Statistical analyses
We analyzed the effect of landscape parameters (arable field cover and landscape heterogeneity) and different measures of the local flowering plant community on wild bee and hoverfly species richness and abundance, i.e. number of caught individuals, of a respective site. Apis mellifera , the European honey bee, was excluded from all analyses. Arable field cover, landscape heterogeneity and plant community attributes were investigated in separate models, but the overall model structure was the same. We used GLMMs with poisson distribution and a log-link function for both response variables species richness and abundance (glmer , R-Package lme4, Bates et al. 2020). As covariates, we included sampling campaign as categorical fixed-effect and study site as random-effect to account for the nested design of the study. About 92% of all hoverfly individuals (n=214) were caught in the third sampling campaign and 3/4 of the sites within the first two campaigns did not contain any hoverflies, which is a typical pattern for agrarian landscapes (Brandt et al. 2017). Due to this highly unequal distribution across the sampling campaigns, we restricted our analyses of of local plant attributes on hoverfly species richness and abundance for the third sampling campaign. For these, we used GLMs with poisson error for species richness and negative binomial for abundance without any covariates, as we had no pseudo-replication in this dataset. We assured that model assumptions (normality and over-/underdispersion of residuals, heteroscedasticity, spatial autocorrelation of response variables and model residuals and zero-inflation) were not violated with R-package DHARMa (Hartig and Lohse 2020).
In order to reveal the effects of arable field cover and landscape heterogeneity on pollinators (hypothesis 1 and 2), we determined the scale of effect, i.e. at which spatial scale a landscape parameter has the largest effect on the response variable (Jackson and Fahrig 2015, hypothesis 3). For this purpose, we used the multifit -function of Huais (2018), which compares a series of models that differ solely in the scale a landscape parameter was quantified (see the specific model structure above). In our case, we compared models of arable field cover and landscape heterogeneity that were quantified at radii between 60m and 3000m. Landscape predictors were standardized (z-scaled) for each radius in this analysis that parameter estimates are comparable (Schielzeth 2010). The model with the lowest AIC was considered to be the best model. Further, we assured that the natural distribution of dry grasslands in clusters did not affect our results due to possible pseudoreplication of the landscape parameters (Supplementary information, Appendix 3). Additionally, we assessed whether the predictors had a statistically clear effect sensu Dushoff et al. (2019).
We predicted that functional diversity of flowers and traits associated with attractiveness have a positive effect on pollinators (hypothesis 4). Functional diversity was estimated with Rao’s quadratic entropy (FDtrait) of the different traits (see above) and the ‘attractiveness’ with community weighted mean (CWMtrait) (Fornoff et al. 2017). The frequency that a forb species occurred within the eight segments of the vegetation survey was used as abundance measure for the weighting of forb species. In a first step, we checked for possible correlations between FDtrait,, CWMtrait and number of flowering forb species (Supplementary information, Appendix 4). Due to multiple correlations, we included only parameters that were not strongly related to each other (|r|<0.6): number of flowering forbs, FDflowering height,, FDcolor,, FDnectar access,, FDUV reflectance, and CWM traits that should attract pollinators, CWMcolor yellow (percentage of yellow flowering species) and CWMflowering height in our full model. In order to find the most likely parameter combination of local plant community attributes that explain pollinator richness and abundance, we applied an information theoretic approach and compared all possible submodels derived from the full model with AICc (Burnham and Anderson 2002). Since several models performed equally well, we performed model averaging over the best models (delta AICc <6, Harrison et al. 2018), in order to get more reliable parameter estimates (Dormann et al. 2018). Local plant community attributes were not correlated to the landscape predictors at any scale. All analyses were carried out in R Version 4.1.0 (R Core Team 2021).