Causal Inference: A Pathway for System Identification using Observational Datasets

Over the last two decades, research in empirical causal inference has been burgeoning rapidly with a wide range of techniques being developed for retrieving the causal structure of systems from observational time series. In this study, we investigate the efficiency of such techniques in system identification using synthetic data generated from a hydrological model. Specifically, we compare the efficiency of the recently developed methods of convergent cross mapping with methods based on the framework of graphical models and contrast their performances with commonly used methods such as Granger causality. In addition, we examine the sensitivity of such techniques to practical problems frequently encountered in climate research such as sample size, presence of noise and strong seasonality. The results show that, at the limit of large samples, methods based on the framework of graphical models are more efficient than others in system identification, however, they are more sensitive to changes in sample size. Overall, the results suggest that none of the methods is superior in all conditions and the selection of a causal inference method should take into consideration the characteristics of the system under study.