Estimation Issues for Precipitation Extreme Quantile Determination

Engineering hydrology is primarily conducted using rainfall frequency atlases and return period design storms. Recent anecdotal experience has suggested that this method may result in extreme event estimates that are underestimates of actual events. In this research, 332 rain gauges covering the southeastern United States were analyzed for trends in the 100-year quantile estimate, using annual maxima values derived from daily precipitation data covering the past 79+ years. A nonparametric trend test was applied by use of a moving cumulative 100-year storm estimate calculated using the GEV distribution and confidence intervals generated by a Monte Carlo simulation. The results show about 40% of the stations exhibited some form of nonstationarity, although any upward or downward trends were difficult to ascertain, as were geographic trends. The cumulative sample aggregation method resulted in large fluctuations in quantile estimates. Furthermore, empirical confidence intervals for the extreme quantiles were very wide for much of the moving cumulative range, because of the inherently small sample sizes.