Inferences on weather extremes and weather-related disasters: a review of statistical methods
- 1PBL Netherlands Environmental Assessment Agency, The Hague, The Netherlands
- 2Centre for the Analysis of Time Series, London School of Economics and Political Science (LSE), London, UK
- 3Institute for Environmental Studies (IVM), VU University, Amsterdam, The Netherlands
Abstract. The study of weather extremes and their impacts, such as weather-related disasters, plays an important role in research of climate change. Due to the great societal consequences of extremes – historically, now and in the future – the peer-reviewed literature on this theme has been growing enormously since the 1980s. Data sources have a wide origin, from century-long climate reconstructions from tree rings to relatively short (30 to 60 yr) databases with disaster statistics and human impacts.
When scanning peer-reviewed literature on weather extremes and its impacts, it is noticeable that many different methods are used to make inferences. However, discussions on these methods are rare. Such discussions are important since a particular methodological choice might substantially influence the inferences made. A calculation of a return period of once in 500 yr, based on a normal distribution will deviate from that based on a Gumbel distribution. And the particular choice between a linear or a flexible trend model might influence inferences as well.
In this article, a concise overview of statistical methods applied in the field of weather extremes and weather-related disasters is given. Methods have been evaluated as to stationarity assumptions, the choice for specific probability density functions (PDFs) and the availability of uncertainty information. As for stationarity assumptions, the outcome was that good testing is essential. Inferences on extremes may be wrong if data are assumed stationary while they are not. The same holds for the block-stationarity assumption. As for PDF choices it was found that often more than one PDF shape fits to the same data. From a simulation study the conclusion can be drawn that both the generalized extreme value (GEV) distribution and the log-normal PDF fit very well to a variety of indicators. The application of the normal and Gumbel distributions is more limited. As for uncertainty, it is advisable to test conclusions on extremes for assumptions underlying the modelling approach. Finally, it can be concluded that the coupling of individual extremes or disasters to climate change should be avoided.