IPCWJK

Jackknife Approach to Estimate the Prediction Uncertainty from Binary Classifiers under Right-Censoring

Provides functions for fitting binary classification models with inverse probability of censoring weights (IPCW) to estimate survival probabilities, and implements jackknife resampling methods for unbiased prediction error estimation; see doi:10.1177/09622802251393626.

We described the methodology implemented in this package in our paper Jahn-Eimermacher et al. (2025).

Please note that all methods implemented in this package are intended to reproduce the results presented in the paper. They may not be suitable for other use cases without further validation, especially the employed hyperparameter settings.

The documentation is available at https://iden-project-uas-darmstadt.github.io/IPCWJK/

Installation

You can install the development version of IPCWJK like so:

# With remotes
remotes::install_github("IDEN-Project-UAS-Darmstadt/IPCWJK")
# With pak (recommended for speed)
pak::pkg_install("IDEN-Project-UAS-Darmstadt/IPCWJK")

Releases of the library can be found here.

Examples

IPCW weights can be calculated with the ipcw_weights function.

library(survival)
tau <- 100
df <- veteran[, c("time", "status", "karno", "age")]
newdata <- data.frame(karno = c(80, 70), age = c(40, 50))

# Calculation of IPCW weights at tau
library(IPCWJK)
w <- ipcw_weights(df, tau, time_var = "time", status_var = "status")
hist(w,
  breaks = 30,
  xlab = "IPCW Weight", col = "lightblue", border = "grey"
)

For survreg models with a log-logistic distribution and logitIPCW we provide convenient access to Wald confidence intervals based on the delta method on the logit scale.

# Fit a log-logistic survival model
survreg_fit <- survreg(Surv(time, status) ~ karno + age,
  data = df,
  dist = "loglogistic"
)
# This provides you with an Wald CI on the logit scale:
pred_fun <- deltamethod_from_model(survreg_fit, tau = tau)
pred_fun(newdata)
#>   prediction     lower     upper         se
#> 1  0.6087168 0.4506212 0.7668123 0.08066101
#> 2  0.4815100 0.3795966 0.5834235 0.05199666

# Fit a logitIPCW model
library(mets)
logipcw_fit <- logitIPCW(Event(time, status) ~ karno + age,
  time = tau,
  data = df
)
pred_fun <- deltamethod_from_model(logipcw_fit, tau = tau)
pred_fun(newdata)
#>   prediction     lower     upper         se
#> 1  0.7406860 0.5672413 0.9141306 0.08849216
#> 2  0.5808678 0.4513883 0.7103472 0.06606092

Models using IPCW are also available. For these, we provide the jackknife-based estimation of the standard error and Wald confidence intervals based on the delta method on the logit scale.

# IPCW Logistic Regression
fit <- ipcw_logistic_regression(df,
  tau = tau, time_var = "time",
  status_var = "status"
)
predict(fit, newdata)
#>   prediction     lower     upper         se
#> 1  0.7477355 0.5387945 0.8826380 0.08960564
#> 2  0.5819723 0.4455359 0.7069205 0.06821779

# IPCW XGBoost Classifier
fit <- ipcw_xgboost(df, tau = tau, time_var = "time", status_var = "status")
predict(fit, newdata)
#>   prediction     lower     upper         se
#> 1  0.7654250 0.5621175 0.8924056 0.08545969
#> 2  0.5937184 0.4579942 0.7164942 0.06741503

Development

Restore the development environment with:

renv::restore()

Then use anything available in the devtools package to develop the package.

library(devtools)
document() # to update documentation and roxygen functionality
load_all() # to load the package functions for development
build_readme() # to update the README
test() # to run tests
check() # to check the package
covr::package_coverage() # to check code coverage
styler::style_pkg() # to style the code
lint() # to check the code for linting issues

References

Jahn-Eimermacher, Antje, Lukas Klein, and Gunter Grieser. 2025. “A Jackknife Approach to Estimate the Prediction Uncertainty from Binary Classifiers Under Right-Censoring.” Statistical Methods in Medical Research, ahead of print. https://doi.org/10.1177/09622802251393626.