Modeling Non-Stationary Temperature Extremes with Level-Dependent Extremal Dependence

• Activity Number: 72 - Methods for Extreme Values in Environmental Data
• Type: Contributed
• Date/Time: Monday, August 3, 2020 : 10:00 AM to 2:00 PM
• Sponsor: Section on Statistics and the Environment
• Author(s): Peng Zhong* and Raphael Huser and Thomas Opitz
• Companies: KAUST and King Abdullah University of Science and Technology (KAUST) and BioSP, INRAE, Avignon
• Keywords: Asymptotic dependence and independence; Block Maximum; Extreme Event; Max-infinitely divisible process; Max-stable process; Sub-asymptotic modeling

Abstract

The statistical modeling of spatio-temporal trends of temperature extremes can help to better understand the structure and frequency of heatwaves in a changing climate. In this work, we model temperature extremes over the southern half of Europe using annual maxima observed at 44 monitoring stations over 100 years. We extend the spectral representation of max-stable processes to construct a novel, highly flexible, max-infinitely divisible model that includes covariates in the dependence structure to capture non-stationarities in space and time. Unlike max-stable processes, our model can capture weakening extremal dependence at increasing quantile levels and remains in the neighborhood of the popular max-stable class of extremal-t models. Parameter estimation is done by pairwise likelihood inference, and we implement a parametric bootstrap to assess the uncertainty of parameter estimates. In our real data application, our model outperforms other natural alternative models. Results show that the spatial extent of heatwaves becomes smaller when very high temperatures occur and in regions with higher altitudes, and hint at a climate change effect.