Assessing changes in clusters of wildlife road mortalities after the construction of wildlife mitigation structures

Abstract Collisions with vehicles can be a major threat to wildlife populations, so wildlife mitigation structures, including exclusionary fencing and wildlife crossings, are often constructed. To assess mitigation structure effectiveness, it is useful to compare wildlife road mortalities (WRMs) before, during, and after mitigation structure construction; however, differences in survey methodologies may make comparisons of counts impractical. Location‐based cluster analyses provide a means to assess how WRM spatial patterns have changed over time. We collected WRM data between 2015 and 2019 on State Highway 100 in Texas, USA. Five wildlife crossings and exclusionary fencing were installed in this area between September 2016 and May 2018 for the endangered ocelot (Leopardus pardalis) and other similarly sized mammals. Roads intersecting State Highway 100 were mitigated by gates, wildlife guards, and wing walls. However, these structures may have provided wildlife access to the highway. We combined local hot spot analysis and time series analysis to assess how WRM cluster intensity changed after mitigation structure construction at fine spatial and temporal scales and generalized linear regression to assess how gaps in fencing and land cover were related to WRM cluster intensity in the before, during, and after construction periods. Overall, WRMs/survey day decreased after mitigation structure construction and most hot spots occurred where there were more fence gaps, and, while cluster intensity increased in a few locations, these were not at fence gaps. Cluster intensity of WRMs increased when nearer to fence gaps in naturally vegetated areas, especially forested areas, and decreased nearer to fence gaps in areas with less natural vegetation. We recommend that if fence gaps are necessary in forested areas, less permeable mitigation structures, such as gates, should be used. Local hot spot analysis, coupled with time series and regression techniques, can effectively assess how WRM clustering changes over time.


| INTRODUC TI ON
The distribution of wildlife road mortalities (WRMs) is often affected by species, road, and landscape attributes (Ascensão et al., 2017;Clevenger et al., 2001), and characterizing spatial patterns of WRMs is often beneficial for developing and assessing mitigation measures (Andis et al., 2017). However, counts of WRMs are not always a good measure of clustering (Teixeira et al., 2017), and clustering and counts of WRMs are often associated with different environmental factors (Bíl et al., 2019;Snow et al., 2014). Additionally, longterm WRM datasets may be affected by variation in detection rates through time due to changes in survey methodology and researcher experience, so examining counts may bias conclusions about how WRM patterns have changed over time. Finally, mitigation structures could cause there to be fewer WRMs along a highway, but because they become more concentrated around gaps in fencing (van der Ree et al., 2015), researchers may draw different conclusions about the effectiveness of mitigation structures depending on whether they examine counts or clustering of WRMs.
Different methods exist to examine how WRM spatial clustering changes through time, including kernel density estimation and time series analyses of clustering algorithms such as hot spot analysis and Moran's I analysis. Kernel density estimation creates a probability surface of a road where hot spots can be identified based on a defined isopleth threshold, while hot spot analysis and Moran's I use location-based nearest neighbor clustering algorithms to identify where hot spots occur (Anselin, 1995;Getis & Ord, 1992;Snow et al., 2014). While both kernel density estimation and a locationbased approach can be used to identify patterns through time, kernel density estimation is more strongly affected by small sample sizes, such as WRM datasets, potentially causing isolated WRMs to have a strong influence on the probability surface generation causing an overestimation of hot spot locations. While a location-based approach is also affected by small sample sizes, it is less affected by isolated WRMs. Using a location-based approach also allows one to explicitly examine how the intensity and distribution of WRM clusters changes through time using time series analysis such as the Mann-Kendall test (Getis & Ord, 1992;Harris et al., 2017).
Local hot spot analysis measures whether block values are high relative to surrounding blocks (Getis & Ord, 1992), while local Moran's I analysis measures whether block values are high relative to all other blocks (Anselin, 1995). Both measures use a weighting factor to determine how much influence neighboring blocks have on a particular block. When studying changes in WRMs, researchers are typically interested in how WRMs in particular locations change over time, and local hot spot analysis is better than both Moran's I and kernel density estimation at identifying how this pattern changes (Getis & Ord, 1992).
Using local hot spot analysis to identify WRM clusters also allows one to examine how the intensity of a cluster is affected by environmental factors and how this relationship changes through time. Factors that influence the distribution of WRM clusters include variation in land cover and land use (Ascensão et al., 2017;Caceres, 2011), highway characteristics Grilo et al., 2015), and the presence of wildlife mitigation structures, especially exclusionary fencing (Cserkész et al., 2013). Fencing restricts access to roadways to narrow gaps at road intersections and private drives which can decrease the overall number of WRMs on the highway (Forman et al., 2003); however, it could increase the intensity of WRM clusters near these locations by funneling animals toward gaps in the fences (Cserkész et al., 2013). The potential for funneling is often a concern in wildlife mitigation structure construction (Huijser et al., 2016), so gaps are often mitigated by various structures including gates, wildlife guards, and wing walls. These structures are not 100% effective at keeping wildlife off roads, and WRMs may still result (Allen et al., 2013;van der Ree et al., 2015). Therefore, examining how fence gaps influence the intensity of WRM clusters may be important in determining how wildlife mitigation structures affect WRMs.
We used local hot spot analysis to assess how WRM clusters changed through time with the construction of wildlife mitigation structures on State Highway 100 (SH100) in Cameron County, Texas, USA. We examined how the intensity of WRM clusters changed with mitigation structure construction at a fine temporal scale and how factors influencing WRM cluster intensity changed from before construction to after construction of wildlife mitigation structures. We expected to see fewer WRM clusters in the after construction period than the before-and during construction periods coupled with increased cluster intensity due to limited access to the road area. We also expected that the intensity of WRM clusters would decrease with increased distance to wildlife mitigation structures in the after construction period only.

