Coronavirus (Covid-19): Latest updates and information
Skip to main content Skip to navigation

Talk Abstracts

SPEAKER: Prof. Willy Aspinal (University of Bristol) (PDF Document)

TITLE: Real world applications of the Classical Model for expert elicitation: experience and insights.

ABSTRACT: In recent years, Cooke’s Classical Model (1991) has been used for a growing number of expert elicitations in an ever-widening range of scientific, engineering and medical problems. We describe some choice case histories and discuss aspects of the approach that underpin success and aspects that give rise to challenges for the problem owner or for the facilitator. We also mention recent validation analyses, and identify methodology research opportunities.

SPEAKER: Martine Barons (University of Warwick) (PDF Document)

TITLE: Subjective probabilistic judgements for decision support to save the bees.

ABSTRACT: Food poverty is on the increase even in developed nations and with a growing population and a finite planet, there is urgent need for action and decision support. However, eliciting the probability distributions to evaluate subjective expected utility scores associated with ameliorating policies that might be enacted, when the underlying process model is extremely large and complex, brings its own peculiar challenges. It is first necessary to elicit the overall, agreed structure describing in broad terms the underlying nature of the system from representatives of all domain experts across the system as a whole. We have now shown that this can be done formally and consistently with probability models if the elicitations concern the elicitation of dependences – formally termed irrelevances (Smith, Barons and Leonelli (2015)). Within a probability model, these irrelevance statements then transform into assertions about various conditional independence statements. These, in turn, can be used to determine how the system can be divided up into (conditionally) independent segments. The quantitative expert judgements associated with each segment of the process can then be delegated to a relevant panel of experts. Under suitable conditions it can then be shown that the elicited overarching structure can compose these judgments together to form a coherent probabilistic model to score different options available to the user, termed an integrating decisions support system (IDSS).

One element of the overarching food poverty models is food supply, and key to parts of this is an abundant and healthy population of pollinating insects to pollination services for food. In 2014 the UK government undertook a consultation and produced their pollinator strategy for the next 10 years 'to see pollinators thrive, providing essential pollination services and benefits for food production, the wider environment and everyone.' However, the evidence base on the complex system driving pollinator vigour and numbers is patchy and held in disparate domains of expertise, making the evaluation of policy options problematic. In this talk I will describe how we are developing an IDSS for policymakers concerned with evaluating policies relating to pollinator abundance, and how this will then form a sub-module of an IDSS for policies relating to household food poverty in the UK.

SPEAKER: Camila Caiado (Durham University) (PDF Document)

TITLE: Bayesian Uncertainty Analysis for Tipping Points in Systems Modelled by Computer Simulators.

ABSTRACT: Physical systems are often modelled by computer simulators based on highly non-linear systems presenting two or more equilibrium states. At equilibrium, such states can be seen as discontinuities on the model's input space. We focus on the use of emulation and history matching to investigate the simulator's behaviour in different equilibrium states and estimate boundaries between such regions.

SPEAKER: Malcolm Farrow (Newcastle University) (PDF Document)

TITLE: Partially specified beliefs and imprecisely specified utilities.

ABSTRACT: From a Bayesian viewpoint, many problems can be seen within a decision-making framework. In this case we need to specify both prior beliefs and a utility function. In Bayes-linear and Bayes-linear-Bayes methods a fully-specified prior probability distribution is not required. Bayes-linear-Bayes methods are illustrated with examples. It may be difficult to elicit a precise utility function, particularly when there is more than one person involved and when there are trade-offs between different attributes. Some methods for imprecise utility functions are illustrated with examples. Can these ideas be combined? A proposed application in health technology assessment is discussed.

SPEAKER: Dr. John Paul Gosling (University of Leeds) (Powerpoint Presentation)

TITLE: Model selection/averaging for subjectivists.

ABSTRACT: When a Bayesian analysis is being conducted with respect to several competing models, the standard approach to setting a prior over model space is to give each model equal weights. This infuriating practice can easily lead to false conclusions and analyses that are not reflective of anyone’s beliefs. For a subjectivist, the solution is to think carefully about the problem and the models and assign probabilities that are concordant with their beliefs. This sounds easy, but it isn’t.

In this talk, I will discuss the difficulties of setting probabilities for models within some predetermined model space alongside the issues experts may have of being able to think about the model space at all. I will also highlight the challenges of the careful elicitation of parameter distributions within the models when several models are being considered.

SPEAKER: Jim Smith (University of Wawrick) (PDF Document)

TITLE: The Subjective Elicitation of Stories.

ABSTRACT: Intrinsic to good subjective inference is the faithful elicitation of what domain experts actually believe. Over the years I have discovered that many beliefs are framed round stories which capture collections of hypotheses about how things might have happened or might happen. When this is so then an embellishment of an event tree - called a chain event graph - often provides an excellent framework to help the transition from domain beliefs into a faithful probabilistic description of these. I will describe how this framework can be used to take a story and use it to help the client explore the implications of her hypotheses and creatively augment these hypotheses if necessary. This process can occur before embedding domain beliefs into a full probabilistic model and so secure a model that is intelligible and faithful before probabilities are elicited. I will use two simple examples - one drawn from forensic science and the other from public health to describe this process.