A comparison of predictive performance of joint species distribution models for presence-absence data

This is a Preprint and has not been peer reviewed. This is version 1 of this Preprint.

Add a Comment

You must log in to post a comment.


There are no comments or no comments have been made public for this article.


Download Preprint

Supplementary Files

David Peter Wilkinson , Nick Golding, Gurutzeta Guillera-Arroita, Reid Tingley, Michael McCarthy


1. While there has been substantial literature on the evaluation of predictions from single species distribution models, the topic of prediction has only recently begun to be addressed for joint species distribution models (JSDMs). These studies have covered only limited aspects of prediction: limited selection of models being compared, limited number of evaluation metrics, and/or not comparing the different prediction types available to JSDMs.
2. In this study, we perform a large-scale comparison of the predictive performance of eight model types: two stacked species distribution models (SSDMs) and six JSDMs. We fit these models to 22 real and simulated datasets, make four types of JSDM predictions, and evaluate up to 32 metrics from five different classes that quantify different aspects of performance of predictions about species distributions and the community assemblage process.
3. We found that likelihood-based metrics indicated the JSDMs were better fit to the data than the standard SSDM, but most other metric classes showed the SSDM outperforming the JSDMs by generally small amounts. The spatial and non-spatial implementations of the hierarchical multivariate probit regression model with latent factors typically performed better than the other JSDMs, but overall still performed worse than the SSDM. The SSDM predictions constrained with the spatially-explicit species assemblage modelling framework (SESAM) consistently outperformed both the standard SSDM and all JSDMs for both species- and community-level metrics.
4. Our results indicate that despite the additional inference they provide about the community assemblage process by accounting for the residual association between species, JSDMs generally yield worse predictions than stacked single species models when evaluated at either the species or community level. The performance of the SESAM framework suggests that exploring similar approaches to constrain JSDM predictions is an interesting future avenue of research.




Life Sciences


biotic interactions, community assemblage, evaluation metrics, joint species distribution models, prediction, species richness


Published: 2023-11-28 11:21

Last Updated: 2023-11-28 16:21


CC BY Attribution 4.0 International

Additional Metadata


Conflict of interest statement:
No conflict of interest is declared.

Data and Code Availability Statement:
The code and data for this analysis can be found on GitHub and is archived by Zenodo (Doi90/JSDM_Prediction; Wilkinson, 2019).