Identifiability of linear stochastic state-space models with application to ecology

TL;DR

This paper develops a spectral density-based diagnostic to assess the theoretical identifiability of linear stochastic state-space models in ecology, revealing that most issues are practical rather than fundamental.

Contribution

It introduces a new exhaustive summary using spectral density to determine model identifiability, addressing limitations of previous methods that ignored noise parameters.

Findings

Most ecological state-space models are theoretically identifiable.

Unobserved compartments lead to non-identifiability.

Practical issues are mainly due to data limitations, not model structure.

Abstract

State-space models are dynamical systems defined by a latent and an observed process. In ecology, stochastic state-space models in discrete time are most often used to describe the imperfectly observed dynamics of population sizes or animal movement. However, several studies have observed identifiability issues when state-space models are fitted to simulated or real data, and it is not currently clear whether those are due to data limitations or more fundamental model non-identifiability. To investigate such theoretical identifiability, a suitable exhaustive summary is required, defined as a vector of parameter combinations which fully determines the model. Previous work on exhaustive summaries has used expectations of the stochastic process, so that noise parameters are unaccounted for. In this paper, we build an exhaustive summary using the spectral density of the observed process, which fully accounts for all mean and variance parameters. This diagnostic is applied to contrasted ecological models and we show that they are generally theoretically identifiable, unless some model compartements are unobserved. This suggest that issues encountered while fitting models are mostly due to practical identifiability.