Analysis
Our analysis was developed in R 4.1.3 (R Core Team 2021). We began by filtering the data, retaining only datasets and patches compatible with our criteria, resulting in 76 metacommunities containing 1190 patches. Then, we implemented the bias control procedure (Appendix), which resulted in 100 unbiased samples for each metacommunity. From these we generated 7600 SLOSS comparisons. The 100 samples for each dataset represent uncertainty in the SLOSS outcomes in each metacommunity due to resampling. We developed an algorithm that automatically classified the SLOSS comparisons into SS > SL (positive responses of biodiversity to fragmentation), SS = SL (no evidence of a biodiversity responses to fragmentation), or SL > SS (negative responses of biodiversity to fragmentation). The algorithm classifies each SLOSS comparison based on the difference between the two species accumulation curves in the interval between the areas of (i) the largest patch on the large-to-small curve and (ii) all patches except the largest patch on the small-to-large curve (Fig. 1-d).
Next, we fitted generalized linear mixed effects models to determine whether the SLOSS outcome depended on the general taxon of the metacommunity. We assumed a binomial distribution of the response in our models, predicting the likelihood of a dataset being SS > SL. Because SL > SS was much rarer than SS > SL, modelling SL > SS would result in poor model fit. We therefore only present models predicting SS > SL, but we note that results for SL > SS are opposite to those for SS > SL, as expected (Fig. 2). We also included patch size evenness in the model because previous work suggested that the likelihood of observing SS > SL strongly increases with decreasing patch size evenness, i.e., as the difference in size between small and large patches increases (Riva & Fahrig 2022). The model included a random effect for metacommunity to account for the dependency among the 100 SLOSS comparisons generated for each metacommunity.