Extending meta-analysis approaches to address routes to Ecological Intensification using agricultural Long-Term Experiments

Ecological Intensification (EI) could help return agriculture to a “safe operating space” for humanity. The research challenge is to use controlled experiments to identify and explore the combinations of EI practices that will be most effective at transforming agricultural system to achieve this, recognising that different combinations of practices might be most effective for different starting points. Agricultural Long-Term Experiments (LTEs) provide a rich data source for exploring the impacts of different practices against the backdrop of weather variation, and combining the data from a collection of LTEs from a range of different cropping environments presents the potential to compare different combinations of practices across these contrasting environments.
In this study we extend and apply a meta-analysis approach, using the metafor R package, to data obtained from LTEs collated within the Global Long Term Experiment Network (GLTEN – www.glten.org), demonstrating the potential for this approach to address our question about EI. A first step is to develop common indices to define key components of EI that appear in different trials. Unusually for a meta-analysis approach, we then start with the raw data from each trial, and apply analyses to each trial to extract contrasts associated with different aspects of ecological intensification, generally using the replication over time to assess the uncertainty associated with these contrasts. We apply separate analyses to different subsets of the generated data to focus on different EI practices, including mainly qualitative moderators to identify the impacts of different combinations of EI practices, and including trial as a random model component. Model selection takes account of the factorial structure defining the combinations of EI practice moderators, using a combination of AIC and the QM and QE test statistics to select the best model. Results can then be presented in terms of both the main effects and interactions amongst the moderating “context” variables, and the relative importance of these terms assessed using individual QM statistics. This approach further allows the assessment of the contributions made from different sources (trials) to the consensus responses, as well as to identify the potential remaining sources of variation.