[R-sig-eco] nmds in 3D

[Jari Oksanen]

On 19/01/10 22:50 PM, "Roman Luštrik" <roman.lustrik at gmail.com> wrote:

> Dear list,
> 
> I have produced an nMDS plot with stress of 2.5 for first two dimensions. It
> is my understanding this is _very_ poor support.
Roman,

It sounds like you have a very small data set, because stress of 2.5 is
exceptionally low, almost suspicious. In general, stress below 10 is very
low and indicates a good fit (or "support").

Stresses of the magnitude of 0.006 is pathologically low, and surely are
overfit[ted]. In no case should you use 3D solution.

If you can reach stresses this low (like 2.5),  you should seriously
consider whether it is at all appropriate to use nonmetric analysis, or
whether your data set is so multidimensional that a multivariate analysis is
needed. How many points do you have?

Please remember that isoMDS function (that you probably use) gives stress as
per cent, so that random data typically have 40% stress, and anything below
10% is very good.

> When I run the procedure in
> three dimensions (k=3), the stress value drops into the bucket and lands at
> 0.006. I have compared both plots and judging by feel alone I would say the
> 3D interpretation doesn't offer much more information than 2D (but see for
> yourself: 
> 2D<http://i53.photobucket.com/albums/g67/romunov/trash/nmds_3d.jpg>and
> 3D <http://i53.photobucket.com/albums/g67/romunov/trash/nmds_3d.jpg>). Would
> it be sensible to use 3D plots for interpretation instead of the "common" 2D
> based on stress value alone?
> 
Most data sets have 2.5 dimensions (Kruskal).

Cheers, Jari Oksanen