Drought and coyotes mediate mesopredator response to human disturbance

Abstract

Mesopredators in western North America are facing major changes to their ecosystems, including drought and the expansion of human disturbance. To balance resource needs and risk-taking on the landscape, mesopredators are likely to shift their habitat use and interspecies interactions. As part of a large-scale study to help evaluate responses of terrestrial wildlife to severe drought, the California Department of Fish and Wildlife surveyed mesopredator presence across 585 sites in the Mojave Desert (MD) and Central Valley (CV) ecoregions of California. This study spanned a drought year (2016) and a post-drought year (2017), providing the opportunity to investigate how drought and interspecific interactions may mediate spatial patterns of mesopredator occurrence across a continuum of human disturbance levels. We used single-season, single-species, and conditional two-species occupancy models to elucidate these relationships in both ecoregions. We examined the occupancy and detection of coyotes (Canis latrans) and smaller mesopredators, including bobcats (Lynx rufus) in both ecoregions, raccoons (Procyon lotor) in the CV, and desert kit foxes in the MD (Vulpes macrotis arsipus). The presence of coyotes influenced the detection probability of all other mesopredator species, and the influence of drought varied by species and ecoregion. Detection of mesopredators in the CV was typically higher in 2016, especially at low disturbance sites, suggesting species may have become more active during the drought to meet resource needs. However, detection of mesopredators in the MD tended to increase after the drought, suggesting a response to an increase in resources (e.g., prey). Coyotes in the MD became more detectable in high human disturbance in 2016 and less detectable in 2017, possibly increasing activity during the drought in human-disturbed areas to obtain anthropogenic resources. Drought not only affects individual species and their relationships to human disturbance, but it can also influence their interspecies interactions and use of different landscape features.

INTRODUCTION

The risk of drought is increasing in western North America as climate change produces higher temperatures and evaporation rates, which result in increased drying (Diffenbaugh et al., 2015; Trenberth, 2011). The outcome of these changes to the hydrologic cycle is droughts that set in more quickly, are more intense, and last longer (Trenberth et al., 2014). Record high temperatures paired with reduced precipitation led to the most severe drought in California state history from 2012 to 2016 (Griffin & Anchukaitis, 2014; Lund et al., 2018). Drought effects moved up the food web over time and were strongest at the bottom (plants) and top (predators; Prugh et al., 2018). The decline in mammalian predator abundance was likely due to the decline of a few keystone prey species (Prugh et al., 2018). Reduced access to resources such as prey and water during periods of drought could lead to increased competition among predator species. However, it is unclear how disturbances such as drought affect competition within a guild; competition could be greater under drought-induced resource shortage, but alternatively, such resource shortages could result in competitive release when more dominant species decline in numbers (Chesson & Huntly, 1997; McCluney et al., 2012; Prugh et al., 2018; Tilman, 1982).

In much of the United States, coyotes (Canis latrans) now function as top predators (Prugh et al., 2009) and compete with a variety of smaller mesopredator species for limiting resources. As an extreme form of interference competition, coyotes will sometimes kill smaller co-occurring mesopredators (Crooks & Soulé, 1999; Palomares & Caro, 1999). As a result, subordinate mesopredators may avoid habitats occupied by coyotes (Linnell & Strand, 2000; Polis et al., 1989). This behavior is termed “spatial partitioning” and plays a central role in structuring predator communities (Gompper et al., 2016). In some areas, smaller mesopredators use more urban or disturbed areas as potential refugia from coyotes (Moll et al., 2018). While coyotes can exploit urban areas and their associated resources, they appear to have a tolerance threshold for urbanization in which they can utilize human-dominated landscapes but require natural areas to persist (Crooks & Soulé, 1999; Ordeñana et al., 2010). However, foraging behaviors are closely linked to water availability in dryland ecosystems (McCluney et al., 2012), and coyotes will increase their use of anthropogenic resources in response to low precipitation levels (Cypher et al., 2018). Therefore, increased coyote activity in anthropogenic landscapes during times of drought may compromise these areas as possible refugia for subordinate mesopredators.

Using camera traps, we collected mesopredator presence data in 2016 and 2017 in the Central Valley (CV) and Mojave Desert (MD) of California. Our objectives were to examine spatial patterns of mesopredator detection, occurrence, and co-occurrence with the coyote across varying levels of human disturbance, and to investigate how drought may mediate these relationships. As the final year of an extreme multi-year drought (Griffin & Anchukaitis, 2014), vegetation green cover was at its lowest in 2016 (Potter, 2019), and prey species' populations were already reduced from pre-drought levels (Cypher et al., 2018; Kelly et al., 2019). At intermediate water availability, predators may increase foraging behavior and utilize prey as a source of energy and water (McCluney et al., 2012), but at times of very low water availability, such as in 2016, predators are thought to decrease hunting activity and instead seek refuge as a water conservation technique (McCluney et al., 2012). From 2016 to 2017, California went from experiencing one of the driest years on record to one of the wettest years (AghaKouchak et al., 2015; Lund et al., 2018). The intense rains in the winter of 2016–2017 (California Department of Water Resources, 2017) stimulated plant production (Potter, 2019) and likely prey population growth as winter annual vegetation is correlated with successful reproduction of rodents (Beatley, 1969). Additionally, a multispecies occupancy analysis using the camera data from this study showed that black-tailed jackrabbits (Lepus californicus), the most-detected prey species in both ecoregions, had higher detection probability in 2017 compared with 2016 (Rich et al., 2018).

We predicted that detection of mesopredators would increase from 2016 to 2017 as water availability and prey detection increased, releasing mesopredators from water conservation techniques employed at low water availability (McCluney et al., 2012). However, while the CV, one of the most productive agricultural regions in the world (Faunt et al., 2016), did experience reduced water deliveries during the drought, irrigation buffered much of the ecoregion to extreme water shortage (Faunt et al., 2016; Lund et al., 2018). Therefore, we predicted that changes in mesopredator detection from 2016 to 2017 would be less extreme in the CV than in the MD. We also predicted that coyotes in both ecoregions would increase their activity in human-disturbed areas in pursuit of anthropogenic food in 2016 (Cypher et al., 2018), resulting in higher detection. Because subordinate mesopredators, such as fox species, sometimes utilize disturbed areas as urban refugia from coyotes (Moll et al., 2018), we predicted that they might decrease their activity in these areas during the drought when coyotes increased their activity.

