Variance Correction Prewhitening Method for Trend Detection in Autocorrelated Data

Abstract

Detecting trends in hydrometerological data through the commonly used Mann-Kendall test is misleading in the presence of data autocorrelation. Autocorrelation seriously interferes with type I errors and power of trend detection. To mitigate this effect, the authors introduce a variance correction prewhitening method. It addresses two important issues that lacked appropriate attention in the past application of trend-free prewhitening method: inflationary variance of slope estimator and deflationary serial variance. After serial and slope variances correction, the new method keeps a better balance between maintaining a low type I error and a relatively strong power of trend detection. In comparison, other methods for the same purpose only address one of these two characteristics. The new method bears some resemblance to the block-bootstrap method; however, it is superior in its simplicity for implementation. Case studies reveal that uncertainties arising from autocorrelation are substantial. Applying more than one test is helpful to interpret results with uncertainties information. The new method provides a robust choice to this strategy.