Abstract
Detecting and quantifying the causal relations of ecosystem functioning is a challenging endeavor. A global study on grasslands illustrates how reasoning about underlying assumptions—from confounding and nonlinearity to fundamental questions of determinism—is key to unlocking the potential of modern causal inference approaches in ecology. This Comment discusses how these frameworks can move beyond simple correlation to reveal the complex drivers of biodiversity-ecosystem functioning (BEF) relationships.
Introduction
A fundamental question in ecology is how biodiversity, often measured as species richness ($R$), affects ecosystem productivity ($P$). While correlations between these variables are frequently observed, establishing a causal link is notoriously difficult. In natural settings, both biodiversity and productivity are influenced by a complex web of confounding factors, including local weather patterns, topography, and historical land-use.
Traditionally, ecologists have relied on controlled experiments to isolate these variables. However, such experiments are often limited by cost, ethical considerations, and a failure to replicate the complexity of natural ecosystems—for instance, by omitting the natural distribution of rare species. Consequently, there is a growing need for robust mathematical frameworks that can extract causal insights from observational data by effectively separating true effects from spurious correlations.
Methodological Frameworks for Causal Inference
Modern causal inference in ecology generally follows two distinct philosophical and mathematical paths: graphical models and dynamical systems.
Graphical Causal Models and Fixed-Effects Regression
This approach relies on mapping qualitative domain knowledge into a directed acyclic graph (DAG). By explicitly stating assumptions about variable interactions, researchers can identify potential confounders. To address unobserved confounding in longitudinal (time-series) data, fixed-effects panel regression is employed. This method uses "dummy" parameters to mathematically account for confounders that are either time-invariant (e.g., soil type) or site-specific but time-varying (e.g., regional climate).
Nonlinear State-Space Perspective
Rooted in Empirical Dynamic Modeling (EDM), this perspective views ecosystems as deterministic, nonlinear dynamical systems. A prominent method here is Convergent Cross-Mapping (CCM), which infers causality by reconstructing the system's "attractor" state-space. If the historical state of one variable can be predicted using the state-space of another, a causal link is suggested.
Case Study: Global Grasslands
The utility of these methods is exemplified by a recent study by Dee et al. (2023), which analyzed a decade of longitudinal data (2007–2017) from 151 unmanipulated plots across 43 global grassland sites. By applying fixed-effects panel regression, the researchers found that higher plot-level species richness actually caused a decline in productivity ($\beta < 0$).
This finding challenges previous observational studies that lacked sufficient control for confounders and experimental studies that did not account for the role of rare species. The results suggest that while native species generally enhance productivity, an increase in rare species can lead to a net decline in ecosystem output.
Insights and Future Directions
The primary strength of modern causal inference lies in the explicit declaration of assumptions. Using causal graphs forces researchers to be transparent about perceived confounding mechanisms, making models more interpretable and subject to rigorous debate. Furthermore, observational approaches can highlight "blind spots" in experimental data, such as the impact of non-native or rare species distributions.
However, both methods have limitations. Fixed-effects regression cannot account for confounders that vary across both space and time simultaneously. Conversely, CCM assumes the existence of a low-dimensional attractor, which may be difficult to reconstruct in noisy or high-dimensional ecological systems. A key frontier for the field is the integration of these two communities—combining the confounder-handling rigor of graphical models with the nonlinear dynamics expertise of state-space approaches.
Summary updated