Selecting Time Series Length to Moderate the Impact of Nonstationarity in Extreme Rainfall Analyses

• Published in: Journal of applied meteorology and climatology Vol. 57; no. 10; pp. 2285 - 2296
• Main Authors: DeGaetano, Arthur T., Castellano, Christopher
• Format: Journal Article
• Language: English
• Published: Boston American Meteorological Society 01.10.2018
• Subjects: Atmospheric precipitations; Atmospheric sciences; Climate change; Confidence intervals; Design engineering; Design specifications; Duration; Extreme value theory; Extreme values; Extreme weather; Infrastructure; Length; Parameters; Precipitation; Rain; Rainfall; Time series; Trends; Brazil; Canada; New York; United States > US
• ISSN: 1558-8424; 1558-8432
• DOI: 10.1175/JAMC-D-18-0097.1

Observed and projected increases in the frequency of extreme rainfall complicate the extreme value analyses of precipitation that are used to guide engineering design specifications, because conventional methods assume stationarity. Uncertainty in the magnitude of the trend in future years precludes directly accounting for the trend in these analyses. While previous extreme value analyses have sought to use as long a record as possible, it is shown using stochastically generated time series that this practice exacerbates the potential error introduced by long-term trends. For extreme precipitation series characterized by a trend in the location parameter exceeding approximately 0.005% yr−1, limiting the record length to fewer than 70 years is recommended. The use of longer time periods results in partial-duration series that are significantly different from their stationary counterparts and a greater percentage of rainfall extremes that exceed the 90% confidence interval corresponding to a stationary distribution. The effect is most pronounced on the shortest (i.e., 2 yr) recurrence intervals and generally becomes undetectable for recurrence intervals of more than 25 years. The analyses also indicate that the practice of including stations with records of limited length that end several decades prior to the present should be avoided. Distributions having a stationary location parameter but trended scale parameter do not exhibit this behavior.