Data analysis
We evaluated changes in plant, insect, and soil microbial community and functional composition in Wyoming in 2019 and Montana in 2020 in several ways. First, we assessed community level richness across plant species, insect families, and soil microbial OTUs. To calculate richness, we used the codyn package (Hallett et al. 2016) in R version 3.6.2 (R Core Team 2019) for plants and insects, and Qiime2 for soil microbial OTUs (level-7 in Qiime2). Second, we used Bray-Curtis dissimilarity matrices (calculated for plants and insects with the vegan package (Dixon 2003) and for microbes with Qiime2) and non-metric multidimensional scaling (NMDS) with two axes to visualize multidimensional community and functional composition of plants, insects, and microbes. For plants, we focused on changes in non-brome composition by excluding B. arvensis from B. arvensis gradients and B. tectorum fromB. tectorum gradients. We chose plots with very little cover ofB. arvensis in B. tectorum plots, and B. tectorumcover in B. arvensis plots. For functional level responses across plants, insects, and microbes, we assigned each plant species to a functional group (C3 annual grass, C3perennial grass, C4 perennial grass, cactus, forb, sub-shrub/shrub), each insect family to a feeding guild (leaf chewing herbivore, parasitoid, pollen/nectar eating herbivore, predator, sap sucking herbivore, other herbivore) (La Pierre and Smith 2016), and each soil microbial OTU to an ecological function (Faprotax prokaryotic environmental function database; OTUs not found in the database were left as unassigned and OTUs could be assigned to multiple functional groups (Louca et al. 2016)). Third, we determined how each plant, insect, and microbial functional group associated with invasion.
For all analyses, we modeled each site, year, and invasive species separately. To assess invasion gradient establishment, we evaluated the relationship between absolute percent brome cover and relative percent brome cover using Type III mixed-model ANOVAs with a random block effect (lme4 and lmerTest package (Bates et al. 2015, Kuznetsova et al. 2017), using Satterthwaite’s method (Satterthwaite 1941)). We used the same model set-up to compare invasion cover across sites within each invasion level, using Tukey’s pairwise comparisons (Tukey 1977) with Benjamini-Hochberg’s correction for multiple comparisons (Benjamini and Hochberg 1995). We then used the same model set-up to assess how community richness and each functional group related to invasion abundance. To assess compositional differences among invasion levels for both community and functional level analyses, we used permutational multivariate ANOVA (PERMANOVA) with a random effect of block included. For significant effects, we calculated pairwise comparisons (RVAideMemoire package (Hervé 2022)).
We then determined how insect herbivory and total biomass changed with invasion in Montana in 2021 and 2022. For this, we tested the effect of invasion level, species status (native or invasive), and their interaction on total plant herbivory using Type III ANOVA with a random effect of block included. We also tested this for leaf level herbivory damage, but the results were very similar, so we present only on total plant herbivory. Last, we used the same model set-up to assess the relationship between invasion abundance and total insect biomass for all 4 study years.
For all regression analyses, we visually assessed plots of the residuals and Autocorrelation Function/Partial Autocorrelation Function to look for evidence of nonlinearity and autocorrelation, respectively, but we did not find evidence of violations of either assumption in any of our results. To test the assumption of homoscedasticity, we used Levene’s test for equality of variances (Levene 1960). We assessed normality of the residuals of all response variables using Shapiro-Wilk, Anderson-Darling, Cramer-von Mises, and Kolmogorov-Smirnov tests as part of the Olsrr package (Hebbali 2020), transforming data when necessary to achieve approximate normality and homoscedasticity. Throughout this study, we use α = 0.05, but we report results with 0.05 < p < 0.1 as marginally significant.