ipcw: Estimation of censoring probabilities

In riskRegression: Risk Regression Models and Prediction Scores for Survival Analysis with Competing Risks

Estimation of censoring probabilities

Description

This function is used internally to obtain inverse of the probability of censoring weights.

Usage

ipcw(
  formula,
  data,
  method,
  args,
  times,
  subject.times,
  lag = 1,
  what,
  keep = NULL
)

Arguments

formula	A survival formula like, Surv(time,status)~1, where as usual status=0 means censored. The status variable is internally reversed for estimation of censoring rather than survival probabilities. Some of the available models (see argument model) will use predictors on the right hand side of the formula.
data	The data used for fitting the censoring model
method	Censoring model used for estimation of the (conditional) censoring distribution.
args	A list of arguments which is passed to method
times	For what="IPCW.times" a vector of times at which to compute the probabilities of not being censored.
subject.times	For what="IPCW.subject.times" a vector of individual times at which the probabilities of not being censored are computed.
lag	If equal to 1 then obtain G(T_i-|X_i), if equal to 0 estimate the conditional censoring distribution at the subject.times, i.e. (G(T_i|X_i)).
what	Decide about what to do: If equal to "IPCW.times" then weights are estimated at given times. If equal to "IPCW.subject.times" then weights are estimated at individual subject.times. If missing then produce both.
keep	Which elements to add to the output. Any subset of the vector c("times","fit","call").

Details

Inverse of the probability of censoring weights (IPCW) usually refer to the probabilities of not being censored at certain time points. These probabilities are also the values of the conditional survival function of the censoring time given covariates. The function ipcw estimates the conditional survival function of the censoring times and derives the weights.

IMPORTANT: the data set should be ordered, order(time,-status) in order to get the values IPCW.subject.times in the right order for some choices of method.

Value

A list with elements depending on argument keep.

times	The times at which weights are estimated
IPCW.times	Estimated weights at times
IPCW.subject.times	Estimated weights at individual time values subject.times
fit	The fitted censoring model
method	The method for modelling the censoring distribution
call	The call

Author(s)

Thomas A. Gerds [e-mail]

Examples

library(prodlim)
library(rms)
dat=SimSurv(30)

dat <- dat[order(dat$time),]

# using the marginal Kaplan-Meier for the censoring times

WKM=ipcw(Hist(time,status)~X2,
  data=dat,
  method="marginal",
  times=sort(unique(dat$time)),
  subject.times=dat$time,keep=c("fit"))
plot(WKM$fit)
WKM$fit

# using the Cox model for the censoring times given X2
library(survival)
WCox=ipcw(Hist(time=time,event=status)~X2,
  data=dat,
  method="cox",
  times=sort(unique(dat$time)),
  subject.times=dat$time,keep=c("fit"))
WCox$fit

plot(WKM$fit)
lines(sort(unique(dat$time)),
      1-WCox$IPCW.times[1,],
      type="l",
      col=2,
      lty=3,
      lwd=3)
lines(sort(unique(dat$time)),
      1-WCox$IPCW.times[5,],
      type="l",
      col=3,
      lty=3,
      lwd=3)

# using the stratified Kaplan-Meier
# for the censoring times given X2

WKM2=ipcw(Hist(time,status)~X2,
  data=dat,
  method="nonpar",
  times=sort(unique(dat$time)),
  subject.times=dat$time,keep=c("fit"))
plot(WKM2$fit,add=FALSE)

riskRegression documentation built on May 29, 2024, 10:59 a.m.