| Study area
The study area was a 15-km section of SH100 in Cameron County, Texas, USA, between the towns of Laguna Vista and Los Fresnos Mitigation structures built included 11.9 km of exclusionary fencing along the entire mitigation area, five wildlife underpasses, 18 wildlife guards, three wing walls, and 16 gates. The mitigation structures were designed to prevent ocelots (Leopardus pardalis), bobcats (Lynx rufus), and other medium to large mammals from accessing the road, while still providing connectivity across the highway (Environmental Affairs Division, 2015). The fencing material was 5.1 cm wide black plastic-coated chain-link, 1.8 m tall, and was buried 30.5 cm into the ground along most of the fence line. In areas where burial was not possible, the fence was secured to the ground away from the highway. Cameron County is characterized by hot summers with an average daily temperature in August of 29.6℃ and mild winters with an average daily temperature in January of 16.2℃ (National Weather Service, 2020). The area receives an average of 69.7 cm of rain per year, and most rainfall occurs during occasional tropical storms between June and October. The primary vegetation types in the study area were cordgrass prairie, salt marsh, and thornscrub forest (Elliott et al., 2014).

| Wildlife road mortality surveys
Wildlife road mortality surveys were conducted by vehicle before, during, and after the construction of the mitigation structures on SH100. The survey transects included the full mitigation area as well as a 1.5 km buffer on both sides. Survey frequency, speed, and marking differed in the three construction periods (Table 1) (Collinson et al., 2014;Santos et al., 2011). However, because SH100 is a high-speed, high traffic road, it would have been unsafe for the researchers to drive any slower.
Only those species for which fencing provided a barrier to movement were used in analyses to assess how fencing changed WRM patterns. These included all mammals larger than rodents as well as turtles and tortoises (

| Land cover classification
To identify land cover types around SH100, we created a classi- bare, paved road, dirt road, water, and bahia. Classification was confirmed by visual inspection of the map. These classes were simplified to three major land cover types: forested (trees, bahia), shrub (shrubs, cactus), and open (open, bare, paved road, dirt road). The water class was excluded because water was identified using a different method, described below.
We identified permanent sources of fresh and saltwater using the National Wetlands Inventory (U.S. Fish & Wildlife Service, 2018).
Saltwater areas were identified as all locations that had the saltwater, tidal regime subgroup and included the subtidal, irregularly exposed, regularly flooded, and irregularly flooded water regimes.
Permanent freshwater areas were those that were classified into the nontidal regime subgroup and had the permanently flooded, intermittently exposed, or semipermanently flooded water regimes.
In addition to these sources of permanent freshwater, the drainage canals around SH100 were included because they had flowing water throughout most of the year. We extracted linear water fea- We combined the water, agriculture, and developed layers with the classified vegetation map using the reclassify and raster

