R code for chapter 2

The prodint() function used throughout this chapter for product integration requires two arguments: a vector of time points and a function.

prodint <- function(time.points, A) {
    times <- c(0, sort(unique(time.points)))
    S <- prod(1 - diff(apply(X = matrix(times), 
                       MARGIN = 1, FUN = A)))
    return(S)
}

A.exp is a function which computes the cumulative hazard with hazard(t) = 0.9 using a vector of time points that you have to enter.

A.exp <- function(time.point) {
    return(0.9 * time.point)
}

The first example (p. 12) uses a sequence of timepoints between 0 and 1.

times <- seq(0, 1, 0.001)
prodint(times, A.exp)
exp(-0.9 * max(times))

The next example uses a vector of uniformly distributed random numbers.

prodint(runif(n = 1000, min = 0, max = 1), A.exp)

Creation of a similar function for Weibull distributed survival times.

A.weibull <- function(time.point){
  return(2 * time.point^0.25)

prodint(times, A.weibull)
## True result
exp(-2 * max(times)^0.25)
}

## finer grid of time points
prodint(seq(0, 1, 0.000001), A.weibull)

Simulate some event times with constant hazard equal to 0.9.

event.times <- rexp(100,0.9)

Computing survival function S(t) with survfit function, which is part of the survival package.

library(survival)
fit.surv <- survfit(Surv(event.times) ~ 1)

Compute the Nelson-Aalen estimator based on what survfitreturns.

A <- function(time.point) {
    sum(fit.surv$n.event[fit.surv$time <= time.point]/
        fit.surv$n.risk[fit.surv$time <= time.point])
}

One can see that using product integration leads to the same result.

prodint(event.times[event.times <= 1], A)

Computation of the empirical survival function:

sum(event.times > 1) / length(event.times)

Now, we simulate censoring times following a uniform distribution.

cens.times <- runif(100,0,5)

 Next, we take the minimum of the event times and the censoring times.

obs.times <- pmin(event.times, cens.times)

and with their help we create an event indicator

event.times <= cens.times

Survival probability estimate:

fit.surv <- survfit(Surv(obs.times, 
                         event.times <= cens.times) ~ 1)

with product integration

prodint(obs.times[obs.times<=1], A)

estimate at time 1 using survfit

S <- fit.surv$surv
S[fit.surv$time <= 1][length(S[fit.surv$time <= 1])]

Figure 2.3:

nt <- cumsum(fit.surv$n.event)

Function that computes the compensator

integral2 <- function(x) { 

    tmp <- rep(0,length(x))
    
    for(i in 1:length(x)){
        if(x[i]<=min(fit.surv$time)){
            tmp[i] <- x[i] * fit.surv$n.risk[1]
        }
        else{
            if(x[i]>=max(fit.surv$time)){
                tmp[i] <- sum(diff(c(0,fit.surv$time)) * fit.surv$n.risk)
            }
            else{
                bla <- diff(c(0,fit.surv$time)[c(0,fit.surv$time) < x[i]])  
                tmp[i] <- sum(bla * fit.surv$n.risk[1:length(bla)]) +
                             (x[i]-max(fit.surv$time[fit.surv$time<x[i]])) * 
                             fit.surv$n.risk[length(bla)+1]
            }
        }
    }
    return(0.9 * tmp)
}

 Plot the compensator and counting process.

plot(x = c(0, 4), y = c(0, 80), xlab = expression(paste(Time, " ", italic(t))), 
     ylab = "", type="n", axes = F, main = "", 
     cex.main = 1.7, cex.lab = 1.5)
axis(1, at = seq(0, 4, 1), cex.axis = 1.5)
axis(2, at = seq(0, 80, 10), cex.axis = 1.25)

box()

lines(fit.surv$time, nt, type="s", lwd=2)
curve(integral2, from = 0, to = 5, add = TRUE, lwd=2, n = 1001)