Williams-Climate-change-refugia-for-terrestrial-biodiversity_0

Atlas of Living Australia (Table 3). Vertebrate data were compiled and vetted by A.
Reside and collaborating researchers. Across NSW, three sources of vascular plant
data were compiled and applied in separate case studies. Across the entire state, we
used the floristic survey data compiled and vetted by the NSW Office of Environment
and Heritage (OEH) for GDM analysis (Logan et al. 2009). In south-eastern NSW, we
used CSIRO’s survey data for tree species capable of reaching the canopy (Austin et
al. 1996). In far southern Western Australia we used the Tingle Mosaic survey data
(Wardell-Johnson and Williams 1996). More detailed descriptions of these datasets are
provided in APPENDIX 7. Compositional turnover modelling.
The biological response variable used for fitting the GDMs was the Sørenson
compositional dissimilarity between pairs of sites. The number of possible site-pairs in
a dataset is frequently beyond the capacity of conventional computing. A Perl script
written as an extension to Biodiverse (Laffan et al. 2010) was used to generate sitepairs
and their Sørenson dissimilarity values (described in Rosauer et al. in review).
We randomly sampled approximately 300 000 site-pairs, evenly stratified by Australian
bioregions (DSEWPAC 2012) for the continental data and subregions for the NSW
data. Equal weighting was applied to each bioregion or subregion with a slight
emphasis (10%) on sampling more site-pairs from within regions relative to between
regions. The sampling of NSW floristic survey data was further weighted by the log of
the total number of species within each subregion. For sampling purposes, a site is
defined by the resolution of the spatial analysis grid (9-second, 3-second, 1-second).
Lists of species were aggregated within each site.
To account for variation in survey adequacy, we used the number of species occurring
at a site as a threshold in generating site-pair samples for the continental case study.
Presence-only data, such as that aggregated by the Atlas of Living Australia, are
dominated by ad hoc observations that under-sample biodiversity at the site level.
Under-sampled sites can lead to inflation of estimated levels of compositional
dissimilarity between sites, and contribute to model error. The threshold number of
species was determined after testing alternative site-pair sample data with the same
predictors in GDM models. As the threshold for the number of species is increased, the
number of occurrence sites and species represented decreases, the model deviance
explained increases and the sum of predictor coefficient values decreases. Where the
sums of predictor coefficient values were similar, but the deviance explained increased,
the model with a lower threshold number of species was selected. This ensured more
sites (and species) were available for site-pair sampling. Any overall inflation of
estimated dissimilarities remaining after applying this threshold was accounted for, and
adjusted, through the inclusion of an intercept term in the fitted models. This term
effectively accounts for the average level of dissimilarity expected to be observed
between two sites with identical environmental attributes, given the average level of
under-sampling exhibited by the model-fitting dataset. Survey adequacy, and the
effects of under-sampling, were much less of an issue for the biological datasets used
in the NSW and Tingle Mosaic case studies, despite the smaller grid-cell sizes
employed. This was because these data were derived from thorough presence–
absence sampling of floristic survey plots rather than from aggregation of presenceonly
records.
Climate change refugia for terrestrial biodiversity 57