Stationary and nonstationary generalized extreme value modelling of extreme precipitation over a mountainous area under climate change

The generalized extreme value (GEV) distribution is often fitted to environmental time series of extreme values such as annual maxima of daily precipitation. We study two methodological issues here. First, we compare criteria for selecting the best model among 16 GEV models that allow nonstationary scale and location parameters. Simulation results showed that both the corrected Akaike information criterion and Bayesian information criterion (BIC) always detected nonstationarity, but the BIC selected the correct model more often except in very small samples. Second, we examined confidence intervals (CIs) for model parameters and other quantities such as the return levels that are usually required for hydrological and climatological time series. Four bootstrap CIsnormal, percentile, basic and bias-corrected and acceleratedconstructed by random-t resampling, fixed-t resampling and the parametric bootstrap methods were compared. CIs for parameters of the stationary model do not present major differences. CIs for the more extreme quantiles tend to become very wide for all bootstrap methods. For nonstationary GEV models with linear time dependence of location or log-linear time dependence of scale, CI coverage probabilities are reasonably accurate for the parameters. For the extreme percentiles, the bias-corrected and accelerated method is best overall, and the fixed-t method also has good average coverage probabilities. A case study is presented of annual maximum daily precipitation over the mountainous Mesochora catchment in Greece. Analysis of historical data and data generated under two climate scenarios (control run and climate change) supported a stationary GEV model reducing to the Gumbel distribution. Copyright (c) 2013 John Wiley & Sons, Ltd.