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

Abstract

Biotic interactions have been rarely included in traditional species distribution models, wherein joint species distribution models (JSDMs) emerge as a feasible approach to incorporate environmental factors and interspecific interactions simultaneously, making it a powerful tool for analyzing the structure and assembly processes of biotic communities. However, the predictability and statistical robustness of JSDMs are largely unknown because of the lack of research efforts for those newly developed models. This study systematically evaluated the performances of five JSDMs in predicting the occurrence and biomass of multiple species, with a particular focus on diverse characteristics of sampling data, including type of response variables, number of sampling sites, and the number of species included in models. In general, most models yielded satisfactory performances on fitting to observed data and on the estimation of environmental effects; however, they showed less well performances in evaluating species associations, and their predictability had large variations. The JSDMs showed inconsistent performances between the goodness-of-fit and predictability in cross-validation, and the Boral model was relatively robust than others. The predictability of JSDMs was less influenced by sample sizes and substantially improved by incorporating rare species. This study contributes to an appropriate model selection and application of JSDMs.

How to predict biodiversity in space? An evaluation of modelling approaches in marine ecosystems

“…In addition, from a technical viewpoint, HMSC is essentially a multispecies generalized linear mixed model. The algorithm makes it convenient to model all species simultaneously but constrains the capacity to deal with nonlinearity, massive zeros and interactions (Zhang, Chen, Xu, Xue, & Ren, ), which can be better handled in GAM and RF. Nevertheless, the model' less performance can still help understand the patterns of biodiversity, as the model provides evidences of limited biotic interactions which could not be assured a prior.…”

Integrating statistical and mechanistic approaches with biotic and environmental variables improves model predictions of the impact of climate and land-use changes on future mosquito-vector abundance, diversity and distributions in Australia

“…This helps to analyse the spatio-temporal structure and correlated distribution of multiple species at multiple hierarchical levels [ 74 – 76 ]. JSDMs can identify and quantify the effects of biotic interactions such as predation, competition and mutualism on species distributions [ 77 ]. Incorporation and specification of biotic factors in JSDMs improves our understanding of the processes underlying species assembly and lowers bias in the prediction of species community structure [ 78 , 79 ].…”

“…The performance of various presence-absence JSDMs has been assessed at length [ 74 ]. JSDMs are considered computationally efficient but memory intensive and poor at evaluating species associations [ 74 , 77 ], as a result these models will not be discussed further.…”

Inference of biogeographic history by formally integrating distinct lines of evidence: genetic, environmental niche and fossil

“…Continued improvements in ENMs are occurring, such as use of non-point occurrence data like range maps (Merow et al 2017), multi-species modeling (Nieto-Lugilde et al 2018, Zhang et al 2018, and accommodations for imperfect detection (Koshikana et al 2017). Another interesting avenue is to incorporate multiple molecular marker types and genomic-scale data.…”