Statistical Methodology

Developing Markov switching autoregressive models to analyse animal movement data

Animal movement data collected at a high temporal frequency can be very rich in information, but extracting this information poses many challenges. Certain issues must be overcome. First, the number of underlying behaviour types is unknown. Second, the correlation structure between observations depends on the types of behaviour being exhibited, with long bouts of one behaviour leading to longrange dependencies among the data, including asymmetric cycles and high autocorrelations at high time lags.

flapper skateWe have adapted Markov switching autoregressive models (MSARMs), developed previously to assess river conditions from water chemistry time series and to identify behavioural states of animals from locational data. These models conjecture that the time series of movements of an individual derive from a latent set of behavioural states, modelled as a Markov chain, between which individuals switch according to some probabilistic rules. Switches between states, and movements within any state, are stochastic outcomes with underlying rates, variances and correlations influenced by diel cycle (day or night), lunar cycle (binary) and lunar phase (4 categories). Fitting these models using Markov chain Monte Carlo techniques allows us to identify the number of behaviour types best supported by the data, along with the associated movement patterns and rates of interchange between behaviours.

fitted MSARMs data Fitting MSARMs to data from flapper skates (Dipturus intermedia), a marine apex predator threatened with extinction, identified evidence for two behavioural states and indicated where these occur in the water column.

Further details from:
Luigi Spezia

Article date 2015


Statistical Genomics and Bioinformatics

Process and Systems Modelling

Statistical Methodology

PhD Opportunities

Meetings & Seminars