MATERIALS AND METHODS Study area Mojave Desert

The MD ecoregion of California covers an area of approximately 66,834 km² and is characterized by its extensive, undulating plains with short mountain ranges, playas, basins, and dunes (Bailey, 1980; McNab et al., 2007). The elevation in the MD varies greatly, from ~85 m below sea level to a high point of ~3353 m (Bailey, 1980); mountains are rocky and mostly devoid of vegetation, and the valleys are primarily characterized by creosote bush (Larrea tridentata) dominated shrub communities (Bailey, 1980; Griffith et al., 2016). High temperatures prevail for long periods of time in the summer while winters are moderate (Bailey, 1980). Rains are widespread and gentle in the winter, and summer rains are characterized by rare but powerful thunderstorms (Bailey, 1980). Annual precipitation in the MD is typically ~5–25 cm in the valleys but can be much higher in the mountains (Bailey, 1980).

Central Valley

The CV of California is a long, flat, alluvial plain, which stretches between the Sierra Nevada Mountains and the Coast Ranges, covering an area of approximately 49,176 km² (Bailey, 1980). Once dominated by native bunchgrasses, the CV is now characterized by flat, intensely farmed plains (Griffith et al., 2016). Since the 1970s, there has been a trend of converting annual cropland into perennial crops such as orchards and vineyards (Faunt et al., 2016; Sleeter, 2008). The CV has long, hot, dry summers and mild winters, during which it receives most of its rain (~15–76 cm; Bailey, 1980; Griffith et al., 2016). Elevation in the CV ranges only slightly, from sea level to ~152 m (Bailey, 1980).

Data collection

The primary objective of our sampling design was to survey many sites spread across the possible range of covariate conditions within each ecoregion to provide model-based inference at the ecoregion scale (Gregoire, 1998). Therefore, we did not revisit sites between years and used total cover of vegetative life-forms (e.g., dunes, grasslands, and riparian) within each ecoregion to develop stratified sampling goals. To select site locations, we first drew a spatially balanced random sample of hexagons (2.6 km radius) from the US Department of Agriculture Forest Inventory and Analysis Program (Brand et al., 2000) and identified the life-form at points spaced 100 m apart within each hexagon. Camera sites were identified by assigning random numbers to each point and selecting the lowest numbered points that met stratified sampling goals and land access restrictions. Typically, 1–3 sites were established 1–2 km apart within a chosen hexagon (Appendix S1: Table S1). Additional natural and man-made water features were also randomly selected to meet stratification goals (Rich et al., 2019). We used random selection whenever possible in the CV and contacted over 1200 landowners by letter or phone to request property access when a site fell within private property (Smith et al., 2021). We received around 100 positive responses from the landowners that we contacted. The success of our private landowner outreach process allowed us to include more private lands, which increased the range of covariate conditions in our modeling and led to more robust modeling of spatial variation in the ecoregions.

At each site, we mounted a Reconyx Hyperfire PC900 (Reconyx, Holmen, WI, USA) camera 1 m above the ground on a T-post, aimed at a shallow angle to a point 3–4 m away where we placed a 250 g salt lick, 100 ml of peanut butter–oat mixture, and 150 g of fishy cat food (Rich et al., 2019). Cameras were baited once, at the time of deployment and were preferentially set near signs of wildlife activity such as game trails or small mammal burrows, and away from vegetation, which might trigger the camera. We set the cameras to high trigger sensitivity and took three photos per trigger event, with no delay between triggers. We deployed cameras at each site for approximately 28 nights. After a camera trap had been deployed for at least 28 nights, cameras could be retrieved from a site and redeployed at a new site. As a result, camera trap data were collected anywhere from March to August depending on the site, as well as the year. All procedures adhered to the animal care and use policy at California Polytechnic State University, Humboldt and were approved by the Institutional Animal Care and Use Committee (IACUC; protocol number 16/17.W.08-A).

Camera images were independently reviewed by two different analysts and images were identified to species, when possible. Each species was counted only once per 24-h period (clock starting at time of deployment, not 12:00 AM). If the camera was not aimed at the bait for more than 12 h in a 24-h period (determined by the time of the camera's header photo), the camera was considered nonfunctional. When results from the two analysts differed, a third analyst also processed the data to resolve any differences. Following the review of images, we built occupancy models for coyotes and bobcats (Lynx rufus) in both the MD and CV, the desert kit fox (Vulpes macrotis arsipus) in the MD, and the raccoon (Procyon lotor) in the CV.

Occupancy models

Because sites were not resurveyed across years, we used single-season occupancy models to address our predictions. Single-season occupancy modeling relies on a sampling scheme wherein N sites are each surveyed T times to establish the presence or absence of a species at a site (MacKenzie et al., 2002). For modeling purposes, we defined a site as a 1-km buffer around a camera location, a survey as a 24-h period, and a survey period as 27 days. Repeated surveys within a survey period were used to estimate the probability of detecting a species at a site if it was present (MacKenzie et al., 2002) and therefore the probability that the site was occupied while accounting for imperfect detection. Occupancy is a measure reflecting behavior at the landscape level, whereas detection probability is a smaller scale measure of behavior, which reflects the magnitude of use as a response to landscape characteristics within the home range (Lewis, Bailey, et al., 2015). The probability of detecting a species is both a function of density (Broadley et al., 2019; Lewis, Logan, et al., 2015; Neilson et al., 2018; Royle & Nichols, 2003) and animal space use (Broadley et al., 2019), such as movement (Neilson et al., 2018) and frequency of use (Lewis, Logan, et al., 2015). Therefore, detection is a more sensitive measure of species' relationships to habitat conditions, which may be masked in occupancy estimates. Consequently, we were primarily interested in the relationship between species' detection probability and covariates.

