Defining and evaluating predictions of joint species distribution models

Abstract

Joint species distribution models (JSDMs) simultaneously model the distributions of multiple species, while accounting for residual co‐occurrence patterns. Despite increasing adoption of JSDMs in the literature, the question of how to define and evaluate JSDM predictions has only begun to be explored. We define four different JSDM prediction types that correspond to different aspects of species distribution and community assemblage processes. Marginal predictions are environment‐only predictions akin to predictions from single‐species models; joint predictions simultaneously predict entire community assemblages; and conditional marginal and conditional joint predictions are made at the species or assemblage level, conditional on the known occurrence state of one or more species at a site. We define five different classes of metrics that can be used to evaluate these types of predictions: threshold‐dependent, threshold‐independent, community dissimilarity, species richness and likelihood metrics. We illustrate different prediction types and evaluation metrics using a case study in which we fit a JSDM to a frog occurrence dataset collected in Melbourne, Australia. Joint species distribution models present opportunities to investigate the facets of species distribution and community assemblage processes that are not possible to explore with single‐species models. We show that there are a variety of different metrics available to evaluate JSDM predictions, and that choice of prediction type and evaluation metric should closely match the questions being investigated.

The Internal Structure of Metacommunities

“…It is possible that more realistic and complex mechanisms driving these processes will weaken associations between pattern and process or create biases in the partitioning of the variation revealed by JSDMs. However, the developments of JSDMs are still progressing, and we anticipate that future developments will solve some of these problems (see Wilkinson et al 2020).…”

Predicting species distributions and community composition using satellite remote sensing predictors

“…Nevertheless, recent studies have suggested that J-SDM do not improve species assemblage predictions 21,52 , in fact, richness predictions from S-SDMs and J-SDMs tend to return similar outcomes 52 , but see 53 . Although J-SDM represent an outstanding alternative for modeling and predicting biodiversity at assemblage level 50,53 , this kind of biodiversity models become untraceable when they are applied to large datasets 52,54 (e.g., NEON or FIA datasets), in other words, the species-pairwise dependencies matrix or residual correlation matrix increases quadratically by adding new species to the dataset 54 . Further research is needed to correctly predict biodiversity at the assemblage level.…”

“…Furthermore, we acknowledge that we evaluated only one type of biodiversity model-S-SDM using two stacking procedures-and that other biodiversity models such as the Joint-SDM (J-SDM) 49,50 , could potentially improve biodiversity predictions by jointly estimating both the species-environment relationships (as in SDMs) and the species-pairwise dependencies-that re ect patterns of co-occurrence -not accounted by the covariates [50][51][52] . Nevertheless, recent studies have suggested that J-SDM do not improve species assemblage predictions 21,52 , in fact, richness predictions from S-SDMs and J-SDMs tend to return similar outcomes 52 , but see 53 . Although J-SDM represent an outstanding alternative for modeling and predicting biodiversity at assemblage level 50,53 , this kind of biodiversity models become untraceable when they are applied to large datasets 52,54 (e.g., NEON or FIA datasets), in other words, the species-pairwise dependencies matrix or residual correlation matrix increases quadratically by adding new species to the dataset 54 .…”

Predicting species distributions and community composition using satellite remote sensing predictors

“…Nevertheless, recent studies have suggested that J-SDM do not improve species assemblage predictions 21 , 53 , in fact, richness predictions from S-SDMs and J-SDMs tend to return similar outcomes 53 , but see Ref. 54 . Although J-SDM represent an outstanding alternative for modeling and predicting biodiversity at assemblage level 51 , 54 , this kind of biodiversity models become untraceable when they are applied to large datasets 53 , 55 (e.g., NEON or FIA datasets), in other words, the species-pairwise dependencies matrix or residual correlation matrix increases quadratically by adding new species to the dataset 55 .…”

“… 54 . Although J-SDM represent an outstanding alternative for modeling and predicting biodiversity at assemblage level 51 , 54 , this kind of biodiversity models become untraceable when they are applied to large datasets 53 , 55 (e.g., NEON or FIA datasets), in other words, the species-pairwise dependencies matrix or residual correlation matrix increases quadratically by adding new species to the dataset 55 . Further research is needed to correctly predict biodiversity at the assemblage level.…”