On the Interpretations of Joint Modeling in Community Ecology

Giovanni Poggiato, Tamara Münkemüller, Daria Bystrova, Julyan Arbel, James S. Clark, Wilfried ThuillerTrends 2021 · DOI

TL;DR

Joint Species Distribution Models (JSDMs) are increasingly used to understand ecological communities, but their ability to disentangle environmental and biotic effects remains unclear. This paper rigorously demonstrates that JSDMs, like their simpler counterparts (SDMs), primarily capture realized niches and do not inherently separate environmental influences from species interactions. The key contribution lies in providing a clear framework for interpreting JSDMs and guiding their appropriate application.

What Problem Does This Paper Address?

Understanding and predicting how biodiversity responds to global change is a critical challenge in ecology. Species Distribution Models (SDMs) are widely used to predict where species occur, but they typically model each species independently, failing to account for interactions between species. Joint Species Distribution Models (JSDMs) were developed to address this by modeling multiple species simultaneously, with the hope of disentangling the effects of the environment from biotic interactions and improving predictions, especially for rare species. However, there has been a lack of clarity on whether JSDMs truly achieve these goals, with some claims about their ability to "account for biotic interactions" remaining unproven. This paper aims to clarify the true capabilities and limitations of JSDMs and related models (MSDMs) in the context of ecological theory.

Methods

The authors provide a rigorous mathematical re-unification of SDMs, MSDMs, and JSDMs under a common notation. They analytically demonstrate the relationships between these models, particularly focusing on how they estimate species' niches (represented by regression coefficients) and how residual correlations are handled. The core of their methodological contribution is a theoretical and mathematical dissection of these models, rather than the introduction of new statistical techniques or empirical data analysis. They use mathematical proofs and conceptual arguments to explain the interpretations of model parameters and predictions.

Datasets

Not discussed in the paper. The paper focuses on theoretical and mathematical interpretations of models, using an empirical example for illustration but not for primary dataset-driven analysis.

Key Insights and Ideas

  • JSDMs don't magically separate environment from interactions: Despite their sophistication, JSDMs, like SDMs, estimate species' "realized niches" – the combination of environmental and biotic factors that determine where a species is found. They do not inherently disentangle these two components.
  • Niche estimates are similar across models (mostly): For many common scenarios, the estimated species' niches (regression coefficients) are mathematically identical between SDMs and JSDMs. The primary difference lies in how they handle uncertainty and correlations among species' residuals.
  • MSDMs offer a different kind of improvement: Multi-species Distribution Models (MSDMs), particularly those with hierarchical coefficients, can improve niche estimation for rare species by "borrowing strength" from related species. However, this is based on trait or phylogenetic similarity, not directly on biotic interactions.
  • Residual correlation is tricky: The residual correlation matrix in JSDMs captures patterns not explained by environmental variables. While it can be influenced by biotic interactions, it also reflects model misspecifications, missing covariates, and other factors. Interpreting it solely as a direct measure of biotic interaction is often misleading.
  • Exploiting, not interpreting, residuals: The true power of JSDMs lies not in interpreting the residual correlation matrix for biotic interactions, but in exploiting it for improved predictions, particularly "joint" and "conditional" predictions. This means using the co-occurrence patterns to predict the presence of one species given the presence/absence of others.

Key Results

  • Mathematically demonstrated that for Gaussian data, the regression coefficient estimates (species' niches) are identical for SDMs and JSDMs.
  • Empirical comparisons (in Box 3) also showed very similar regression coefficient estimates between SDMs and JSDMs.
  • Marginal predictions (predicting each species independently) are also very similar between SDMs and JSDMs, as they depend on the same niche estimates.
  • The residual correlation matrix from JSDMs and correlations derived from independent SDMs are highly similar in the provided empirical example (R² = 0.862).
  • JSDMs can provide more accurate estimates of uncertainty around predictions (e.g., species richness) compared to independent SDMs, due to their ability to account for correlated residuals.

Limitations and Future Work

  • Interpretation of Residuals: The paper strongly cautions against interpreting the residual correlation matrix as a direct measure of biotic interactions. Its signal is context-dependent and can be confounded by many factors.
  • Fundamental Niche: JSDMs and MSDMs do not directly estimate the "fundamental niche" (the niche based solely on physiological suitability) but rather the "realized niche" (which includes biotic influences).
  • Hierarchical Models: While MSDMs and JSDMs with hierarchical coefficients can improve niche estimation for rare species, this relies on assumptions about trait or phylogenetic similarity, which may not always hold true for truly rare species.
  • Future Work: The paper highlights "Outstanding Questions" such as:
  • How to distinguish biotic interaction imprints from environmental effects in co-occurrence patterns.
  • How to better harness temporal data for dynamic JSDMs.
  • How to best utilize conditional predictions from JSDMs for applications like invasion risk assessment.
  • How to truly account for biotic interactions when predicting distributions and community composition.

Summary updated

BibTeX
@inproceedings{poggiato2021interpretations,
  author = {Giovanni Poggiato and Tamara Münkemüller and Daria Bystrova and Julyan Arbel and James S. Clark and Wilfried Thuiller},
  title = {{On the Interpretations of Joint Modeling in Community Ecology}},
  booktitle = {Trends},
  year = {2021},
  doi = {10.1016/j.tree.2021.01.002},
}
Figure