prodlim (version 1.3.7)

jackknife: Compute jackknife pseudo values.

Description

Compute jackknife pseudo values based on marginal Kaplan-Meier estimate of survival, or based on marginal Aalen-Johannsen estimate of cumulative incidence.

Usage

jackknife(object, times,keepResponse=FALSE, ...)
jackknife.survival(object, times,keepResponse=FALSE, ...)
jackknife.competing.risks(object, times, cause, keepResponse=FALSE, ...)
leaveOneOut(object, times, ...)
leaveOneOut.survival(object, times, lag=FALSE, ...)
leaveOneOut.competing.risks(object, times, cause, ...)

Arguments

object

Object of class "prodlim".

times

Time points at which to compute pseudo values.

cause

For competing risks the cause of failure.

keepResponse

If TRUE add the model response, i.e. event time, event status, etc. to the result.

lag

Logical. If TRUE lag the result, i.e. compute S(t-) instead of S(t).

...

References

Andersen PK & Perme MP (2010). Pseudo-observations in survival analysis Statistical Methods in Medical Research, 19(1), 71-99.

See Also

prodlim

Examples

## pseudo-values for survival models

d=SimSurv(100) 
f=prodlim(Hist(time,status)~1,data=d) 
jackknife(f,times=c(30,50))

## in some situations it may be useful to attach the
## the event time history to the result
jackknife(f,times=c(30,50),keepResponse=TRUE)

# pseudo-values for competing risk models
d=prodlim:::SimCompRisk(100) 
f=prodlim(Hist(time,cause)~1,data=d) 
jackknife(f,times=c(3,10),cause=1)
jackknife(f,times=c(3,10,17),cause=2)

# comparison to pseudoci
# make sure we get the same
# results with both packages

library(prodlim)
library(pseudo)

set.seed(17)
N <- 200
ddd <- data.frame(time=1:N,cause=rbinom(N,2,.5),X=rbinom(N,1,.5))
ttt <- c(3,5,10)
# ttt <- ddd$time
fff <- prodlim(Hist(time,cause)~1,data=ddd)
system.time(jack <- with(ddd,pseudoci(time,cause,ttt))[,-c(1:2,seq(4,N+2,2))])
system.time({jack2 <- jackknife.competing.risks(fff,times=ttt)})

## check individual 1
all(round(jack2[,1],9)==round(jack[,1],9))

## check all individuals
all(sapply(1:N,function(x){
a <- round(jack[x,],8)
b <- round(jack2[x,],8)
# all(a[!is.na(a)]==b[!is.na(b)])
all(a[!is.na(a)]==b[!is.na(a)])
}))


## the pseudoci function seems only slightly slower
## for small sample sizes (up to ca. 200) but
## much slower for large sample sizes:

set.seed(17)
N <- 200
ddd <- data.frame(time=1:N,cause=rbinom(N,2,.5),X=rbinom(N,1,.5))
ttt <- c(3,5,10)
# ttt <- ddd$time
fff <- prodlim(Hist(time,cause)~1,data=ddd)
system.time(jack <- with(ddd,pseudoci(time,cause,ttt))[,-c(1:2,seq(4,N+2,2))])
system.time({jack2 <- jackknife.competing.risks(fff,times=ttt)})
all(round(jack2[,1],9)==round(jack[,1],9))

set.seed(17)
N <- 2000
ddd <- data.frame(time=1:N,cause=rbinom(N,2,.5),X=rbinom(N,1,.5))
ttt <- c(3,5,10)
# ttt <- ddd$time
fff <- prodlim(Hist(time,cause)~1,data=ddd)
system.time(jack <- with(ddd,pseudoci(time,cause,ttt))[,-c(1:2,seq(4,N+2,2))])
system.time({jack2 <- jackknife.competing.risks(fff,times=ttt)})
all(round(jack2[,1],9)==round(jack[,1],9))