资源环境科技发展态势分析平台

Serial correlation in a hydrometeorological time series can have deleterious effects on trend detection tests. To account for significant autocorrelation detected in datasets, various techniques have been developed over time, each having their own assumptions and accuracy. Furthermore, the existence of positive or negative serial correlation has dissimilar effects on these statistical techniques. This research compares the power and Type I error rates of various well‐known and several newer techniques to account for positive and negative serial correlation in combination with the Mann‐Kendall nonparametric trend test. The paper additionally explores the application of these techniques in the presence of higher‐order dependence structures. Through a case study of southern Ontario watersheds, it is determined that the block maxima series (BMS) data are more likely to have significant negative lag‐1 serial correlation. Peaks‐over‐threshold (POT) data are more likely to be serially correlated and this autocorrelation is more likely to be positive. It is determined that in the case of positively serially correlated AR(1) data, block bootstrap (BBS), Hamed and Rao (1998) variance correction (VCCF1), Yue and Wang (2004) variance correction (VCCF2), and sieve bootstrap (SBS) are the most robust. Alternatively, in the case of negative AR(1) autocorrelation, the corrected trend‐free prewhitening approach (CTFPW), modified trend‐free prewhitening (MTFPW), bias corrected prewhitening (BCPW), and VCCF1 are recommended. In the presence of higher order dependence structures VCCF1 (with all significant lags included) and VCCF2 (with all lags included) should be applied cautiously. Lastly, an assessment of the causality of the serial correlation is provided.

This article is protected by copyright. All rights reserved.