estimatrTools: Data-Driven Covariate Adjustment for the 'estimatr' Estimators

Companion tools for 'estimatr'. Selects covariates by LASSO within each treatment arm and on each treatment-arm indicator (post-double selection, following Belloni, Chernozhukov and Hansen (2014) doi:10.1093/restud/rdt044 ), then fits the selected set with the Lin (2013) interacted estimator doi:10.1214/12-AOAS583 , an additive robust regression, or a treatment-by-moderator interaction model. Each fitting function falls back to an unadjusted specification when selection returns nothing or the adjusted fit is degenerate, so a fit is always returned and the fallback is reported rather than hidden.

Selecting covariates

lasso_select_covariates chooses a covariate set by running LASSO of the outcome on candidates within each treatment arm, and of each arm indicator on candidates. The first is what makes the selection appropriate for an estimator with a separate slope per arm; the second is the double-selection step that retains covariates predicting assignment even when they barely predict the outcome.

Fitting

lm_lin_lasso fits the Lin (2013) interacted estimator on the selected set, lm_robust_lasso fits them additively, and lm_moderator_lasso fits a treatment-by-moderator model with the selected covariates entered additively alongside.

Knowing what actually ran

Each fitting function falls back to a simpler specification when selection returns nothing or the adjusted fit is degenerate. This is usually the right response, but it changes what the estimate is, so it is recorded rather than hidden. How to read it back depends on what you still have:

• one fit in hand: adjustment, selected_covariates, fallback_reason

• a list of fits: fallback_summary

• fits that were built and discarded, which is the usual case inside a map() pipeline: fallback_log

The last matters more than it sounds. Selection returning nothing is common, not exceptional, and when it happens the "adjusted" estimate is an unadjusted difference in means. A pipeline that never reads the log has no way of knowing what fraction of its estimates that describes.

See also

Useful links:

• https://alexandercoppock.com/estimatrTools/

• https://github.com/acoppock/estimatrTools

• Report bugs at https://github.com/acoppock/estimatrTools/issues

Author

Maintainer: Alexander Coppock acoppock@gmail.com

Authors:

• Alexander Coppock acoppock@gmail.com