The presence of a species at a site is reliant upon site-specific conditions, and many factors at a site may influence the behavior of a species and its associated detectability (Lewis, Bailey, et al., 2015). Therefore, occupancy modeling allows both occupancy and detection probability to vary by site-specific covariates, and detection probability by survey-specific covariates (MacKenzie et al., 2002). To address our hypotheses and variation introduced by confounding factors, we included a series of environmental and survey covariates in our occupancy models (Table 1), which were first tested for multicollinearity using Pearson correlation tests (Appendix S1) in Program R (version 3.5.2, www.r-project.org, accessed 4 February 2018). A pair of covariates with |r| ≥ 0.50 were considered correlated, and we would select one covariate to be retained for model selection. All covariates retained for analysis in an ecoregion were included in model selection for all species within that ecoregion. Continuous covariates were standardized prior to analysis.

TABLE 1 Covariates used in single-season occupancy modeling along with their definition and source, ecoregion in which they were used, and the parameter the covariate explains.

Covariate	Definition	Ecoregion	Parameter
Year	Year when site was sampled (2016 = 0, drought; 2017 = 1, post-drought). Used in interactions with other covariates.	MD, CV	p, ψ
Human disturbance	Measure (1–10) of human footprint averaged across 1-km buffer. Used in interactions with year. MD: low = 1, average = 3.4, high = 6.8; CV: low = 4, average = 8.1, high = 10. Source: Human Footprint of the West (Leu et al., 2008; U.S. Geological Society, 2016a).	MD, CV	p, ψ
Guzzler	Presence of artificial water catchment. Used in interactions with year.	MD	p
Water at site	Presence of water at site. Used in interactions with year.	CV	p
Distance to water (km)	Distance to water from site. Used in interactions with year. Source: National Hydrography Dataset (U.S. Geological Society, 2016b) and CropScape (U.S. Department of Agriculture, 2016).	MD, CV	ψ
Ordinal date	Represents date of survey. Only used first ordinal date in MD.	MD, CV	p
Bait age	Represents time since baiting and associated strength of lure.	MD	p
Elevation (km)	Elevation of site. Source: LANDFIRE (LANDFIRE, 2016)	MD	ψ
Latitude	Latitude of site.	CV	ψ
Percent tree cover	Percentage of tree cover including orchards averaged across 1-km buffer. Source: CropScape (U.S. Department of Agriculture, 2016)	CV	ψ

Note: All covariates in an ecoregion were used in modeling for all species in that ecoregion: MD species = coyote, bobcat, and kit fox; CV species = coyote, bobcat, and raccoon.

Abbreviations: CV, Central Valley; MD, Mojave Desert; p, detection; ψ, occupancy.

To address our predictions, we included year of data collection, human disturbance, and the interaction of the two as covariates for detection and occupancy in both ecoregions (Table 1). Year was a binary covariate used to represent drought (2016 = 0, drought; 2017 = 1, post-drought) and to create interaction terms. Tables 2 and 3 show average monthly precipitation levels from 2012 to 2017 and 30-year (1981–2010) normal averages for all sites in the MD and CV, respectively. Our measure of human disturbance used an index of human “footprint” based on 14 landscape factors including human habitation, roads and other linear features, agriculture, landfills, oil and gas development, and human-induced fires (Leu et al., 2008; U.S. Geological Society, 2016a). We calculated the mean human footprint within a 1-km buffer of all camera sites to represent the human disturbance likely experienced by a mesopredator at that location. Mean human footprint at our sites in both years was representative of their respective ecoregions (Appendix S1: Figure S1).

TABLE 2 Average monthly precipitation levels (4-km resolution) and 30-year (1981–2010) normal precipitation averages (in millimeters) at our 319 study sites in the Mojave Desert from the beginning of the drought (2012) until the year after the drought (2017), extracted from PRISM (Oregon State University, 2017).

Month	2012	2013	2014	2015	2016	2017	30-year average
Jan	3.59	12.55	0.16	22.64	22.98	56.90	21.91
Feb	5.42	2.55	3.04	9.87	8.29	27.08	28.48
Mar	8.33	5.59	12.67	8.64	6.53	0.85	19.13
Apr	12.59	0.48	3.64	3.96	20.05	0.31	6.58
May	0.04	3.40	1.58	2.91	4.78	1.79	3.51
Jun	0.00	0.01	0.00	1.22	1.06	0.00	1.45
Jul	19.54	10.24	8.14	22.67	5.37	3.22	8.17
Aug	29.72	22.12	20.77	4.35	10.28	8.31	11.59
Sep	7.47	8.77	9.08	5.64	3.61	9.48	7.65
Oct	8.29	1.30	0.21	24.22	8.68	0.00	7.56
Nov	0.54	18.41	0.91	2.90	3.25	0.02	10.04
Dec	18.42	2.14	25.28	3.95	37.60	0.03	18.34
Water year (previous Oct–Sep)		92.96	80.93	108.29	114.03	157.49	144.40
Partial water year (previous Oct–Mar)		47.94	37.72	67.53	68.87	134.37	105.45

Note: Partial water year shows the water year at the time sampling began for our study each year.

TABLE 3 Average monthly precipitation levels (4-km resolution) and 30-year (1981–2010) normal precipitation averages in millimeter at our 266 study sites in the Central Valley from the beginning of the drought (2012) until the year after the drought (2017), extracted from PRISM (Oregon State University, 2017).

