|
Abstract:
|
Spatial capture-recapture (SCR) has recently gained popularity as a method for incorporating spatial information into capture-recapture population models. By explicitly modelling individual centers of activity and movement, SCR provides more accurate estimation of population density and information about movement patterns and space use. However, even for modest-sized datasets, fitting these models using Markov chain Monte Carlo (MCMC) can have prohibitively large computational requirements, requiring weeks or longer to achieve convergence. In this talk, we present approaches to efficient MCMC sampling of SCR models in ways that can be applied quite generally. These include analytical integration over discrete latent states to reduce the dimensionality of the MCMC sampling, formulating the model for efficiently batched computations, and applying customized samplers that take advantage of the model's structure. These techniques are possible using the MCMC engine of the NIMBLE package for R. Using real data examples of varying complexity, we show performance gains of one or more orders of magnitude relative to common MCMC engines such as WinBUGS and JAGS.
|
