A novel Bayesian Network modelling approach that can ideally represent the information contained in a set of sample data

The ultimate goal of statistical modelling is to best represent the information contained in a set of sample data. A sampling subject (i.e., an experimental unit or an observational unit on which measurements were taken) in a data collection scheme can be a person, a plant, an animal, or even an event, etc. In this study, Bayesian Network (BN) models were fitted to four data sets to explore the potential for ideal representation of a data set. BN models following both the conventional regression approach and the novel sampling-subject-oriented approach were built for comparison of the model fitting and predictive performance, using both research project and textbook data sets. The sampling-subject-oriented approach treated the sampling subject as the target variable for specification of the optimal model structure via Tree Augmented Naive Bayes algorithm. The results showed clear superiority of the sampling-subject-oriented approach models. A BN model following the sampling-subject-oriented approach had great potential for quantifying interrelationships between variables of mixed types in a complex model, hence it was particularly suitable for analysing data sets from a survey study with large numbers of numeric and categorical interrelated variables.