Data-driven confounder selection via Markov and Bayesian networks.

Area of Science:

• Causal Inference
• Statistics
• Machine Learning

Background:

• Unconfoundedness is crucial for estimating causal effects but often requires knowledge of the underlying causal structure.
• Selecting appropriate subsets of covariates (X) for unconfoundedness is challenging when the causal structure is unknown.

Purpose of the Study:

• To develop and evaluate a method for estimating target subsets of covariates sufficient for unconfoundedness when the causal structure is unknown.
• To compare the performance of the proposed method against existing techniques like random forests and LASSO.

Main Methods:

• Modeling the causal structure using probabilistic graphical models (e.g., Bayesian networks).
• Estimating the graph from observed data.
• Selecting target subsets based on the estimated graph.
• Evaluating performance via simulation in high-dimensional settings.

Main Results:

• The proposed method successfully identifies target subsets for unconfoundedness.
• It outperforms random forests and LASSO in selecting these subsets.
• The subset including all causes of the outcome yielded the smallest Mean Squared Error (MSE) in average causal effect estimation.

Conclusions:

• Probabilistic graphical models offer a robust approach to confounder selection in causal inference when the causal structure is unknown.
• This method is effective even in high-dimensional settings and improves the accuracy of average causal effect estimation.