds
that these sets are as different as you can get. If you want to analyse the
mean dissimilarities (that are identical above), then vegan function
betadisper() is a tool. In general, adonis() studies differences in
multivariate location, betadisper() differences in multivariate dispersion
(approximately similar to mean dissimilarity). However, you should be
careful in interpreting adonis(): although it is used to study multivariate
location, it is sensitive to other differences in the data, such as
differences in multivariate dispersion.

There are many things you can do in vegan, but that does not mean you should
do them.

We do not know what was your problem, so we have no idea if you had a
meaningful analysis.

Cheers, Jari Oksanen