Data extraction and effect size calculation
When two different landscape categories were compared; natural or rural versus urban sites, we selected the most extreme category comparison (natural versus urban; Fenoglio et al. 2020). We obtained mean values, sample sizes, and standard deviation from texts or tables (mean value -type data), for each of the two contrasting ecosystems: control (natural, forest, rural, or suburban sites) versus urban (urban sites). A meta-analysis may produce spurious results and further exacerbate publication bias when excluding studies with missing information. Therefore, we converted or imputed data from relevant studies that report incomplete information on means, correlations, variances and sample sizes (Koricheva et al. 2013).
When the effect of urbanisation was measured using a continuous variable (i.e., impervious surfaces, distance to the city centre or green area), we extracted Pearson’s correlation coefficients (r) or the coefficients of determination (R2; r -type data). When none of these values was reported, we used statistical values of parametric tests (e.g., ANOVAs, Chi-square, t-tests; statistic values -type data). If these parameter values were only presented in graphs, we estimated the values from the figures using WebPlotDigitizer (Burdaet al. 2017). If the standard deviation was not shown in graphs; but instead using a boxplot of minimum, maximum, first quartile or third quartile, we estimated it according to Wan et al. (2014). Moreover, when all the above information was not available in the main text, we calculated means and standard deviation or correlation coefficients from supporting material and/or original datasets.
If a publication reported the results of several taxonomic groups or cities separately, each was considered to be a separate observation (Aguilaret al. 2006). When abundance, species richness, traits, or plant reproductive success were reported at multiple time points (months or years), we selected the time point with the higher sample size; if multiple time points had equal sample sizes, we chose the most recent period of sampling, or if possible, we chose the sampling period of maximum pollinator activity (Fenoglio et al. 2020). For pollination services, in addition to the fruit and seed set, the number, rate and duration of visits were also extracted and used as proxies of plant reproductive success (Kleijn et al. 2015).