Ultra-high dimensional variable selection for doubly robust causal inference.

Area of Science:

• Causal inference
• Statistical modeling
• High-dimensional data analysis

Background:

• Observational studies increasingly rely on rich covariate data for causal inference.
• Extracting key features from high-dimensional data is crucial for developing causal procedures like doubly robust estimators.
• Existing methods may not optimally handle ultra-high dimensional confounder selection.

Purpose of the Study:

• Propose a novel confounder selection method, causal ball screening, for ultra-high dimensional data.
• Enhance the efficiency of causal effect estimates by controlling for confounding.
• Maintain double robustness in causal effect estimation.

Main Methods:

• Develop causal ball screening for confounder selection in ultra-high dimensional datasets.
• Utilize an outcome model-free procedure for propensity score model selection.
• Incorporate predictors of the outcome into both propensity score and outcome regression models.

Main Results:

• Theoretical analysis demonstrates model selection consistency and pointwise normality.
• Synthetic and real-world data analyses show favorable performance compared to existing methods.
• The method effectively controls for confounding while improving causal effect estimate efficiency.

Conclusions:

• Causal ball screening is an effective method for confounder selection in ultra-high dimensional settings.
• The proposed approach maintains the double robustness of causal effect estimators.
• This technique offers advantages in efficiency and robustness for causal inference using observational data.