Learning Bayesian Networks from Correlated Data.

Area of Science:

• Statistics
• Probabilistic Modeling
• Bioinformatics

Background:

• Bayesian networks are widely used for modeling complex probability distributions.
• Standard methods for building Bayesian networks assume independent and identically distributed data.
• Clustered or longitudinal sampling introduces correlations, leading to inflated false positive rates if ignored.

Purpose of the Study:

• To develop a novel Bayesian network parameterization capable of handling correlated data.
• To enable accurate structure and parameter learning from clustered or longitudinal observations.
• To prevent the inflation of Type I error rates in analyses of correlated data.

Main Methods:

• Introduced a new parameterization for Bayesian networks incorporating random effects.
• Modeled within-sample correlations using these random effects.
• Evaluated different structure and parameter learning metrics for correlated data via simulations.

Main Results:

• The proposed method effectively models correlations within sample units.
• Structure and parameter learning from correlated data can be performed without inflating the Type I error rate.
• Demonstrated the method's utility in analyzing human longevity and sickle cell anemia risk factors.

Conclusions:

• The novel Bayesian network parameterization offers a robust solution for analyzing correlated observational data.
• This approach enhances the reliability of findings from family-based and longitudinal studies.
• It provides a valuable tool for fields dealing with complex, correlated datasets.