Causal discovery from ecological time-series with one timestamp and multiple observations

Daria Bystrova,Charles Assaad,Sara Si-moussi,Wilfried Thuiller

DOI: https://doi.org/10.1101/2024.10.10.608447

2024-10-12

Abstract:Ecologists frequently seek to establish causal relations between entities of an ecological system, such as species interactions, ecosystem functions or ecosystem services, using observational data. Despite this, many studies still primarily rely on correlation-based methods, which lack the capability for causal interpretation. Recently, causal discovery methods have gained traction in analysing ecological time-series. However, the scarcity of ecological time-series data presents a challenge due to the demanding and time-consuming nature of collecting consistent measurements over extended periods. In this paper, we delve into the applicability of causal discovery methods when applied to point-in-time (or snapshot-like) observational data obtained from ecological dynamic systems. Specifically, we examine the PC algorithm, which holds theoretical validity assuming the causal Markov condition, faithfulness and causal sufficiency. Additionally, we explore the FCI algorithm, an extension of the PC algorithm designed to handle cases where causal sufficiency is violated. Through a combination of theoretical reasoning and simulation experiments, we elucidate the scenarios in which both algorithms are expected to yield meaningful results. We demonstrate that even in situations where causal sufficiency is not satisfied, the PC algorithm - characterized by its comparatively simpler interpretability - can still deduce specific types of relationships between ecological entities. Furthermore, we illustrate our theoretical findings on simulated data as well as on real data containing records of the abundance of various bird species as well as climatic and land-cover conditions.

Ecology

What problem does this paper attempt to address?

The paper attempts to address the problem of discovering causal relationships in ecological research from ecological time series data with multiple observations at a single time point. Specifically, due to the scarcity and difficulty of collecting ecological time series data, researchers often rely on correlation methods to analyze species interactions, ecosystem functions, or services within ecosystems. However, correlation methods cannot provide causal explanations. Therefore, the paper explores the application effects and limitations of several causal discovery algorithms (such as the PC algorithm and the FCI algorithm) in the context of data available at only a single time point.

- The paper first examines the performance of the PC algorithm under different settings, including acyclic and cyclic summary graphs, the presence or absence of self-causal relationships, and immediate or lagged causal relationships.
- For the case where all time series are self-causal, a new simplified algorithm called RestPC is introduced to reduce computation time without sacrificing result accuracy.
- It explores how to enhance the accuracy and interpretability of causal discovery methods by utilizing ecological background knowledge (such as possible or forbidden links, and information on species groups with similar interaction patterns).
- Finally, the effectiveness of these causal discovery methods is validated through simulation experiments and real datasets.