Improving estimates of species distribution change by incorporating local trends

Abstract

A common goal in ecology and its applications is to better understand how species' distributions change over space and time, yet many conventional summary metrics (e.g. center of gravity) of distribution shifts may offer limited inference because such changes may not be spatially homogenous. We develop a modeling approach to estimate a spatially explicit temporal trend (i.e. local trend), alongside spatial (temporally constant) and spatiotemporal (time‐varying) components, to compare inferred spatial shifts to those indicated by conventional metrics. This method is generalizable to many data types including presence–absence data, count data and continuous data types such as density. To demonstrate the utility of this new approach, we focus on the application of this model to a community of well‐studied marine fish species on the US west coast (19 species, representing a wide range of presence–absence and densities). Results from conventional model selection indicate that the use of the model accounting for local trends is clearly justified for over 89% of these species. In addition to making more parsimonious and accurate predictions, we illustrate how estimated spatial fields from the local trend model can be used to classify regions within the species range where change is relatively homogenous. Conventional summary metrics, such as center of gravity, can then be calculated on each such region or within previously defined biogeographic boundaries. We use this approach to illustrate that change is more nuanced than what is expressed via global metrics. Using arrowtooth flounder Atheresthes stomias as an example, the observed southward shift over time in the center of gravity is not reflective of a uniform shift in densities but local trends of decreasing density in the northern region and rapidly increasing density at the southern edge of the species' range. Thus, estimating local trends with spatiotemporal models improves interpretation of species distribution change.

Using spatial distance sampling models to optimize survey effort and address violations of the design assumption

“…Similarly, more complex structures may be useful in addressing changes in spatial distribution in both space and time (e.g. Barnett et al., 2020; Camp et al., 2020).…”

The shadow model: how and why small choices in spatially explicit species distribution models affect predictions

“…Several recent advances have allowed for maximum likelihood estimation to be done using Template Model Builder ( Kristensen et al, 2016 ) using SPDE matrices output by INLA. This approach often allows for faster estimation than using INLA for model fitting ( Osgood-Zimmerman & Wakefield, 2021 ) and allows for additional model flexibility ( e.g ., Barnett, Ward & Anderson, 2021 ). R packages facilitating such estimation with pre-specified models include VAST ( Thorson, 2019 ) and sdmTMB ( Anderson, Keppel & Edwards, 2019 ; Anderson et al, 2020 ).…”

“…For our sablefish case study, we used the latter approach for all model fitting ( ). This package is useful for large and complex datasets that can be computationally taxing ( Anderson et al, 2020 ), it allows for many different response distributions, and optionally includes spatially ( e.g ., Barnett, Ward & Anderson, 2021 ) and temporally ( e.g ., English et al, 2022 ) varying coefficients. While not thoroughly explored in this paper, practitioners could use model selection tools to compare alternative models ( e.g ., with or without spatiotemporal processes, time-varying depth effect, etc .).…”

“…Data with zeros and positive continuous values can be represented with a single distribution, such as the Tweedie distribution ( Dunn & Smyth, 2005 , 2008 ), or the model may separately estimate effects on the probability of occurrence and density in a delta- or hurdle framework ( Pennington, 1983 ). For our SDMs developed for the sablefish case study, we used a Tweedie distribution ( Dunn & Smyth, 2005 , 2008 ) to account for zeros in our continuous response variable (catch per unit effort, CPUE, measured in kg per km 2 swept by the sampling gear) since it has been shown to perform well in previous applications to such data ( Anderson, Keppel & Edwards, 2019 ; Barnett, Ward & Anderson, 2021 ; Thorson et al, 2021 ; English et al, 2022 ). While the majority of hurdle models construct separate spatial fields for presence-absence and positive biomass density ( i.e ., processes assumed to be independent), an advantage of using the Tweedie distribution is that it involves modeling a single spatial field (processes dependent), simplifying interpretation (as an example of sharing elements across fields, see Martínez-Minaya et al, 2018 ).…”

sdmTMB: An R Package for Fast, Flexible, and User-Friendly Generalized Linear Mixed Effects Models with Spatial and Spatiotemporal Random Fields

“…While measured variables can be accounted for with predictors in a model (e.g., measuring and modelling temperature effects on species abundance), unmeasured variables (e.g., everything influencing species abundance but not explicitly modelled) can cause residual spatial correlation. Accounting for this residual correlation is important because doing so allows for valid statistical inference (Legendre & Fortin 1989;Dormann et al 2007), can improve predictions (e.g., Shelton et al 2014), and can be of ecological interest itself by, for example, identifying locations with similar population responses (e.g., Thorson 2019b; Barnett et al 2021).…”