Probabilistic graphical models have been used to support decisions in various fields and have the following advantages: 1) they can model and combine different sources of uncertainty; 2) they can model dependencies in complex systems or hypotheses; 3) they provide a clear visualisation of these dependencies that can be understood by non-experts; 4) further analysis can be applied, such as sensitivity analysis, to identify variables and events which have the greatest impact on outcomes; 5) they can be used to identify knowledge gaps where additional data or expert knowledge are needed to support robust decision making by stakeholders and policy makers.

Graphical models can be used in legal processes to represent arguments proposed by the prosecution and defence in complex cases. They allow arguments made by both sides to be converted from natural language into an unambiguous graph, providing a framework for determining the extent to which evidence supports the prosecution and defence hypotheses. 

Of the many different graphical models developed, three types are commonly used in forensics or legal statistics: the Wigmore Chart (WC); the Bayesian Network (BN); and recently, the Chain Event Graph (CEG). In this study, we (the Silent Witnesses) explored the usefulness of these three graphical methodologies in understanding and representing judicial proceedings in a complex criminal case, the first trial of Amanda Knox and Raffaele Sollecito for the murder of Meredith Kercher and compare how each representation performed.

Creating the graphs

The first stage of the study was to read (and translate!) the Italian trial transcript document (Corte Assise 2009) including the list of propositions, evidence, testimony and facts given in the trial. Due to the complexity of the case and the multiple types of evidence we decided to examine a single but vital piece of evidence, the knife found in Sollecito’s kitchen that, according to the prosecution, was used in the murder. We split into three teams, one creating a WC and the other two creating a BN and CEG respectively to graphically model the knife-related evidence.

Conclusions

Each graphical representation has its own advantages and disadvantages. As these methods were taken together in this study, we show how they can be used to provide a better overview of extremely complex arguments used in a criminal case. In the future, one could develop formal ways of moving between the different representations so that understandings from one could help structure and develop another.

Impact

This research has the potential to aid forensic science expert witnesses in making decisions about the extent to which evidence supports prosecution and defence propositions in complex cases, reducing the risk of miscarriages of justice 

This work was completed by the Silent Witnesses (Philip Dawid, Francesco Dotto, Maxine Graves, Jay Kadane, Julia Mortera, Gail Robertson, Jim Smith and Amy Wilson) and was funded by the Engineering and Physical Sciences Research Council via the Alan Turing Institute under wave one of the “Criminal Justice System” theme of the UK Research and Innovation Strategic Priorities Fund. (EP/W006022/1)

Read the paper: Dawid et al. (2025) A comparison of graphical methods using the case of the murder of Meredith Kercher as an example. Law Probability and Risk 24: https://doi.org/10.1093/lpr/mgaf002