hte3

hte3 provides R workflows for estimating conditional average treatment effects (CATE) and conditional risk ratios (CRR). Use the high-level wrappers for standard analyses, the GRF wrappers for forest-based fits, and the lower-level sl3 interface when you need direct control over nuisance learners or candidate HTE models.

Installation

if (!requireNamespace("remotes", quietly = TRUE)) {
  install.packages("remotes")
}

remotes::install_github("Larsvanderlaan/hte3", dependencies = TRUE)

Install grf separately if you want the causal-forest wrappers:

install.packages("grf")

Quickstart

library(hte3)
library(sl3)

data <- hte3_example_data(n = 150, seed = 1)

task <- hte_task(
  data = data,
  modifiers = c("W1", "W2", "W3"),
  confounders = c("W1", "W2", "W3"),
  treatment = "A",
  outcome = "Y",
  propensity_learner = Lrnr_mean$new(),
  outcome_learner = Lrnr_mean$new(),
  mean_learner = Lrnr_mean$new(),
  cross_fit = FALSE
)

fit <- fit_cate(
  task,
  method = "dr",
  base_learner = Lrnr_mean$new(),
  cross_validate = FALSE
)

pred <- predict(fit, data)
head(pred)
summary(fit)

Choose Your Workflow

High-level wrappers

Use hte_task(), fit_cate(), and fit_crr() when you want the standard package interface for difference-scale or risk-ratio targets.

GRF / causal-forest path

Use grf_cate() or grf_crr() when you want the optional grf engine and a wrapper that accepts raw data plus optional nuisance estimates.

Advanced / low-level sl3

Use make_hte3_Task_tx(), cross_validate_cate(), cross_validate_crr(), and the Lrnr_* classes when you want to assemble learner portfolios directly.

Where to Find Documentation

• Quickstart: one end-to-end example with the high-level wrappers.
• CATE Workflow: choosing among dr, r, t, and ep for conditional mean differences.
• CRR Workflow: fitting conditional risk ratios with fit_crr().
• Causal Forests with GRF: grf_cate(), grf_crr(), direct nuisance inputs, and GRF-specific arguments.
• Advanced sl3 Integration: low-level task construction, learner control, and manual cross-validation.
• Reference: exported functions, helper constructors, and learner classes.

Important Constraints

• fit_crr() and grf_crr() require a non-negative outcome.
• grf_cate() and grf_crr() currently support binary treatment only.
• Continuous-treatment CATE currently goes through the R-learner path via fit_cate(..., method = "r").