Model output
To analyse the output of the simulation model, we ran linear models of defaunation parameters and species traits (predictor variables) against ‘defaunation response’ (response variable; log10-transformed to account for extreme skew). We assessed a global model as well as models representing all subsets of predictor variables, and selected the model with the lowest BIC value as our top ‘additive model’ using the MuMIn package (Barton 2019) in R (R Core Team 2018). We used BIC rather than AIC because it more strongly penalizes complex models, and so is better for identifying a few variables with strong effect (Burnham & Anderson 2004). We also created ‘interaction models’ to assess whether the inclusion of statistical interaction terms between main-effect variables in the additive model improved model fit.
For the additive model and each interaction model, we used therelaimpo package (Grömping 2006) in R to calculate the relative importance of each variable, measured as the contribution of each variable to the model’s R2, following the Lindeman-Merenda-Gold (Lindeman et al. 1980) method.