An Efficient Doubly-Robust Test for the Kernel Treatment Effect

About

The average treatment effect, which is the difference in expectation of the counterfactuals, is probably the most popular target effect in causal inference with binary treatments. However, treatments may have effects beyond the mean, for instance decreasing or increasing the variance. We propose a new kernel-based test for distributional effects of the treatment. It is, to the best of our knowledge, the first kernel-based, doubly-robust test with provably valid type-I error. Furthermore, our proposed algorithm is computationally efficient, avoiding the use of permutations.

Diego Martinez-Taboada, Aaditya Ramdas, Edward H. Kennedy · 2023

Related benchmarks

Task	Dataset	Result
Two-sample testing	Synthetic high-dimensional location-learning dataset n=3000 v1	Null Rate0.07	7
Distributional treatment effect testing	IHDP Scenario III Hill (test)	TPR0.34	4
Distributional treatment effect testing	IHDP Scenario IV Hill (test)	True Positive Rate53	4
Distributional treatment effect testing	IHDP Scenario VI Hill (test)	TPR3	4
Distributional treatment effect testing	IHDP Scenario V Hill (test)	True Positive Rate0.99	4
Distributional treatment effect testing	IHDP Scenario II Hill (test)	TPR44	4
Hypothesis Testing	Synthetic Null setting	Empirical Rejection Rate (n=300)7.5	2
Hypothesis Testing	Synthetic Mean shift alternative	Empirical Rejection Rate (N=300)0.155	2
Hypothesis Testing	Synthetic Variance shift alternative	Empirical Rejection Rate (n=300)6.5	2
Hypothesis Testing	Synthetic Localized bump alternative	Empirical Rejection Rate (N=300)0.035	2