Gamma imputation - simulation study"

#replication of simulation study performed by Jackson et al DOI: 10.1002/sim.6274
#we perform 100 simulations per value of gamma, whereas Jackson et al used 1000

#load informative censoring package
library(InformativeCensoring)

runSim <- function(nSim=100,n=100,gamma=0) {

  ICEst <- array(0, dim=c(nSim,2))
  ICCI <- array(0, dim=c(nSim,4))
  miEst <- array(0, dim=c(nSim,2))
  miCI <- array(0, dim=c(nSim,4))
  trueEst <- array(0, dim=c(nSim,2))
  for (sim in 1:nSim) {
    u <- runif(n)
    z <- rep(0, n)
    z[(u>0.5) & (u<0.8)] <- 1
    z[u>0.8] <- 2

    #generate censoring time
    c <- rexp(n, rate=0.3)
    lambda <- 0.03+0.02*(z==1)+0.06*(z==2)
    t <- rexp(n, rate=lambda)

    y <- t
    y[c<t] <- c[c<t]
    y[y>3] <- 3

    delta <- 1*(y==t)

    #note that thus far T and C are independent
    ICmod <- coxph(Surv(y,delta)~factor(z))
    ICEst[sim,] <- coef(ICmod)
    ICCI[sim,1:2] <- log(summary(ICmod)$conf.int[1,3:4])
    ICCI[sim,3:4] <- log(summary(ICmod)$conf.int[2,3:4])

    #now we apply the gamma imputation approach
    data <- data.frame(y,delta,z=factor(z))
    imputed <- gammaImpute(formula=Surv(y,delta)~z, data=data, m=10, gamma=rep(gamma,n), DCO.time = 3)
    fits <- ImputeStat(imputed)
    s <- summary(fits)
    miEst[sim,] <- s[,1]
    miCI[sim,1:2] <- s[1,6:7]
    miCI[sim,3:4] <- s[2,6:7]

    #now we simulate what would have been observed in the absence of censoring
    tstar <- t
    a <- rexp(n, rate=lambda*exp(gamma))
    tstar[t>c] <- c[t>c]+a[t>c]
    #in their paper, Jackson et al again censor tstar at 3
    delta <- 1*(tstar<3)
    tstar[delta==0] <- 3

    truemod <- coxph(Surv(tstar,delta)~factor(z))
    trueEst[sim,] <- coef(truemod)
    #in the way Jackson et al have done it, these now become the 'true' log hazard ratios
  }

  list(ICEst=ICEst, ICCI=ICCI, miEst=miEst, miCI=miCI, trueEst=trueEst)
}

gammaseq <- seq(-2,5,1)
resultsTable <- array(0, dim=c(length(gammaseq), 9))
resultsTable[,1] <- gammaseq

for (i in 1:length(gammaseq)) {
  gammaval <- gammaseq[i]
  print(paste("Gamma = ", gammaval, sep=""))
  results <- runSim(nSim=100,n=1000,gammaval)
  #calculate bias, as defined in Jackson paper
  truth <- colMeans(results$trueEst)

  ICbias <- colMeans(results$ICEst)-colMeans(results$trueEst)
  MIbias <- colMeans(results$miEst)-colMeans(results$trueEst)

  ICCI1 <- ((results$ICCI[,1]<truth[1]) & (results$ICCI[,2]>truth[1]))
  ICCI2 <- ((results$ICCI[,3]<truth[2]) & (results$ICCI[,4]>truth[2]))
  miCI1 <- ((results$miCI[,1]<truth[1]) & (results$miCI[,2]>truth[1]))
  miCI2 <- ((results$miCI[,3]<truth[2]) & (results$miCI[,4]>truth[2]))

  #save to results table, mirroring formatting in Jackson paper
  resultsTable[i,2:9] <- c(MIbias[1], ICbias[1], mean(miCI1), mean(ICCI1), MIbias[2], ICbias[2], mean(miCI2), mean(ICCI2))
}

colnames(resultsTable) <- c("Gamma", "MI bias 1", "IC bias 1", "MI CI 1", "IC CI 1",  "MI bias 2", "IC bias 2", "MI CI 2", "IC CI 2")

format(round(resultsTable, 2), nsmall=2)


Try the InformativeCensoring package in your browser

Any scripts or data that you put into this service are public.

InformativeCensoring documentation built on July 24, 2020, 5:07 p.m.