Bayesian Networks for Probabilistic Inference and Decision Analysis in Forensic Science, 2nd Edition

13.2 Sensitivity analyses

13.2.1 Preliminaries

When a scientist has set up a graphical model for structuring a given inferential task, the model typically conveys structural assumptions, that is, relationships believed to hold true amongst particular variables of interest. In Bayesian networks, probability theory is used as a concept to characterize the nature of the relationships assumed to hold amongst variables. Most often, however, a given model is considered only as a tentative approach for a given problem at hand and represents, as such, a starting point for further investigations. In particular, it is important for scientists to study the properties of a given model. This leads to a topic broadly called sensitivity analysis, a term that brings together several related aspects.

One important aspect concerns sensitivity to findings. In specialized literature on Bayesian networks, this is also referred to as evidence sensitivity analysis [e.g. Kjærulff and Madsen (2008)]. This kind of analysis focuses on question of the following kind: ‘What is the support offered by a given item of information on the belief in a given hypothesis of interest—versus the belief in another?’, ‘Which item of information (in a large set of findings) provides the best discrimination between the hypotheses of interest?’ Such sensitivity analyses may help understanding the conclusions reached when using a given model for inference. The provision of answers to questions as mentioned above is crucial ...