prune.modelparty: Post-Prune `modelparty` Objects

Description

Post-pruning of modelparty objects based on information criteria like AIC, BIC, or related user-defined criteria.

Usage

# S3 method for modelparty
prune(tree, type = "AIC", ...)

Value

An object of class modelparty where the associated tree is either the same as the original or smaller.

Arguments

tree: object of class modelparty.
type: pruning type. Can be "AIC", "BIC" or a user-defined function (details below).
...: additional arguments.

Details

In mob-based model trees, pre-pruning based on p-values is used by default and often no post-pruning is necessary in such trees. However, if pre-pruning is switched off (by using a large alpha) or does is not sufficient (e.g., possibly in large samples) the prune method can be used for subsequent post-pruning based on information criteria.

The function prune.modelparty can be called directly but it is also registered as a method for the generic prune function from the rpart package. Thus, if rpart is attached, prune(tree, type = "AIC", ...) also works (see examples below).

To customize the post-pruning strategy, type can be set to a function(objfun, df, nobs) which either returns TRUE to signal that a current node can be pruned or FALSE. All supplied arguments are of length two: objfun is the sum of objective function values in the current node and its child nodes, respectively. df is the degrees of freedom in the current node and its child nodes, respectively. nobs is vector with the number of observations in the current node and the total number of observations in the dataset, respectively.

For "AIC" and "BIC" type is transformed so that AIC or BIC are computed. However, this assumes that the objfun used in tree is actually the negative log-likelihood. The degrees of freedom assumed for a split can be set via the dfsplit argument in mob_control when computing the tree or manipulated later by changing the value of tree$info$control$dfsplit.

Examples

Run this code

set.seed(29)
n <- 1000
d <- data.frame(
  x = runif(n),
  z = runif(n),
  z_noise = factor(sample(1:3, size = n, replace = TRUE))
)
d$y <- rnorm(n, mean = d$x * c(-1, 1)[(d$z > 0.7) + 1], sd = 3)

## glm versus lm / logLik versus sum of squared residuals
fmla <- y ~ x | z + z_noise
lm_big <- lmtree(formula = fmla, data = d, maxdepth = 3, alpha = 1)
glm_big <- glmtree(formula = fmla, data = d, maxdepth = 3, alpha = 1)

AIC(lm_big)
AIC(glm_big)

## load rpart for prune() generic
## (otherwise: use prune.modelparty directly)
if (require("rpart")) {

## pruning
lm_aic <- prune(lm_big, type = "AIC")
lm_bic <- prune(lm_big, type = "BIC")

width(lm_big)
width(lm_aic)
width(lm_bic)

glm_aic <- prune(glm_big, type = "AIC")
glm_bic <- prune(glm_big, type = "BIC")

width(glm_big)
width(glm_aic)
width(glm_bic)

}