Invariant Image Reparameterisation: A Unified Approach to Structural and Practical Identifiability and Model Reduction (2502.04867v2)

Abstract: Both structural and practical parameter non-identifiability present fundamental challenges when using mathematical models to interpret data. This issue is particularly acute in complex, applied areas such as the life sciences or engineering, where determining appropriate model complexity is challenging. While several approaches exist for diagnosing and resolving parameter non-identifiability, including symbolic methods, profile likelihood analysis, and sloppiness analysis, these approaches have distinct limitations and are rarely combined. We present an integrated approach called Invariant Image Reparameterisation (IIR) that incorporates key elements of these methods in a new way. Our approach replaces symbolic computations with numerical calculations at a single reference estimate and an invariance condition that determines when this local calculation holds globally. Parameter combinations determined by this method are naturally ordered by degree of identifiability, and this supports model reduction by replacing a practically non-identified model with a structurally non-identified approximate model. This approximate model can be further parameterised in terms of identified parameters only. By treating parameter combinations determined by our approach as interest parameters within our established likelihood-based Profile-Wise Analysis (PWA) framework, we incorporate uncertainty quantification in terms of likelihood profiles and confidence sets. We provide a Julia library on GitHub (https://github.com/omaclaren/reparam) demonstrating our methodology across a range of mathematical models.