In this paper we have put forward a Bayesian framework for the analysis and testing of possible non-stationarities in extreme events. We use the extreme value theory to model temperature and precipitation data in the Dar es Salaam region, Tanzania. Temporal trends are modeled writing the location parameter of the generalized extreme value distribution in terms of deterministic functions of explanatory covariates. The analyses are performed using synthetic time series derived from a Regional Climate Model. The simulations, performed in an area around the Dar es Salaam city, Tanzania, take into account two Representative Concentration Pathways scenarios from the Intergovernmental Panel on Climate Change. Our main interest is to analyze extremes with high spatial and temporal resolution and to pursue this requirement we have adopted an individual grid box analysis approach. The approach presented in this paper is composed of the following key elements: (1) an advanced Bayesian method for the estimation of model parameters, (2) a rigorous procedure for model selection, and (3) uncertainty assessment and propagation. The results of our analyses are intended to be used for quantitative hazard and risk assessment and are presented in terms of hazard curves and probabilistic hazard maps. In the case study we found that for both the temperature and precipitation data, a linear trend in the location parameter was the only model performing better than the stationary one in the areas where evidence against the stationary model exists.
Read full abstract