Causal Clustering for 1-Factor Measurement Models.

Area of Science:

• Causal inference
• Statistical modeling
• Machine learning

Background:

• Scientific research often requires understanding causal structures of complex phenomena.
• Latent (unobserved) variables pose a challenge, necessitating the use of measurable indicator variables.
• Accurate measurement models are crucial for linking indicators to latent variables, often represented as Bayesian networks.

Purpose of the Study:

• To develop a reliable method for inferring measurement models from indicator variables.
• To address the challenge of unknown causal relations and the number of latent variables.
• To introduce a novel algorithm for this specific type of Bayesian network inference.

Main Methods:

• Introduction of a novel algorithm, FindOneFactorClusters (FOFC).
• The algorithm infers measurement models without prior knowledge of causal relations or latent variable count.
• Development of the first correctness proofs for this inference problem, not assuming linearity or acyclicity.

Main Results:

• The FindOneFactorClusters (FOFC) algorithm is proven to be correct.
• FOFC demonstrates superior performance compared to existing state-of-the-art algorithms.
• The algorithm exhibits increased speed, scalability for larger indicator sets, and reliability at small sample sizes.

Conclusions:

• FOFC provides a robust solution for inferring measurement models in causal discovery.
• The algorithm advances the field by offering a more efficient and reliable approach.
• The theoretical proofs extend the applicability of causal inference methods to more complex latent variable structures.