Causal Inference: Joel Zeder
From Belle Taylor on July 2nd, 2021
This talk has been automatically captioned. You can remove these by pressing CC on the video toolbar.
Name: Joel Zeder
Talk Title: The value of regularisation and model robustness in the context of climate extremes
Abstract: Single-model initial condition large ensembles – an multitude of climate simulations based on the same model with a minor perturbation in the initial conditions – provide novel opportunities to study the physical drivers of large-scale climate extremes. The severity and conditional probability of extreme events, such as heatwaves, can be estimated with a non-stationary generalised extreme value distribution GEV that accounts for variability in the climate system.
In this work we explore the potential of applying established and recently developed methods of statistical learning theory to address the challenge of building a robust model and selecting relevant covariates from (often highly correlated) climatic fields. Based on millennial simulations of stationary and transient climate we use regularisation to obtain an objective subset of explanatory physical predictors (ranging from large-scale thermodynamic to local short-term dynamic climate variables) determining model parameters, and the application of an anchor regression approach (Rothenhäusler et al., 2021) ensures model stability also under decadal internal climate variability.
Rothenhäusler, D., Meinshausen, N., Bühlmann, P., & Peters, J. (2021). Anchor regression: Heterogeneous data meet causality. Journal of the Royal Statistical Society. Series B: Statistical Methodology, (June 2018), 1–32. https://doi.org/10.1111/rssb.12398
This talk is a contributed talk at EVA 2021.