Causal Forests with GRF

Read this article if you want the optional grf wrappers rather than the standard hte_task() plus fit_*() workflow. The GRF path keeps the same hte3_model output interface, but it uses grf as the forest engine and accepts raw data plus optional nuisance estimates directly.

library(hte3)

When to Use This Workflow

Use the GRF wrappers when all of the following are true:

• you want a forest-based workflow built around the optional grf package
• you want to call a wrapper directly on raw data instead of building hte_task() first
• you still want an hte3_model that works with predict() and summary()

Choose the Wrapper from the Estimand

• Use grf_cate() when the target is a conditional mean difference.
• Use grf_crr() when the target is a conditional risk ratio and the outcome is non-negative.

Choose the CATE Method

grf_cate() supports method = c("dr", "r", "ep").

• method = "r" is the direct causal-forest path and trains grf::causal_forest().
• method = "dr" uses the DR-learner with GRF nuisance and base learners.
• method = "ep" uses the EP-learner with GRF nuisance and base learners.

Within the EP family, ep_targeting_style = "dr" selects EP-DR and ep_targeting_style = "r" selects EP-R. If you pass ep_targeting_style = c("dr", "r"), wrapper-level cross-validation can compare both EP variants.

If you omit both method and cross_validate, grf_cate() compares the default EP-R candidate and the direct GRF R-learner.

data <- hte3_example_data(n = 200, seed = 1)

fit <- grf_cate(
  data = data,
  modifiers = c("W1", "W2"),
  confounders = c("W1", "W2", "W3"),
  treatment = "A",
  outcome = "Y",
  method = "r",
  cross_fit = FALSE
)

head(predict(fit, data))
summary(fit)

grf_crr()

grf_crr() supports method = c("ep", "ipw", "t"). If you omit method, the wrapper fits one EP model.

fit <- grf_crr(
  data = data,
  modifiers = c("W1", "W2"),
  confounders = c("W1", "W2", "W3"),
  treatment = "A",
  outcome = "Y_binary",
  method = "ep",
  cross_fit = FALSE
)

head(predict(fit, data))
summary(fit)

Direct Nuisance Inputs

The GRF wrappers accept direct nuisance estimates:

• mu.hat: an n x 2 matrix ordered as (control, treatment)
• pi.hat: a length-n vector for P(A = treatment_level | X)
• m.hat: an optional length-n vector of marginal outcome means

If m.hat is omitted and both mu.hat and pi.hat are supplied, hte3 computes m.hat internally.

mu.hat <- cbind(data$mu0, data$mu1)
pi.hat <- data$pi1

fit <- grf_cate(
  data = data,
  modifiers = c("W1", "W2"),
  confounders = c("W1", "W2", "W3"),
  treatment = "A",
  outcome = "Y",
  mu.hat = mu.hat,
  pi.hat = pi.hat,
  cross_fit = FALSE
)

Important Arguments

The GRF-specific arguments worth understanding are:

• cross_fit and folds for nuisance estimation
• cross_validate and cv_control for outer learner selection
• ep_targeting_style and ep_r_targeting_basis for the GRF-backed EP variants
• treatment_level and control_level when treatment labels are not the default 1 and 0
• grf_params for GRF hyperparameters such as num.trees or min.node.size
• tune for the amount of GRF tuning: "light", "none", or "all"

In the GRF learners, tune = "light" tunes sample.fraction, mtry, and min.node.size with a modest search budget. tune = "none" disables tuning, and tune = "all" requests GRF’s full tuning path.

Important Constraints

• grf_cate() and grf_crr() currently support binary treatment only.
• grf_crr() requires a non-negative outcome.
• If you want the standard hte_task() plus fit_*() interface or continuous-treatment CATE, use the non-GRF workflow instead.

Helper Constructors

When the wrappers are too narrow, these helpers expose the forest-backed building blocks directly:

• make_grf_base_learner()
• make_grf_nuisance_learners()
• make_grf_cate_learners()
• make_grf_crr_learners()

Use these when you want manual cross-validation or a custom sl3 learner library rather than the default wrapper behavior.