Spatial Analysis of Plague in California - Recent Advances and Issues in Environmental Science - page 124

were set aside for model testing and independent validation. Ten thousand random
background points (pseudo-absences) were used to evaluate commission. A regular-
ization setting of two was used for data smoothing and to address spatial autocorre-
lation. Model results were compared and validated using area under the ROC curve
(AUC) statistics. The AUC statistic is similar to the MannWhitney U test and com-
pares the likelihood that a random presence site will have a higher predicted value in
the model than a random absence site [42, 43]. One of the appeals of ROC curves is
that they do not depend on a user-defi ned threshold for determining presence versus
absence. However, because using a geographical extent that goes beyond the presence
environmental domain can lead to infl ated AUC scores [44, 45], we limited the study
area to the rough geographic extent of the sampling distribution (i.e., the California
state boundary). The four most predictive models were used as the fi nal models, and
mapped as a cumulative probability output.
To explore the spatial relationship between model predictions and serologic
samples of carnivores, we compared the fi nal model results to data on positive and
negative specimens from California coyotes. We used prediction values extracted for
negative and positive coyote specimens using Hawth's point intersect tool [46]. A
one-tailed t-test was performed using JMP (SAS Institute) to test the hypothesis that
model predictions at positive coyote points would be signifi cantly higher than model
predictions at negative coyote points.
In order to simulate the distribution of plague under possible future climate con-
ditions, we ran Maxent using coupled global climate model data from the IPCC 3rd
Assessment (available at http://www.worldclim.org/futdown.htm). These data were
originally produced by three different global climate models: CCCma [47], HadCM3
[48, 49], and CSIRO [50], and had been further processed using downscaling proce-
dures in order to match current climate data from Worldclim [39]. We implemented an
ArcInfo AML script (freely available at http://www.worldclim.org/mkBCvars.amL)
to reformat and substantively convert these future temperature and precipitation data
into the same bioclimatic variables that had been used as inputs for current-conditions
modeling.
For each model we tested for two different time horizons, 2020 and 2050, and two
different emissions scenarios (A2 and B2). The A2 scenario assumes that population
growth does not slow down and reaches 15 billion by 2100 [51], with an associated
increase in emissions and implications for climate change. The B2 scenario assumes a
slower population growth (10.4 billion by 2100) and that precautionary environmental
practices are implemented [51], yielding more conservative predictions of anthropo-
genic emissions. To simulate plague response to climate change, we used the fi nal
models that had been developed based on the rodent/ground squirrel data, and ran
them with the future climate data.
Four models were selected as the fi nal candidate models predicting plague distri-
bution based on climate variables (Table 3). In all four cases, models based only on
California ground squirrel specimens had higher AUC values than their counterpart
models that used all rodent samples as case points. Biologically meaningful variables
used in these models included two temperature variables (Maximum Temperature of