Whose Judgement? Reflections on Elicitation in Bayesian Analysis

Abstract

Bayesian statistical, risk, and decision analyses require that one addresses many uncertainties and preferences, modelling those that can be with subjective probabilities and utilities, perhaps supported by sensitivity explorations. Subjective probabilities need eliciting either in their entirety or partially via prior distributions that are updated in the light of data during the analysis. Some uncertainties, however, are not easily modelled probabilistically, either because they are deep or because they relate to uncertainties in the modelling process itself. Preferences also require elicitation, a process which in many cases constructs these by contextualising broader values to the issues at hand. We discuss broader issues of elicitation without getting into specific details of the elicitation process. We also briefly discuss communication because elicitation sets the context for all subsequent communications to the problem owners and stakeholders. In particular, we emphasise the need for the problem owners to be fully acquainted with all the residual uncertainties at the end of the analysis, not just those captured quantitatively within the modelling. Moreover, we also consider whose uncertainties and preferences should be elicited and addressed by the analysis, arguing that the answer may be different in the varied contexts of Bayesian statistical, risk, and decision analyses. Moreover, the model may be constructed from a synthesis of several people’s judgements.