Month	2012	2013	2014	2015	2016	2017	30-year average
Jan	60.78	22.65	5.40	2.03	127.88	192.90	80.41
Feb	18.53	8.34	77.27	47.73	20.12	145.96	75.56
Mar	91.81	23.75	57.95	5.39	110.15	47.54	61.43
Apr	58.77	14.71	28.15	25.17	28.69	47.97	27.05
May	0.34	4.73	2.39	4.14	13.01	2.85	15.31
Jun	2.21	5.04	0.02	0.97	2.42	5.52	4.29
Jul	0.03	0.06	0.05	1.28	0.00	0.00	0.48
Aug	0.00	0.08	1.38	0.21	0.00	0.02	1.40
Sep	0.07	10.60	9.17	1.20	0.01	2.57	6.78
Oct	15.48	0.62	4.52	8.21	59.08	4.75	21.54
Nov	67.42	20.46	42.06	44.98	38.28	42.76	46.02
Dec	136.70	6.44	166.60	51.23	68.15	1.54	68.64
Water year (previous Oct–Sep)		309.55	209.30	301.30	406.71	610.84	408.92
Partial water year (previous Oct–Mar)		274.32	168.14	268.33	362.57	551.92	353.60

Note: Partial water year shows the water year at the time sampling began for our study each year.

Ordinal date was included as a detection covariate in both ecoregions (Table 1) and was correlated with daily maximum temperature (Appendix S1: Table S2), which was extracted from PRISM (Oregon State University, 2017). We chose to retain ordinal date as it likely accounted for more variability between surveys than maximum temperature. In the MD, we only used the first ordinal date and used bait “age” as our only survey-specific covariate. Bait age represented day of survey and was used to reflect the declining strength of our lure over time. Water at site was also included as a detection covariate in both ecoregions (Table 1). Pictures taken at the time of camera deployment were used to determine whether water was present at a site in the CV, while the presence of artificial water catchments termed “guzzlers” was used in the MD. We also included the interaction of year and water at site as a detection covariate in both ecoregions (Table 1).

While detection covariates were the same in both ecoregions, occupancy covariates varied between the two (Table 1; Appendix S1). An initial analysis of National Landcover Database (NLCD; Dewitz, 2019) cover types in the two ecoregions revealed that 80% of the MD is classified as shrub/scrub, with an additional 13% as barren (Appendix S1: Table S3). Therefore, rather than using landcover type as an occupancy covariate in the MD, we tested collinearity of a 1-km average of NLCD percent shrubland cover (U.S. Geological Society, 2016b; Xian et al., 2015), a 1-km average of Normalized Difference Vegetation Index (NDVI), an indicator of the amount of green vegetation in an area (eMODIS, 2018), elevation (LANDFIRE, 2016), and mean temperature (Oregon State University, 2017). All four covariates were correlated with one another (Appendix S1: Table S4), but elevation had a consistently higher correlation with the other covariates and was retained as the only covariate of the set (Table 1).

We included distance to water and the interaction of year and distance to water as occupancy covariates in both the MD and CV (Table 1). We used the National Hydrography Dataset (NHD; U.S. Geological Society, 2016c) to identify locations of perennial sources of water in both ecoregions, which included perennial rivers, streams, lakes, and springs/seeps in the MD and all perennial water bodies, which held drinkable water, such as decorative pools, in the CV. In the MD, we also included the location of guzzlers, and in the CV, we included the locations of aquaculture, open water, and herbaceous wetlands from CropScape (U.S. Department of Agriculture, 2016). On average, our sites were closer to water than a random sample of points within each ecoregion (Appendix S1: Figure S2).

Because landcover was more variable in the CV (Appendix S1: Table S3), we extracted either the 2016 or 2017 mean percent cover within a 1-km buffer of a camera location for water/wetlands, forest, grass/pasture, natural habitat, and “trees” from CropScape (U.S. Department of Agriculture, 2016). Water/wetlands were composed of aquaculture, open water, and herbaceous wetlands; forest was composed of deciduous forest, evergreen forest, mixed forest, shrubland, and woody wetlands; natural habitat was composed of open water/wetlands, forest, and grass/pasture; and the “trees” category was composed of forest and all tree crops, to account for the many orchards that provide cover in the CV. Correlation testing revealed that percent natural cover was correlated with both percent grass and percent water/wetland cover (Appendix S1: Table S5). However, it was also correlated with human disturbance, so we chose to retain only human disturbance as covariate (Table 1). Additionally, percent forest and percent tree were correlated (Appendix S1: Table S5), so we retained only percent tree cover (Table 1) because research has shown that mesopredators use orchards as habitat (Nogeire et al., 2013).

Finally, because the CV does not have much variation in elevation, but does stretch a long-distance north to south, we included latitude as a covariate, as well as mean temperature and precipitation (Oregon State University, 2017). Latitude was correlated with both mean temperature and precipitation (Appendix S1: Table S5), so it was retained as the only occupancy covariate of the three (Table 1). Appendix S1: Figures S3–S6 in show mean covariate values for sites by ecoregion, year, and life-form. Mean covariate values were statistically similar between 2016 and 2017 in both ecoregions apart from ordinal date and latitude in the CV.

Single-season, single-species occupancy models assume that sites are closed to changes in occupancy during the survey period, species are never falsely detected, there is no unmodeled heterogeneity, and that detection of a species at one site is independent of detection at other sites (MacKenzie et al., 2017). Goodness-of-fit testing diagnoses whether the underlying assumptions of the models being fit to data are met and identifies how much extra variation there is, termed overdispersion. Therefore, if data show overdispersion, they are likely violating one of these underlying assumptions of occupancy modeling. Because our survey periods lasted only 27 days and photos were reviewed by three independent observers, we believe we met the first two assumptions of occupancy modeling and attributed any overdispersion to unmodeled heterogeneity and lack of independence between our sites.

To assess the fit of occupancy models, we calculated a Pearson χ² statistic and used a parametric bootstrap procedure to determine whether the observed statistic was unusually large (MacKenzie & Bailey, 2004). This procedure resulted in the estimation of an overdispersion parameter, ĉ (c-hat). A c-hat value of 1 indicates that the model fits the data and Akaike information criterion (AIC; Akaike, 1974) can be used for model selection, a c-hat value over 1 suggests that there is more variation in the data than expected by the model and c-hat is then used to compute quasi-AIC for model selection (MacKenzie & Bailey, 2004).

