Speaker
Description
Inferring causal relationships from ecological time series enables the study of ecosystem dynamics without direct observation. Methods such as Unified Information-Theoretic Causality (UIC) make this possible but assume independent abundance measurements. In many ecological applications, particularly eDNA studies, data are compositional and contain only relative abundances. The common remedy is log-ratio transformation to restore Euclidean structure, but this introduces shared noise across taxa that may obscure causal signals.We evaluated UIC’s ability to recover known causal relationships from simulated ecological communities under varying process and observational noise. Time series were generated using generalized Lotka–Volterra dynamics for systems with 3, 10, or 30 interacting species. Each dataset was analyzed as original abundances, closed compositions, and CLR-transformed compositions. As expected, causal inference performed best on original abundance data. However, analyses using compositional data consistently outperformed CLR-transformed data. This advantage increased with community size, with compositional data even approaching the performance of raw abundance data at high dimensionality. These findings indicate that CLR transformation can degrade rather than improve causal signal detection in compositional ecological data, suggesting that untransformed compositional data may be preferable for causal inference.
