Poor performance of regime shift detection methods in marine ecosystems

Abstract

Regime shifts have been reported as ubiquitous features across the world’s oceans. Many regime shift detection methods are available, but their performance is rarely evaluated, and the supporting evidence for regime shifts may be thin because of the nature of marine ecological time series that are often short, autocorrelated, and uncertain. In the Norwegian Sea, a regime shift has been reported to have occurred in the mid-2000s, with simultaneous changes in oceanography, plankton, and fish. Here, we evaluate the evidence for this regime shift using four commonly used regime shift detection methods (Strucchange, STARS, EnvCpt, and Chronological Clustering) on 32 annual time series that describe the main components of the Norwegian Sea ecosystem, from hydrography and primary production up to fish population metrics. We quantify the performance of each method by measuring its false-positive rate, i.e. the proportion of times the method detects a regime shift that was not present in simulated control time series. Our results show that all methods have high to very high false-positive rates. This challenges the evidence for a regime shift in the Norwegian Sea and questions earlier reviews of regime shifts across the world’s oceans.