Peter Miersch

Sensitivity analysis of causal discovery on simulated river flood data using non-linear conditional independence testing

22-Apr-24

Analyzing river floods with causal inference is a promising pursuit, as it holds the potential to unravel the compounding drivers contributing to them, such as intense precipitation, snowmelt, and elevated antecedent soil moisture. Modern causal inference methods, like the PCMCI (PC algorithm Momentary Conditional Independence) framework, are able to identify such drivers from complex multivariate time series through causal discovery and build causally aware statistical models. However, causal inference tailored to extreme events remains a challenge due to data length limitations, proper conditional independence testing tailored to non-linear relationships and a high level of uncertainty in the resulting causal graphs. In this study, we generate a dataset of simulated high runoff events for many different catchments and perform a sensitivity analysis on these data. We go beyond the typically used linear models by applying Gaussian Process Regression in the conditional independence test to identify the causal graph. We illustrate that Gaussian Process Regression is more suitable to represent the relationship between drivers and response. However, we also show that, even using non-linear independence testing and very long time series, the results do not converge towards a single causal graph. Ultimately, this work will help establish best practices in causal inference for flood research to identify meteorological and catchment specific flood drivers in a changing climate.