[Paper Review] Nonparametric Estimation of Heterogeneous Treatment Effects: From Theory to Learning Algorithms

Alicia Curth, Mihaela van der Schaar|arXiv (Cornell University)|Jan 26, 2021

Advanced Causal Inference Techniques33 references31 citations

TL;DR

The paper provides a theoretical classification and analysis of four meta-learners for CATE estimation (one-step plug-in and three two-step learners: RA, PW, DR), and evaluates neural-network architectures for nuisance estimation across synthetic and IHDP-like data.

ABSTRACT

The need to evaluate treatment effectiveness is ubiquitous in most of empirical science, and interest in flexibly investigating effect heterogeneity is growing rapidly. To do so, a multitude of model-agnostic, nonparametric meta-learners have been proposed in recent years. Such learners decompose the treatment effect estimation problem into separate sub-problems, each solvable using standard supervised learning methods. Choosing between different meta-learners in a data-driven manner is difficult, as it requires access to counterfactual information. Therefore, with the ultimate goal of building better understanding of the conditions under which some learners can be expected to perform better than others a priori, we theoretically analyze four broad meta-learning strategies which rely on plug-in estimation and pseudo-outcome regression. We highlight how this theoretical reasoning can be used to guide principled algorithm design and translate our analyses into practice by considering a variety of neural network architectures as base-learners for the discussed meta-learning strategies. In a simulation study, we showcase the relative strengths of the learners under different data-generating processes.

Motivation & Objective

• Assess conditions under which different meta-learners for CATE estimation perform better.
• Classify CATE meta-learners into meaningful categories to guide algorithm design.
• Investigate neural-network architectures for nuisance estimation within meta-learners.
• Provide practical implementations and simulations to compare learners under various data-generating processes.

Proposed method

• Introduce a four-class taxonomy of meta-learners: one-step plug-in and three two-step learners (RA, PW, DR).
• Derive theoretical error bounds and remainders for each learner under nonparametric regression with smoothness assumptions.
• Analyze asymptotic and finite-sample behavior, highlighting when DR and PW may achieve oracle rates.
• Propose and evaluate neural-network based nuisance estimators, including a general SNet architecture that unifies existing SNet variants and TNet.
• Provide public code implementations for CATENets and related architectures.

Experimental results

Research questions

• RQ1 What are the theoretical properties (rates and remainder terms) of RA-, PW-, and DR- meta-learners compared to plug-in learners?
• RQ2 Under which data-generating conditions do DR- and PW-learners outperform plug-in or RA-learners asymptotically and in finite samples?
• RQ3 How do neural-network architectures for nuisance estimation affect CATE meta-learner performance?
• RQ4 Can a unified SNet architecture outperform specialized SNet variants and TNet across DGPs?
• RQ5 How do overlap, confounding, and treatment effect sparsity influence estimator performance?

Key findings

• DR-learner tends to achieve better finite-sample performance under confounding with no treatment effect.
• RA-learner can outperform others when there is confounding and a non-trivial treatment effect.
• PW-learner generally exhibits high variance and poor performance in the settings considered.
• One-step plug-in estimators can match two-step learners asymptotically when nuisance functions are similarly complex.
• A general SNet architecture that shares representations can outperform or match specialized architectures, depending on the DGP and sample size.
• Two-step learners may incur higher variance in small samples, making RA- and DR-learner advantageous mainly in larger samples or appropriate regimes.