I have a Bayesian Network whose structure was defined by expert consensus (Delphi) and whose conditional probability tables were calibrated using the EM algorithm with data from a pilot group. My goal is not to predict missing values, since all variables are fully observed, but rather to identify and compare reasoning patterns and alternative conceptions between a traditional group and an experimental group in an educational study.
In this context, what type of inference, network-based measure, or derived indicator could the Bayesian Network provide that is not already contained in the observed response trajectories themselves?
Or, if all variables are fully observed, is the Bayesian Network mainly useful as a calibrated probabilistic model, while the comparison of reasoning patterns should instead be performed directly from the observed data?