A ROBUST AND EFFICIENT APPROACH TO CAUSAL INFERENCE BASED ON SPARSE SUFFICIENT DIMENSION REDUCTION

Area of Science:

• Causal inference
• Observational data analysis
• Statistical modeling

Background:

• Estimating treatment effects from observational data relies on the assumption that treatment assignment is ignorable given measured confounders.
• Including numerous baseline covariates is often necessary due to unknown confounders, but existing methods struggle with model misspecification and variable selection.
• Bias from incorrect model specification or confounder selection can lead to misleading results in treatment effect estimation.

Purpose of the Study:

• To develop a robust and efficient approach for causal inference of average treatment effects using observational data.
• To overcome limitations of existing methods that require restrictive parametric models and sensitive variable selection.
• To provide reliable treatment effect estimation even when dealing with a large number of potential confounding variables.

Main Methods:

• Proposed a flexible modeling strategy incorporating penalized variable selection for estimating average treatment effects.
• Developed an estimator based on an efficient influence function involving propensity score and outcome regression.
• Introduced a novel sparse sufficient dimension reduction method to estimate propensity score and outcome regression without restrictive parametric assumptions.

Main Results:

• The proposed estimator for the average treatment effect is asymptotically normal and semiparametrically efficient.
• The method does not require variable selection consistency, enhancing its robustness.
• Demonstrated the utility of the proposed methods through simulation studies and a biomedical application.

Conclusions:

• The novel approach offers a robust and efficient solution for causal inference with observational data, particularly when many covariates are present.
• Flexible modeling and sparse sufficient dimension reduction mitigate issues related to model misspecification and variable selection.
• The method provides reliable estimation of average treatment effects, applicable in various research settings including biomedical applications.