Abstract (EN)

Within the Bayesian paradigm, expert knowledge is typically used to construct informative priors, with an emphasis on combination with empirical data to form posterior estimates. In pioneering research, however, the initial step of represent- ing the current state of knowledge may deserve more attention, due to its pivotal role in helping focus scientific investigation and guide data collection. In this paper we concentrate on situations we have encountered in pioneering research where it is worthwhile to invest considerable e!ort in both design and analysis of expert opinions. A recurring example from our experience involves discerning habitat preferences, in order to map potential spatial distribution, of rare and threatened species. Expert opinion is valuable in these contexts since empirical data is typically limited due to sparsity of patterns occurring across broad spatio- temporal extents. This problem arises both in landscape ecology where the aim is to conserve key species, and in biosecurity where the aim is to estimate risk that a pest species establishes. Many mathematical methods for collating expert opinion have been developed. In this work we demonstrate the flexibility of a Bayesian statistical modelling framework, through careful consideration of how variability among and within experts enters into the model. Most commonly, expert opinions are collated once encoded into individual priors. It is also possible to aggregate posteriors resulting from an analysis informed by each expert individually via a form of Bayesian model averaging. When indirect elicitation methods are used then expert opinion may also be combined in intermediate steps, either by combining the elicited information or waiting until it is encoded to combine it. We compare these four methods of aggregating expert opinions using a case study on eliciting habitat preferences. This allows us to examine the choice of the level of aggregation from the perspective of the sources of variation addressed, and the summaries they provide. In addition, aggregation of expert opinion provides a useful analogy for aggregation of models more generally. We show that this aggregation can be flexibly accommodated within the Bayesian framework: by aggregating posteriors, priors, or inputs to priors.