| Changes in wildlife road mortalities through time
We assessed changes in WRM cluster intensity through time by coupling local hot spot analysis and a time series analysis. We divided our WRM location dataset into space time blocks that were 100-m × 4 months. We used 100-m space blocks because fence gaps are highly localized features, and this block size best represented the spatial relationship between blocks and gaps. We tried smaller and larger block sizes, but the 100-m block performed the best. We used 4-month time blocks (June-September, October-January, February-May) because this block size fits both the construction periods and seasonal rainfall patterns and movement of wildlife in South Texas.
To assess changes in clustering through time, we ran a local hot spot analysis using ArcMap 10.6 on each time block (4-month pe-

| Impact of fence gaps on wildlife road mortality cluster intensity
We also tested how the presence of gaps in the fence influenced the intensity of WRM clusters. We were interested in comparing cluster intensity in the three construction periods, instead of time blocks, so we ran local hot spot analysis on each of the three construction periods (before, during, and after) to create a comparable measure of WRM clustering among the three periods. We measured the distance from each space block to three different types of fence gaps (gates, wildlife guards, and wing walls) and recorded whether there was continuous fencing within each space block. In space blocks at the edges of the mitigation area, fencing was determined by whether or not the majority of the block had fencing. Distances to each fence gap type and fence presence were highly correlated to each other (r = 0.72-0.88) so we performed a principal components analysis (PCA) using the "prcomp" function in R to develop an index representing distance to fence gaps. The first principal components (PC) axis, hereafter fence gap index, explained 85% of the variation in distance to fence gaps, so it was the only axis used in the regression. Positive values of the first axis represented locations that were closer to gaps and unfenced areas ( Figure 2).
To assess how local land cover was related to clustering intensity, we created 100-m buffers around each space block. We performed this analysis at the local scale because WRM risk has been shown to be associated with the presence of specific habitat features such as freshwater sources, access to roads, or movement corridors (Červinka et al., 2015;Grilo et al., 2016), and we expected that this distance would be small enough to assess these local scale effects.
Additionally, at larger spatial scales the influence of fence gaps is overshadowed by larger scale landscape effects such as habitat type.
We calculated the proportion of each cover type within the buffer using an iterative version of the tabulate area tool in ArcMap 10.6.
We conducted a generalized linear regression with a Gaussian error distribution to assess how cluster intensity was related to fence gap index, the proportions of forested, shrub, open, agriculture, developed, and freshwater, and the interactions between the fence gap index and land cover variables. No saltwater was located within any of the buffers. We did not include distance to wildlife crossing in the final models because the variable was never significant and did not improve model fit. While road characteristics such as traffic volume, road size and type, and speed limit may also impact WRMs (Clevenger et al., 2001;Grilo et al., 2015), the variations in these characteristics were minor along SH100, so they were excluded.
We used the MuMIn package in R to perform AICc model selection and model averaging to model the relationship between cluster intensity and fence gap index and land cover (Barton, 2013;Burnham & Anderson, 2002). The relevant main effects were always included in models containing interactions. Models that were within two ΔAICc values of the best model were used for averaging. We calculated the McFadden pseudo-R 2 values for individual models included in the averaged model using the pscl package in R (Jackman, 2012).

| Change in wildlife road mortalities through time
In total, we surveyed 3,360 km of road and identified 391 target species WRMs (13-44 per time block) and 376 nontarget WRMs (10-60) (  (Figure 3). There was greater variation in WRMs/survey day in the before construction period when only two surveys were conducted per month than in either of the other periods when more surveys were conducted ( Figure 3).
Visually, the majority of WRMs occurred on the western side of the survey transect, an area with most of the wildlife crossings and fence gaps ( Figure 4).
We identified hot spots in all time blocks, although the majority of these were not significant after applying the FDR correction Additionally, the Mann-Kendall trend test revealed several increasing and decreasing trends in WRM hot spot intensity; however, none of these were statistically significant after applying the FDR correction ( Figure 6).

| Impacts of fence gaps on mortality trends
The PCA of distance to fence gaps indicated that approximately 85% of the variation among fence gap types was explained along the first PC axis (PC1), 8.0% on the second axis, 4.1% on the third, and 2.5% on the fourth (Figure 2). Distance to gates, wildlife guards, and wing walls were negatively correlated with PC1 (r = −0.96, −0.93, −0.93 respectively), and fencing was positively correlated with PC1 (r = 0.88; Figure 2).
Seven main effects and six interactions were included in the global model, giving a total of 793 possible models. The number of models included in the averaged model ranged from 3 (before construction) to 19 (during construction; Table 3

