Skip to content

Comparing glmnet/RF vs balnet for doubly robust debiasing weights #33

Description

@erikcs

For estimating an ATE under unconfoundedness, the debiasing/Riesz weights are simply the difference in IPW weights for the treated and control

$$ \frac{W_i}{e(X_i)} - \frac{1 - W_i}{1 - e(X_i)}. $$

Standard approaches uses off-the-shelf "ML algorithms", like glmnet, that targets e(X). balnet targets the two ratios directly (but separately!), and should in theory do better in estimating these debiasing weights. It could be nice with simulation example(s) showing exactly how much there is to gain.

For comparing against estimators beyond glmnet, like forests: can simply pass these weights to the debiasing.weights argument of grf's average_treatment_effect (can maybe add an example of this to the grf docs).

I guess the DR-learner could also improve with balnet's balancing weights, though wonder how much you actually gain in practice since cross-fitting kills the fit balance.

Metadata

Metadata

Assignees

No one assigned

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions