Reproduce “Comparison of meta-learners for estimating multi-valued treatment heterogeneous effects” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Naoufal Acharki, Ramiro Lugo, Antoine Bertoncello, Josselin Garnier
This paper rigorously investigates and compares meta-learners for estimating heterogeneous treatment effects in settings with multiple treatment levels, providing a novel theoretical analysis of how the number of treatments (K) and propensity score (r_min) affect estimator error bounds. Empirical validation on synthetic and semi-synthetic datasets demonstrates the generalized X-learner's superior robustness, while highlighting the M-learner's sensitivity and the R-learner's non-identifiability in this complex setting.
Key insights
- The number of treatment levels (K) significantly impacts meta-learner performance, with T- and M-learners being highly sensitive to decreasing sample sizes per treatment and the inverse of the propensity score lower bound, respectively.
- The generalized X-learner's error bound does not directly depend on the propensity score lower bound (r_min), indicating its theoretical robustness to poor overlap in multi-valued treatment settings.
- The generalized R-learner suffers from non-identifiability in the multi-treatment context, leading to non-unique solutions and poor practical performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.