Causal Effect Inference for Structured Treatments

About

We address the estimation of conditional average treatment effects (CATEs) for structured treatments (e.g., graphs, images, texts). Given a weak condition on the effect, we propose the generalized Robinson decomposition, which (i) isolates the causal estimand (reducing regularization bias), (ii) allows one to plug in arbitrary models for learning, and (iii) possesses a quasi-oracle convergence guarantee under mild assumptions. In experiments with small-world and molecular graphs we demonstrate that our approach outperforms prior work in CATE estimation.

Jean Kaddour, Yuchen Zhu, Qi Liu, Matt J. Kusner, Ricardo Silva · 2021

Related benchmarks

Task	Dataset	Result
pancytopenia onset prediction	Claims	Rate0	60
Causal effect prediction	Medical claims pancytopenia occurrence (test)	Rate @ 0.9990	15
CATE estimation	Claims dataset (test)	Rate @ 0.9990	15
Causal effect prediction	LINCS 20 DEGs (unseen perturbation/cell-line)	PEHE4.06	8
Causal effect prediction	LINCS 50 DEGs (unseen perturbation cell-line)	PEHE3.78	8
CATE estimation	SW (In-sample)	WPEHE@623	5
CATE estimation	SW (Out-sample)	WPEHE@623.19	5
CATE estimation	TCGA (In-sample)	WPEHE@610.98	5
CATE estimation	TCGA (Out-sample)	WPEHE@68.15	5