Extensions of Dynamic Factor Analysis for Better Modeling of Noisy Ecological Time Series

• Activity Number: 131 - Methods for Spatial, Temporal, and Spatio-Temporal Data
• Type: Contributed
• Date/Time: Monday, August 9, 2021 : 1:30 PM to 3:20 PM
• Sponsor: Section on Statistics and the Environment
• Author(s): Eric Ward* and Sean Anderson and Mary Hunsicker and Mike Litzow
• Companies: NOAA and Fisheries and Oceans Canada and NOAA and NOAA
• Keywords: environmental science; ecology; Bayesian; multivariate time series; DFA; state-space

Abstract

Dynamic factor analysis (DFA) is a statistical tool for understanding the latent processes that may be shared amongst a collection of observed multivariate time series. DFA has been widely applied to a number of fields, including economics, medicine, and environmental sciences. We extend conventional DFA to (1) model deviations drawn from a Student-t distribution to better capture extremes, (2) model latent processes (trends) with autoregressive and moving-average components, (3) develop alternative models for the latent trends. While DFA generally models trends as a random walk process, we introduce a new class of models, where latent trends are modeled as smooth functions. These models are applied to two long-term datasets from the west coast of the United States: (1) a 35-year dataset of juvenile rockfishes from the west coast of the United States, and (2) a 39-year dataset of fisheries catches. Estimation and model selection is done in a Bayesian framework using Stan, with code provided as the ‘bayesdfa’ R package.