cfTrunc

The goal of cfTrunc is to generate robust prediction intervals using conformal prediction approach for time-to-event data subject to truncation with or without censoring. Currently, cfTrunc supports left, right, and double truncation. More truncation types will be added later.

Installation

You can install the development version of cfTrunc like so:

if (!require("devtools")){
    install.packages("devtools")
}
devtools::install_github("wwd007/cfTrunc")

Example

This is a basic example which shows you how to generate prediction intervals for left truncated and right censored data:

# load the package
library(cfTrunc)

# generate data
set.seed(42)
N <- 1000
Z <- runif(N)
X <- exp(Z)
C <- rexp(N)
L <- rgamma(N, 0.25, 0.25)
dat <- data.frame(X, Z, C, L)
dat$delta <- as.integer(dat$X > dat$C)
dat$X <- pmin(dat$X, dat$C)
dat <- dat[dat$L < dat$X, ]
dat_tr <- dat[1:(nrow(dat) %/% 2), ]
dat_ca <- dat[(nrow(dat) %/% 2 + 1):nrow(dat), ]
tau <- quantile(c(dat_tr$X), 0.9)

N_te <- 500
Z <- runif(N_te)
X <- exp(Z)
C <- rexp(N_te)
L <- rgamma(N_te, 1, 1)
dat_te <- data.frame(X, Z, C, L)
dat_te$delta <- as.integer(dat_te$X > dat_te$C)

# conformal prediction
pred <- conformal_pred(dat_tr, dat_ca, dat_te,
                       X = "X", Z = c("Z"), L = "L", delta = "delta",
                       trunc_type = "left", target = "RMST",
                       tau = tau, model = "aft", alpha = 0.1)

# calculate coverage
coverage <- mean(pred$y_pred_hi > pmin(dat_te$X, tau) &
                   pred$y_pred_lo < pmin(dat_te$X, tau))
print(paste0("The coverage rate is ", coverage*100, "%."))

Acknowledgment

The implementation of the distribution function estimation under double truncation is adapted from cdfDT() function from SurvTrunc R package.