Comparison of abundances between 1982 and 2018:
In order to compare abundances and biomasses between the two time periods, we used two approaches. First, we fitted the log-normal model (Oksanen et al. 2013) to the rank abundance distribution of species in 1982 and 2018 (McGill et al. 2007) and compared the parameter estimates using 1000 parametric bootstrap replicates. An overlap in log-normal SADs’ confidence intervals suggested no change in community structure. Second, we regressed the abundances and biomasses in 2018 against those in 1982 using GLMs (Supplementary Information). To determine if the slope of the abundance regression was equal to one, we used a Likelihood Ratio Test with a null model constraining intercept=0 and slope=1 and the alternative with a free slope. We partitioned the data by foraging guild, foraging strata, sociality and habitat specialists and re-ran regressions using the same approach as for the entire data set. We excluded outlier species from each test using Cook’s distance (Venables & Ripley 2002). We performed a non-parametric bootstrap test to evaluate the number of outliers expected by chance using alpha=0.01. To account for phylogenetic dependency, we computed the linear model with a correlation structure based on the community’s phylogeny and compared the BIC of the latter model with the model without phylogenetic correlation structure. We repeated this procedure for 1000 phylogenetic trees from BirdTree.org and report the mean and the 2.5% and 97.5% quantiles difference in BIC between the non-phylogenetic and phylogenetically corrected models.