We used 2000 bootstraps to get final c-hat estimates. Large initial c-hat values led to a re-examination of our data where we identified unmodeled temporal autocorrelation. To account for this, we created a series of binomial covariates representing Markov dependency between survey days at each site for each species and tested which “lag” time most improved model fit (Furnas et al., 2017; Hines et al., 2010). In the CV, we discovered that a 3-day dependency most improved model fit for coyotes (ĉ: 1.4311), and that a Markov dependency covariate did not improve fit for bobcats (ĉ: 1.9887) or raccoons (ĉ: 1.93). We were unable to find a dependency covariate that improved fit for MD species. Instead, we aggregated detection histories for MD species so that 3 days of data collection were compressed into one survey occasion, which resulted in a detection history of nine surveys rather than 27 for each species and removed detectable autocorrelation in the time series (coyote ĉ: 1.2738, bobcat ĉ: 1.4753, and kit fox ĉ: 1.2368).

We fit single-species occupancy models in program PRESENCE (version 2.12.6, https://www.mbr-pwrc.usgs.gov/software/presence.html, accessed 1 October 2017; Hines, 2006) and implemented a two-stage model selection process. Using a constant global occupancy model, we tested all additive combinations of detection covariates, including interactions between year and human disturbance and presence of water, when their additive components were present, and used AIC corrected for small sample size (AIC_c; Hurvich & Tsai, 1989) to evaluate relative support for alternative models. Null detection models were also considered. We compared 52 detection models for mesopredator species in the MD and the coyote in the CV and compared 26 detection models for bobcats and raccoons in the CV. We considered any model within two delta (∆) AIC_c of the top model to have plausible support (Burnham & Anderson, 2004) and defined any model within that set with informative parameters (Arnold, 2010) as competitive. We identified informative parameters as those that reduced the deviance of the model by 2 per additional parameter.

We selected the competitive model with the most parameters to be held as a constant detection model in occupancy model selection. Again, we tested all additive combinations of occupancy covariates, including interactions between year and human disturbance and distance to water, when applicable. Null occupancy models were also considered. We tested 26 occupancy models for all MD mesopredators and 52 models for all CV mesopredator species. We used the same criteria to select a top occupancy model as we used for detection models. Finally, we fit top models in program MARK (version 9.0, http://www.phidot.org/software/mark/downloads/, accessed 5 October 2018; White & Burnham, 1999) to estimate detection and occupancy probabilities after specifying covariate values pertinent to our predictions.

We also used conditional two-species single-season occupancy modeling in PRESENCE to investigate the relationship between mesopredators and coyotes. These models estimate the probability of detection and occupancy of a subordinate species (Species B) conditional upon the presence of a dominant species (Species A; MacKenzie et al., 2017; Richmond et al., 2010), and can be formulated differently to test hypotheses about the relationships between the two species. Occupancy of Species A, the dominant species $$$ \left({\uppsi}^A\right), $$$ is always an unconditional parameter. In this study, we considered Species A to be the coyote. We tested whether the occupancy of a subordinate species (bobcat, raccoon, and kit fox) was conditional upon the presence of a coyote by estimating occupancy of Species B when Species A is present ($$$ {\uppsi}^{BA} $$$) and occupancy of Species B when Species A is absent ($$$ {\uppsi}^{Ba} $$$) separately or when $$$ {\uppsi}^{BA} $$$ = $$$ {\uppsi}^{Ba} $$$ and determined which model performed better according to AIC_c (Richmond et al., 2010). We also tested whether the detection of the subordinate species was conditional on the presence of coyotes by estimating the detection of Species B when Species A is absent ($$$ {p}^B $$$) and the detection of Species B when Species A is present ($$$ {r}^B $$$) separately or when $$$ {p}^B $$$ = $$$ {r}^B $$$ and comparing model performance.

We used covariates from the top single-species occupancy model for coyotes as constant covariates for Species A in all two-species occupancy models. However, we used the same two-stage model selection process used for single-species modeling for all subordinate species in conditional two-species modeling. When testing possible detection models for Species B, we kept the parameters conditional ($$$ {p}^B $$$ and $$$ {r}^B $$$ estimated separately and $$$ {\uppsi}^{BA} $$$ and $$$ {\uppsi}^{Ba} $$$ estimated separately). Once a top Species B detection model was selected, we then tested the performance of the model when Species B detection parameters were set conditional or unconditional to Species A.

We then held the top Species B detection model constant as we tested all possible Species B occupancy models. We repeated the same process of selecting a top model as was used for single-species modeling, and then tested the formulation of the top occupancy model to see if occupancy of Species B was conditional upon the occupancy of Species A. We then refit all top models in MARK to estimate detection and occupancy probabilities for scenarios of interest involving drought and human disturbance.

RESULTS

Between March–August 2016 and March–July 2017, we deployed camera traps at 319 sites in the MD (2016: 217 sites, 2017: 102 sites) and 266 sites in the CV (2016: 86 sites, 2017: 180 sites; Figure 1; Appendix S1: Table S6). Approximately 36% of CV sites were on private property (n = 95). We collected presence data over 15,795 camera trap surveys from 585 cameras operating for 27 days each; 2.25% of surveys (n = 356) were classified as “nonfunctional.” There were 2871 survey occasions in the MD following the aggregation of 27 survey days into nine survey occasions. Camera analysis differed between the first and second reviewer 10% of the time, primarily driven by 2016 analyses. Most differences were the result of missed observations and incorrect dates, rather than inconsistencies in species identification. All differences were resolved in the third analysis.

FIGURE 1. Map of all 585 study sites (black circles = 2016 sites, white circles = 2017 sites) in the Mojave Desert (white outline, 319 sites) and Central Valley (black outline, 266 sites) of California and the Human Footprint of the West (Leu et al., 2008).

In the MD, coyotes were detected at 99 of 319 sites surveyed, bobcats were detected at 53 sites, and kit foxes were detected at 133 sites. In the CV, coyotes were detected at 108 of 266 sites surveyed, bobcats were detected at 24 sites, and raccoons were detected at 117 sites. Odds ratios for detection ($$$ p $$$) and occupancy (ψ) covariates in final single-species and two-species occupancy models for CV and MD mesopredators are provided in Tables 4 and 5, respectively. Occupancy estimates for all species using both single- and two-species models are provided in Table 6 and model selection tables are available in Appendix S1: Tables S7–S16. The following results and interpretation focus on year and human disturbance.

