In stefkruegel/CPsurv: Nonparametric Change Point Estimation for Survival Data

The R-package CPsurv provides a function 'cpsurv' for estimating the changepoint in an L-shaped hazard rate of survival data, where a declining hazard rate is assumed until the change point and a constant hazard rate after the change point. Estimation is based on p-values for testing the null-hypothesis of a specific hazard rate in time intervals of an appropriate length and a regression to this p-values.

Installing the package

CPsurv is published on CRAN, so you can use

install.packages("CPsurv")

to install the package.

Installation from GitHub is done by

devtools::install_github("stefkruegel/CPsurv")

Getting started

Here's a short demo how to estimate the change point of the example lifetime data which are included in CPsurv.

library(CPsurv)
data("survdata")

cp <- cpsurv(time = survdata$time, event = survdata$event, cpmax = 360, intwd = 20,
             censoring = "random", boot.ci = TRUE)

We choose a value for 'cpmax' of 360. That means we expect the change point to be within the first year. 'intwd' is a tuning parameter, which has to be chosen appropriate based on given graphical tools (plot.cpsurv).

Now we can check the summary and use the plot function.

summary(cp)
plot(cp, ask = FALSE)

By means of the plot of p-values you can check, if the regression line has an appropriate adjustment to the p-values. Otherwise you could try different grid length. But you have to take care about the number of events within one interval (check the barplot). If you choose a small interval length, it's possible that there are too low or even no events within an interval.
Within the first phase we assume a high mortality rate which means a high number of events within one interval. But if only a few of these intervals contain a small number of events, caused by a too small interval width, the estimated change point is biased downwards.