Experimental Design ------------------- I worked in a 25 ha study area at the *rare* research reserve ([http://www.raresites.org/][1]) near Cambridge, Ontario, Canada (43°22’18”N – 80°21’44”W). The area consisted of two parts: a 13 ha array of prairie islands constructed in Fall 2011 in area field cropped intensively for 50+ years, and a 12 ha adjacent ‘mainland’ oldfield. The islands were of three sizes (25, 100, and 400 m²) randomly located on an isolation gradient from the nearest edge of the 12 ha oldfield mainland (see [fig. S1][2]). This mainland field was dominated by plant species ubiquitous to oldfields of eastern North America (grasses: *Festuca rubra*, *Poa pratensis*, forbs: *Solidago* spp., *Daucus carota*, *Cirsium* spp.), and served as the main source of insect colonization to the islands (Harvey & MacDougall 2014). The south and east borders of the study area contained agricultural fields (corn, soy and wheat); a scattered line of trees borders the area between the prairie islands and the crop fields ([Fig S1][3]). At the beginning of the experiment, the 13 ha island area was bare ground. Plant colonization of the islands likely came from two main sources – oldfield species recruiting naturally from wind dispersal or the seed bank (*i.e.*, seeds deposited prior to island establishment). The matrix between the islands was mowed to ~10 cm height throughout the growing season to maintain isolation. Starting in spring 2012, the islands were subjected to one of four perturbation treatments: addition of nitrogen, defoliation by mowing, nitrogen and defoliation combined, and control. The treatments were replicated three times per island size (4 treatments × 3 sizes × 3 replicates = 36 islands). Enriched islands were treated with 10 g N m⁻² of urea pellets [46-0-0] once per year in early June. Patch perturbation was done annually at the end of the growing season (November), by mowing and raking off all aboveground biomass to ground level. Arthropod and Plant Sampling ---------------------------- Sampling of arthropods and plants was conducted at peak plant biomass at the end of July 2013. For insects, we sweep-netted 4 m² quadrats randomly located on each island. There were two quadrats on the 25 m² islands, three on the 100 m² islands, and five on the 400 m² islands, for a total of 120 quadrats. All specimens were stored on ice before being frozen in the lab the same day. Each individual was identified to the family level, following Marshall (2007). We also sampled insects in nine randomly located 4 m² quadrats in the mainland oldfield. Overall, the most abundant herbivores were Cercopoidea (spittlebugs), Chrysomelidae (leaf beetles), Cicadellidae (leafhoppers), Gryllidae (crickets) and Miridae (plant bugs); the most abundant predators were Dictynidae (sac spiders), Salticidae (jumping spiders) and Thomisidae (crab spiders). In all 4 m² quadrats, we measured the composition, richness, and relative abundance (percent cover to 1%) of all plants. The final data set, including the mainland, contained 74 plant species, and 4848 arthropod individuals from 24 primary consumer families and 23 predator families (see tables [S1][4] and [S2][5] for a description of sampled plant and insect taxa in mainland and island areas). Statistical Analyses -------------------- The analytical approach occurred in three complementary steps. I first tested the individual and interactive effects of the treatments (island size, spatial constraints, fertilization, defoliation) on insect herbivore richness, abundance, evenness, and spatial turnover in composition. Then I used a model comparison approach to evaluate whether the detected effects of the stressors on insect herbivores were mainly mediated by changes in plants, predators or both. Finally, I contrasted my findings with insect diversity and composition data from adjacent untreated mainland grassland. ***Treatment Effects*** To determine the individual and interactive effects of the treatments on insect herbivores, I used three-way ANOVA testing the effects of island size, fertilization, and defoliation with spatial constraints as a co-variate. I tested for changes in herbivore taxa richness, abundance and evenness, given that not all community metrics will respond the same way to the treatments. To integrate spatial constraints related to distance between islands and isolation I performed a Principal Coordinates of Neighbourhood Matrix analysis (PCNM) on the geographical coordinates of each sampling quadrat (Borcard & Legendre 2002). For the SEM models (see below) and the ANOVA analyses, I selected the PCNM axis with the highest explanatory power (axis 3, ***R***² = 13%) based on a forward selection procedure (Legendre & Legendre 2012). To test for treatment effects on herbivore among-islands spatial turnover in composition I used PERMANOVA-PERMDISP analyses on square-root transformed Bray-Curtis abundance matrices (Anderson 2005). PERMANOVA analyses are more robust than other similar tests (e.g., Mantel) but still tend to confound differences