OutcomeWeights

This R package calculates the outcome weights derived in Proposition 1 of Knaus (2024). Its use is illustrated in the vignettes of the package website.

The core functionality is the get_outcome_weights() method that can be called with objects of the package internal Double ML implementation and the grf package.

In the future it should be compatible with as many estimated R objects as possible.

The package can be downloaded from CRAN:

install.packages("OutcomeWeights")

The package is work in progress. Find here the current state (suggestions welcome):

In progress

• Compatibility with grf package
  • causal_forest() outcome weights for CATE
  • instrumental_forest() outcome weights CLATE
  • causal_forest() outcome weights for ATE from average_treatment_effect()
  • All outcome weights for average parameters compatible with average_treatment_effect()
• Package internal Double ML implementation handling the required outcome smoother matrices
  • Nuisance parameter estimation based on honest random forest (regression_forest() of grf package)
  • dml_with_smoother() function runs for PLR, PLR-IV, AIPW-ATE, and Wald_AIPW and is compatible with get_outcome_weights()
  • Add more Double ML estimators
  • Add support for more smoothers

Envisioned features

• Compatibility with DoubleML (this is a non-trivial task as the mlr3 environment it builds on does not provide smoother matrices)
  • Extract the smoother matrices of mlr3 available, where possible
  • Make the smoother matrices of mlr3 accessible within DoubleML
  • Write get_outcome_weights() method for DoubleML estimators
• Collect packages where weights could be extracted and implement them
  • lm()
  • lm_robust() of the estimatr package
  • ivreg() of the AER package

The development version is available using the devtools package:

library(devtools)
install_github(repo="MCKnaus/OutcomeWeights")

References

Knaus, M. C. (2024). Treatment effect estimators as weighted outcomes, arXiv:2411.11559