Construct binary action scores from independent nuisance predictions

Computes unweighted doubly robust control and treatment scores. The caller supplies predictions estimated without the scored observation's outcome; evaluation predictions must respect the policy-development boundary.

Usage

margot_policy_action_scores(
  outcome,
  treatment,
  outcome_mean,
  propensity,
  treatment_effect
)

Arguments

outcome

Numeric outcome vector, already on the intended oriented scale.

treatment

Binary numeric vector, zero for control and one for treatment.

outcome_mean

Predicted conditional outcome mean under the observed exposure distribution.

propensity

Predicted treatment probabilities strictly between zero and one. No truncation or clipping is performed.

treatment_effect

Predicted conditional treatment-minus-control effect.

Value

A two-column numeric matrix named `control` and `treated` on the supplied outcome scale.

Details

Conditional action means are recovered as `outcome_mean - propensity * treatment_effect` for control and `outcome_mean + (1 - propensity) * treatment_effect` for treatment. The observed action's residual is corrected by its inverse propensity. Correct interpretation requires the causal identification and nuisance-estimation conditions for the supplied design. This function verifies numerical inputs, not independence or those conditions. It applies neither analysis weights nor a benefit threshold. Apply each once in the subsequent learning and evaluation procedure.