Causal discovery from ecological time-series with one timestamp and multiple observations
Ecologists often aim to uncover causal relations within ecological systems using observational data. However, many studies still rely mainly on correlation-based approaches, which do not support causal interpretation. Causal discovery methods have recently gained interest, particularly for ecological time-series data. Yet such data are often scarce, as collecting repeated and consistent measurements over long periods is costly, require strict continuation of protocols and is time-consuming. Consequently, ecologists have often access only to single-time-point observational data generated by dynamical systems. In this setting, unobserved past states can induce substantial unmeasured confounding, limiting the ability of standard algorithms such as PC and FCI to recover micro-level causal relations between observed variables. We show that PC and FCI can nevertheless recover meaningful causal information. In particular, they can identify specific cluster-level structures, which we call clustered super-unshielded colliders, which provide information about the partial causal ordering of macro-level variables. We further show that both algorithms can be reduced to a simple procedure, which we call RestPC, that yields the same identifiable information. We illustrate our results using simulated data and two real-world datasets: one on bird abundance, climate, and land cover, and another on soil microbial communities, environmental, terrain, and geochemical variables.