Essential ingredients in Joint Species Distribution Models: influence on interpretability, explanatory and predictive power

Abstract

AbstractJoint Species Distribution Models (jSDM) are increasingly used to explain and predict biodiversity patterns. jSDMs account for species co-occurrence patterns and can include phylogeny or functional traits to better capture the processes shaping communities. Yet, several factors may limit or affect the interpretability and predictive ability of jSDMs : missing abiotic predictors, omitting ecologically-important species, or increasing the number of model parameters by adding phylogeny and/or trait information.We assessed how interpretability, explanatory and predictive power of jSDM varied across four alternative models focusing on 99 coastal benthic marine polychaete species: (1) a baseline jSDM with no additional information sources other than abiotic predictors and residual co-occurrence patterns, (2) a jSDM including phylogeny alone or (3) in combination with traits data and (4) a jSDM including monitoring information related to additional species sampled alongside the target assemblage (i.e. non-target species that are not of direct interest but potentially interact with the target assemblage). The four models fitted on both presence/absence and abundance data from a regional monitoring programme were assessed using complementary metrics. We compared performance at both species- and community-level, considering multiple facets of species responses and assemblage diversity.For both presence/absence and abundance data, all models displayed good and similar explanatory power but varied in their interpretability and predictive power. Considering trait data provides insights on species response along environmental gradients, which is a decisive element for model interpretability. Relative to the baseline model, predictive power increased by 26% when including data on additional species, whereas only marginal changes were detected for the two other models. These patterns are explained by changes in the species-environment relationships and residual co-occurrence patterns inferred by these models.Overall, this study highlights that adequate strategy to fit jSDM depends on data at hand, modelling objective and research question. To understand observed community space-time variability, adding phylogenetic or trait information is most effective. Inclusion of non-target species is however a better strategy to predict how the target species assemblage responds to environmental changes. Importantly, we provide a comprehensive toolbox for the comparative assessment of jSDM performance.