Causal Representation Learning for Heterogeneous Treatment Effect Estimation

Abstract

Estimating Heterogeneous Treatment Effects (HTEs) is challenging, primarily due to issues related to confounding factors, especially unobserved confounders and selection bias. This thesis proposes novel methodologies to enhance the accuracy and robustness of Conditional Average Treatment Effect (CATE) estimation through advanced representation learning techniques.

First, we introduce DIGNet, an innovative representation balancing method designed to mitigate the inherent trade-off problem between factual outcome prediction and distribution balancing. By deriving theoretical upper bounds based on H-divergence, we establish a crucial connection between propensity scores and representation balancing. DIGNet incorporates two key components that work synergistically to improve CATE estimation through complementary pathways: Patterns Decomposed with Individual propensity confusion and Group distance minimization (PDIG), and Patterns of Pre-balancing and Balancing Representations (PPBR). PDIG transforms the balancing patterns into a decomposition that captures more effective balancing representations by focusing on individual propensity confusion and group distance minimization, thereby enhancing the model's ability to learn balanced representations across treated and control groups. Complementarily, PPBR decomposes features into pre-balancing and balancing representations, preserving patterns beneficial for outcome modeling while facilitating the learning of balanced representations. This dual focus allows the model to maintain strong predictive power for factual outcomes while achieving better balance in the representation space. The integration of PDIG and PPBR in DIGNet improves representation balancing, addressing the inherent trade-off issues found in previous methods.

Second, we propose DR-HiVAE, a novel doubly robust two-stage framework designed for scenarios with proxy variables for unobserved confounders. This method addresses the critical challenge of estimating HTEs in the presence of hidden confounders, a common issue in real-world observational studies. In the first stage, we employ an identifiable Variational Auto-Encoder (iVAE) structure to recover the representation of latent confounders from proxy variables. This process not only mitigates bias caused by unobserved confounders but also simultaneously trains nuisance functions based on the learned representation. The use of iVAE ensures the identifiability of the learned latent confounders, a crucial aspect for reliable causal inference. In the second stage, we apply a doubly robust approach through pseudo-outcome regression to estimate HTEs, leveraging the recovered confounders and trained nuisance functions. To enhance the robustness of our method, we introduce a consistent regularizer that improves the quality of nuisance training, ensuring consistency between the training and testing processes. This results in a doubly robust estimator capable of handling both proxy variables and selection bias effectively. Our theoretical analyses provide insights into two key aspects: the identifiability of the first-stage model, demonstrating that the learned latent confounders and nuisance functions recover the true ones up to a certain degree, and the convergence properties of the estimator, highlighting its robustness against errors in nuisance estimation. This comprehensive approach allows DR-HiVAE to provide more reliable and accurate HTE estimates.

Extensive experiments on benchmark datasets demonstrate the superior performance of our proposed methods compared to existing approaches. This work contributes to the advancement of causal inference methodologies, offering more accurate and reliable tools for estimating heterogeneous treatment effects.

Date of Award	24 Apr 2025
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Qi WU (Supervisor)