Causal Discovery with Generalized Linear Models through Peeling Algorithms.

Area of Science:

• Causal inference
• Statistical genetics
• Machine learning

Background:

• Causal discovery is challenged by unmeasured confounders, complicating relationship identification.
• Generalized structural equation models (GSEMs) offer flexibility for diverse data types but require robust causal discovery methods.
• Existing methods often struggle with unmeasured confounding and mixed data outcomes.

Purpose of the Study:

• To develop a novel causal discovery method for GSEMs applicable to discrete, continuous, and mixed data.
• To address the challenge of unmeasured confounders in identifying causal relationships.
• To provide a robust framework for constructing gene regulatory networks.

Main Methods:

• Developed two peeling algorithms (bottom-up and top-down) for causal discovery and instrument identification.
• Reconstructed a super-graph of ancestral relationships using nodewise GLM regressions.
• Estimated parent-child effects by deconfounding child models with parent information.

Main Results:

• Established theoretical conditions for model identifiability and statistical guarantees for discovering parent-child relationships.
• Numerical experiments demonstrated superior performance compared to state-of-the-art methods.
• Successfully applied the method to construct gene-to-gene and gene-to-disease networks in Alzheimer's disease using SNPs.

Conclusions:

• The proposed peeling algorithms effectively perform causal discovery in the presence of unmeasured confounders.
• The method is robust for analyzing diverse data types within GSEMs.
• Demonstrated practical utility in unraveling complex biological networks, such as those in Alzheimer's disease.