Date of Award

Fall 2016

Project Type


Program or Major


Degree Name

Doctor of Philosophy

First Advisor

Ernst Linder

Second Advisor

Jennifer Jacobs

Third Advisor

Linyuan Li


Downscaling of precipitation extremes, while providing essential information for impact assessment of climate change and the development of adaptation strategies, is subject to significant uncertainty, particularly in connection with long-term predictions. The focus of this dissertation is to quantify the uncertainty and reduce the variability in probabilistic downscaling of precipitation extremes, especially with regard to the prediction of long-term extreme events such as 25-year return levels. Two major sources of downscaling uncertainty were addressed: uncertainty arising from parametric distribution modeling for precipitation extremes and uncertainty due to the particular predictive model employed for estimation of their return levels. Both sources were examined and quantified using credibility intervals of downscaled 25-year return levels of precipitation extremes. This was done by a probabilistic downscaling procedure, linking regional climate model (referred to as RCM_GCM) output obtained from the North American Regional Climate Change and Assessment Program (NARCCAP) to daily precipitation data of 58 weather stations in the New England region. A total of eight RCM_GCM models were considered in our study.

The uncertainty from parametric distribution modeling was reduced by extending the traditional method of separate analysis for the individual stations to the spatial hierarchical modeling for all stations. In our study, a latent spatial process is assumed for the scale parameters of the generalized Pareto distributions (GPDs) for all the stations over the study region. This spatial process then includes both a regional effect explained by geographic covariates and a local effect captured by an exponential spatial covariance structure. Thus, a spatial pattern of precipitation extremes over the entire study region was obtained by simultaneously estimating the GPD distributions for all the stations. The empirical results reduced the 90% confidence intervals for the future parameters, as well as the downscaled 25-year return levels, relative to the results obtained from station-specific downscaling. Such reduction in uncertainty was found for all 58 weather stations and all eight RCM_GCM models.

Moreover, downscaling uncertainty could also arise from the model used for downscaling, including choosing the type of translation applied to the downscaling and the choice of predictive model used for the return level estimation. In particular, two types of translation (“model-to-station” translation (MST) and “current-to-future translation” (CFT)) were employed, and both a restricted model and a full model were used for parametric estimation. Empirical comparisons of 90% confidence intervals of the return levels show that the differences between MST and CFT translations are surprisingly small for all 58 stations and all eight regional climate models, and the differences between the two predictive models depend largely on the discrepancies in the estimated shape parameters.

Based on the confidence intervals of return levels, we made regional inferences on the 25-year return levels for precipitation extremes in New England. The results indicate that at the 90% confidence level, no decreasing trend in the 25-year return levels is found in the comparison of the period 2040–2070 with the calibration period of 1970–2000, and that within the region, significant increases in the return levels are more likely to occur in the southern part of New England and along the coast of Maine.

To sum up, our study highlights the importance of uncertainty when drawing inferences on downscaled return levels. In particular, we find that the use of spatial hierarchical modeling can significantly reduce such uncertainty in the downscaling of precipitation extremes by taking into account the spatial dependence among data series. Furthermore, we also find that the uncertainty is not largely affected by the type of translation, but can vary substantially for some cases when different predictive models are used.

Available for download on Tuesday, September 01, 2116