Abstract
dc:description.abstractBayesian methods rely on prior distributions to express information that is external to the observed data. When data are limited or scarce, expert judgement becomes an important source of prior information. To use this information in a Bayesian analysis, it must be expressed in a numerical form through a clear and structured elicitation process. This thesis aims to address some gaps in existing elicitation techniques for models where parameters lie on the unit interval and for multivariate models in which several parameters must be elicited jointly. The work places particular emphasis on producing methods that are practical and user-friendly for researchers who may not have specialist training in Bayesian analysis. Methods for eliciting prior distributions for parameters restricted to the interval [0,1] have mainly been limited to the Beta distribution. Although flexible, the Beta distribution cannot capture certain shapes or tail behaviours that experts may wish to express. This motivates the development of elicitation procedures for a wider set of one-, two- and three-parameter distributions on [0,1]. A similar issue appears in regression models with responses restricted to the unit interval. Experts rarely find it natural to express beliefs about regression coefficients, but they can usually give clear probability summaries for the response at selected covariate values. The thesis utilises this feature to develop a quartile-based elicitation method applicable to both univariate priors and regression models, such as the Beta, Kumaraswamy, Unit-Weibull, and Unit-Chen models, among others. Experts provide quartile summaries and conditional medians for the response at selected covariate values, and these are used to obtain the means and variance-covariance matrices of the multivariate normal prior distributions for the regression parameters. This provides a unified approach with greater flexibility in distributional form and produces multivariate normal priors that reflect expert judgement more accurately. For multinomial models, elicitation methods are developed for the Dirichlet distribution and for several generalisations that permit more flexible forms of variation and dependence. These include scaled and Shifted–scaled Dirichlet forms and Dirichlet type 3 and further variants built using mixtures formed from standard Dirichlet distributions, such as the Flexible Dirichlet and Extended Flexible Dirichlet families. A second approach constructs multivariate priors through Gaussian and vine copulas, with marginal distributions taken from the univariate families introduced earlier. Separating marginal and dependence information using copulas provides a clearer route for expressing expert assessments of joint behaviour. All elicitation methods developed in this thesis are supported by Shiny R software that is freely available at https://profiles.open.ac.uk/nayana-unnipillai. The software guides experts through each assessment step, checks the internal coherence of their inputs, and produces the corresponding prior distributions. The aim is to provide tools that can be used without specialised statistical training. The procedures rely on simple probability summaries, such as quartiles and conditional quartiles, rather than complex probability statements, as these are generally easier for experts to supply. The developed software performs the necessary calculations, offers clear graphical feedback, allows adjustments when required and ensures that the resulting priors are coherent and suitable for Bayesian analysis. Taken together, the methods form a unified collection of elicitation tools for univariate priors on the unit interval [0,1], and also for multinomial models and regression models with outcomes on [0,1]. They extend the range of priors that can be obtained directly from expert opinion and provide a practical framework for applied Bayesian analyses in situations where expert knowledge plays an essential role. The contributions also lay the groundwork for further research, including extensions to multivariate regression models and new elicitation procedures for dispersion parameters in multivariate prior distributions.
Degree
thesis:*- Name dc:type.qualificationname
- phd
- Level dc:type.qualificationlevel
- doctoral
- Grantor dc:publisher.institution
- The Open University
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Unnipillai, Nayana
Rights
- Language dc:language
- en