Speaker
Description
The increasing availability of experimental data has motivated interest in calibrating stochastic models, raising fundamental questions about parameter identifiability. Structural identifiability analysis determines whether parameters can be uniquely inferred from idealised, noise-free data; however, existing methods are not suited to partially-observed stochastic processes with hidden internal states.
We develop a framework to analyse structural identifiability in stochastic models with unobserved internal dynamics and investigate how identifiability depends on the type of data available, namely single-particle trajectories or total particle densities.
For density measurements, we derive a partial differential equation describing the population-level dynamics and apply the differential algebra approach. We show that the initial condition plays a crucial role and cannot be fully characterised using the differential algebra techniques alone. To address this, we introduce a method to determine identifiable parameter combinations arising from the initial condition.
We apply the framework to a model in which individuals alternate between constant-velocity motion and pauses governed by a hidden state, and we show that parameters are structurally identifiable from trajectory data but only locally identifiable from density data.
