Simulation-based expert prior elicitation

Abstract

This thesis focuses on the specification of prior distributions for parameters in Bayesian models. In particular, it addresses the formulation of informative priors derived from expert knowledge, a research area referred to as expert prior elicitation. This thesis contributes to the field by introducing a novel simulation-based method that builds on recent advances in predictive prior elicitation. The proposed method is hybrid and model-agnostic, developed within a modular framework that facilitates flexible adaption and extension. Its performance is systematically evaluated across several simulation studies. In addition, we introduce elicito, an open-source Python package that implements the proposed method. The software adheres to the FAIR principles of research software engineering and is distributed via PyPI and conda-forge. Comprehensive documentation, including API references and tutorials, is hosted on a project website. The source code is openly available on GitHub, where it is actively maintained and version-controlled. A continuous integration testing framework is incorporated for quality assurance. Future research directions include validating the method on complex, real-world case studies and expanding its implementation into a fully interoperable software ecosystem for expert prior elicitation.