Evaluating the Robustness of PCMCI+ for Causal Discovery of Flood Drivers

Estimating causal drivers of high-impact extreme events such as floods from data is an aspiring pursuit. Time series causal discovery methods, such as the conditional-independence-based PCMCI framework, are designed to identify causal relationships from complex multivariate observational time series. However, the application to extreme event data remains a challenge due to, by the nature of extremes, data length limitations, conditional independence testing for nonlinear relationships, and potential violations of the methods’ assumptions. So far these challenges have mostly been explored on synthetic data with limited transferability to real-world applications. In this study, we evaluate causal discovery on real and pseudo-real data generated with a hydrological model across 45 catchments with varying flood-generating processes. Because no detailed causal ground truth exists, we focus on the robustness of output graphs. To this end, we simulate a large sample, identify discharge peaks, and investigate the robustness of the causal discovery algorithm PCMCI+ when applied to different realizations of the same setting for various sample sizes. We find that the robustness generally increases with sample size, yet a significant proportion of inferred causal edges remain inconsistent even for large datasets. Notably, while some flood drivers are reliably identified, other key hydrological mechanisms are systematically missed even for very large sample sizes, highlighting methodological limitations. Our study provides a blueprint for investigating the real-world performance of causal discovery methods and illustrates their current limitations for identifying causal drivers of floods.