2 Methods and Materials
2.1 Modeling the resource-driven contacts and comparing with
individual-level resource selection
To test the hypothesis that the landscape features driving contacts are
the same as those that drive individual-level space use, we apply a
resource selection functions (RSF) framework to contact locations of
animal pairs (hereafter:
contact-RSF model) and compared it with the habitat selection RSF of the
individual animals involved in a contact pair (individual-RSF model).
2.1.1 individual-RSF model for contact pairs
Since the habitat selection of contact pairs are considered as the
reference to compare with to test our hypothesis, we first developed an
individual-level RSF model for all contact pairs and aggregated them to
estimate the spatial distributions of their habitat selection. Before
the development of individual-RSF models, we subsampled movement data
for each contact pair so as to only include the time period when both
animals were tracked, as an animal may change the area used over time.
We then combined the subsampled movement data for individuals involved
in each contact pair. This allowed time matched comparison with the
contact-RSF model predictions aggregated by all contact pairs (Figure
S1).
We applied the used-available framework, as described in Manly et al.
(2007), to develop individual-RSF models for each individual involved in
each contact pair. Specifically, the available area is defined as the
95% home range for each individual in the contact pair for the period
when movement data collection overlapped with another in the pair (i.e.,
individual availability at pair level, e.g., HR1-2, HR1-3 in Figure 1).
We defined used points as the interpolated GPS fixes and generated 30
random (available) points per used point within the individual’s home
range (Northrup et al. 2013).
For the development of RSFs, we conducted logistic regressions (Manly et
al. 2007), and implemented a model selection procedure to evaluate
candidate models using the cumulative log-likelihoods for all animals to
calculate Akaike information criteria (AIC) and select the most
parsimonious model (Burnham and Anderson 2004; Bastille‐Rousseau and
Wittemyer 2019). We then calculated the population average and
confidence intervals of coefficients by weighting both the number of
times that each individual is detected in a unique contact pair and
their sample size following Murtaugh (2007). We assessed the predictive
power of the top-selected model by performing a 5-fold cross validation
to calculate the Spearman rank correlation coefficient
(rs) for each individual. We withheld 20% of the data,
assessed the model fit, and repeated the process five times for each
individual (Boyce et al. 2002). Development of the individual RSF models
was conducted using “IndRSA” R-package (Bastille‐Rousseau and
Wittemyer 2019).
2.1.2 pair-level contact-RSF model
To estimate the RSF for contact, we adopted a similar use-available
framework but define the location of contact for each pair of
individuals as the used points. We define the available area for contact
as the home range overlaps between a pair during an overlapping tracking
period. The available points in contact-RSF model are defined as the GPS
fixes that were not contact locations within the home range overlaps.
For pairs that have numerous available points, a subsample of available
points was randomly selected to limit the used: available points ratio
to 1:30. We used logistic regression for each contact pair to compare
the landscape features at the location of contact with those at
locations that animals used without contact.
To test the hypothesis that contact RSFs are similar to individual-level
RSFs, we assume the individual-RSF model is the null hypothesis and
reference to be compared to. Thus, we fit the logistic model for
contact-RSF using the same model structure as that of the top-selected
individual-RSF model. Similarly, we follow Murtaugh (2007) to aggregate
the pair-wise contact-RSF models to a weighted population average to
predict the spatial distribution of the occurrence of the
resource-driven contact for the animal population on the landscape.
Finally, we compare the weighted population-level contact-RSF prediction
and the weighted population-level individual RSF to examine whether
predictions of resource-based contact are equivalent to individual-RSF
patterns.
2.2 Case studies
We implemented the framework of modeling the resource-driven contacts
and comparing with individual-level resource selection in two wild pig
populations in two different ecosystems. Wild pigs, are a socially
structured species that maintain matrilineal, multigenerational social
groups of female adults with their offspring (Podgorski et al. 2014).
These groups spend most of their time moving together as a unit. Male
adults often move alone but join female groups for short periods. Based
on these behaviors, we examined contact between individuals from
different family groups or between adult males and females because these
events are distinct and rarer compared to the many within group contacts
and potentially independent of individual-level resource selection.
2.2.1 GPS Data
Our case study includes two sites, one on the Archbold’s Biological
Station - Buck Island Ranch (ABIR) in Florida and one on a private
ranchland in north-central Texas. ABIR is a 42.3 km2commercial beef cow-calf operation managed at commercial production
levels with an average standing inventory of ~3,000 head
of cattle. In FL, we deployed GPS collars (Catlog GPS device and Lotek
LMRT3 VHF Collars, Lotek ©, WA, US) on 17 adult wild pigs (12 females
and 5 males) from Dec 13, 2019 – July 13, 2020. During the capture, we
intended to cover most social groups of the wild pigs across the focal
pastures (given pre-collaring camera survey) and avoid deploying
multiple collars in the same social group. Such study design aimed to
measure between-group contacts to understand potential disease
transmission in the population (Yang et al. 2021a). Collars were
programmed to record GPS fixes every 10 minutes with locational errors
of 6 – 10 meters on average. In coordination with these collar
deployments, anthropogenic cattle feed and water troughs within the
study pastures were mapped and time available recorded.
