One of the most severe limitations of global general circulation models (GCMs) is that the smoothing associated with the relatively coarse spatial resolution makes it difficult to perform meaningful analysis of future climate extremes. In recent years, however, the modelling community has devoted considerable effort towards the development of much higher resolution regional climate models (RCMs). Using forcing data from a global model, RCMs provide climatologies for a relatively small domain at a spatial resolution of 0.5° latitude/longitude, or better. Because of the computing overhead associated with the higher resolution, it is not possible to run RCMs for the typical 250 years of a GCM. This study presents an analysis of an indices of extreme temperatures over Europe, derived from daily temperature date from four RCMs, comparing the periods 1961-1990 and 2070-2099.

The prime focus of the study is to determine an appropriate analytical strategy for examining extremes of model data. As the data cannot be assumed to be normally distributed, many commonly available statistical tests are inappropriate. On the other hand, the non-parametric tests provided in statistics packages tend to have low statistical power. The first part of the analysis, therefore, compares five non-parametric tests, both in terms of general statistical power and in the context of the distributions of the data to be compared. It is found that the Anderson-Darling and Cramer von Mises tests used in tandem provide the most effective comparison, with the more familiar Kolmogorov-Smirnov test, for example, being shown to be completely inappropriate for comparing the distributions of extremes.

The analysis concludes by applying the Anderson-Darling and Cramer von Mises tests to the Generalised Extreme Value (GEV) parameters of the temperature indices, demonstrating their effectiveness at identifying parts of Europe likely to experience statistically significant changes in temperature extremes under global warming. The use of four different models enables the development of meaningful confidence limits in the results of the statistical tests.