TABLE 4 Odds ratios and their 95% CIs for detection ($$$ p $$$, top) and occupancy (ψ, bottom) covariates in final single-species occupancy models for Central Valley and Mojave Desert mesopredators.

	Central Valley			Mojave Desert
Covariate	Coyote	Bobcat	Raccoon	Coyote	Bobcat	Kit fox
$$$ p $$$: Bait						0.57 (0.49, 0.66)
$$$ p $$$: Guzzler				2.19 (1.43, 3.38)	1.72a (0.97, 3.05)	0.54a (0.26, 1.11)
$$$ p $$$: Human (2016)	0.73a (0.48, 1.10)		0.48 (0.35, 0.66)	1.29 (1.02, 1.64)
$$$ p $$$: Human					1.44 (1.03, 2.02)	1.12a (0.96, 1.30)
$$$ p $$$: Human (2017)	1.27 (1.00, 1.60)		1.42 (1.11, 1.81)	0.54 (0.33, 0.90)
$$$ p $$$: Ordinal date		0.41 (0.25, 0.69)	0.71 (0.59, 0.85)	1.21a (0.98, 1.49)
$$$ p $$$: Lag (3 days)	2.24 (1.51, 3.31)
$$$ p $$$: Water	1.61 (1.07, 2.44)
$$$ p $$$: Water (2016)			2.03 (1.16, 3.57)
$$$ p $$$: Water (2017)			1.22a (0.86, 1.74)
$$$ p $$$: Year (2017)	0.94a (0.62, 1.42)	0.21 (0.07, 0.65)	0.36 (0.22, 0.59)	0.83a (0.52, 1.33)	1.93 (1.02, 3.64)	2.03 (1.47, 2.80)
ψ: Distance to water		0.66a (0.30, 1.46)	0.47 (0.27, 0.83)		0.44 (0.23, 0.82)	1.90 (1.39, 2.60)
ψ: Elevation					2.33 (1.45, 3.75)	0.41 (0.28, 0.60)
ψ: Human (2016)			1.94a (0.82, 4.59)			0.84a (0.60, 1.20)
ψ: Human				1.26a (0.92, 1.74)
ψ: Human (2017)			0.93a (0.48, 1.82)			1.58a (0.94, 2.67)
ψ: Latitude	0.59 (0.38, 0.92)	1.59a (0.73, 3.46)	3.42 (1.87, 6.27)
ψ: Tree cover		2.40 (1.34, 4.30)
ψ: Year (2017)			3.61 (1.31, 9.97)			0.69a (0.36, 1.30)

Note: Covariates with “2016” or “2017” are from an interaction with year within the model.

^a95% CI for odds ratio crosses 1.

TABLE 5 Odds ratios and their 95% CIs for covariates in final two-species occupancy models for subordinate mesopredators in the Central Valley and Mojave Desert: Detection without coyotes ($$$ {p}^B $$$, top), detection with coyotes ($$$ {r}^B $$$, middle), and occupancy (ψ, bottom).

	Central Valley		Mojave Desert
Covariate	Bobcat	Raccoon	Bobcat	Kit fox
$$$ {p}^B $$$: Bait				0.61 (0.52, 0.72)
$$$ {p}^B $$$: Human (2016)		0.36 (0.17, 0.76)	0.84a (0.38, 1.87)
$$$ {p}^B $$$: Human	1.64a (0.83, 3.25)			1.14a (0.95, 1.37)
$$$ {p}^B $$$: Human (2017)		2.02a (0.66, 6.19)	2.43 (1.30, 4.56)
$$$ {p}^B $$$: Ordinal date			0.377 (0.22, 0.64)	1.05a (0.88, 1.24)
$$$ {p}^B $$$: Water	9.06 (2.83, 28.96)
$$$ {p}^B $$$: Water (2016)		7.40 (1.57, 34.82)
$$$ {p}^B $$$: Water (2017)		6.91 (1.48, 32.25)
$$$ {p}^B $$$: Year (2017)		5.16a (0.561, 47.36)	4.55 (1.26, 16.41)	1.41a (0.92, 2.15)
$$$ {r}^B $$$: Bait				0.47 (0.36, 0.62)
$$$ {r}^B $$$: Human (2016)		0.34 (0.25, 0.46)	4.18 (2.15, 8.13)
$$$ {r}^B $$$: Human	0.27 (0.12, 0.62)			1.41 (1.01, 1.97)
$$$ {r}^B $$$: Human (2017)		1.17a (0.92, 1.50)	3.07 (1.24, 7.58)
$$$ {r}^B $$$: Ordinal date			2.44 (1.56, 3.83)	0.36 (0.24, 0.54)
$$$ {r}^B $$$: Water	0.06 (0.01, 0.31)
$$$ {r}^B $$$: Water (2016)		7.02 (3.92, 12.55)
$$$ {r}^B $$$: Water (2017)		0.70 (0.47, 1.03)
$$$ {r}^B $$$: Year (2017)		0.39 (0.25, 0.60)	4.69 (1.73, 12.70)	4.57 (2.62, 7.96)
$$$ {\uppsi}^B $$$: Distance to water	0.65a (0.36, 1.17)	0.49 (0.32, 0.75)	0.40 (0.23, 0.70)	1.80 (1.33, 2.45)
$$$ {\uppsi}^B $$$: Elevation			2.78 (1.71, 4.53)	0.31 (0.20, 0.48)
$$$ {\uppsi}^B $$$: Human (2016)			0.88a (0.44, 1.77)	0.74a (0.52, 1.06)
$$$ {\uppsi}^B $$$: Human		1.21a (0.84, 1.74)
$$$ {\uppsi}^B $$$: Human (2017)			0.401 (0.17, 0.96)	1.65 (1.01, 2.69)
$$$ {\uppsi}^B $$$: Latitude	1.76a (0.99, 3.13)	3.67 (2.28, 5.90)
$$$ {\uppsi}^B $$$: Tree cover	2.36 (1.49, 3.75)
$$$ {\uppsi}^B $$$: Year (2017)		2.02a (0.88, 4.62)	0.41a (0.12, 1.35)	0.49 (0.26, 0.93)

Note: Covariates with “2016” or “2017” are from an interaction with year within the model.

