# Quantifying statistical uncertainty in the attribution of human influence on severe weather

## TL;DR

This paper develops a rigorous frequentist statistical framework for quantifying sampling uncertainty in climate event attribution, introducing new methods that outperform traditional bootstrap approaches and providing software tools for implementation.

## Contribution

It introduces novel statistical methods for event attribution, improving accuracy and robustness over existing bootstrap techniques, and offers software for practical application.

## Key findings

- New statistical methods outperform bootstrap in simulations
- Methods are more robust with small probabilities
- Software available in climextRemes package

## Abstract

Event attribution in the context of climate change seeks to understand the role of anthropogenic greenhouse gas emissions on extreme weather events, either specific events or classes of events. A common approach to event attribution uses climate model output under factual (real-world) and counterfactual (world that might have been without anthropogenic greenhouse gas emissions) scenarios to estimate the probabilities of the event of interest under the two scenarios. Event attribution is then quantified by the ratio of the two probabilities. While this approach has been applied many times in the last 15 years, the statistical techniques used to estimate the risk ratio based on climate model ensembles have not drawn on the full set of methods available in the statistical literature and have in some cases used and interpreted the bootstrap method in non-standard ways. We present a precise frequentist statistical framework for quantifying the effect of sampling uncertainty on estimation of the risk ratio, propose the use of statistical methods that are new to event attribution, and evaluate a variety of methods using statistical simulations. We conclude that existing statistical methods not yet in use for event attribution have several advantages over the widely-used bootstrap, including better statistical performance in repeated samples and robustness to small estimated probabilities. Software for using the methods is available through the climextRemes package available for R or Python. While we focus on frequentist statistical methods, Bayesian methods are likely to be particularly useful when considering sources of uncertainty beyond sampling uncertainty.