Comparing estimates of abundance trends and distribution shifts using single- and multispecies models of fishes and biogenic habitat

Abstract

Several approaches have been developed over the last decade to simultaneously estimate distribution or density for multiple species (e.g. “joint species distribution” or “multispecies occupancy” models). However, there has been little research comparing estimates of abundance trends or distribution shifts from these multispecies models with similar single-species estimates. We seek to determine whether a model including correlations among species (and particularly species that may affect habitat quality, termed “biogenic habitat”) improves predictive performance or decreases standard errors for estimates of total biomass and distribution shift relative to similar single-species models. To accomplish this objective, we apply a vector-autoregressive spatio-temporal (VAST) model that simultaneously estimates spatio-temporal variation in density for multiple species, and present an application of this model using data for eight US Pacific Coast rockfishes (Sebastes spp.), thornyheads (Sebastolobus spp.), and structure-forming invertebrates (SFIs). We identified three fish groups having similar spatial distribution (northern Sebastes, coastwide Sebastes, and Sebastolobus species), and estimated differences among groups in their association with SFI. The multispecies model was more parsimonious and had better predictive performance than fitting a single-species model to each taxon individually, and estimated fine-scale variation in density even for species with relatively few encounters (which the single-species model was unable to do). However, the single-species models showed similar abundance trends and distribution shifts to those of the multispecies model, with slightly smaller standard errors. Therefore, we conclude that spatial variation in density (and annual variation in these patterns) is correlated among fishes and SFI, with congeneric fishes more correlated than species from different genera. However, explicitly modelling correlations among fishes and biogenic habitat does not seem to improve precision for estimates of abundance trends or distribution shifts for these fishes.

Comparing the prediction of joint species distribution models with respect to characteristics of sampling data

“…The different approaches might partially contribute to the divergence from our results, which suggested that sample sizes had minor influence on model predictability. Many studies have demonstrated that JSDMs consistently over-perform single-species SDMs implemented by GLMs (Hui et al 2013, Tikhonov et al 2017, Clark et al 2017, boosted regression trees, neutral network (Harris 2015), and spatial delta models (Thorson and Barnett 2017). As the marginal predictions usually involve extensive averaging among simulations, the predicted values tend to be stable (small standard error in Fig.…”

“…However, handling abundance data is not an easy task for JSDMs, as abundance data are commonly sparse for the whole community with irregular distributions. In particular, spatial structure may be considered to narrow down the variation of site-specific LVs for improving predictability (Thorson et al 2016, Ovaskainen et al 2017b, Thorson and Barnett 2017. Boral and Gjam), the large proportion of zero observations may hinder model performances (Clark et al 2017).…”

“…Currently, most JSDMs are focused on species occurrence, and rare models aim to address species abundance (but see Thorson and Barnett 2017). Fisheries surveys usually collect data of species abundance or biomass, whereas it is also common to record presence-absence or ordinal data in ecological studies.…”

“…However, species are closely associated with each other through biological processes such as predation, competition and mutualism, and the misspecification of biotic interactions may limit the understanding of assembly processes, leading to substantial bias in predicting community structure. Therefore, multispecies distribution modelling has drawn broad interests in classical and recent community ecology (Legendre and Legendre 2012, Clark et al 2014, Warton et al 2015a, Ovaskainen et al 2017b, Thorson and Barnett 2017. On the other hand, when species distributions are related, one species can provide information for predicting the distribution of other species, implying that modelling the distribution of multiple species simultaneously is an efficient way of using available data (Latimer et al 2009, Ovaskainen et al 2010, Pollock et al 2014.…”

“…In the early development of JSDMs, the models were formulated as multivariate generalized linear mixed model (GLMM), in which the random effect was specified as an unstructured variance-covariance matrix (Ovaskainen et al 2010, Kissling et al 2012, Pollock et al 2014. The latent variable (LV) method enables JSDMs to include hundreds of species, as well as multiple hierarchical levels (Ovaskainen et al 2017b) and spatio-temporal structure (Thorson and Barnett 2017). The limitation has been overcome by the introduction of latent variable models (LVMs) (Skrondal and Rabe-Hesketh 2004, Warton et al 2015a, Ovaskainen et al 2016b, which use unobserved (latent) variables as predictors in the form of random effects to specify species correlations.…”