Hint If you don’t want the screen display of the intermediate results of ordistep()
and ordiR2step() during the computation, add argument trace = FALSE
to the ordistep() call. At the end, display the table of results by typing nameof-the-ordistep-result-object$anova.
The selected variables are the same four in this example as found with forward.sel() without the R
2
adj stopping criterion of Blanchet et al. (2008a) (not
shown): ele, oxy, bod and slo. Its R
2
adj exceeds the R
2
adj of the model
containing all variables, although not by much (0.5947).
Backward elimination with function ordistep()
ordistep() allows backward elimination. In this case, only the model containing
all candidate explanatory variables must be provided to the function. Backward
elimination returns the same result as forward selection with our data; this will not
necessarily be the case for other data:
# Backward elimination using vegan's ordistep()
step.backward permutations = how(nperm = 499)
)
RsquareAdj(step.backward)
Hint When the 'scope' argument is missing, direction = "backward"
becomes the default.
Forward selection with function ordiR2step()
Function ordiR2step(), which is limited to RDA and db-RDA and does not
allow backward elimination, uses the same two criteria as forward.sel() (α
level and R
2
adj of the global model) with the added bonus that it accepts factor
variables. An application with the same quantitative environmental data matrix as
the one used with forward.sel() returns, of course, the same result:
# Forward selection using vegan's ordiR2step()
step2.forward scope = formula(spe.rda.all),
direction = "forward",
R2scope = TRUE,
permutations = how(nperm = 199)
)
RsquareAdj(step2.forward)
6.3 Redundancy Analysis (RDA)
229