Fit development-only forest nuisances for policy action scores

Fits outcome, exposure and causal forests on development participants. Uses out-of-bag predictions for development scores and development-trained predictions for evaluation scores. Evaluation outcomes enter only the latter residual corrections. Preparation and supplied weights remain the caller's declared inputs.

Usage

margot_policy_development_scores(
  development_X,
  development_Y,
  development_W,
  evaluation_X,
  evaluation_Y,
  evaluation_W,
  development_weights = NULL,
  forest_args = list(num.trees = 2000),
  seed = 42L,
  num_threads = 1L,
  save_models = FALSE
)

Arguments

development_X, evaluation_X

Numeric covariate matrices with identical named columns, already prepared on compatible scales.

development_Y, evaluation_Y

Oriented outcome vectors.

development_W, evaluation_W

Binary exposure vectors.

development_weights

Positive development analysis weights, or `NULL` for equal weights. Evaluation weights enter the later policy evaluator.

forest_args

Named list of causal-forest settings. Settings also accepted by `grf::regression_forest` are used for the nuisance forests. Data, nuisance predictions, seeds and thread counts cannot be overridden.

seed

Integer seed. The three forests use this seed plus zero, one and two respectively.

num_threads

Positive integer native thread count, default one.

save_models

Whether to retain the three fitted forests.

Value

A list of development and evaluation action scores and predictions, requested forest settings, per-forest seeds and a preparation qualification. Optionally includes the development-trained forests.

Details

This supplies a nuisance-estimation boundary, rather than verifying causal identification or removing bias in the supplied preparation or analysis weights. Pointwise policy-score intervals require appropriate nuisance rates and sampling assumptions. Propensities at zero or one cause an error; the function does not select a clipping rule from the results.