Variable selection for doubly robust causal inference.

Area of Science:

• Statistics
• Causal Inference
• Observational Studies

Background:

• Confounding control is critical but difficult in observational studies for causal inference.
• Augmented inverse probability weighting (AIPW) is a popular method for estimating average causal effect (ACE) due to its double robustness.
• Variable selection is essential for ensuring the unconfoundness assumption and for efficient estimation.

Purpose of the Study:

• To investigate the impact of variable selection strategies on the double robustness property of AIPW estimators.
• To propose a novel variable selection approach that preserves the double robustness of AIPW.
• To provide a robust method for causal effect estimation in observational studies.

Main Methods:

• Demonstrated that variable selection for efficient estimation can compromise AIPW's double robustness.
• Proposed a new principle: control the propensity score model for any predictor of treatment or outcome.
• Developed a two-stage procedure involving penalized variable selection and AIPW estimation.

Main Results:

• The proposed method preserves the desirable double robustness property of the AIPW estimator.
• Variable selection targeted for efficient estimation can lead to loss of double robustness.
• The proposed procedure shows favorable finite-sample performance in simulations and applications.

Conclusions:

• The proposed variable selection strategy ensures the reliability of AIPW for causal inference.
• This approach offers a robust solution for confounding control and accurate ACE estimation in observational data.
• The findings are validated through simulation studies and a real-world data application.