Summarise policy-tree leaves with treatment-control contrasts and sample shares

Computes leaf-level summaries for a stored policy tree. Leaf contrasts are estimated from doubly robust action scores and use a fixed signed comparison: treatment minus control. Positive values favour treatment and negative values favour control. The legacy action-conditional advantage columns are retained as compatibility aliases.

Usage

margot_policy_leaf_summary(
  object,
  model_name,
  depth = 1L,
  weights = NULL,
  digits = 3L,
  ci_level = 0.95,
  label_mapping = NULL
)

Arguments

object: A margot_causal_forest()-style object containing results, covariates, and optionally weights.
model_name: Outcome/model name, with or without the model_ prefix.
depth: Integer policy-tree depth, usually 1 or 2.
weights: Optional evaluation weights. Defaults to object$weights.
digits: Integer; rounding used in formatted labels.
ci_level: Numeric confidence level for approximate row-level score intervals. Defaults to 0.95.
label_mapping: Optional named list used to label actions.

Value

A tibble with one row per leaf and columns for node id, action, unweighted count, weighted sample share, signed treatment-control contrast, approximate interval columns, legacy action-conditional advantage aliases, selected-action metadata, and policy-value contributions.

Details

Let $\Gamma_{ja}$ denote the action score for observation $j$ under action $a$, let $L$ denote a policy-tree leaf, and let $E_L$ denote the evaluation observations routed to that leaf. For binary actions $C$ and $T$, this helper reports the signed evaluation-sample contrast $$ \Delta_L = \frac{\sum_{j \in E_L} w_j\{\Gamma_{jT} - \Gamma_{jC}\}} {\sum_{j \in E_L} w_j}. $$ The selected action is reported separately. For the fitted tree, the selected action is the action stored in the terminal node, learned on the training observations $S_L$ routed to that node: $$ \pi(L) = \arg\max_{a \in \{C,T\}} \frac{\sum_{j \in S_L} w_j \Gamma_{ja}}{\sum_{j \in S_L} w_j}. $$ The reported $\Delta_L$ is computed on the evaluation rows and does not reselect the action from evaluation-row means. Between-leaf differences in $\Delta_L$ describe variation in magnitude. They are not the decision rule used by the policy tree.