| D ISCUSS I ON
Overall, we found that at a fine temporal scale, the intensity of WRM clusters increased or decreased in few locations after construction of the mitigation structures on SH100, but none of these changes Statistically significant hot spots are those that were significant after applying the false discovery rate correction, while nonsignificant hot spots were those that were only significant without the correction. The survey transect blocks represent road segments and increase from west to east. To better relate this to the study area map, the approximate locations of wildlife crossings and fence ends are also indicated by vertical lines and the construction periods are indicated by horizontal lines were significant after applying the FDR correction. Interestingly, the fence gap index showed a negative relationship with intensity in all three construction periods, although this effect was only significant in the before and during construction periods. Perhaps unsurprisingly, as forest proportion increased, WRM cluster intensity increased when closer to fence gaps in the during-and after construction periods. Generally, our two analyses agreed, indicating that, as of 1.5 years after construction of mitigation structures on SH100, WRM intensity has locally increased. However, although these locations were near fence gaps, they were not directly at fence gap locations. While intensity did increase in some locations, only one of these locations was associated with a statistically significant hot spot, indicating that WRMs are decreasing overall along SH100.
Thus, with more time, we may expect to see additional decreasing trends in WRM clustering across most of the study area. Previous studies have shown that it may take years for wildlife to regularly use wildlife crossings (Clevenger & Waltho, 2005). Many of the wildlife crossings on SH100 occur near fence gaps, so as wildlife become familiar with wildlife crossings, we may see fewer animals attempting to cross on the road surface and fewer WRMs as a result.
We can draw several conclusions from these analyses. First, there appeared to be a geographical disparity between WRM clusters along the length of the transect. Second, when access to the highway is limited, habitat strongly affected how WRMs were related to distance to fence gaps. Finally, conducting local hot spot analysis at fine spatial and temporal scales can provide a unique picture of how WRM patterns change over time.

| Wildlife road mortality distribution along SH100
As and salt marsh (Elliott et al., 2014) which tended to have fewer species and fewer individuals than forested habitats in Cameron County (Yamashita, 2020). The western side of the transect was primarily agricultural and forested habitat, and both land cover types have been shown to be associated with greater WRM rates (Ascensão et al., 2017;Puglisi et al., 1974;Smith-Patten & Patten, 2008).
Therefore, while we could not measure this, it is possible that WRM

F I G U R E 6
Trends in the intensity of wildlife road mortality (WRM) clusters along State Highway 100, Cameron County, Texas, from the Mann-Kendall trend test. Decreasing trends indicate that the intensity of WRM clusters decreased over time while increasing trends indicate that intensity of WRM clusters increased over time. No trends were statistically significant (at α = 0.05) after the false discovery rate (FDR) correction was applied rates may be similar along the length of the survey transect. It is also likely that wildlife living in disturbed habitats (such as those near agricultural lands) may be more willing to use road rights of way than individuals living in more natural habitats, thus increasing their risk of vehicle caused mortality (Forman et al., 2003).
In 2018  Another contributing factor may be that there were more fence gaps on the western side of the survey transect than the eastern side. While this does not explain the high numbers of WRMs before or during construction, it may have contributed to the lack of decrease in WRMs seen after construction. The western side of the transect had 12 of 18 wildlife guards, 10 of 16 gates, and two of three wing walls offering multiple places for wildlife to access the road. The effects of different types of fence gaps were not examined in the present study, so it is possible that WRM cluster intensity may be higher around more permeable gaps such as wing walls or wildlife guards. Therefore, these mitigated fence gaps may not be as effective as gates at reducing wildlife access to the road.

