Sina Mews (pictured) from Bielefeld University, Rachel McCrea from University of Kent, Richard Glennie from University of St Andrews, and Takis Besbeas from Athens University gathered on 3 September 2019 for an RSS conference session on applications of hidden Markov models (HMMs) in ecology, part of the environmental and spatial statistics stream.
Sina Mews’ work started as an applied project analysing the movements of endangered dolphins along the Scottish coast and lead to her developing a model for multistate capture-recapture data where capture history is a realisation of a HMM. Data collection is irregular, it depends on boat trips occurring. A seasonal pattern of dolphin movement was found, with dolphins spending more time north in summer and more time south in autumn. It was noted that the survival rate is biased if observations assumed to be dead are excluded.
Rachel McCrea also considered capture-recapture data and presented a test to determine the number of latent states needed in a model with partial observations. Simulations as well as application to a real data set of Canada geese were showcased. Determining the number of latent states needed in a model requires that the properties of each latent state are sufficiently different from other latent states to be detected.
Richard Glennie considered the movement of animals when they are surveyed. An application to jaguars in Belize was shown. He found that ignoring movement causes bias in something we care about: abundance or survival probabilities. He used an advection-diffusion model to examine animal movement over time. Starting with Brownian motion he ended up with a Markov chain determining an animal’s location in a grid at a given time.
Takis Besbeas showcased various options for modelling population dynamics. He started with models of a single index of abundance and classical growth models. From there he introduced state-space versions of the growth models and how model-fitting might not be straightforward as the likelihood is intractable in general. He showed how such likelihoods can be approximated using a HMM approach.
Extensions to models were discussed, such as including additional geographical information to the dolphin model, considering large numbers of unknown states in the latent state model and different models of movement in the animal movement model. Some parts of the discussion were focused on computational time concerns.