Bagged random causal networks for interventional queries on observational biomedical datasets

Area of Science:

• Causal inference
• Machine learning
• Biomedical informatics

Background:

• Estimating causal effects from observational data (e.g., electronic health records) is prone to bias from unmeasured confounders.
• Directed acyclic graphs and do-calculus can identify adjustment sets for bias elimination, but require complete causal structure knowledge.
• Causal structure discovery algorithms can be computationally intractable and yield ambiguous solutions for large feature sets.

Purpose of the Study:

• To introduce and evaluate bagged random causal networks as an ensemble method for approximating causal effects from observational data.
• To address the limitations of partial causal structure knowledge and high computational complexity in causal inference.

Main Methods:

• Bagged random causal networks are constructed by ensembling subnetworks from sampled feature subspaces.
• Conditional dependencies are drawn within subnetworks to infer adjustment sets.
• Causal effects are estimated using regression functions of the outcome on the query and adjustment sets.

Main Results:

• The bagged estimator is consistent with true causal effects when the causal structure is known.
• It demonstrates a favorable variance/bias trade-off with heuristically estimated structures.
• The method exhibits lower computational complexity compared to full network learning and outperforms boosted regression.

Conclusions:

• Bagged random causal networks provide a robust and computationally efficient approach for estimating query-target causal effects from high-dimensional observational data.
• This method is well-suited for applications in electronic health records and other biomedical databases.