Creation of Probabilistic Models by Experts: A Case Study

Several approaches for the data-based construction of probabilistic models have been described in the literature. The present paper presents an alternative approach based on experts’ knowledge. Naturally, the knowledge must be somehow transformed into probabilistic notions, which are, in our case, two-dimensional contingency tables. The basic idea is that even non-mathematicians can easily understand simple probabilistic elements (like two-dimensional distributions), from which one can assemble non-trivial complex models. For this, we use decomposable compositional models. The resulting models may be equivalent to a classical Bayesian network or a causal model of Pearl’s type if the experts can determine which variables are causes and which are their consequences.