Kernel-Weighted Cox Regression for Treatment Effect Heterogeneity

Explores treatment effect heterogeneity and candidate predictive biomarkers by re-fitting weighted Cox proportional hazards models on a biomarker grid. Kernel weights centred at each grid point produce local coefficient estimates that can be visualised across biomarker space. Builds on ideas related to graphical Cox treatment-covariate interaction methods [see Bonetti and Gelber (2004) and local partial-likelihood approaches Fan, Lin and Zhou (2006) ].

chestR

Kernel-weighted Cox regression for exploring treatment effect heterogeneity and candidate predictive biomarkers.

Installation

# install.packages("devtools")
devtools::install_github("richJJackson/chestR")

Or install from a local checkout:

devtools::install("path/to/chestR")

Quick start

library(survival)
library(chestR)

# Fit a global Cox model
base <- coxph(Surv(time, status) ~ treatment + covariate, data = mydata)

# Local estimates over a biomarker grid
cr <- chestr(base, mydata[, c("biom1", "biom2")], grid.size = 25,
             treat_term = "treatment")

# Visualise local treatment effect
plot(cr, trt.param = "treatment")

# Optional permutation test of constant treatment effect
# tst <- chestr_test(cr, B = 99, seed = 1)

See vignette("chestr-workflow", package = "chestR") after install, or inst/examples/simulation.R for a longer simulation script.

Development

Open chestR.Rproj in RStudio, then:

devtools::load_all()
devtools::test()
devtools::document()
devtools::check()

Related work

• Bonetti M, Gelber RD (2000). Statistics in Medicine.
• Liu Y, Lu W, Chen G (2015). Local partial-likelihood test. Statistics in Medicine.

License

MIT