Major Advantages of CARTs Major Disadvantages of CARTs
CART models are nonparametric and flexible in that
they don’t assume a functional form between outcome
and predictors.
Single trees are likely to have sub-optimal
predictive performance compared to other methods
[@1377].
CART models naturally detect higher level
interactions among the predictors.
CART models are based on splits that depend on
previous splits; so if an error is made in a
higher split it can propagate down the tree.
CART models produce a series of rules for
classification that are easy to interpret and
implement in field work (e.g. adaptive survey
interviewing protocols).
Because of the conditional nature of the fitting,
CART models can also be very sensitive to changes
in the underlying data set.
CART models can handle missing data through the
use of surrogate predictors.
CART models generally consider all predictor
variables at each step of branching and thus
cannot “force” variables to be included
a priori. (e.g. cannot create a model to
predict nonresponse that must include demographics
first, for example).
CART models are computationally fast