The Texas site is ~52 km2 and located
within Southwest Plateau and Plains Dry Steppe and Shrub ecoregion of
North America, with vegetation communities dominated by a mosaic of
wheat croplands, grasslands, mesquite, and oak woodlands (Bailey 1998).
We deployed GPS satellite-transmitting collars (VERTEX PLUS-2 Collar,
VECTRONIC Aerospace GmbH, Berlin, Germany) to 36 adult wild pigs (22
females and 14 males) during Jan 2018. We programmed the GPS collars to
record locations every 15 min from Jan 28 – Feb 24, with locational
errors of 5–10 m. The study was initially designed to estimate the
efficacy of toxic baiting on controlling wild pigs, therefore we
deployed bait sites targeting the collared animals. Starting on Feb 13,
the baiting was commenced with whole-kernel corn at a maximum baiting
density of 1 bait site per 0.75 × 0.75 km2. This grid
size was selected to expose 90 – 100% of wild pigs to bait within the
study area. Given the behavioral and movement changes of wild pigs after
exposed to toxicants, we only used data collected during the period
before deploying toxicants. See further details about study site and
design in Texas in Snow and VerCauteren 2019.
2.2.2 Environmental variables
The environmental factors that we tested related to contact occurrence
in two wild pig populations are presented in Table 1. For both study
sites, the wetland variable is a binary layer, with freshwater
emergent wetland and woody wetland classified as 1, and all other land
cover types classified as 0. Similarly, the water variable is a
binary layer, with freshwater pond, riverine, lake, and water trough
classified as 1, and all other land cover types classified as 0.
Vegetation greenness was measured using the daily Normalized Difference
Vegetation Index (NDVI ) accessed from NASA’s MODIS MOD09GA
product. Tree canopy cover and daily meteorological measurements were
accessed from the U.S. Forest Service tree canopy cover product and
GridMET, respectively. In the FL site, we also included binary variables
for road, ditch, fence , and food , indicating whether a
grid cell includes a road, ditch, fence, or cattle supplement.
Similarly, in TX we included binary variables for road andtrail (e.g., 2-track road) , and the food layer represents
the availability of pig baits. The variable cattle in the FL site
measures the daily cattle density on each pasture. All environmental
layers were calculated or resampled to 30 * 30-meter grids.
2.2.3 Extract resource selection of contacts based on a continuous-time
movement model of GPS data
We converted the movement data from both study areas to continuous-time
movement trajectories discretized to 5-min intervals using the “ctmm”
R-package (Calabrese et al. 2016). The 5-min time frame was chosen based
on evidence that wild pigs have small home ranges and low rates of daily
movement (Kay et al. 2017). We followed Yang et al. (2023) to define
direct contact as the colocation of two individuals at the same time
with a spatial buffer of 10 meters to consider the GPS locational
errors.
Because wild pigs in the same group do not move independently, we
expected that contact locations among them would track their movement
patterns and thus not test our hypothesis. Since only adult wild pigs
were studied, we assumed that all female-male and male-male contacts to
be between-group contacts which may be driven by reproductive processes
and landscape features. For female-female contacts, we examined the
weekly home range core area overlap using kernel density home range
estimator to exclude pairs that were in the same family group. We
assumed that pairs with core area overlaps less than 0.5 over the
tracking period were between-group pairs, while pairs with core area
overlaps greater than 0.5 were considered within-group pairs. For pairs
with core area overlaps over 0.5 for part of the subsequential tracking
period (e.g., more than 12 weeks, a season), we assumed that the pairs
were temporarily in the same social group. We included all pairs or
timeframes of pairs (female-female, female-male, and male-male) that
were considered separate groups at each site in the following analyses.
2.2.4 Application of contact-RSF modeling framework and comparison with
individual-RSF
To ensure reliable statistical inference, we only included the wild pig
pairs with more than 10 direct contacts over the time period that
movement data from each individual overlapped. Following the modeling
framework, we first fit the individual-RSF models for each individual in
the contact pairs. The available areas were defined as the 95% home
range estimated by the kernel density estimator. The used and available
points were defined as CTMM interpolated points and randomly generated
points at a ratio of 30 per used points, respectively., We screened
environmental variables for multicollinearity (Pearson’s correlation
coefficient |r| ≥ 0.6) and standardized the continuous
variables (i.e., NDVI, tmax, tmin, vp, rhum, prcp, and tree canopy)
using the scale function in R (R Core Team 2023), before the development
of logistic regression for individual-RSF models. For contact-RSF
models, we defined contact locations as used points and available points
as the CTMM interpolated movement locations that were not the contact
locations within the home range overlaps. We then followed the modeling
framework described above to implement analyses for modeling, scaling
up, and validating both individual- and contact-RSFs in the empirical
systems.