DNA transfer experiments are typically conducted within individual laboratories, yet forensic casework often requires interpretation of evidence generated under different laboratory conditions. Here we present a statistical framework that enables transfer data generated across multiple laboratories to be integrated within a hierarchical Bayesian framework by explicitly modelling inter-laboratory variability. The framework departs from traditional Bayesian Network approaches by avoiding generalisations about background DNA from unknown contributors, and instead modelling DNA from unknown contributors explicitly on a per-case basis.

The development of this framework was motivated by the findings of the ReAct study (a collaboration of 20 labs), which demonstrated substantial inter-laboratory variability in DNA transfer experiments. Lessons from miscarriages of justice such as Amanda Knox and Birgitte Tengs highlighted the importance of evaluating DNA evidence given activity level propositions, accounting explicitly for unknown contributors, and considering the evidential significance of the absence of DNA.

The statistical framework is implemented as HaloGen (Hierarchical Activity Level Orchestrator: next Generation software), a new open-source framework. It addresses and provides a solution to three key challenges in interpretation given activities. First, that laboratory-specific transfer models based on quantitative DNA measurements are embedded within a multi-laboratory calibration framework. This enables DNA quantity observations from complex mixtures, generated using different experimental systems and analytical methods, to be evaluated coherently. 

Second, the framework implements an exhaustive hypothesis approach for evaluating activity-level propositions across multiple stains. All admissible combinations of contributors are enumerated and evaluated directly, providing a principled method for analysing complex patterns of DNA recovery encountered in casework.

Third, HaloGen introduces an explicit treatment of unknown contributors (background DNA) and the absence of DNA from a person of interest. Since these features always provide important support for the defence proposition, it considerably reduces risks of confirmation bias.

Simulation studies implemented in HaloGen show stable likelihood-ratio behaviour across multiple laboratory datasets and illustrate how conditioning on the number of offenders and the treatment of unknown contributors can materially affect evidential strength. Using ground-truth data, Tippett plots are used to calibrate the likelihood-ratio system by calculating the distributions of likelihood ratios under competing propositions.  This shows how effectively the model separates true prosecution cases from true defence cases. The framework is modular and extendable, providing a transparent basis for future integration of additional activity-level information, including body-fluid context and shedder-related effects.

Authors

• Peter Gill (Oslo University Hospital, Norway)
• Elida Fonneløp (Oslo University Hospital, Norway)
• Helen Johannessen (Oslo University Hospital, Norway)
• Øyvind Bleka (Oslo University Hospital, Norway)