Repeated Split Diagnostic for Policy Trees — margot_policy_split

Repeatedly fits policy trees on a training split and evaluates their action summaries on the complementary split. This diagnostic describes instability in the policy-learning procedure; it is not the displayed full-data tree.

Usage

margot_policy_split_diagnostic(
  object,
  model_names = NULL,
  depths = c(1L, 2L),
  n_splits = 50L,
  train_proportion = 0.5,
  seed = 12345,
  covariate_mode = c("original", "custom", "add", "all"),
  custom_covariates = NULL,
  exclude_covariates = NULL,
  tree_method = c("fastpolicytree", "policytree"),
  min_node_size = NULL,
  weights = NULL,
  verbose = TRUE
)

Arguments

object: A list returned by margot_causal_forest with saved covariates, not_missing, and per-model dr_scores.
model_names: Optional model names, with or without the model_ prefix. Defaults to all models.
depths: Integer vector containing one or two permitted branching levels. Defaults to c(1L, 2L); these are not node sizes.
n_splits: Integer. Number of repeated train/test splits. Defaults to 50.
train_proportion: Numeric in (0,1). Fraction assigned to the training split. Defaults to 0.5.
seed: Integer or NULL. Seed for reproducibility. Defaults to 12345.
covariate_mode: Character; one of "original", "custom", "add", or "all". Defaults to "original".
custom_covariates: Optional character vector used with "custom" or "add".
exclude_covariates: Optional character vector of exact names or patterns to exclude from the policy-tree covariates.
tree_method: Character; "fastpolicytree" when available, otherwise "policytree".
min_node_size: Integer or NULL. Smallest permitted policy-tree terminal node. It is separate from depths and causal-forest node size. NULL retains the compatibility option fallback, then 1.
weights: Optional weights for held-out summaries. Defaults to object$weights.
verbose: Logical; print progress messages. Defaults to TRUE.

Value

A tibble with one row per model, split, and depth. The returned object has class margot_policy_split_diagnostic and a summary attribute with mean and standard deviation of key metrics.