Uncertainty Analysis for Non-identifiable Dynamical Systems: Profile Likelihoods, Bootstrapping and More

Dynamical systems are widely used to describe the behaviour of biological systems. When estimating parameters of dynamical systems, noise and limited availability of measurements can lead to uncertainties. These uncertainties have to be studied to understand the limitations and the predictive power of a model. Several methods for uncertainty analysis are available. In this paper we analysed and compared bootstrapping, profile likelihood, Fisher information matrix, and multi-start based approaches for uncertainty analysis. The analysis was carried out on two models which contain structurally non-identifiable parameters. We showed that bootstrapping, multi-start optimisation, and Fisher information matrix based approaches yield misleading results for parameters which are structurally non-identifiable. We provide a simple and intuitive explanation for this, using geometric arguments.