Efficient methods for fitting nonlinear non-Gaussian state-space models of wildlife population dynamics

点击翻译按钮获取中文摘要

Recent years have seen an enormous growth in interest in and methods for fitting mechanisticmodels of wildlife population dynamics to survey data on animal numbers, survival and birth rates.Various statistical methods have been proposed, based both on maximum likelihood and Bayesianapproaches. Examples include the Kalman filter (and extensions) for maximum likelihood estimationand Markov chain Monte Carlo (MCMC) or particle filtering for Bayesian models. Each hasadvantages and disadvantages - for example the Kalman filter is designed for linear Gaussian models(but seems to do remarkably well in other circumstances); MCMC is an excellent omnibus methodbut it can be difficult to derive efficient samplers (i.e., those that produce reliable answers in areasonable amount of computer time); particle filtering is easy to program but very inefficient forsome models (e.g., those with random effects). In this project, we aim to blend aspects of theseapproaches to increase the efficiency of the estimation. For example, we will investigate usingKalman filter estimates as importance sampling starts in a particle filter algorithm, and using theparticle filter to provide proposals in an MCMC algorithm.