Skip to contents

Run Multiple Generalized Random Forest (GRF) Causal Forest Models with Enhanced Qini Cross-Validation

Source: R/margot_causal_forest.R
margot_causal_forest.Rd

This function runs multiple GRF causal forest models with enhanced features. In addition to estimating causal effects, it can compute the Rank-Weighted Average Treatment Effect (RATE) for each model. It also gives you the option to train a separate "Qini forest" on a subset of data and compute Qini curves on held-out data, thereby avoiding in-sample optimism in the Qini plots.

Usage

margot_causal_forest(
  data,
  outcome_vars,
  covariates,
  W,
  weights,
  grf_defaults = list(),
  save_data = FALSE,
  compute_rate = TRUE,
  top_n_vars = 15,
  save_models = TRUE,
  train_proportion = 0.5,
  train_prop = NULL,
  qini_train_prop = NULL,
  compute_conditional_means = TRUE,
  compute_marginal_only = FALSE,
  verbose = TRUE,
  qini_treatment_cost = 1,
  seed = 12345,
  use_train_test_split = TRUE,
  flip_outcomes = NULL,
  flip_method = "zscore",
  flip_scale_bounds = NULL
)

Arguments

data

A data frame containing all necessary variables.

outcome_vars

A character vector of outcome variable names to be modelled.

covariates

A matrix of covariates to be used in the GRF models.

W

A vector of binary treatment assignments.

weights

A vector of weights for the observations.

grf_defaults

A list of default parameters for the GRF models.

save_data

Logical indicating whether to save data, covariates, and weights. Default is FALSE.

compute_rate

Logical indicating whether to compute RATE for each model. Default is TRUE. Note: Direct computation of RATE, QINI, and policy trees within this function may be deprecated in future versions. Use margot_rate(), margot_qini(), and margot_policy_tree() instead.

top_n_vars

Integer specifying the number of top variables to use for additional computations. Default is 15.

save_models

Logical indicating whether to save the full GRF model objects. Default is TRUE.

train_proportion

Numeric value between 0 and 1 indicating the proportion of data to use for training set, or NULL. Used for policy trees and QINI evaluation when compute_marginal_only=FALSE. Must be NULL when compute_marginal_only=TRUE. Default is 0.5.

train_prop

Deprecated. Use train_proportion instead.

qini_train_prop

Deprecated. Use train_proportion instead.

compute_conditional_means

Logical indicating whether to compute conditional means using policytree::conditional_means(). These represent expected outcomes under each treatment arm. Default is TRUE.

compute_marginal_only

Logical indicating whether to compute only marginal (average) treatment effects, skipping heterogeneity metrics (QINI curves, RATE, policy trees). When TRUE, train_proportion must be NULL. When FALSE (default), heterogeneity metrics are computed using train/test splits.

verbose

Logical indicating whether to display detailed messages during execution. Default is TRUE.

qini_treatment_cost

Scalar treatment cost per unit for QINI calculations. Default 1. Lower values (e.g., 0.2) represent cheap treatments creating steeper QINI curves; higher values (e.g., 5) represent expensive treatments creating shallower curves.

seed

Integer. Random seed for reproducibility of train/test splits for policy trees and QINI evaluation. Default is 12345.

use_train_test_split

Logical. If TRUE (default), uses a consistent train/test split where: - The main causal forest is trained on all data (following GRF best practices for honesty) - All reported results (ATE, E-values, combined_table) are computed on the TEST SET - Policy trees and QINI also use the same train/test split - All-data results are stored in split_info for reference If FALSE, the main forest uses all data for estimation and policy trees/QINI get their own splits.

flip_outcomes

Character vector or list specifying outcomes to flip (reverse score). - Character vector: applies default flip_method and flip_scale_bounds to all listed outcomes - List: allows per-outcome specification, e.g., list(anxiety = list(method = "ordinal", scale_bounds = c(1, 7))) Flipped outcomes are appended with "_r" suffix in results.

flip_method

Character string specifying default inversion method when flip_outcomes is provided: - "zscore": Simple negation (default). Assumes data is already standardized (e.g., outcomes ending in "_z"). - "ordinal": Invert on ordinal scale using bounds. Use for Likert scales or bounded measures.

flip_scale_bounds

Numeric vector of length 2 [min, max] or named list of bounds per outcome. Required when flip_method = "ordinal" unless bounds should be inferred from data. Example: c(1, 7) for a 7-point Likert scale, or list(anxiety = c(1, 7), depression = c(0, 21)).

Value

A list containing: * `results` - per-outcome diagnostics and objects * `combined_table` - combined e-value table across outcomes * `outcome_vars` - vector of outcome names * `not_missing` - indices of complete-case rows * (`data`, `covariates`, `weights`, `W`) when `save_data = TRUE` * `full_models` when `save_models = TRUE` * `flip_info` when `flip_outcomes` is used - contains mapping and metadata for flipped outcomes