goodness and inertcomp can
be used to assess the goodness of fit for individual sites or
species. Function vif.cca and alias.cca can be used to
analyse linear dependencies among constraints and conditions. In
addition, there are some other diagnostic tools (see 'Details').
"goodness"(object, display = c("species", "sites"), choices, model = c("CCA", "CA"), statistic = c("explained", "distance"), summarize = FALSE, addprevious = FALSE, ...)
inertcomp(object, display = c("species", "sites"), statistic = c("explained", "distance"), proportional = FALSE)
spenvcor(object)
intersetcor(object)
vif.cca(object)
"alias"(object, names.only = FALSE, ...)"species" or "sites". Species
are not available in capscale. "model". "CCA") or unconstrained
("CA") results. "explained" gives the cumulative
percentage accounted for, "distance" shows the residual
distances. statistic="explained". For model="CCA" add
conditionened (partialled out) variation, and for model="CA"
add both conditioned and constrained variation. This will give
cumulative explanation. The argument has no effect when
statistic="distance", but this will always show the residual
distance after current axis and all previous components.
goodness gives the diagnostic statistics for species
or sites. The alternative statistics are the cumulative proportion of
inertia accounted for up to the axes, and the residual distance left
unaccounted for. Function inertcomp decomposes the inertia into partial,
constrained and unconstrained components for each site or
species. Instead of inertia, the function can give the total
dispersion or distances from the centroid for each component.
Function spenvcor finds the so-called “species --
environment correlation” or (weighted) correlation of
weighted average scores and linear combination scores. This is a bad
measure of goodness of ordination, because it is sensitive to extreme
scores (like correlations are), and very sensitive to overfitting or
using too many constraints. Better models often have poorer
correlations. Function ordispider can show the same
graphically.
Function intersetcor finds the so-called “interset
correlation” or (weighted) correlation of weighted averages scores
and constraints. The defined contrasts are used for factor
variables. This is a bad measure since it is a correlation. Further,
it focuses on correlations between single contrasts and single axes
instead of looking at the multivariate relationship. Fitted vectors
(envfit) provide a better alternative. Biplot scores
(see scores.cca) are a multivariate alternative for
(weighted) correlation between linear combination scores and
constraints.
Function vif.cca gives the variance inflation factors for each
constraint or contrast in factor constraints. In partial ordination,
conditioning variables are analysed together with constraints. Variance
inflation is a diagnostic tool to identify useless constraints. A
common rule is that values over 10 indicate redundant
constraints. If later constraints are complete linear combinations of
conditions or previous constraints, they will be completely removed
from the estimation, and no biplot scores or centroids are calculated
for these aliased constraints. A note will be printed with default
output if there are aliased constraints. Function alias will
give the linear coefficients defining the aliased constraints, or
only their names with argument names.only = TRUE.
Gross, J. (2003). Variance inflation factors. R News 3(1), 13--15.
cca, rda, capscale,
vif. data(dune)
data(dune.env)
mod <- cca(dune ~ A1 + Management + Condition(Moisture), data=dune.env)
goodness(mod, addprevious = TRUE)
goodness(mod, addprevious = TRUE, summ = TRUE)
# Inertia components
inertcomp(mod, prop = TRUE)
inertcomp(mod, stat="d")
# vif.cca
vif.cca(mod)
# Aliased constraints
mod <- cca(dune ~ ., dune.env)
mod
vif.cca(mod)
alias(mod)
with(dune.env, table(Management, Manure))
# The standard correlations (not recommended)
spenvcor(mod)
intersetcor(mod)
Run the code above in your browser using DataLab