# # Test out the "return.all" argument of xpred # The data set has the virtue of continuous, categorical, and missings # library(rpart) require(survival) fit1 <- rpart(Surv(pgtime, pgstat) ~ age + eet + g2+grade+gleason +ploidy, stagec, method='poisson') xgrp <- rep(1:3, length=nrow(stagec)) # explicitly set the xval groups xfit1 <- xpred.rpart(fit1, xval=xgrp, return.all=T) xfit2 <- array(0, dim=dim(xfit1)) cplist <- as.numeric(dimnames(xfit1)[[2]]) for (i in 1:3) { tfit <- rpart(Surv(pgtime, pgstat) ~ age + eet + g2+grade+gleason +ploidy, stagec, method='poisson', subset=(xgrp !=i)) # xvals are actually done on the absolute risk (node's risk /n), not on # the rescaled risk ((node risk)/ (top node risk)) which is the basis # for the printed CP. To get the right answer we need to rescale. cp2 <- cplist * (fit1$frame$dev[1] / fit1$frame$n[1]) / (tfit$frame$dev[1] / tfit$frame$n[1]) for (j in 1:length(cp2)) { tfit2 <- prune(tfit, cp=cp2[j]) temp <- predict(tfit2, newdata=stagec[xgrp==i,], type='matrix') xfit2[xgrp==i, j,] <- temp } } all.equal(xfit1, xfit2, check.attributes=FALSE)