Skip to content
Prev 2242 / 7420 Next

envfit and adonis restricted permutations

On Mon, 2011-06-27 at 16:39 -0700, Steve Leonard wrote:
Yep; Notice that the r2 values don't change whether you use `strata =
site` in the `envfit()` examples. What changes is the null distribution
of r2 values against which the observed r2 is compared. Random
permutation appears to produce results with far lower r2, in general
than the observed, whilst permuting within `site` results in many r2
values larger than the observed (~51 and ~27% more respectively for
easting and northing). The run without strata is not appropriate as the
samples are not freely exchangeable /between/ sites under the null
hypothesis if you have clustering at the site level. Using `strata` is
the more appropriate null hypothesis as samples are only freely
exchangeable **within** site. Depending on whether you have sufficient
sites, 999 permutations might be too many (there are only two
permutations within each level of `site`, so what is 2 * number of sites
in your data set?), but as yet we don't have a facility for iterating
over the full set of permutations. The general facility is available in
the `permute` package on r-forge but won't get fully integrated into
vegan until later in the summer.

As for `adonis()`, AFAIUI, it is decomposing the **entire**
dissimilarity matrix into "variance" explained by each covariate. In
nMDS, you have reduced the entire dissimilarity matrix to three
dimensions and then look for correlations in those three dimensions with
covariates. The "discrepancy" is due to the two techniques working with
different "data" and assessing different versions of "explanatory
power". These results, at first blush, would suggests to me that easting
and northing effects are in the parts of the dissimilarity matrix not
represented well by the 3-d nMDS solution.

Does this help?

G