[R-sig-eco] Envfit function

[Jari Oksanen]

Stephen,

On 16/09/11 04:12 AM, "Stephen LeDuc" <sdleduc at gmail.com> wrote:

> Dear R-sig-ecology list members,
> 
> I'm using the envfit function in the vegan package to fit environmental
> variables as vectors onto an NMDS orientation.  If there are large scale
> differences between variables, do I need to normalize the data first?

No.

> And,
> as a somewhat related matter, do I need to worry about the distribution of
> the data?  Any help or advice on these questions would be most appreciated!

That depends on your inclination, and what are the things that get you
worried. If is is about "significances", then many people say that you need
not get worried because envfit uses permutation tests. However, if your
variable is skewed, or has some extreme values, those exceptional values
will be influential both for the assessment of the significance, and in
particular for direction of your fitted vectors. There are no influence
statistics available for envfit, but you must use your judgment. Visual
inspection for each variable is possible by using ordisurf(..., knots=1,
bubble=4) which plots linear trend surface where the isoclines are
equidistant lines perpendicular to the envfit vectors (that you can have in
the same graph), and data values are shown by point sizes. If you save the
ordisurf result, you can get some diagnostic statistics from that.

Here a (warning) example using Bryce Canyon data from the labdsv package:

library(vegan)
library(labdsv)
data(bryceveg)
data(brycesite)
m <- metaMDS(bryceveg)
ordisurf(m ~ slope, brycesite, knots=1, bubble=4)
plot(envfit(m ~ slope, brycesite))

Cheers, Jari Oksanen