How do I extract hazards from survfit in R?

Question

I have a survfit object. A summary survfit for my t=0:50 years of interest is easy enough.

summary(survfit, t=0:50)

It gives the survival at each t.

Is there a way to get the hazard for each t (in this case, the hazard from t-1 to t in each t=0:50)? I want to get the mean and confidence interval (or standard error) for the hazards relating to the Kaplan Meier curve.

This seems easy to do when a distribution is fit (eg. type="hazard" in flexsurvreg) but I can't figure out how to do this for a regular survfit object. Suggestions?

Answer 1

For sake of performance, we could use a more streamlined bootstrap function.

## define custom times
t0 <- 0
t1 <- 744

## bootstrap fun
boot_fun <- function(x) {
  n <- dim(x)[1]
  x <- x[sample.int(n, n, replace=TRUE), ]
  muhaz::muhaz(x$futime, x$fustat, min.time=t0, max.time=t1)
}

# bootstrap
set.seed(42)
R <- 999
B <- replicate(R, boot_fun(ovarian))

The results may be calculated by hand.

## extract matrix from bootstrap
r <- `colnames<-`(t(array(unlist(B[3, ]), dim=c(101, R))), B[2, ][[1]])

## calculate result
library(matrixStats)  ## for fast matrix calculations
r <- cbind(x=as.numeric(colnames(r)), 
           y=colMeans2(r),
           shape=(colMeans2(r)/colSds(r))^2, 
           scale=colVars(r)/colMeans2(r))
r <- cbind(r[, 1:2], 
           lower=qgamma(0.025, shape=r[, 'shape'] + 1, scale=r[, 'scale']),
           upper=qgamma(0.975, shape=r[, 'shape'], scale=r[, 'scale']))

Gives

head(r)
#          x            y        lower       upper
# [1,]  0.00 0.0003836816 9.359267e-05 0.001400539
# [2,]  7.44 0.0003913992 9.746868e-05 0.001387551
# [3,] 14.88 0.0003997275 1.018897e-04 0.001374005
# [4,] 22.32 0.0004087439 1.069353e-04 0.001360212
# [5,] 29.76 0.0004178464 1.123697e-04 0.001346187
# [6,] 37.20 0.0004275531 1.184685e-04 0.001332237

range(r[, 'y'])
# [1] 0.0003836816 0.0011122910

Plot

matplot(r[, 1], r[, -1], type='l', lty=c(1, 2, 2), col=4, 
        xlab='Time', ylab='Hazard Rate', main='Hazard Estimates')
legend('topleft', legend=c('estimate', '95% CI'), col=4, lty=1:2, cex=.8)

---

Data

data(cancer, package="survival")  ## loads `ovarian` data set

Answer 2

As a supplement to Mike's answer, one could model the number of events by a Poisson distribution instead of a Normal distribution. The hazard rate can then be calculated via a gamma distribution. The code would become:

library(muhaz)
library(data.table)
library(rGammaGamma)
data(ovarian, package="survival")
attach(ovarian)
fit1 <- muhaz(futime, fustat)
plot(fit1)

#Function to bootstrap hazard estimates
haz.bootstrap <- function(data,trial,min.time,max.time){
  library(data.table)
  data <- as.data.table(data)
  data <- data[sample(1:nrow(data),nrow(data),replace=T)]
  fit1 <- muhaz(data$futime, data$fustat,min.time=min.time,max.time=max.time)
  result <- data.table(est.grid=fit1$est.grid,trial,haz.est=fit1$haz.est)
  return(result)
}

#Re-run function to get 1000 estimates
haz.list <- lapply(1:1000,function(x) haz.bootstrap(data=ovarian,trial=x,min.time=0,max.time=744))
haz.table <- rbindlist(haz.list,fill=T)

#Calculate Mean, gamma parameters, upper and lower 95% confidence bands
plot.table <- haz.table[, .(Mean=mean(haz.est),
                            Shape = gammaMME(haz.est)["shape"],
                            Scale = gammaMME(haz.est)["scale"]), by=est.grid]
plot.table[, u95 := qgamma(0.95,shape = Shape + 1, scale = Scale)]
# The + 1 is due to the discrete character of the poisson distribution.
plot.table[, l95 := qgamma(0.05,shape = Shape, scale = Scale)]

#Plot graph
ggplot(data=plot.table) + 
  geom_line(aes(x=est.grid, y=Mean),col="blue") + 
  geom_ribbon(aes(x=est.grid, y=Mean, ymin=l95, ymax=u95),alpha=0.5, fill= "lightblue")

As can be seen the negative estimates for the lower bound of the hazard rate are now gone.

Answer 3

It is a bit tricky since the hazard is an estimate of an instantaneous probability (and this is discrete data), but the basehaz function might be of some help, but it only returns the cumulative hazard. So you would have still have to perform an extra step.

I have also had luck with the muhaz function. From its documentation:

library(muhaz)
?muhaz
data(ovarian, package="survival")
attach(ovarian)
fit1 <- muhaz(futime, fustat)
plot(fit1)

I am not sure the best way to get at the 95% confidence interval, but bootstrapping might be one approach.

#Function to bootstrap hazard estimates
haz.bootstrap <- function(data,trial,min.time,max.time){
  library(data.table)
  data <- as.data.table(data)
  data <- data[sample(1:nrow(data),nrow(data),replace=T)]
  fit1 <- muhaz(data$futime, data$fustat,min.time=min.time,max.time=max.time)
  result <- data.table(est.grid=fit1$est.grid,trial,haz.est=fit1$haz.est)
  return(result)
}

#Re-run function to get 1000 estimates
haz.list <- lapply(1:1000,function(x) haz.bootstrap(data=ovarian,trial=x,min.time=0,max.time=744))
haz.table <- rbindlist(haz.list,fill=T)

#Calculate Mean,SD,upper and lower 95% confidence bands
plot.table <- haz.table[, .(Mean=mean(haz.est),SD=sd(haz.est)), by=est.grid]
plot.table[, u95 := Mean+1.96*SD]
plot.table[, l95 := Mean-1.96*SD]

#Plot graph
library(ggplot2)
p <- ggplot(data=plot.table)+geom_smooth(aes(x=est.grid,y=Mean))
p <- p+geom_smooth(aes(x=est.grid,y=u95),linetype="dashed")
p <- p+geom_smooth(aes(x=est.grid,y=l95),linetype="dashed")
p