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

Joint Species Distribution Models (jSDM) are increasingly used to explain and predict biodiversity patterns. By accounting for species co-occurrence patterns and potentially including species-specific information, jSDMs capture the processes that shape ecological communities. Yet, factors like missing covariates or omitting ecologically-important species may alter the interpretability and effectiveness of jSDMs. Additionally, while the specific formulation of a jSDM directly affects its performances, the effects of choices related to model structure, such as inclusion, or not of phylogeny or trait information, are not well-explored. Here, we developed a multifaceted framework to comprehensively assess performances of alternative jSDM formulations at both species and community levels. We applied this framework to four alternative models fitted on presence/absence and abundance data of a polychaete assemblage sampled in two coastal habitats over 500 km and 8 years. Relative to a benchmark jSDM only capturing the effects of abiotic predictors and residual co-occurrence patterns, we explored the performance of alternative formulations that also included species phylogeny, traits, or some additional 179 non-target species, which were sampled alongside the species of interest. For both presence/absence and abundance data, explanatory power was good for all models but their interpretability and predictive power varied. Relative to the benchmark model, predictive errors on species abundances decreased by 95% or 53%, when including non-target species, or phylogeny, respectively. These differences across models relate to changes in both species-environment relationships and residual co-occurrence patterns. While considering trait data did not improve explanatory or predictive power, it facilitated interpretation of trait-mediated species response to environmental gradients. This study demonstrates trade-offs in jSDM formulation for explaining or predicting species data, highlighting the importance of using a comprehensive framework to compare models. Furthermore, our study provides some guidance for model selection tailored to specific objectives and available data.