Hint See argument R2scope = TRUE, which asks ordiR2step() to compute the
R
2
adj of the global model to be used as a second stopping criterion. Actually TRUE
is the default; we added it to emphasize its importance.
A forward selection on object env3, which contains a factor variable instead of
the quantitative slo, would return a different result because the transformation of
our quantitative slo variable into a four-level factor produced a loss of information,
which resulted in the R
2
adj of the global RDA being smaller. Note also that a factor
represents a single variable in this procedure. Its levels are not selected separately.
Forward selection within a partial RDA with functions ordistep()
and ordiR2step()
It could happen that a researcher would like to forward-select explanatory variables
while keeping others constant. This is possible with ordistep() and
ordiR2step(). The conditioning variables (i.e., variables to be held constant)
must be specified in the two models provided to the functions. As an illustration, let
us run a forward selection of environmental variables while holding the slope
constant, using ordiR2step().
# Partial forward selection with variable slo held constant
mod0p <- rda(spe.hel ~ Condition(slo), data = env2)
mod1p <- rda(spe.hel ~ . + Condition(slo), data = env2)
step.p.forward
direction = "forward",
permutations = how(nperm = 199)
)
As you can see, the result is different from those obtained above, because the
influence of the slope is taken into account when starting the forward selection.
Parsimonious RDA
These analyses show that the most parsimonious attitude would be to settle for a
model containing only three explanatory variables: elevation, oxygen and biological
oxygen demand. What would such an analysis look like? Although it is already made
and stored in object step2.forward, let us run it again with an explicit formula:
230
6 Canonical Ordination
R
2
adj of the global model to be used as a second stopping criterion. Actually TRUE
is the default; we added it to emphasize its importance.
A forward selection on object env3, which contains a factor variable instead of
the quantitative slo, would return a different result because the transformation of
our quantitative slo variable into a four-level factor produced a loss of information,
which resulted in the R
2
adj of the global RDA being smaller. Note also that a factor
represents a single variable in this procedure. Its levels are not selected separately.
Forward selection within a partial RDA with functions ordistep()
and ordiR2step()
It could happen that a researcher would like to forward-select explanatory variables
while keeping others constant. This is possible with ordistep() and
ordiR2step(). The conditioning variables (i.e., variables to be held constant)
must be specified in the two models provided to the functions. As an illustration, let
us run a forward selection of environmental variables while holding the slope
constant, using ordiR2step().
# Partial forward selection with variable slo held constant
mod0p <- rda(spe.hel ~ Condition(slo), data = env2)
mod1p <- rda(spe.hel ~ . + Condition(slo), data = env2)
step.p.forward
direction = "forward",
permutations = how(nperm = 199)
)
As you can see, the result is different from those obtained above, because the
influence of the slope is taken into account when starting the forward selection.
Parsimonious RDA
These analyses show that the most parsimonious attitude would be to settle for a
model containing only three explanatory variables: elevation, oxygen and biological
oxygen demand. What would such an analysis look like? Although it is already made
and stored in object step2.forward, let us run it again with an explicit formula:
230
6 Canonical Ordination