[R] Problem with metaMDS in vegan

[Jari Oksanen]

Fabian B <fabian.boetzl <at> freenet.de> writes:

> 
> Thanks for the fast response!
> 
> My data is actually just the percentages of the substances found in the
> cuticular layer of the flies. The quantitative and qualitative composition
> of this layer differs in gender, locality and so on. 
> My 70 „species“ are the chemical compounds we identified and my 48 „sites“
> are the individuals (location and gender are encrypted in the individuals'
> recognition code). Therefore, I have the percentages of one chemical
> compound over all tested individuals in one row (as numeric data). You could
> say that these are the abundances of the species (the compounds) at the
> different sites (individuals). So my data is quite similar to community
> structure. I have done this before and it worked just fine. I used:
> 
> metaMDS(data, distance=“bray“, k=2, trymax=10000, autotransform=F)
> 
> As I mentioned before, this time the stress is very low (around 9.5 * 10^-5)
> and R gives me that warning message. Though I have always had data in the
> same structure, I've never been in this situation. The NMDS just stops after
> the first run, because of the low stress. But there are obviously some
> differences in my data – I can see them with my eye. If I discard the
> 'distance=”bray”', nothing changes.
> May it be possible that these are simply just to little to be recognised
> correctly? How could I change that?
> 
> If I add 'maxit=10', I get quite what I assumed. Two completely separated
> clusters with cloudlike appearance (stress around 0.02). The clusters are
> the two sexes and in each cluster, the localities group together. But I'm
> not quite sure, if I may set a maxit-level that low – or if I just create
> some false relations with that. 
> 
> I also recognised , that the stress gets lower, the higher I set 'maxit' and
> the lower I set 'trymax'. 
> 
> What could be going wrong? I really have no clue … 
> 

I think nothing is going wrong: you just have that kind of data, and
there is nothing wrong there.

There is nothing to reproduce here, but it *seems* that you may have
a disjoint data set with two clusters sharing nothing between them.
The metaMDS is obsessed with that difference and does its best to
separate these clusters. If you set 'maxit=10', it does not have time
to finish the job, and you get two clouds. Just let it finish, and the
clouds are squeezed to singularities. 

If you want to see the structure within these singularities, you
could analyses these groups of points separately. Then the major
difference between clouds won't obscure you.

I am not quite convinced that Bray-Curtis (distance='bray') is the
best choice of dissimilarity measure for these kind of data. If you
have scaled values (percentages), choices like 'distance = "manhattan"'
or 'distance = "euclidean"' or 'distance = "gower"' could work better.

As we have no reproducible example, this is all guessing.

Cheers, Jari Oksanen