LOOL: Flexible & Robust Estimator of Heterogeneous Treatment Effects

Pham, Duy

Etd

LOOL: Flexible & Robust Estimator of Heterogeneous Treatment Effects

Öffentlich Deposited

This thesis presents a two-stage approach – dubbed “Leave-OneOut Learner” (LOOL) – for estimating heterogeneous treatment effects – with the conditional average treatment effect (CATE) as the estimand of interest, as an extension of the “Leave-One-Out Potential Outcomes” (LOOP) Estimator proposed by Wu and Gagnon-Bartsch (2018). Researchers can first obtain unbiased estimates of the individual treatment effects (ITE) from the LOOP estimator and can then estimate the CATE by regressing on these estimated ITE. This estimator is robust as it guarantees unbiased estimates of the ITE in the first stage. It is also flexible since researchers can use any off-the-shelf machine learning methods for either stage without additional modification. Moreover, researchers can utilize parametric approaches like simple least-squares for drawing inferences on the impacts of each covariate or leverage more flexible non-parametric ones for more accurate predictions of the effects. We first compared this estimator to the meta-learner algorithms proposed by Kunzel et al. (2019) in initial simulation studies, where its performance is competitive. We then implemented and tested this method on data from a study examining the effectiveness of three educational platforms in teaching middle school algebra. The applied portion of this thesis is adapted from a paper in submission at the Seventeenth International Conference on Educational Data Mining (EDM 2024) by Duy Pham (the thesis’s author), Kirk Vanacore, Adam Sales (the thesis’s advisor), and Johann Gagnon-Bartsch.

Creator