^a95% CI for odds ratio crosses 1.

TABLE 6 Single-species ($$$ \uppsi $$$) and two-species occupancy probabilities for subordinate species ($$$ \uppsi $$$^B) and coyotes ($$$ \uppsi $$$^A) at average covariate values.

Species	Ecoregion	Occupancy	Estimate	SE
Bobcat	Central Valley	$$$ \uppsi $$$	0.066	0.026
		$$$ \uppsi $$$B	0.088	0.026
	Mojave Desert	$$$ \uppsi $$$	0.128	0.032
		$$$ \uppsi $$$B (2016)	0.213	0.081
		$$$ \uppsi $$$B (2017)	0.099	0.035
Coyote	Central Valley	$$$ \uppsi $$$	0.546	0.058
		$$$ \uppsi $$$A (B: bobcat)	0.542	0.047
		$$$ \uppsi $$$A (B: raccoon)	0.682	0.064
	Mojave Desert	$$$ \uppsi $$$	0.374	0.038
		$$$ \uppsi $$$A (B: bobcat)	0.390	0.035
		$$$ \uppsi $$$A (B: kit fox)	0.384	0.035
Kit fox	Mojave Desert	$$$ \uppsi $$$ (2016)	0.443	0.047
		$$$ \uppsi $$$ (2017)	0.352	0.061
		$$$ \uppsi $$$B (2016)	0.516	0.051
		$$$ \uppsi $$$B (2017)	0.345	0.056
Raccoon	Central Valley	$$$ \uppsi $$$ (2016)	0.239	0.079
		$$$ \uppsi $$$ (2017)	0.532	0.072
		$$$ \uppsi $$$B (2016)	0.361	0.078
		$$$ \uppsi $$$B (2017)	0.532	0.062

Note: If occupancy probabilities differed by year, both estimates were included, and year is indicated in parentheses.

Coyotes

The interaction of year and human disturbance influenced detection probability of coyotes in both the MD and CV (Table 4; Figure 2). Detection of CV coyotes was negatively associated with human disturbance in 2016, while detection of MD coyotes was positively associated (Table 4). In 2017, the detection of CV coyotes was positively associated with human disturbance, while the detection of MD coyotes was negatively associated (Table 5). Overall, detection probability of both CV and MD coyotes decreased from 2016 to 2017, although this relationship was uncertain for both (Table 4; Figure 2). Additionally, the occupancy of MD coyotes was positively associated with human disturbance (Table 5).

FIGURE 2. Single-species detection estimates and 95% CI (shaded regions) for coyotes (Canis latrans) in the Mojave Desert and Central Valley of California along a human disturbance gradient in 2016 and 2017. Circles represent 2016 estimates and triangles represent 2017 estimates.

Subordinate mesopredators Occupancy

Coyotes impacted the detection probability, but not occupancy, of all subordinate species. The relationships and magnitude of influence between environmental covariates (e.g., elevation) and subordinate mesopredators' occupancy probability did not change between single- and two-species modeling (Tables 4 and 5). However, relationships between human disturbance, year, and mesopredator occupancy did vary between the two modeling frameworks.

In single-species modeling, raccoon occupancy was positively associated with human disturbance in 2016, but negatively associated in 2017 and overall raccoon occupancy increased from 2016 to 2017 (Table 4). In two-species modeling, raccoon occupancy was positively associated with human disturbance regardless of year, and occupancy still increased from 2016 to 2017 (Table 5). Kit fox occupancy was negatively associated with human disturbance in 2016, and positively associated in 2017, in both single- and two-species modeling (Tables 4 and 5). Additionally, kit fox occupancy decreased from 2016 to 2017 in both models. Following single-species modeling, human disturbance was not in the top occupancy model for MD bobcats, but this changed after two-species modeling. The occupancy of MD bobcat decreased from 2016 to 2017 and was negatively associated with human disturbance in both 2016 and 2017, although this relationship was stronger in 2017 (Table 5).

Detection

Following single-species modeling, year influenced the detection of all subordinate species (Table 4). The detection of CV mesopredators (bobcat and raccoon) decreased from 2016 to 2017 (negatively associated with year), while the detection of MD mesopredators (bobcat and kit fox) increased from 2016 to 2017 (Table 4; Figure 3). Year had to be removed from CV bobcat two-species modeling due to few detections of bobcats and coyotes in these contexts. Following two-species modeling, raccoon detection was still negatively associated with year in the presence of coyotes, but positively associated in the absence of coyotes; however, this relationship was uncertain (Table 5; Figure 4). Detection of MD mesopredators remained positively associated with year regardless of the presence of coyotes; however, the relationship between kit fox detection and year in the absence of coyotes was uncertain (Table 5; Figures 4 and 5).

FIGURE 3. Single-species detection estimates and 95% CIs (shaded regions) for bobcats (Lynx rufus) and kit foxes (Vulpes macrotis arsipus) in the Mojave Desert (MD) and raccoons (Procyon lotor) in the Central Valley of California along a human disturbance gradient in 2016 and 2017. Circles represent 2016 estimates and triangles represent 2017 estimates.

FIGURE 4. Two-species detection estimates and 95% CIs (shaded regions) for kit foxes (Vulpes macrotis arsipus) in the Mojave Desert and raccoons in the Central Valley of California at sites with and without coyotes (Canis latrans) present along a human disturbance gradient in 2016 and 2017. Circles represent 2016 estimates and triangles represent 2017 estimates.

FIGURE 5. Two-species detection estimates and 95% CIs (shaded regions) for bobcats (Lynx rufus) at sites with and without coyotes (Canis latrans) present in the Mojave Desert and Central Valley of California along a human disturbance gradient in 2016 and 2017. Circles represent 2016 estimates, triangles represent 2017 estimates, and squares represent estimates for both years together.

