Lrnr_cate_EP Class

This class defines the EP-learner of van der Laan et al. (2023) for estimation of the conditional average treatment effect.

Format

An R6 class with public methods to initialize the learner, create a regression task, and access the base learner.

Details

The EP-learner is a robust and doubly-robust meta-learner that inherits desirable properties of both T-learner and DR-learner. The current implementation exposes two CATE EP variants through targeting_style: "dr" is the plug-in analogue of the DR-learner and "r" is the plug-in analogue of the R-learner. The "r" variant uses an overlap-weighted targeting objective that can be more stable under severe overlap violations, but it generally targets an overlap-weighted projection in reduced-modifier settings.

Super classes

sl3::Lrnr_base -> hte3::Lrnr_hte -> Lrnr_cate_EP

Methods

Public methods

• Lrnr_cate_EP$new()

• Lrnr_cate_EP$get_pseudo_data()

• Lrnr_cate_EP$clone()

Inherited methods

• sl3::Lrnr_base$assert_trained()
• sl3::Lrnr_base$base_chain()
• sl3::Lrnr_base$base_predict()
• sl3::Lrnr_base$base_train()
• sl3::Lrnr_base$chain()
• sl3::Lrnr_base$custom_chain()
• sl3::Lrnr_base$get_outcome_range()
• sl3::Lrnr_base$get_outcome_type()
• sl3::Lrnr_base$predict()
• sl3::Lrnr_base$predict_fold()
• sl3::Lrnr_base$print()
• sl3::Lrnr_base$process_formula()
• sl3::Lrnr_base$reparameterize()
• sl3::Lrnr_base$retrain()
• sl3::Lrnr_base$sample()
• sl3::Lrnr_base$set_train()
• sl3::Lrnr_base$subset_covariates()
• sl3::Lrnr_base$train()
• sl3::Lrnr_base$train_sublearners()
• hte3::Lrnr_hte$check_treatment_type()
• hte3::Lrnr_hte$get_modifiers()
• hte3::Lrnr_hte$make_metalearner_task()

---

Method new()

Usage

Lrnr_cate_EP$new(
  base_learner,
  sieve_num_basis = NULL,
  sieve_interaction_order = 3,
  screen_basis_with_lasso = FALSE,
  targeting_style = c("dr", "r"),
  r_targeting_basis = c("v_plus_propensity", "full_w"),
  treatment_level = NULL,
  control_level = NULL,
  ...
)

Arguments

base_learner

A sl3 learner object inheriting from Lrnr_base that specifies the base supervised learning algorithm used by the meta-learner.

sieve_num_basis

The number of trigonometric basis functions to construct the EP-learner sieve space. This argument is passed as the argument basisN to the function sieve_preprocess. By default, sieve_num_basis = ceiling((n^(1/3)*d) where n denotes the sample size and d is the dimension of the first-stage EP sieve covariates. For wrapper-level EP cross-validation, use sieve_basis_grid in fit_cate() or fit_crr() to compare multiple basis sizes.

sieve_interaction_order

The maximum interaction degree of tensor-product basis functions in the EP-learner sieve basis. Default is 3. sieve_interaction_order = 1 corresponds to an additive sieve model and sieve_interaction_order = 2 corresponds to a bi-additive sieve model. This argument is passed as the argument interaction_order to the function sieve_preprocess.

screen_basis_with_lasso

EXPERIMENTAL. Whether to use a lasso-based DR-learner algorithm to screen sieve basis functions for EP-learner. There are no theoretical guarantees for EP-learner with screen_basis_with_lasso = TRUE. However, this argument may be useful in high dimensional settings. It also reduces the computational complexity of EP-learner, as the argument sieve_num_basis does not need to be externally cross-validated.

targeting_style

One of "dr" or "r". The default "dr" path keeps the current EP debiasing update. The "r" path uses an R-learner-style overlap-weighted targeting update and currently supports binary-treatment CATE tasks only. High-level wrappers such as fit_cate() compare both variants by expanding multiple Lrnr_cate_EP fits rather than by passing a vector here.

r_targeting_basis

First-stage basis construction for targeting_style = "r". The default "v_plus_propensity" builds the first-stage sieve on the modifier set V plus the treated-arm propensity score e(W); "full_w" builds the first-stage EP-R sieve on the full confounder set W.

...

Additional arguments to pass to the initialization function.

---

Method get_pseudo_data()

Usage

Lrnr_cate_EP$get_pseudo_data(
  hte3_task,
  treatment_level = NULL,
  control_level = NULL,
  train = TRUE,
  ...
)

Arguments

...

Additional arguments to pass to the initialization function.

---

Method clone()

The objects of this class are cloneable with this method.

Usage

Lrnr_cate_EP$clone(deep = FALSE)

Arguments

deep

Whether to make a deep clone.