Bayesian Modelling of Dependence Between Experts: Some Comparisons with Cooke’s Classical Model

A Bayesian model for analysing and aggregating structured expert judgement (sej) data of the form used by Cooke’s classical model has been developed. The model has been built to create predictions over a common dataset, thereby allowing direct comparison between approaches. It deals with correlations between experts through clustering and also seeks to recalibrate judgements using the seed variables, in order to form an unbiased aggregated distribution over the target variables. Using the Delft database of sej studies, compiled by Roger Cooke, performance comparisons with the classical model demonstrate that this Bayesian approach provides similar median estimates but broader uncertainty bounds on the variables of interest. Cross-validation shows that these dynamics lead to the Bayesian model exhibiting higher statistical accuracy but lower information scores than the classical model. Comparisons of the combination scoring rule add further evidence to the robustness of the classical approach yet demonstrate outperformance of the Bayesian model in select cases.