A Hybrid Causal Search Algorithm for Latent Variable Models.

Area of Science:

• Causal inference
• Machine learning
• Statistical modeling

Background:

• Score-based causal discovery algorithms (e.g., GES, FGS) assume no unmeasured confounders.
• Constraint-based algorithms (e.g., RFCI, FCI, FCI+) handle unmeasured confounders but struggle with small sample sizes.

Purpose of the Study:

• To introduce a novel causal discovery algorithm, GFCI, that combines score-based and constraint-based approaches.
• To address the limitations of existing algorithms in the presence of unmeasured confounders and small sample sizes.

Main Methods:

• Developed a hybrid algorithm, GFCI, integrating score-based and constraint-based causal discovery techniques.
• The algorithm's asymptotic correctness was mathematically proven.

Main Results:

• GFCI demonstrates asymptotic correctness.
• On synthetic datasets, GFCI achieved higher accuracy than FCI, RFCI, and FCI+.
• GFCI exhibited only a marginal increase in computational speed compared to RFCI.

Conclusions:

• GFCI offers a robust and accurate method for causal model search, particularly when unmeasured confounders are present.
• The algorithm provides a significant improvement over existing methods for causal discovery in challenging scenarios.