Eltonian Niche Modelling: Applying Joint Hierarchical Niche Models to Ecological Networks

Abstract

There is currently a dichotomy in the modelling of Grinnellian and Eltonian niches. Despite similar underlying data, Grinnellian niches are modelled with species-distribution models (SDMs), whereas Eltonian niches are modelled with ecological-network analysis, mainly because the sparsity of species-interaction data prevents the application of SDMs to Eltonian-niche modelling. Here, we propose to adapt recently developed joint species distribution models (JSDMs) to data on ecological networks, functional traits, and phylogenies to model species' Eltonian niches. JSDMs overcome sparsity and improve predictions for individual species by considering non-independent relationships among co-occurring species; this unique ability makes them particularly suited for sparse datasets such as ecological networks. Our Eltonian JSDMs reveal strong relationships between species' Eltonian niches and their functional traits and phylogeny. Moreover, we demonstrate that JSDMs can accurately predict the interactions of species for which no empirical interaction data are available, based solely on their functional traits. This facilitates prediction of new interactions in communities with altered composition, for example, following climate-change induced local extinctions or species introductions. The high interpretability of Eltonian JSDMs will provide unique insights into mechanisms underlying species interactions and the potential impacts of environmental changes and invasive species on species interactions in ecological communities.

Keywords: Grinnellian niche; community ecology; frugivorous birds; functional traits; interaction networks; joint species distribution models; phylogeny; species interactions.

FIGURE 1

Parallels between the modelling of Grinnellian and Eltonian niches. (a) Grinnellian niches are modelled by assessing the environmental conditions characterising species' ranges, either across the entire range or at point samples of known occurrences. (b) Typically, the occurrence of species across the study area is recorded in a presence–absence matrix; the environmental conditions at the different sites are recorded as a site $\times$ environment matrix. Grinnellian niches are then modelled based on the environmental conditions found within a species' range. In joint hierarchical niche models, additional data such as the species' functional traits or their phylogenetic relationships can be taken into account to improve the modelling of Grinnellian niches. (c) Grinnellian niches can then be plotted in a multidimensional space representing the differences in environmental conditions found at the sites at which species occur. (d) In contrast, Eltonian niches are modelled by assessing species' resource use or interactions with other species, usually recorded in an interaction network, and the traits of the resources with which each species interacts. (e) Typically, the resource use of different consumers is recorded in a consumer $\times$ resource interaction network; the functional traits of resources are recorded in a resource $\times$ trait matrix. These two matrices are analogous to the presence–absence matrix and site $\times$ environment matrix in (b), respectively. Eltonian niches are then modelled based on the traits of the resources with which a consumer interacts. In joint hierarchical niche models, additional data such as the consumers' functional traits or their phylogenetic relationships can be taken into account to improve the modelling of Eltonian niches. (f) Eltonian niches can then be plotted in a multidimensional space representing the differences in the traits of the resources used.

FIGURE 2

Interaction probabilities between frugivorous birds and fleshy‐fruited plants determined by bird and plant traits. Bird traits (rows) include bill width, body mass, and Kipp's index; plant traits (columns) include fruit diameter, crop mass, and tree height. Posterior‐mean interaction probabilities predicted by model 2 ($y$‐axis) are shown along a gradient of a focal plant trait ($x$‐axis) for four hypothetical bird species with different values of a focal bird trait (blue, 0.025 percentile; black, 0.5 percentile; grey, midpoint between the 0.025 and the 0.975 percentiles; red, 0.975 percentile). For all predictions, non‐focal plant traits and non‐focal bird traits were set to the respective community median.

FIGURE 3

Using joint hierarchical niche models to predict unobserved interactions based on proxies. (a) The Eltonian niches of three bird species in a multi‐dimensional plant trait space (only first two axes shown). Plant species are projected into a multidimensional trait space using PCoA. The colour of the dots corresponds to the predicted interaction probability between a bird species and a plant species in the network using model 2. For each bird species, interaction probabilities were modelled once from observed interactions (left) and, after removing all observed interactions from the network, predicted based solely on their functional traits and the relationship between traits and Eltonian niches of the other bird species in the network (right). Interactions in the network were marked as observed (1), unobserved but likely (NA), and unobserved and unlikely (0; see Box 1 for details), differences are represented by dot size. (b) Correlations between the interaction probabilities modelled from observed interactions and those predicted based on birds' functional traits for the three bird species. Bird images by DMD.