Fast fitting of non-Gaussian state-space models to animal movement data via Template Model Builder

Christoffer Moesgaard Albertsen, Kim Whoriskey, David Yurkowski, Anders Nielsen, Joanna Mills Flemming

Research output: Contribution to journalJournal articleResearchpeer-review

Abstract

State-space models (SSM) are often used for analyzing complex ecological
processes that are not observed directly, such as marine animal movement. When outliers are present in the measurements, special care is needed in the analysis to obtain reliable location and process estimates. Here we recommend using the Laplace approximation combined with automatic differentiation (as implemented in the novel R package Template Model Builder; TMB) for the fast fitting of continuous-time multivariate non-Gaussian SSMs. Through Argos
satellite tracking data, we demonstrate that the use of continuous-time t-distributed measurement errors for error-prone data is more robust to outliers and improves the location estimation compared to using discretized-time t-distributed errors (implemented with a Gibbs sampler) or
using continuous-time Gaussian errors (as with the Kalman filter). Using TMB, we are able to estimate additional parameters compared to previous methods, all without requiring a substantial increase in computational time. The model implementation is made available through the R package argosTrack.
Original languageEnglish
JournalEcology
Volume96
Issue number10
Pages (from-to)2598-2604
ISSN0012-9658
Publication statusPublished - 2015

Cite this

Albertsen, Christoffer Moesgaard ; Whoriskey, Kim ; Yurkowski, David ; Nielsen, Anders ; Flemming, Joanna Mills. / Fast fitting of non-Gaussian state-space models to animal movement data via Template Model Builder. In: Ecology. 2015 ; Vol. 96, No. 10. pp. 2598-2604.
@article{ee732265a4d54810bbfa8f75e6d911e5,
title = "Fast fitting of non-Gaussian state-space models to animal movement data via Template Model Builder",
abstract = "State-space models (SSM) are often used for analyzing complex ecologicalprocesses that are not observed directly, such as marine animal movement. When outliers are present in the measurements, special care is needed in the analysis to obtain reliable location and process estimates. Here we recommend using the Laplace approximation combined with automatic differentiation (as implemented in the novel R package Template Model Builder; TMB) for the fast fitting of continuous-time multivariate non-Gaussian SSMs. Through Argossatellite tracking data, we demonstrate that the use of continuous-time t-distributed measurement errors for error-prone data is more robust to outliers and improves the location estimation compared to using discretized-time t-distributed errors (implemented with a Gibbs sampler) orusing continuous-time Gaussian errors (as with the Kalman filter). Using TMB, we are able to estimate additional parameters compared to previous methods, all without requiring a substantial increase in computational time. The model implementation is made available through the R package argosTrack.",
author = "Albertsen, {Christoffer Moesgaard} and Kim Whoriskey and David Yurkowski and Anders Nielsen and Flemming, {Joanna Mills}",
year = "2015",
language = "English",
volume = "96",
pages = "2598--2604",
journal = "Ecology",
issn = "0012-9658",
publisher = "JohnWiley & Sons, Inc.",
number = "10",

}

Albertsen, CM, Whoriskey, K, Yurkowski, D, Nielsen, A & Flemming, JM 2015, 'Fast fitting of non-Gaussian state-space models to animal movement data via Template Model Builder', Ecology, vol. 96, no. 10, pp. 2598-2604.

Fast fitting of non-Gaussian state-space models to animal movement data via Template Model Builder. / Albertsen, Christoffer Moesgaard; Whoriskey, Kim; Yurkowski, David; Nielsen, Anders; Flemming, Joanna Mills.

In: Ecology, Vol. 96, No. 10, 2015, p. 2598-2604.

Research output: Contribution to journalJournal articleResearchpeer-review

TY - JOUR

T1 - Fast fitting of non-Gaussian state-space models to animal movement data via Template Model Builder

AU - Albertsen, Christoffer Moesgaard

AU - Whoriskey, Kim

AU - Yurkowski, David

AU - Nielsen, Anders

AU - Flemming, Joanna Mills

PY - 2015

Y1 - 2015

N2 - State-space models (SSM) are often used for analyzing complex ecologicalprocesses that are not observed directly, such as marine animal movement. When outliers are present in the measurements, special care is needed in the analysis to obtain reliable location and process estimates. Here we recommend using the Laplace approximation combined with automatic differentiation (as implemented in the novel R package Template Model Builder; TMB) for the fast fitting of continuous-time multivariate non-Gaussian SSMs. Through Argossatellite tracking data, we demonstrate that the use of continuous-time t-distributed measurement errors for error-prone data is more robust to outliers and improves the location estimation compared to using discretized-time t-distributed errors (implemented with a Gibbs sampler) orusing continuous-time Gaussian errors (as with the Kalman filter). Using TMB, we are able to estimate additional parameters compared to previous methods, all without requiring a substantial increase in computational time. The model implementation is made available through the R package argosTrack.

AB - State-space models (SSM) are often used for analyzing complex ecologicalprocesses that are not observed directly, such as marine animal movement. When outliers are present in the measurements, special care is needed in the analysis to obtain reliable location and process estimates. Here we recommend using the Laplace approximation combined with automatic differentiation (as implemented in the novel R package Template Model Builder; TMB) for the fast fitting of continuous-time multivariate non-Gaussian SSMs. Through Argossatellite tracking data, we demonstrate that the use of continuous-time t-distributed measurement errors for error-prone data is more robust to outliers and improves the location estimation compared to using discretized-time t-distributed errors (implemented with a Gibbs sampler) orusing continuous-time Gaussian errors (as with the Kalman filter). Using TMB, we are able to estimate additional parameters compared to previous methods, all without requiring a substantial increase in computational time. The model implementation is made available through the R package argosTrack.

M3 - Journal article

VL - 96

SP - 2598

EP - 2604

JO - Ecology

JF - Ecology

SN - 0012-9658

IS - 10

ER -