Skip to content

Capscale: Centroids ignored with Condition() in formula

4 messages · Wilson, Michael E, Jari Oksanen

Original
Jari Oksanen
# 
Michael,

Sorry for top posting, but with this structure of email it seems to be the
only sensible way of continuing.

I cannot see any problem, but I only see a situation, and quite a normal
situation. You have a model of type capscale(Y ~ A + Condition(A %in% B)).
Here A appears twice: its effectsare removed first in Condition() and then
you try analyse its residul effects, but no effect of A remains because they
were already removed. In this case, no constrained component is calculated
and you won't get explicit information on skipping this stage. Situation is
different if some constrained variation is left: in capscale(Y ~ A + x +
Condition(A %in% B)) you will get a constrained component showing effects of
x on residual and information that A was aliased. So no effect of A, but
info that it cannot be calculated. All that we could do for the first case
is to return an empty constrained component with information why it is
empty. 

The same situation continues with partially aliased models. For instance, in
vegan standard example capscale(dune ~ Management + Condition(Manure)), the
effects of Management level NM (Natural Management) cannot be displayed
because you have partialled out Manure, and Manure level = 0 (no manure)
only occurs with Management = NM. So you alias away ManagementNM. This seems
to be the case with your later message update below.

If you remove something by partialling out, then you remove it. If you
removed it, you cannot analyse its effects. That is how these models work
and how they were designed to work.

Cheers, Jari Oksanen
On 17/12/10 03:27 AM, "Wilson, Michael E" <mwilso14 at utk.edu> wrote:

            

      
      
      
    

        

      
      
    

  
  


  


      

    

    

      
      

        


  

        

      Wilson, Michael E
    

    

      
      #
    

  


  

      
Thanks Jari,

I've still much to learn. I'm working on my master's thesis.

Within a context of mixed models, I thought that a model where Y ~ A + Random(B within A)
would only remove random variance explained by different levels of B in each level of A,
leaving A to be analyzed as fixed without the variance created by different levels of B within it.

I guess that is either wrong, not completely correct, or in ordination it functions differently.

After your reply, I tried conditioning on locations (different farms) of the transects, and that seems to work well.

Thanks much

Your tutorial is very helpful as well.

Michael Wilson.
University of Tennessee

        

      
      
    

  
  


  


      

    

    

      
      

        


  

        

      Jari Oksanen
    

    

      
      #
    

  


  

      
On 17/12/10 17:30 PM, "Wilson, Michael E" <mwilso14 at utk.edu> wrote:

        

            

      
      
      
    

        
Michael,

You're correct: Y ~ A + Random(B within A) would work like you expect, but
such a model doesn't exist for capscale (or rda or cca). These ordination
models only have fixed effects, and Condition is *not* Random. If you want
to have anything like random effects then it is up to designing a
permutation test that would mimic random clustering. Restricted clustering
can achieve something resembling random effects models in significance
tests. Currently the only thing we have is the 'strata' argument in
permutations, but Gavin Simpson is working with more powerful permutation
routines.

Cheers, Jari