[R-sig-eco] Choice of constrained ordination method on complex data

[Tim Richter-Heitmann]

Good Morning,

i need to once more ask you for advice.
My dataset includes 360 samples, 60 samples per date, which are randomly 
distributed among a square grid. I measured bacterial abundance as my 
outcome.
My final aim is to calculate the variance partitioning for a number of 
classes of constraints, including spatial (via PCNM or cubic polynomes) 
and temporal components.

Here is what i tried so far in order to visualize the data, for this 
example without taking into account any spatial or temporal variance 
(for the sake of the example). They key point of this dataset is the 
presence of few dramatically different plots. CCA does a good job in 
showing those, but fail to separate any other samples (due to to 
presence of basically two types of plots with many "0" introduced in the 
respective other type). RDA and capscale do not find these differences, 
but find better gradients in the general dataset.

The question is if there are better options to ordinate my data, 
especially with my main aim in mind (variance partitioning).

Here is what i have done:

1. CCA on non-transformed (but relative) species and environmental data
2. RDA on hellinger-transformed total species counts and standardized 
environmental data
3. capscale on hellinger-transformed total species counts and 
standardized environmental data

Capscale was done on the gower metric, as rankindex from vegan showed 
its far superiority to bray-curtis and a slight one compared to 
euclidian (R=0.15 > 0.12 >>> 0.02).

Here are the results from the R-output:
http://s15.postimg.org/r1bq7863v/Untitled.png
Here are the procrustes rotation between all three ordinations:
http://s13.postimg.org/ainxqsljr/Rplot.png

These are just basic ordinations without forward selection or variance 
inflation.

The presence of a few dramatically shifted sites is shown in CCA, but 
not in RDA or CAP, but the latter two explain more of the general 
variance in the dataset. The big question is if i am doing it wrong on 
any level, or if there are better ways to visualize the data. NMDS with 
envfit is similar to CCA, by the way.

Here are some more questions:
1) Can i apply vegan's toolbox for RDA also to objects created by 
capscale (including goodness.cca)?
2) Can i apply forward selection of "packfor" to vegan's 
capscale-derived objects?
3) Are spatial coordinates (PCNM or cubic polynomes) suitable for 
capscale, or shall i go back to RDA?
4) Connected to that, vegan's Varpart function only works with RDA, not 
with capscale?

5) What could be an explanation why bray-curtis distances of my species 
data show no correlation (0.02) at all with the predictor variables, but 
a somewhat stronger correlation with euclidian (0.12) or gower (0.15) 
distances, as calculated with rankindex()? I am surprised by it, as i 
have fairly typical species data, for which bray-curtis was designed.

Thank you very much. I am not so sure what directions i should take at 
this point.

Cheers,


-- 
Tim Richter-Heitmann (M.Sc.)
PhD Candidate



International Max-Planck Research School for Marine Microbiology
University of Bremen
Microbial Ecophysiology Group (AG Friedrich)
FB02 - Biologie/Chemie
Leobener Straße (NW2 A2130)
D-28359 Bremen
Tel.: 0049(0)421 218-63062
Fax: 0049(0)421 218-63069