In this post Craig Allen and Hannah Birge discuss a paper from Sarah Burthe and colleagues ‘Do early warning indicators consistently predict nonlinear change in long-term ecological data?’

Complex systems of humans and nature rarely change in gradual, expected ways. Instead, changes often occur suddenly, with major, non-linear losses of human and ecological capital. Once these unexpected changes occur, it can be difficult to restore the system –even with intensive intervention. As a result, there is a growing desire among managers, ecologists, and policymakers to understand how surprising, non-linear system shifts can be avoided.

One way to better understand non-linear system changes is through the identification of early warning indicators. Increased variance and temporal autocorrelation are two early warning indicators that have emerged from theory and are supported by laboratory manipulations. They are also easily identifiable in large data sets, making them attractive to those who study complex systems. However, the application of these early warning indicators in larger, more complex ecosystems is mixed. This is likely due to the lack of carefully executed studies designed to isolate and test whether increased variation and autocorrelation precede non-linear system changes (which the authors assumed to be indicators of regime change), and due to the difficulty in separating stochastic environmental variation from actual trends.

Burthe et al. set out to do just this, undertaking one of the largest ever investigations of early warning indicators from real world data sets. The authors analyzed process rate data from multiple trophic levels across six different types of aquatic ecosystems. They searched the data for (1) non-linear changes, (2) periods of increasing variance and autocorrelation, and (3) whether and where (i.e. ecosystem type, trophic level) variance and autocorrelation preceded a non-linear change. This thorough approach allowed Burthe et al. to identify whether early warning indicators existed in the data and, if so, whether they occurred within a trophic level (e.g. plankton scale indicators preceded a plankton scale non-linear change) and/or across trophic levels (e.g. high variability and autocorrelation at the plankton scale preceded a non-linear change at the apex predator scale). By including multiple types of aquatic systems, the authors also hoped to elucidate universal aspects of early warning indicators transferable across complex systems.

The authors found repeated examples of changing variance and autocorrelation, and non-linear changes at different trophic levels and in every system. Yet they found very little evidence that the two were linked. Instead, variance and autocorrelation would often increase without a subsequent system shift, or a system shift would occur without preceding indicators. Although the authors did find a case of increased variance and autocorrelation preceding a non-linear change not attributed to chance alone, this was the exception, and not the rule.

Burthe et al.’s work highlights the need to probe excellent ideas with well-designed empirical studies. Increased variation and autocorrelation are relatively easy to identify in data sets, and resemble what should precede a threshold according to our growing understanding of complex systems. Remaining challenges for the interpretation of these analyses is whether or not non-linear changes actually signify a regime change in real world systems, and whether the variables analyzed are likely to be mechanistically linked with basic system processes of interest. This work reveals that variability and autocorrelation are far from straightforward, general indicators of non-linear change. Unsurprisingly, reality is much messier than theory permits us to presume.