CoxFlexBoost-package: CoxFlexBoost: Boosting Flexible Cox Models (with Time-Varying...

In CoxFlexBoost: Boosting Flexible Cox Models (with Time-Varying Effects)

Description

Likelihood-based boosting approach to fit flexible, structured survival models with component-wise linear or P-spline base-learners. Variable selection and model choice are built in features.

Details

Package:	CoxFlexBoost
Type:	Package
Version:	0.7-0
Date:	2009-03-XX
License:	GPL
LazyLoad:	yes

Author(s)

Benjamin Hofner <[e-mail]>

References

Benjamin Hofner, Torsten Hothorn and Thomas Kneib (2008), Variable Selection and Model Choice in Structured Survival Models. Department of Statistics, Technical Report No. 43. http://epub.ub.uni-muenchen.de/7901/

See Also

Important functions include: rSurvTime for simulations of survival times, cfboost for the boosting algorithm, bols and bbs for the base-learners and methods for methods to be used with results.

Examples

## a simple example

#################
# simulate data #
#################

set.seed(6102)

## sample covariates first
X <- matrix(NA, nrow=400, ncol=3)
X[,1] <- runif(400, -1, 1)
X[,2] <- runif(400, -1, 1)
X[,3] <- runif(400, -1, 1)

## time-constant hazard rate
lambda <- function(time, x){
  exp(0 * time + 0.7 * x[1] + x[2]^2)
}

## specify censoring function
cens_fct <- function(time, mean_cens){
  censor_time <- rexp(n = length(time), rate = 1/mean_cens)
  event <- (time <= censor_time)
  t_obs <- apply(cbind(time, censor_time), 1, min)
  return(cbind(t_obs, event))
}
data <- rSurvTime(lambda, X, cens_fct, mean_cens = 5)

##################
# estimate model #
##################

ctrl <- boost_control( mstop = 100, risk="oobag")
weights <- c(rep(1, 300), rep(0, 100))

## fit (a simple) model
model <- cfboost(Surv(time, event) ~ bbs(x.1) + bbs(x.2) + bbs(x.3),
                 control = ctrl, data = data, weights = weights)

#########################
# processing the output #
#########################

## determine appropriate stopping iteration
mstop(model)

## summary (with mstop)
summary(model)
## or (almost) same result
model

## plot the risk
plot(risk(model))

## proceed with optimal model
model_opt <- model[mstop(model)]

## extract coefficients of optimal model
## (not very meaningfull for P-spline base-learner)
coef(model_opt)

## plot optimal model
plot(model_opt)

## extract fitted values
fitted(model_opt)

## make prediction for the first out-of-bag observation
predict(model_opt, newdata = data[301,])

## look at the selection frequencies for optimal model)
freq.sel(model_opt)

###################################
# model with time-varying effects #
###################################

## Not run: 
## time-dependent hazard rate
lambda <- function(time, x){
  exp(log(time) + 0.7 * x[1] + x[2]^2)
}

data <- rSurvTime(lambda, X, cens_fct, mean_cens = 5)

## fit model (quite time-consuming)
model <- cfboost(Surv(time, event) ~ bbsTime(time) + bbs(x.1) + bbs(x.2) + bbs(x.3),
                 control = ctrl, data = data, weights = weights)

## End(Not run)

##################################################
# estimate model with model choice decomposition #
##################################################

## the next model estimation is very time-consuming

## Not run: 
model_mod_choice <- cfboost(Surv(time, event) ~ bolsTime(time)
    + bbsTime(time, df = 1, center = TRUE)
    + bbsTime(time, df = 1, center = TRUE)
    + bols(x.1) + bbs(x.1, df = 1, center = TRUE)
    + bolsTime(time, z = x.1)
    + bbsTime(time, z = x.1, df = 1, center = TRUE)
    + bols(x.2) + bbs(x.2, df = 1, center = TRUE)
    + bolsTime(time, z = x.2)
    + bbsTime(time, z = x.2, df = 1, center = TRUE)
    + bols(x.3) + bbs(x.3, df = 1, center = TRUE)
    + bolsTime(time, z = x.3)
    + bbsTime(time, z = x.3, df = 1, center = TRUE),
  data = data, control = ctrl, weights = weights)

## End(Not run)

## proceed as above

CoxFlexBoost documentation built on May 2, 2019, 6:53 p.m.