Learning Bayesian networks from demographic and health survey data

Area of Science:

• Computational epidemiology
• Public health informatics
• Bayesian network modeling

Background:

• Child mortality from preventable diseases like pneumonia and diarrhea is a significant global challenge, particularly in low and middle-income countries.
• Understanding the complex factors contributing to childhood diarrhea is crucial for developing effective interventions.

Purpose of the Study:

• To construct Causal Bayesian Networks (CBNs) using Indian Demographic and Health Survey (DHS) data to investigate factors associated with childhood diarrhea.
• To evaluate the performance of different structure learning algorithms and the impact of data characteristics on CBN construction.

Main Methods:

• Utilized freeware tools for score-based, constraint-based, and hybrid structure learning algorithms on DHS data.
• Investigated the influence of missing values, sample size, and knowledge-based constraints on algorithm accuracy.
• Assessed algorithm performance using multiple scoring functions and compared generated CBNs.

Main Results:

• No definitive CBN could be learned due to data limitations and algorithm variability.
• Knowledge-based constraints effectively reduced graph variation and produced more plausible relationship models.
• Score-based algorithms TABU and FGES showed desirable qualities, including robustness to missing values and good performance with constraints.

Conclusions:

• While definitive CBNs were not achievable, knowledge-based constraints enhance the reliability of causal discovery from survey data.
• Specific algorithms like TABU and FGES show promise for analyzing complex public health data.
• Further investigation using DHS data and understanding algorithm behavior in real-world settings are recommended.