Skip to contents

Recompute Average Treatment Effects with Different Target Samples

Source: R/margot_recompute_ate.R
margot_recompute_ate.Rd

This function allows recomputation of average treatment effects (ATEs) from fitted causal forest models using different target samples or estimation methods without refitting the models. This is useful for sensitivity analysis and exploring treatment effects in different subpopulations.

Usage

margot_recompute_ate(
  causal_forest_output,
  target_sample = c("all", "treated", "control", "overlap"),
  method = "AIPW",
  scale = c("RD", "RR"),
  delta = 1,
  sd = 1,
  subset = NULL,
  respect_train_test_split = TRUE
)

Arguments

causal_forest_output

Output from `margot_causal_forest()` or `margot_causal_forest_parallel()`. Must have been run with `save_models = TRUE`.

target_sample

Character string specifying the target sample for ATE estimation. Options are:

  • "all" - All units (default in original models)

  • "treated" - Only treated units (ATT)

  • "control" - Only control units (ATC)

  • "overlap" - Units with propensity scores away from 0 and 1

method

Character string specifying the estimation method. Currently only "AIPW" (Augmented Inverse Probability Weighting) is supported by grf.

scale

Character string specifying the scale of the estimate. Either "RD" (risk difference) or "RR" (risk ratio). Default is "RD".

delta

The hypothesized increase in outcome for RD scale E-value calculations. Default is 1.

sd

The standard deviation of the outcome for RD scale E-value calculations. Default is 1.

subset

Optional logical vector for subsetting the data. Default is NULL.

respect_train_test_split

Logical. If TRUE and the original analysis used use_train_test_split=TRUE, computes treatment effects on the test set only for consistency with policy evaluation. Default is TRUE.

Value

A list with the same structure as the original causal forest output, but with:

  • Updated `ate` values in each model's results

  • Recomputed `custom_table` with new E-values and renamed effect columns (ATE/ATT/ATC/ATO)

  • Updated `combined_table` with all new estimates

  • A new element `ate_params` documenting the parameters used

Details

This function requires that the original causal forest models were saved (using `save_models = TRUE`). It extracts the fitted models and recomputes ATEs using the specified target sample.

The different target samples represent different estimands: - "all": Average Treatment Effect (ATE) for the entire population - "treated": Average Treatment Effect on the Treated (ATT) - "control": Average Treatment Effect on the Control (ATC) - "overlap": ATE for units with good overlap in propensity scores

Examples

if (FALSE) { # \dontrun{
# First fit causal forest models with saved models
cf_results <- margot_causal_forest(
  data = mydata,
  outcome_vars = outcomes,
  covariates = X,
  save_models = TRUE
)

# Later, get ATEs for different target populations
ate_overlap <- margot_recompute_ate(cf_results, target_sample = "overlap")
ate_treated <- margot_recompute_ate(cf_results, target_sample = "treated")

# Compare estimates
cf_results$combined_table # Original (all units)
ate_overlap$combined_table # Overlap units only
ate_treated$combined_table # Treated units only
} # }