Predicting spatial and temporal variations: maps
We mapped the effect of future climate change on the magnitude of the
effect of ecosystem engineers on arthropods using the effect sizes
(Hedges’ d) obtained at the experiment scale.
The magnitudes of the engineering effects on arthropod communities
observed at each site and their association with climatic variables (see
Table S1 for variable definitions) were tested using linear models
(Gaussian error distribution). We used the function predict.lm to
obtain the predicted relationship between independent variables and the
Hedges’ d effect size (abundance, richness and biomass) (n = 52).
Temperature (bio1, bio2, bio4, bio7) and precipitation (bio12, bio14,
bio15) were included as independent continuous variables in separate
models (see Table S1 for variable definitions). Separate climatic
variables rather than the principal component scores identified in our
meta-analyses were used because the correlations structure among
variables may shift with climate change, and thus parameter estimation
may be biased in different models (Hadi and Ling, 1998). In addition,
using the climatic variables in the predictive modelling allowed us to
inspect the reliability of the predict.lm function. However, for
both temperature and precipitation climatic models most variables were
collinear (see Figs. S2 and S3). We therefore first eliminated the
influence of collinearity using the variance inflation factor (VIF)
(Zuur et al., 2010). We then reduced the model through backward
selection, retaining only the significant variables and those that
improved the model fit (i.e., higher R2 values). We
modelled the current climate scenario and those projected to 2070 using
rcp8.5 (MIROC5) as a general circulation model. We checked normality,
homoscedasticity and outliers through graphical inspections (e.g.,
qq-plots, Cook’s d, and influence), and residuals of the response
variables were log transformed when necessary. Previous studies have
used this approach to modelling distribution maps to provide a
geographical interpretation of global ecological patterns (Tallavaaraet al ., 2018, Gusmão et al ., 2020).
Model performance was tested using a k -fold cross-validation
approach (see Appendix S1).