Summarise Policy-Tree Value and Oracle Regret — margot_policy_regret

Computes policy-tree value contrasts from stored doubly robust action scores. The summary compares a learned policy tree with universal control, universal treatment, and an oracle that chooses the better action score for each observation. The oracle is an optimistic benchmark for describing remaining model-implied heterogeneity; it is not a deployable policy.

Usage

margot_policy_regret_summary(
  object,
  model_names = NULL,
  depths = c(1L, 2L),
  weights = NULL,
  R = 499L,
  seed = 42L,
  ci_level = 0.95,
  label_mapping = NULL
)

policy_regret_summary(
  object,
  model_names = NULL,
  depths = c(1L, 2L),
  weights = NULL,
  R = 499L,
  seed = 42L,
  ci_level = 0.95,
  label_mapping = NULL
)

Arguments

object: A list returned by margot_causal_forest, margot_policy_tree, or a compatible policy-tree object with results and either top-level covariates or per-model plot_data.
model_names: Optional character vector of model names to include, with or without the model_ prefix. Defaults to all available models.
depths: Integer vector of tree depths to evaluate. Defaults to c(1, 2).
weights: Optional numeric vector of evaluation weights. If NULL, the function uses object$weights when available.
R: Integer. Bootstrap replicates for uncertainty intervals. Defaults to 499.
seed: Integer or NULL. Seed for bootstrap reproducibility.
ci_level: Numeric confidence level for bootstrap percentile intervals. Defaults to 0.95.
label_mapping: Optional named character vector for display labels.

Value

A tibble with one row per model and depth, including policy value, universal-action values, oracle value, gains relative to control-all and treat-all, and regret relative to the oracle benchmark.

Examples

if (FALSE) { # \dontrun{
policy_regret_summary(causal_forest_results, depths = 2)
margot_policy_regret_summary(policy_tree_results, model_names = "model_y")
} # }