Resolving large-scale pressures on species and ecosystems: propensity modelling identifies agricultural effects on streams.

Although agriculture is amongst the world's most widespread land uses, studies of its effects on stream ecosystems are often limited in spatial extent. National monitoring data could extend spatial coverage and increase statistical power, but present analytical challenges where covarying environmental variables confound relationships of interest.Propensity modelling is used widely outside ecology to control for confounding variables in observational data. Here, monitoring data from over 3000 English and Welsh river reaches are used to assess the effects of intensive agricultural land cover (arable and pastoral) on stream habitat, water chemistry and invertebrates, using propensity scores to control for potential confounding factors (e.g. climate, geology). Propensity scoring effectively reduced the collinearity between land cover and potential confounding variables, reducing the potential for covariate bias in estimated treatment-response relationships compared to conventional multiple regression.Macroinvertebrate richness was significantly greater at sites with a higher proportion of improved pasture in their catchment or riparian zone, with these effects probably mediated by increased algal production from mild nutrient enrichment. In contrast, macroinvertebrate richness did not change with arable land cover, although sensitive species representation was lower under higher proportions of arable land cover, probably due to greatly elevated nutrient concentrations. Synthesis and applications. Propensity modelling has great potential to address questions about pressures on ecosystems and organisms at the large spatial extents relevant to land-use policy, where experimental approaches are not feasible and broad environmental changes often covary. Applied to the effects of agricultural land cover on stream systems, this approach identified reduced nutrient loading from arable farms as a priority for land management. On this specific issue, our data and analysis support the use of riparian or catchment-scale measures to reduce nutrient delivery to sensitive water bodies.