Hartley, David Stephen (2020) Bayesian hierarchical modelling for structured expert judgement. PhD thesis, University of Warwick.

Preview

PDF WRAP_Theses_Hartley_2020.pdf - Submitted Version - Requires a PDF viewer. Download (2845Kb) | Preview

Request Changes to record.

Abstract

Decision makers will often approach experts to help understand uncertainty when their problems cannot be analysed through empirical data alone. When formalised, this process is known as Structured Expert Judgement (SEJ).

Despite the fundamental premise of SEJ being about updating belief, which is the core of Bayesian statistics, SEJ studies often do not consider the Bayesian paradigm. Most SEJ studies utilise techniques which essentially take a pragmatic view of probability (e.g. Cooke's Classical model). Bayesian models have been proposed historically but are used rarely in practice.

This thesis outlines a Bayesian framework for SEJ. The research details the structure of an SEJ study and notes the benefits and limitations of traditional expert aggregation techniques. A collection of recently proposed Bayesian models are highlighted, before presenting a new model which aims to combine and enhance the best of these existing frameworks. In particular, clustering, calibrating and aggregating experts' judgements utilising a Supra-Bayesian parameter updating approach combined with either an agglomerative hierarchical clustering or an embedded Dirichlet process mixture model.

The new approach is assessed by analysing data from existing studies in a variety of domains including healthcare, climatology, volcanology and environmental management. These studies highlight significant overconfidence in expert assessments and consequently a wider range of uncertainty when considering the Bayesian approach. Cross-validation of over twenty studies demonstrates that the Bayesian approach generates higher statistical accuracy than performance weighting but at the cost of lost information.

Key process considerations when implementing a Bayesian model within a broader study facilitation protocol are outlined. A mechanism to embed the new Bayesian model into the popular idea protocol is proposed. A new tool, beam - (B)ayesian (E)xpert (A)ggregation (M)odel, to allow easy deployment of Bayesian thinking into idea is presented. Finally, some areas for further research are recommended.

Item Type:	Thesis or Dissertation (PhD)
Subjects:	Q Science > QA Mathematics
Library of Congress Subject Headings (LCSH):	Bayesian statistical decision theory, Multilevel models (Statistics), Decision making -- Mathematical models, Judgment
Official Date:	September 2020
Dates:	Date Event September 2020 UNSPECIFIED
Institution:	University of Warwick
Theses Department:	Department of Statistics
Thesis Type:	PhD
Publication Status:	Unpublished
Supervisor(s)/Advisor:	French, Simon
Extent:	xiii, 212 leaves
Language:	eng

Request changes or add full text files to a record

Repository staff actions (login required)

View Item

Downloads