Predicting microcystin concentrations in lakes and reservoirs at a continental scale: A new framework for modelling an important health risk factor

• Published in: Global ecology and biogeography 2017-06, Vol.26 (5/6), p.625-637
• Main Authors: Taranu, Zofia E., Gregory-Eaves, Irene, Steele, Russell J., Beaulieu, Marieke, Legendre, Pierre
• Format: Article
• Language: English
• Subjects: Algae; Algal blooms; boosted regression tree [BRT]; cross‐scale interaction [CSI]; Cyanobacteria; ecoregion; Environmental changes; eutrophication; Fresh water; Health risks; Heterogeneity; Impairment; lake; Lakes; land use; Microcystins; Modelling; NGOs; Nongovernmental organizations; Predictions; Regression analysis; reservoir; Reservoirs; Risk factors; Scale models; Spatial heterogeneity; Thresholds; Toxicity; Toxins; Water quality; zero‐altered hurdle model
• ISSN: 1466-822X; 1466-8238
• DOI: 10.1111/geb.12569

Aim: Scientists, governments and non-governmental organizations are increasingly moving towards the collection of large, open-access data. In aquatic sciences, this effort is expanding the scope of questions and analyses that can be performed to further our knowledge of the global drivers of water quality. Cyanotoxin concentration is one variable that has received considerable attention, and although strong local-scale models have been described in the literature, modelling cyanotoxin concentrations across broader spatial scales has been more difficult. Commonly used statistical frameworks have not fully captured the complex response of toxic algal blooms to global change, limiting our ability to predict and mitigate the impairment of freshwaters by toxic algae. Here, we advance our understanding of emergent drivers of cyanotoxins across a structured landscape by applying a hierarchical "hurdle" model. Location: Lakes and reservoirs in the conterminous United States [n = 1127]. Methods: We studied cyanobacteria and their toxins [microcystins] during the 2007 summer period. We applied a hierarchical zero-altered model to test the importance of multi-scale interactions among environmental features in driving microcystin concentrations above the limit of detection. We then used boosted regression trees [BRTs] to identify environmental thresholds associated with severe impairment by microcystins. Results: Accounting for numerous non-detections, spatial heterogeneity and cross-scale interactions substantially improved continental-scale predictions of bloom toxicity. Our model accounted for 55% of the variance in the probability of detecting microcystins across the United States, and 26% of the variability in microcystin concentrations once detected. BRTs further showed that although both local and regional drivers were associated with microcystin concentrations at low to intermediate provisional guidelines, only local drivers came into play when predicting higher limits. Main conclusions: Identifying the interaction between local and regional processes is key to understanding the heterogeneous responses of microcystins to environmental change. Our framework could increase the effectiveness of continental-scale analyses for many different water variables.