Hierarchical partitioning (HP)
We conducted HP analyses to evaluate the relative importance of each predictor together with the shared variation (Chevan & Sutherland, 1991). HP can overcome problems related to model-selection procedures, which may fail to provide a valid means for ranking the relative importance of the predictors (Whittingham et al., 2006; Mundry & Nuun, 2009). HP uses all combinations of predictors (2N for N predictors) to determine the amounts of variation explained by each predictor (Lai et al., 2022). This is achieved by calculating the individual shared percentage of variation explained by the model (semi-partialR 2) (Chevan & Sutherland, 1991). If a predictor had a negative semi-partial R 2 due to the strong and complex correlation among predictors (Peres-Neto et al., 2006), we set the value to 0. We evaluated the statistical significance in HP analyses based on 999 permutations (Lai et al., 2022). We applied HP for each taxonomic group with only the relative measures (HPD, and GDPc), only the absolute measures (HP and GDP), and both. We used the R package rdacca.hp to estimate the relative importance of each predictor in multivariate models through canonical analysis, without limiting the number of predictors (Lai et al., 2022).