Communication-Efficient Distributed Estimation of Causal Effects With High-Dimensional Data.

Area of Science:

• Statistics
• Econometrics
• Machine Learning

Background:

• Estimating average treatment effect (ATE) is crucial in causal inference.
• Distributed data settings and high-dimensional covariates pose significant challenges.
• Existing methods often struggle with communication efficiency and model misspecification.

Purpose of the Study:

• To develop a communication-efficient algorithm for ATE estimation in distributed settings.
• To address challenges posed by a large number of covariates relative to site-specific sample sizes.
• To ensure robust estimation under potential model misspecification.

Main Methods:

• Proposed a distributed covariate balancing propensity score (disthdCBPS) estimator.
• Utilized surrogate loss functions for calibrating propensity score and outcome models.
• Leveraged approximate covariate balancing for distributed data.

Main Results:

• The disthdCBPS estimator approximates the global estimator (pooled data) at a fast rate.
• The estimator remains consistent and asymptotically normal under model misspecification.
• Achieved semi-parametric efficiency bounds when models are correctly specified.

Conclusions:

• The proposed method offers a communication-efficient and robust approach to ATE estimation.
• Demonstrated strong empirical performance in simulations and real-world data.
• Provides a valuable tool for causal inference in decentralized data environments.