| Fence gaps and wildlife road mortality
Interestingly, our regression models indicated that WRM intensity increased with increasing distance to fence gaps across all three construction periods. However, we found statistically significant interactions with different habitat variables in all three construction periods which may have affected the identified relationship.
Generally, WRM cluster intensity increased when nearer to (future) fence gaps when in areas with a high proportion of natural habitat (forested, shrubs, open), while intensity decreased when in areas of low natural habitat. Forested habitat had the strongest effect on the relationship between fence gaps, especially in the during-and after construction periods. Intensity of WRM clusters increased with increasing distance from fence gaps when forest proportion was low, and intensity increased with decreasing distance from fence gaps when forest proportion was high.
While we did document increases in WRM cluster intensity over time in some locations, we did not see evidence that fencing funneled animals onto SH100. Our documented locations of increased WRM cluster intensity did not occur at fence gaps; rather, they occurred 200-300 m from a gap. It is possible either that animals moved from fence gaps toward those locations while in the right of way before getting hit or that animals were climbing over or digging underneath the fence to get to the road at those locations. Cserkész et al. (2013) examined how WRM counts on a fenced highway were affected by distance to highway interchanges and demonstrated fencing funneled animals toward fence gaps. Fence gaps along SH100 occurred at high rates ( Our study indicated that there was limited change in WRM clustering with construction and that fence gaps were important, but not always significant, predictors of intensity in all three construction periods, thus indicating that fence gaps, especially in unforested areas, may be located in places previously used as wildlife travel corridors. In the after construction period, fence gaps TA B L E 3 Summary of the averaged regression models for the effect of land cover and fence gaps on the intensity of wildlife road mortality clustering along State Highway 100, Cameron County, Texas   (Červinka et al., 2015). At broader scales, the influence of access points to the highway may become masked by landscape-level effects such as land cover and the presence of freshwater (Yamashita, 2020).
Finally, this study was conducted less than 2 years after the completion of mitigation structure construction, and it has been shown that wildlife may take several years to adjust to the presence of wildlife crossings (Clevenger, 2005;Clevenger & Waltho, 2005). It is possible that animals along SH100 were still in the "learning" phase and WRMs, especially around wildlife crossings, may begin to decrease as time passes. There is some visual evidence of this already with only three WRMs occurring within 200 m of four of the five wildlife crossings in the final two time blocks (8 months; Figure 4),

| Using wildlife road mortality clusters to examine road mortality patterns
Using a location-based clustering method to examine patterns of WRMs allowed us to determine the statistical significance of visually identified WRM hot spots. Knowing whether or not a cluster is significant can have important management implications because wildlife crossings can be expensive when they are built as a standalone project (Huijser et al., 2009). Solely using counts of WRMs may miss important clustering of fewer WRMs which may benefit more from a wildlife crossing (Teixeira et al., 2017). visualizations of WRM data that can help display patterns hidden at larger scales. Generally, WRMs need to be examined at broad spatial and temporal scales due to sample size limitations. These analyses can miss important patterns occurring at finer scales (Levin, 1992 Grubesic et al., 2014). For analysis purposes, medium to large mammal WRM rates tend to be fairly low (Ascensão et al., 2017).
Therefore, the power of local hot spot analysis may be too low to detect significant WRM hot spots in medium to large mammals without very high WRMs or access to long-term datasets. However, spatial indices of WRM rates, such as intensity of clusters, are essential to comparing long-term WRM datasets where data collection and researcher experience may change over time. These sources of bias are likely to be consistent along an entire survey transect (Collinson et al., 2014), so they would not affect clustering patterns derived from WRM counts.
We assumed that WRM clustering was not affected by survey frequency and vehicle speed, but both sources of bias likely affected overall detections of WRMs and may have contributed to the reduced number of WRMs detected in the after construction period.
By examining cluster intensity instead of WRM numbers, we focused on the relative distribution of WRMs through time and having fewer WRMs overall is unlikely to have a significant impact on hot spot intensity. It is possible that WRMs may be easier to detect along some parts of the survey transect when driving slower or that some areas may have lower carcass persistence times. Therefore, we believe that, because the locations of WRMs changed little through time (Figure 4), survey frequency and vehicle speed likely did not affect detection probability along different sections of the transect although more research is needed into how highway properties interact with vehicle speed and survey frequency to influence WRM detection probability.
The Mann-Kendall test requires a minimum of 10 time blocks to run (Harris et al., 2017;Hipel & Mcleod, 2005 et al., 2014). Therefore, an assessment of how sample size affects the power of local hot spot analysis will be required before this method can be applied more broadly. indicating that the wildlife crossings were placed appropriately and animals may be learning that wildlife crossings provide a safer passageway across roads than the road surface. Additionally, fence gaps in forested areas may facilitate increased WRM cluster intensity, so reducing the number of gaps and mitigating necessary gaps with more effective structures, such as gates, will likely help reduce WRM rates. Therefore, local hot spot analysis, coupled with time series and regression techniques, can provide useful insights into how changes in the roadway impact wildlife use of the road area.

ACK N OWLED G M ENTS
The authors thank the Texas Department of Transportation for funding this work. T. Miles Hopkins, Anna Rivera Roy, Tiffany Cogan, Ivonne Cano, Jenny Baez, and Victoria Rodriguez were instrumental in helping to collect the road mortality data. Two anonymous reviewers provided useful feedback on previous versions of this manuscript.

CO N FLI C T O F I NTE R E S T
None declared.