in species composition across-treatments (location) versus withintreatments (dispersion) differences in community dissimilarity (Warton et al. 2012). In order to separate these two effects, for each significant term from the PERMANOVA, I used a post-hoc multivariate homogeneity of group dispersions analysis (PERMDISP) to test for a change in withintreatment variance (Anderson 2005; Anderson & Walsh 2013). To preserve between-object distance accurately (Legendre & Legendre 2012), I illustrated the results using non-metric multidimensional scaling (MDS). ***Mediated Effects*** To evaluate the relative importance of direct versus trophically mediated 33 effects of the experimental treatments on herbivores, I built four different candidate structural equation models (SEM or path model) that I compared using the Akaike’s Information Criterion corrected for sample size (AICc, Burnham & Anderson 2002). Each model described different possible outcomes covering current hypotheses from the literature on consumer diversity regulation (see [Appendix S1][6] for more details on each candidate model). The first and second models represented the effects of the treatments on herbivores either by a direct effects on plants (bottom-up model) or predators (top-down model). The third model represented interactive top-down/bottom-up effects with the treatments affecting herbivores via effects on both plants and predators (mixed model). The fourth model represented the direct effects of the treatments on herbivores without mediated effects by plants or predators (set to zero, keystone model). For each candidate SEM I included the effect of richness, abundance, evenness and composition at each trophic level. To integrate changes in community composition at each trophic level I reduced the number of variables using principal component analyses (PCA). The first PCA axis was used in each model to represent changes in plant (PCA1 explained 56% of the variation for the island design and 52% in the mainland), herbivore (PCA1 explained 58% of the variation for the island design and 42% in the mainland), and predator (PCA1 explained 21% of the variation for the island and 70% in the mainland) community composition. Community metrics that were strongly correlated (e.g., predator richness and abundance), and thus generated co-variance matrix issues, were removed to keep only one of the two, since they are considered redundant (Grace 2006). To facilitate comparison between selected models, I present standardized path coefficients where the effect of any endogenous (dependent) variable *i* on another potentially co-varying endogenous variable *j* is noted ***β** _{**j : i**}* and the effect of any exogenous (explanatory) variable *k* on an endogenous variable *j* is noted ***γ** _{**j : k**}* (following Grace 2006). ***Mainland Islands*** To test how the among-island dynamics differed from an unperturbed, continuous adjacent habitat, I used the same model selection approach but with all the treatment effects set to a value of zero (i.e., because this area was not treated with N or perturbations). I also used one-way ANOVA to test the effect of fragmentation (i.e., transition from the mainland to the islands) on plant, herbivore and predator richness, abundance, and evenness. To test for the effects of fragmentation on changes in spatial turnover in composition I used PERMANOVA-PERMDISP analyses (see above). One issue that could potentially have affected the analysis is the sampling effort imbalances across different island sizes. The issue derived from the necessary sampling differences in plot number by island size, with smaller islands being relatively over-sampled and larger ones under-sampled (as described above). Both for plants and insects, the sampling covered 32%, 12%, 5% of total island area, from small to large islands respectively. This sampling protocol represented a trade-off between the need to increase the sampling coverage of the larger islands versus the difficulty in matching the sampling intensity of the smaller islands (only 5 m x 5 m) in the larger patches (20 m x 20 m). To account for this issue and evaluate my capacity to correctly assess the effects of island size, I used a re-sampling method akin to rarefaction curves (for details on the method and all linear modeling result tables see [Appendix S2][7]). All analyses were conducted with R 2.15 (R Development Core team), using the vegan package (Oksanen et al. 2013) for the PERMANOVA and PERMDISP, and the lavaan package (Rosseel 2012) for the structural equation analyses. We provide [the entire R code][8] to see the structure and to reproduce each SEM model. The data used for this study are available from the Dryad Digital Repository (Harvey and MacDougall 2015). [1]: http://www.raresites.org/ [2]: https://osf.io/pmxyc/ "figure S1" [3]: https://osf.io/pmxyc/ "figure S1" [4]: https://osf.io/a35nc/ "table S1" [5]: https://osf.io/vx9mg/ "table S2" [6]: https://osf.io/yc947/ [7]: https://osf.io/j9sxq/ "Appendix S2" [8]: https://osf.io/d2z9a/ "R scripts"