Best practices for moving from correlation to causation in ecological research

Hannah Correia, Laura E. Dee, Jarrett E. K. Byrnes, John Fieberg, Marie‐Josée Fortin, Clark Glymour, Jakob Runge, Bill Shipley, Ilya Shpitser, Katherine Siegel, George Sugihara, Betsy Von Holle, Paul J. Ferraro2025 · DOI

Abstract

In ecology, causal questions are ubiquitous, yet the literature describing systematic approaches to answering these questions is vast and fragmented across different traditions (e.g., randomization, structural equation modeling, convergent cross mapping). In our Perspective, we connect the causal assumptions, tasks, frameworks, and methods across these traditions, thereby providing a synthesis of the concepts and methodological advances for detecting and quantifying causal relationships in ecological systems. Through a newly developed workflow, we emphasize how ecologists’ choices among empirical approaches are guided by the pre-existing knowledge that ecologists have and the causal assumptions that ecologists are willing to make.

Introduction: The Causal Crisis in Ecology

Ecological research often grapples with a fundamental tension: the gold standard for causal inference—randomized experiments—is frequently impractical at the scales required for meaningful ecological insights. Conversely, while observational data is increasingly abundant, the methods to extract causal meaning from it are often siloed within specific academic traditions. This fragmentation leaves researchers unsure of which tools to apply and when.

The core issue isn't a lack of data, but a lack of transparency regarding the causal assumptions necessary to interpret statistical patterns as evidence of cause-and-effect relationships. This paper provides a conceptual bridge, offering a unified workflow to move beyond simple correlations toward rigorous causal claims.

A Unified Causal Workflow

The centerpiece of the paper is a four-step workflow designed to guide researchers through critical decision points in causal analysis:

  1. Summarizing Knowledge: Researchers should begin by visualizing presumed relationships using Causal Directed Acyclic Graphs (DAGs) or "Target Trial" thought experiments. This process helps identify potential confounders and clarifies the underlying theory.
  2. Defining the Task: A critical distinction is made between Causal Discovery (identifying if a relationship exists when prior knowledge is low) and Causal Inference (quantifying the magnitude of an effect when prior knowledge is high).
  3. Selecting a Framework: Depending on the data and assumptions, researchers might choose Potential Outcomes (often used in experimental design), Structural Causal Models (for complex networks), or Dynamical Systems Causality (specifically for time-series data).
  4. Implementation and Sensitivity Analysis: After selecting specific study designs (e.g., Before-After-Control-Impact) or algorithms (e.g., PC, CCM), researchers must perform sensitivity analyses to determine how robust their conclusions are to violations of their initial assumptions.

Key Insights: Assumptions as the "Glue"

Causal claims are built on untestable assumptions, such as "Causal Sufficiency"—the belief that all major confounders have been accounted for. Because data alone cannot prove these assumptions, they must be justified through ecological theory and field expertise.

A fundamental concept in this process is the Common Cause Principle. Statistical dependence between two variables, $X$ and $Y$, can only arise from three scenarios: $X$ causes $Y$, $Y$ causes $X$, or a third variable $C$ causes both. Causal research is the systematic effort to rule out the non-causal explanations.

Statistical dependence between X and Y can be explained by direct causation in either direction or by a shared common cause C.

Illustrating the Workflow: The Pesticide Example

To demonstrate how common causes can lead to "spurious" correlations, the authors use the example of broad-spectrum pesticide application. In this scenario, pesticide use acts as a common cause ($C$) that reduces the abundance of both ladybird beetles ($X$) and earthworms ($Y$). An observer might find a strong correlation between beetle and worm populations, but there is no direct causal link between them; both are simply responding to the same external pressure.

Pesticide use acts as a common cause, creating a statistical correlation between ladybird beetles and earthworms despite the lack of a direct causal interaction.

Limitations and the Path Forward

The paper emphasizes that even the most sophisticated models are limited by their underlying assumptions. A "perfect" model fit does not guarantee a correct causal conclusion if the initial assumptions are flawed. Furthermore, causal discovery algorithms can be highly sensitive to hyperparameter settings, requiring rigorous robustness checks.

Looking ahead, the authors highlight Causal Data Fusion—the integration of experimental and observational datasets—as a promising frontier. By combining the strengths of different data types and frameworks, ecologists can build more robust and transparent models of the natural world.

Summary updated

Abstract

In ecology, causal questions are ubiquitous, yet the literature describing systematic approaches to answering these questions is vast and fragmented across different traditions (e.g., randomization, structural equation modeling, convergent cross mapping). In our Perspective, we connect the causal assumptions, tasks, frameworks, and methods across these traditions, thereby providing a synthesis of the concepts and methodological advances for detecting and quantifying causal relationships in ecological systems. Through a newly developed workflow, we emphasize how ecologists’ choices among empirical approaches are guided by the pre-existing knowledge that ecologists have and the causal assumptions that ecologists are willing to make.
BibTeX
@misc{correia2025best,
  author = {Hannah Correia and Laura E. Dee and Jarrett E. K. Byrnes and John Fieberg and Marie‐Josée Fortin and Clark Glymour and Jakob Runge and Bill Shipley and Ilya Shpitser and Katherine Siegel and George Sugihara and Betsy Von Holle and Paul J. Ferraro},
  title = {{Best practices for moving from correlation to causation in ecological research}},
  year = {2025},
  doi = {10.32942/x2dw7z},
}
Figure