Extreme events under climate change: a causal extreme value approach

The past few decades have witnessed a surge in academic literature related to climate change as a response to the numerous extreme climate events with tragic consequences on human society and economy, that occurred in various places and due to different physical phenomena, e.g., the deadliest and most destructive 2018 wildfires in California, the major floods in Venice that occurred in November 2019, and the recent bushfires in Australia with economic cost set to exceed all-time records. While global warming is widely agreed on, climate models also project that an increase in greenhouse-gas concentration implies alteration of the frequency and intensity of extreme weather events of different types (Meehl et al., 2007; Seneviratne et al., 2012). The assumption of stationarity, commonly made in climate extreme value analysis, is therefore challenged and the focus of research has shifted towards the identification of environmental covariates driving extremes and the assessment of their effects on the dynamics of multivariate and spatial extreme events.This research plan is directly related to these crucial issues. Relying on multivariate extreme value theory and causal inference, I intend to develop novel statistical approaches to learn the structure of the causal drivers of climate extremes (Project 1) and study the non-stationary behaviour of spatio-temporal extreme events: their frequency, intensity, duration, and spatial pattern (Project 2). Project 1 focuses on causal inference for multivariate extreme events. More precisely, the tail dependence in a random vector that might represent a physical phenomenon observed at multiple locations in space, is assumed to be driven by a set of covariates. The aim is then to learn the dependence structure of this set of drivers and quantify the perturbation of this structure that led to concurrent extreme events at the observed locations. This would result in an improved understanding of the causal effects that multiple contributing variables (climate factors) have on extremes and a characterization of the associated risk by taking into account the dependence structure of these variables. Project 2 puts the spotlight on the dynamic spatio-temporal characteristics of the extremal dependence. Here, the dependence between spatio-temporal threshold exceedances of a continuous process is represented by a latent geometric object moving through space and time, similarly to the hierarchical model of Bacro et al. (2019). Although the latter model is flexible and highly interpretable (storm events), it is unable to capture the above-mentioned changes in the patterns of extremes under an evolving climate. I will thus endow this model with temporal non-stationarity in the characteristics of the latent object describing the extremal dependence. Standard inference approaches for the tail dependence structure rely on composite and/or censored likelihood methods. To allow for time- varying parameters (of the latent geometric object), I will adopt both a local likelihood approach (Davison and Ramesh, 2000) and a spline-based penalised likelihood approach (Wood, 2017). While the applicability of the proposed methodological framework is broad, its practical usefulness is paramount in hurricane and flood-related impact assessment and risk management.