3. Sampling weights and clustering

3.1. Sampling weights

Set gen_weighted and provide var_w_name to use sampling weights. The weight column is required in both training and prediction data. These weights are distinct from the forest weights used to estimate potential outcomes.

3.2. Clustering

Set p_cluster_std and supply var_cluster_name for cluster-robust standard errors. gen_panel_data also enables clustered inference and requires a cluster identifier. A row identifier (var_id_name) identifies an observation; a cluster identifier groups related observations, such as repeated measurements of the same person.

With panel mode enabled, gen_panel_in_rf samples clusters within each tree’s construction subsample. The initial construction/filling split and forest chunking still operate on rows, so clusters do not necessarily remain intact through every split. Bootstrap defaults depend on clustering; see Inference.

3.3. Parameter overview

Parameter	Purpose
var_w_name, gen_weighted	Sampling-weight column and activation.
var_id_name	Observation identifier.
var_cluster_name, p_cluster_std	Cluster identifier and cluster-robust inference.
gen_panel_data	Enables panel mode and cluster-robust inference.
gen_panel_in_rf	Cluster sampling within trees when panel mode is enabled.

3.4. Example

This configuration fragment assumes your data contain row_id and person_id, it does not create these identifiers.

from mcf.mcf_main import ModifiedCausalForest

my_mcf = ModifiedCausalForest(
    var_y_name='outcome', var_d_name='treat',
    var_x_name_ord=['age', 'income'],
    var_id_name='row_id', var_cluster_name='person_id',
    p_cluster_std=True
)

For sampling weights, additionally set gen_weighted and var_w_name='weight' and supply that column in both datasets. See the API for supported combinations.