On the Attribution of Weather Events to Climate Change Using Empirically Fit Extreme Value Distributions

Abstract Changes in extreme weather events are a potentially important aspect of anthropogenic climate change (ACC), yet they are difficult to attribute to ACC because the record length is often similar to, or shorter than, extreme-event return periods. This study is motivated by the “World Weather Attribution” (WWA) initiative and, specifically, their approach of fitting extreme value distribution functions to local observations. They calculate the dependence of distribution parameters on global mean surface temperature (GMST) and use this dependence to attribute extreme events to ACC. Applying this method to preindustrial climate simulations with no time-varying greenhouse gas forcing, we still find a strong dependence of distribution parameters on GMST. This dependence results from internal climate variability (e.g., ENSO) affecting both extreme events and GMST. Therefore, dependence on GMST does not necessarily imply an effect of ACC on extremes. We further consider whether an extreme value, normal, or lognormal distribution better represents the data, if a GMST dependence of distribution parameters is justified using a likelihood ratio test, and if a meaningful attribution is possible given uncertainties in GMST dependence. We find, for example, that an attribution of Australia’s 2020–21 bushfires to ACC is difficult due to the effects of internal variability. For the 2019–21 drought in Madagascar, we find that the small number of available data points precludes a meaningful attribution analysis. Overall, we find that the effects of internal climate variability on GMST and the uncertain relationship between GMST and regional extremes may lead to inaccurate attribution conclusions using the part of the WWA approach examined here.