[R-sig-eco] orthogonal variables aliased by rda(vegan)

[Andrea Bertolo]

Dear Jari,

this is exactly the structure of my data. 
I in fact forgot to mention that I was using the rda fonction as a
regression here (thanks to Karl Cottenie to let me see it...), i.e. to
explain the abundance of a single fish species with 14 MEMs. 

There is only one thing that I don't understant. 
What do you mean exactly when you say that "these extra constraints are
taken into account, but not reported in all cases (the aliasing is
explicit: there is code to turn down displaying the results that were
calculated)." ? 
Does this mean that the rda is calculated with all the 14 contraints (in
this case), but only the first constraint is plotted ? I ran the rda with
the only "un-aliased" MEM and I obtained the same results, showing that the
aliased constraints are effectively eliminated. Also, If I remeber well,
when I run the same rda with Canoco I can keep the all 14 constraints in
the analysis, getting a different result. Are these differences only a
graphic matter or there are more substantial differeces ?

many thanks  
andrea




At 20:51 2008-06-04 +0300, you wrote:
>Quoting Andrea Bertolo <andrea.bertolo at uqtr.ca>:
>
>> Hi to all,
>>
>> I am running some RDAs on Vegan using Moran eigenvector Maps (MEM) to
>> explain the variation in the abundance of a fish species.
>>
>> In several occasions I noticed that, despite MEM are othogonal by
>> definition, they were aliased by the rda fonction and thus not taken into
>> account in the analysis, as if they were collinear !
>>
>> I cannot understand why, since the MEMs used in the analysis showed not
>> only to be orthogonal (I post-verified it by calculating a VIF, in case
>> of), but also to add each a significant fraction to the explained variation
>> of the dependent variable (MEMs were selected by the forward.sel fonction
>> available in the packfor package).
>>
>> Could anybody tell me if I am doing something wrong or there is a real bug
>> here ?
>
>I'll try.
>
>This is something that really may be a problem in vegan, and may  
>change in future versions. The problem here is (probably) that the  
>number of your PCNMs is higher than min(nrow(x), ncol(y)) in your  
>dependent data. There really are two kind of ranks in rda/cca etc. in  
>vegan: the rank of the dependent community data, and the rank of the  
>constraints. It seems that in your case of spatial filtering the rank  
>of the MEM's is higher than the rank of the dependent matrix. The  
>current practice in vegan is the take the rank of the solution as the  
>min of these two ranks, and the extra constraints are aliased.  
>However, these extra constraints are taken into account, but not  
>reported in all cases (the aliasing is explicit: there is code to turn  
>down displaying the results that were calculated). This will change in  
>the future, but I'm not yet quite sure how this should be changed --  
>this needs research, and discussion among developers and other  
>interested parties. However, the main results will not change. Things  
>that will change are the display of the same results, and some derived  
>functions (predict, calibrate.cca and perhaps some others, but I don't  
>know yet which). This is a change that has wide reaching effects, and  
>therefore you cannot expect this before autumn -- or "fall" on that  
>side of the pond).
>
>Vegan is not yet ready for the full blown spatial analysis, but we are  
>working for that (and here "we" is really plural and implies several  
>persons). If you need detailed advice on the possible consequences of  
>the upcoming change, you may directly contact me or us at the  
>http://vegan.r-forge.r-project.org/.
>
>Best wishes, Jari Oksanen
>
>_______________________________________________
>R-sig-ecology mailing list
>R-sig-ecology at r-project.org
>https://stat.ethz.ch/mailman/listinfo/r-sig-ecology
>