Inferential analysis
ISARs were fitted using the logarithmic transformation of the Arrhenius
(1921) power model. We compared the models using
adjusted-R2 values as a measure of their
goodness-of-fit. As the models have the same number of fitted
parameters, the R2 are directly comparable, without
modification (Triantis et al. 2005). We also calculated the mean
area-adjusted species richness for all angiosperm and centipede
chorotypes and distinguished between true and land-bridge islands. To
investigate the impact of past geographical changes on current
diversity, we compared the performance of three alternative generalized
linear mixed models (GLMMs) in explaining species richness. The models
correspond to the geographical setting of the Aegean at three stages:
present-day, LGM and MSL. The present-day model (null model) consisted
of present-day area, distance to the mainland, and spatial
autocorrelation as fixed effect and phytogeographical province as random
effect. The models for the paleogeographical settings (LGM or median)
included the same predictors as the null model plus the following
metrics of change in insular geography compared to the present-day
setting: area loss, increase of distance to the mainland and whether an
island was a land-bridge island or remained a true island. We fitted
these three sets of predictors on the following response variables: the
species richness of all six chorotypes of angiosperms, the total
angiosperm species richness, total angiosperm endemics richness and
total centipede species richness. All models were fitted using a
log-link function and Poisson distribution for the error term (lmerTest
package, Kuznetsova et al. 2017). All variables with p-values
> 0.1 were excluded from the models to obtain a set of
“suggestive, but inconclusive” predictors (Murtaugh 2014) before
refitting with the remaining variables. Multicollinearity was addressed
by computing the variance inflation factors (VIF) of the predictors and
removing variables with VIF > 2.5 (Dormann et al. 2013).
Finally, we identified the models that provided most explanatory
potential using multiple measures of goodness-of-fit: the corrected
Akaike Information Criterion (AICc), Bayesian Information Criterion
(BIC), conditional (R2c) and marginal
(R2m) R2 (MuMIn package, Barton
2009) and leave-one-out cross-validated (loocv)
pseudo-R2. All analyses were performed in R version
3.4.2 (R core team 2018). Crete was excluded from all models fitted as
diagnostic residual plots of residuals vs. leverage systematically
designated it as an outlier (not shown). Because of the low number of
endemic centipede species and their over-dispersion in the islets
adjacent to Crete, we restrained our analysis of the centipede fauna to
native centipede diversity.