Fewer sites but better data? Optimising the representativeness and statistical power of a national monitoring network

Abstract

Indicators of large-scale ecological change are typically derived from long-term monitoring networks. As such, it is important to assess how well monitoring networks provide evidence for ecological trends in the regions they are monitoring. In part, this depends on the network’s representativeness of the full range of environmental conditions occurring in the monitored region. In addition, the statistical power to detect trends and ecological changes using the network depends on its structure, size and the intensity and accuracy of monitoring. This paper addresses the optimisation of representativeness and statistical power when re-designing existing large-scale ecological monitoring networks, for example due to financial constraints on monitoring programmes. It uses a real world example of a well-established river monitoring network of 254 sites distributed across Scotland. We first present a novel approach for assessing a monitoring network’s representativeness of national habitat and pressure gradients using the multivariate two-sample Cramér’s T statistic. This compares multivariate gradient distributions among sites inside and outside of the network. Using this test, the existing network was found to over-represent larger and more heavily polluted sites, reflecting earlier research priorities when it was originally designed. Network re-design was addressed through stepwise selection of individual sites to remove from or add to the network to maximise multivariate representativeness. This showed that combinations of selective site retention and addition can be used to modify existing monitoring networks, changing the number of sites and improving representativeness. We then investigated the effect of network re-design on the statistical power to detect long-term trends across the whole network. The power analysis was based on linear mixed effects models for long-term trends in three ecological indicators (ecological quality ratios for diatoms, invertebrates and macrophytes) over a ten-year period. This revealed a clear loss of power in smaller networks with less accurate sampling, but sampling schedule had a smaller effect on power. Interestingly, more representative networks had slightly lower trend detection power than the current unrepresentative network, though they should give a less biased estimate of national trends. Our analyses of representativeness and statistical power provide a general framework for designing and adapting large-scale ecological monitoring networks. Wider use of such methods would improve the quality of indicators derived from them and improve the evidence base for detecting and managing ecological change.