Using Bayesian Networks to Model Hierarchical Relationships in Epidemiological Studies

OBJECTIVES: To propose an alternative procedure, based on a Bayesian network (BN), for estimation and prediction, and to discuss its usefulness for taking into account the hierarchical relationships among covariates. METHODS: The procedure is illustrated by modeling the risk of diarrhea infection for 2,740 children aged 0 to 59 months in Cameroon. We compare the procedure with a standard logistic regression and with a model based on multi-level logistic regression. RESULTS: The standard logistic regression approach is inadequate, or at least incomplete, in that it does not attempt to account for potentially causal relationships between risk factors. The multi-level logistic regression does model the hierarchical structure, but does so in a piecewise manner; the resulting estimates and interpre-tations differ from those of the BN approach proposed here. An advantage of the BN approach is that it en-ables one to determine the probability that a risk factor (and/or the outcome) is in any specific state, given the states of the others. The currently available approaches can only predict the outcome (disease), given the states of the covariates. CONCLUSION: A major advantage of BNs is that they can deal with more complex interrelationships be-tween variables whereas competing approaches deal at best only with hierarchical ones. We propose that BN be considered as well as a worthwhile method for summarizing the data in epidemiological studies whose aim is understanding the determinants of diseases and quantifying their effects.