Statistical analyses
We used linear mixed models (LMM), as in Dawson et al. (2017), which arguably is the commonest approach used in ecology and evolution (Table S1). We used RF and HP as alternative techniques to estimate measures of relative variable importance. The three techniques followed designs similar to the mixed models in Dawson et al. (2017) to facilitate comparisons among methods. For the HP and LMM (but not in RF because the outcome is invariant to log transformation; Breiman, 2001), we log-transformed the region area, and the relative and absolute measures of economy and population, and we used a square-root transformation for precipitation to satisfy assumptions of linearity, normality of residuals and homoscedasticity and to have the same LMMs as in Dawson et al. (2017).