How important is accounting for serial correlation and field significance in trend detection of extreme rainfall occurrences?

A number of studies have shown that the ability of statistical tests to detect trends in hydrologic extremes is negatively affected by (i) the presence of autocorrelation in the time series, and (ii) field significance. Here, we investigate these two issues and evaluate the power of several trend tests using time series of frequencies (or counts) of precipitation extremes from long-term (100 years) precipitation records of 1087 gauges of the Global Historical Climate Network database. For this aim, we design several Monte Carlo experiments based on simulations of random count time series with different levels of autocorrelation and trend. We find the following. (1) The observed records are consistent with the hypothesis of autocorrelation induced by the presence of trends, indicating that the existence of serial correlation does not significantly affect trend detection. (2) Tests based on the linear and Poisson regressions are more powerful that nonparametric tests, such as Mann Kendall. (3) Accounting for field significance improves the interpretation of the results by limiting the rejection of the false null hypothesis. We then use these results to investigate the presence of trends in the observed records. We find that, depending on the quantiles used to define the frequency of precipitation extremes, 34-47% of the selected gages exhibit a statistically significant trend, of which 70-80% are positive and located mainly in United States and Northern Europe. The significant negative trends are mostly located in Southern Australia.