Evaluation of Data-Driven Causality Discovery Methods among Dominant Climate Modes

Identification of causal networks in atmospheric teleconnection patterns has applications in varieties of climate studies. We evaluate and compare three data-driven causal discovery methods in locating and linking causation of well-known climatic oscillations. Four climate variables in the ERA-Interim reanalysis data (1979-2018) were employed in the study. We first employ dimension reduction to derive the time-series for selected climate variables. Then time-series of dominant modes were processed using three different causal discovery methods: 1) Granger causality discovery, 2) Convergent cross-mapping (CCM), and 3) PCMCI. Discovered causal links were different for different methods as well as for different variables. However, slightly similar causal links were observed between the Granger causality and CCM methods. Comparison of these three methods was discussed based on the El Niño-Southern Oscillation (ENSO) and its linkage with other oscillations. Causal discovery methods were able to capture the linkage between the ENSO, North Atlantic Oscillation (NAO), and Pacific Decadal Oscillation (PDO), for some of the variables. Overall, this study identifies the usage of these statistical models in locating the direct and indirect causal links among the oscillations. Application of these data-driven causal discovery methods in terms of mediation and direct relationships between the observed teleconnection patterns suggest that the data-driven statistical methods are efficient in locating the regimes of climate patterns and their 12 observed real connections to some extent. We will present the evaluation results of the three causal discovery methods we employed.