Statistical Methods for Detection & Attribution of Climate Change

Extreme event attribution refers to analyses of extreme climate events that attempt to relate such events to long-term changes in earth's climate in which greenhouse gas emissions play a large role. The statistical methods used for these attributions include a combination of extreme value theory to assess the probabilities of extreme events under a variety of climate scenarios, combined with analyses that relate observational data to climate model outputs. In this talk, we shall review these techniques and discuss applications to several recent extreme climate events, including the heatwave in the Pacific northwest in 2021 and the wildfires that it the Los Angeles area at the beginning of this year. 

Attribution of climate impacts to a source forcing is critical to understanding, communicating, and addressing the effects of human influence on the climate. Traditional attribution methods are "single-step", directly relating a source forcing and final impact. These methods are unable to utilize additional climate information to improve attribution certainty. To address this shortcoming, we developed a novel multi-step attribution approach which is capable of analyzing multiple climate variables conditionally. That is, a connected series of climate effects are treated as dependent. Significant relationships found in intermediary steps of a causal pathway are leveraged to better characterize the forcing impact. This enables attribution of the forcing level responsible for the observed impacts, while equivalent single-step approaches fail.

This multi-step approach utilizes a carefully designed scalar feature describing the forcing impact, simple forcing response models, and a conditional Bayesian formulation to identify the correct source forcing magnitude. We demonstrate this method for the 1991 eruption of Mt. Pinatubo, which is an example of a short-term, high-variance forcing. Results indicate that including stratospheric and surface temperature and radiative flux measurements increase attribution certainty compared to analyses derived solely from temperature measurements. This conditional framework has broad potential impacts on improving climate attribution assessments in which standard attribution methods may fail. 

While physical understanding predicts a causal relationship between greenhouse gas emissions and warming in the global climate, estimation of the exact magnitude of this causal effect is notoriously difficult to constrain. One of the reasons for this high degree of uncertainty is that the climate system's overall sensitivity depends on how the spatial pattern of temperature changes causally affects outgoing temperature. While climate model simulations provide dynamically informed estimates, performing inference on the observations is challenging due to the lack of suitable counterfactuals and the high-dimensional nature of the global climate system. We propose to address these difficulties through the causal inference framework of stochastic interventions, where the interventions are modeled as continuous spatial Gaussian processes on the domain. Representing the interventions a spatial stochastic process allows for the causal effects to be consistently estimated from the limited observational record. Using a Bayesian framework, prior information in the form of climate model simulations is incorporated into the form of the stochastic interventions in order to relax the underlying assumptions with physical information. The robustness of the results are assessed through sensitivity analyses and validation studies using climate models. 

Climate models are the primary tools for scientists to study climate change. Evaluating whether a climate model simulates the actual climate becomes critical in improving climate models. While most climate model evaluations focused on the mean and dependency of climate process, we focus on marginal extreme behavior, including return levels that often have devastating impacts on our ecosystems and societies. In particular, we aim to identify where the two climate extreme fields exhibit different marginal behavior, by simultaneously evaluating the differences over all spatial locations through multiple testing techniques. However, the large variation inherited in extreme model fitting makes this evaluation more challenging than that for mean and dependency structure. We propose a new multiple testing procedure, bivariate conditional local FDR (BiCLfdr), to efficiently detect signals from highly variable but spatially correlated hypotheses. Our method takes advantage of both the smoothness of large scale spatial variability and the local spatial correlation to enhance the power of comparing the marginal extreme distribution of two spatial extremes. We apply our method to identify where a regional climate model fails to represent the actual extreme behavior of winter precipitation. Our results provide climate scientists with more insights on improving climate models.