If only the following arguments are used, a forward pass
is unnecessary, and update.earth
will perform only the pruning pass.
This is usually much faster for large models.
object
glm
trace
nprune
pmethod
Eval.model.subsets
Print.pruning.pass
Force.xtx.prune
Use.beta.cache
Endspan.penalty
Get.leverages
This automatic determination to do a forward pass can be overridden
with the ponly
argument.
If ponly=TRUE
the forward pass will be skipped and only the pruning pass will be executed.
This is useful for doing a pruning pass with new data.
(Use earth's data
argument to specify the new data.)
Typically in this scenario you would also specify penalty=-1
.
This is because with sufficient new data, independent of the original
training data, the RSS not the GCV should be used for evaluating model
subsets
(The GCV approximates what the RSS would be on new data --- but here
we actually have new data, so why bother approximating.
This "use new data for pruning" approach is useful in situations where
you don't trust the GCV approximation for your data.)
By making penalty=-1
, earth will calculate the RSS, not the GCV.
See also the description of penalty
on the
earth
help page.
Another (somewhat esoteric) use of ponly=TRUE
is to do subset
selection with a different penalty
from that used to build the
original model.
With trace=1
, update.earth
will tell you if earth's
forward pass was skipped.
If you used keepxy=TRUE
in your original call to earth
, then
update.earth
will use the saved values of x
, y
, etc.,
unless you specify otherwise by arguments to update.earth
.
It can be helpful to set trace=1
to see which x
and y
is used by update.earth
.