74 lines
3.0 KiB
R
74 lines
3.0 KiB
R
|
#
|
||
|
|
# Simplest weight test: treble the weights
|
||
|
|
#
|
||
|
|
# By using the unshrunken estimates the weights will nearly cancel
|
||
|
|
# out: frame$wt, frame$dev, frame$yval2, and improvement will all
|
||
|
|
# be threefold larger, other things will be the same.
|
||
|
|
# The improvement is the splits matrix, column 3, rows with n>0. Other
|
||
|
|
# rows are surrogate splits.
|
||
|
|
library(rpart)
|
||
|
|
require(survival)
|
||
|
|
set.seed(10)
|
||
|
|
|
||
|
|
tempc <- rpart.control(maxsurrogate=0, cp=0, xval=0)
|
||
|
|
fit1 <- rpart(Surv(pgtime, pgstat) ~ age + eet + g2+grade+gleason +ploidy,
|
||
|
|
stagec, control=tempc,
|
||
|
|
method='poisson', parms=list(shrink=0))
|
||
|
|
wts <- rep(3, nrow(stagec))
|
||
|
|
fit1b <- rpart(Surv(pgtime, pgstat) ~ age + eet + g2+grade+gleason +ploidy,
|
||
|
|
stagec, control= tempc, parms=list(shrink=0),
|
||
|
|
method='poisson', weights=wts)
|
||
|
|
fit1b$frame$wt <- fit1b$frame$wt/3
|
||
|
|
fit1b$frame$dev <- fit1b$frame$dev/3
|
||
|
|
fit1b$frame$yval2[,2] <- fit1b$frame$yval2[,2]/3
|
||
|
|
fit1b$splits[,3] <- fit1b$splits[,3]/3
|
||
|
|
zz <- match(c("call", "variable.importance"), names(fit1))
|
||
|
|
all.equal(fit1[-zz], fit1b[-zz]) #all but the "call" and importance
|
||
|
|
all.equal(fit1b$variable.importance/fit1$variable.importance, rep(3,4),
|
||
|
|
check.attributes = FALSE)
|
||
|
|
|
||
|
|
#
|
||
|
|
# Compare a pair of multiply weighted fits
|
||
|
|
# In this one, the lengths of where and y won't match
|
||
|
|
# I have to set minsplit to the smallest possible, because otherwise
|
||
|
|
# the replicated data set will sometimes have enough "n" to split, but
|
||
|
|
# the weighted one won't. Use of CP keeps the degenerate splits
|
||
|
|
# (n=2, several covariates with exactly the same improvement) at bay.
|
||
|
|
# For larger trees, the weighted split will sometimes have fewer
|
||
|
|
# surrogates, because of the "at least two obs" rule.
|
||
|
|
#
|
||
|
|
# Create a reproducable psuedo random order using the logisic attractor
|
||
|
|
pseudo <- double(nrow(stagec))
|
||
|
|
pseudo[1] <- pi/4
|
||
|
|
for (i in 2:nrow(stagec)) pseudo[i] <- 4*pseudo[i-1]*(1 - pseudo[i-1])
|
||
|
|
|
||
|
|
wts <- rep(1:5, length=nrow(stagec))
|
||
|
|
temp <- rep(1:nrow(stagec), wts) #row replicates
|
||
|
|
xgrp <- rep(1:10, length=146)[order(pseudo)]
|
||
|
|
xgrp2<- rep(xgrp, wts)
|
||
|
|
# Direct: replicate rows in the data set, and use unweighted
|
||
|
|
fit2 <- rpart(Surv(pgtime, pgstat) ~ age + eet + g2+grade+gleason +ploidy,
|
||
|
|
control=rpart.control(minsplit=2, xval=xgrp2, cp=.025),
|
||
|
|
data=stagec[temp,], method='poisson')
|
||
|
|
|
||
|
|
# Weighted
|
||
|
|
fit2b<- rpart(Surv(pgtime, pgstat) ~ age + eet + g2+grade+gleason +ploidy,
|
||
|
|
control=rpart.control(minsplit=2, xval=xgrp, cp=.025),
|
||
|
|
data=stagec, method='poisson', weight=wts)
|
||
|
|
|
||
|
|
all.equal(fit2$frame[-2], fit2b$frame[-2]) # the "n" component won't match
|
||
|
|
all.equal(fit2$cptable, fit2b$cptable)
|
||
|
|
#all.equal(fit2$splits[,-1],fit2b$splits[,-1]) #fails
|
||
|
|
toss <- c(49, 64)
|
||
|
|
all.equal(fit2$splits[-toss,-1],fit2b$splits[-toss,-1]) #ok
|
||
|
|
all.equal(fit2$csplit, fit2b$csplit)
|
||
|
|
# Line 49 is a surrogate split in a group whose 2 smallest ages are
|
||
|
|
# 47 and 48. The weighted fit won't split there because it wants to
|
||
|
|
# send at least 2 obs to the left; the replicate fit thinks that there
|
||
|
|
# are several 47's.
|
||
|
|
|
||
|
|
|
||
|
|
|
||
|
|
|
||
|
|
|