Following single-species modeling, detection of MD mesopredators was positively associated with human disturbance, regardless of year (Table 4; Figure 3). Following two-species modeling, this relationship remained consistent for kit foxes in both the presence and absence of coyotes (Table 5; Figure 4). However, MD bobcat detection was negatively associated with human disturbance in 2016 in the absence of coyotes, but positively associated in the presence of coyotes (Table 5; Figure 5). MD bobcat detection remained positively associated with human disturbance in 2017, regardless of the presence or absence of coyotes (Table 5; Figure 5).

The detection of CV mesopredators in response to human disturbance was not consistent. The relationship between raccoon detection and human disturbance did not change between single- and two-species modeling. Raccoon detection was negatively associated with human disturbance in 2016 and positively associated in 2017 (Table 5; Figure 4). However, human disturbance did not influence CV bobcat detection in single-species modeling but was positively associated with human disturbance in the absence of coyotes and negatively associated in the presence of coyotes (Table 5; Figure 5).

DISCUSSION

Our research suggests that drought influenced both how species utilized the landscape and how they responded to a top predator. MD mesopredators appeared to use areas of high human disturbance as drought refugia, while CV mesopredators did not. Human disturbance in the MD is significantly lower than in the CV, which is dominated by high levels of human development and agricultural lands; therefore, the prolonged drought may have had greater effects in the MD without the highly managed water systems in the CV (Lund et al., 2018). Coyotes in both ecoregions impacted the detection of all mesopredator species, indicating that, while coyotes do not directly influence the spatial distribution of other mesopredators, they do influence those species' spatial activity patterns and behavior. Coyotes' increased activity in areas of higher human disturbance in the MD during the drought correlated with reduced kit fox occupancy in these areas: a pattern which reversed following the heavy rains in the winter of 2016–2017. This suggests that the safety of urban shields may depend upon how drought influences the behavior of other species within the mesopredator guild.

Drought and mesopredator detection

As predicted, the detection probability of bobcats and kit foxes in the MD increased from 2016 to 2017, responses that may have been driven by prey availability. Rodents are a frequent prey base of mesopredators and rely on winter precipitation and annual vegetation for successful reproduction in arid ecosystems (Beatley, 1969). Due to the prolonged drought that persisted into 2016, rodent populations were reduced from pre-drought levels during our first year of data collection (Kelly et al., 2019). However, rodent populations have been shown to rapidly respond to drought conditions ending (Bradley et al., 2006), and increased requests for rodent pest control after rains began in the winter of 2016 in Alameda County, California (Wilson, 2018) suggest that rodent populations throughout California could have increased following the end of the drought. Mesopredator densities, reproductive success, and spatial overlap have all been tied to prey availability and density (Fuller & Sievert, 2001; White & Ralls, 1993). In 2016, reduced rodent availability could have decreased reproduction in mesopredators, resulting in lower densities. Similarly, White and Ralls (1993) found that kit foxes reduced the overlap of home ranges during prey scarcity in a drought. Both situations could influence detection probability.

Unlike bobcats in the MD, bobcat detection in the CV decreased from 2016 to 2017. While there was a 10% shortage statewide in water available for agricultural use (Lund et al., 2018), groundwater pumping increased in the CV during the drought (Faunt et al., 2016), leading to higher water availability relative to the MD. At intermediate water availability, prey can function as both a source of energy and water, which may lead to increased hunting behavior in predators to meet resource needs (McCluney et al., 2012). It is possible that CV bobcats increased their activity to acquire more prey during the drought in 2016 and reduced hunting activities in 2017 when surface water was more readily available. Additionally, bobcats are known to diversify their diets in times of food scarcity (McKinney & Smith, 2007), which could have caused CV bobcats to forage more broadly in 2016 to meet their resource needs.

Drought-mediated responses to human disturbance

Coyote occupancy was positively correlated with human disturbance in the MD, regardless of drought conditions. Nevertheless, coyotes were more likely to be detected at sites of higher human disturbance in the MD during the drought. By contrast, in the CV, coyote detection was highest in low disturbance sites during the drought, while detection in more disturbed sites was similar during and after the drought. These differences between ecoregions may be due to variation in resource availability. Prey populations were likely reduced during the drought, and coyotes are known to diversify their diets when prey is scarce (Neale & Sacks, 2001). Coyotes in the CV could have spent more time foraging in natural habitats during the drought while maintaining the same level of activity in agricultural areas where resource levels were more stable. In the MD, however, coyotes may have increased their activity in human-disturbed areas as they diversified their diet to include more anthropogenic items (Cypher et al., 2018).

The responses to human disturbance by the other three species we examined also appeared to be mediated by drought and, in some cases, the presence of coyotes. Detection of bobcats in the MD during the drought was negatively correlated with human disturbance at sites without coyotes, while detection at sites with coyotes was positively correlated with human disturbance. After the drought, bobcat detection was positively associated with human disturbance regardless of the presence or absence of coyotes. One possible explanation is that road verges can function as refugia supporting greater abundances of small mammals (Ruiz-Capillas et al., 2013) and may be preferred hunting grounds for bobcats when prey are more abundant, such as in 2017. However, in 2016, when prey abundance was low (Cypher et al., 2018) and where coyotes were absent, bobcats could have diversified their diet beyond preferred prey like rodents to include more lagomorphs (Neale & Sacks, 2001), a primary prey item for coyotes (Cypher et al., 2018) whose abundance is not associated with roads (Cypher et al., 2009). When coyotes were present and there was increased competition for lagomorphs during the drought, bobcats may have sought out anthropogenic resources, such as fruits and seeds (McKinney & Smith, 2007), to supplement their diversified diet. Unlike the bobcats in the MD, CV bobcats were more likely to be detected at higher human disturbance sites if coyotes were absent and less likely to be detected at disturbed sites if coyotes were present. High human disturbance radiates from agricultural areas (Leu et al., 2008) and bobcats' main source of prey are rodents (McKinney & Smith, 2007), which are agricultural pests (Gebhardt et al., 2011). Bobcats likely utilize agricultural fields as hunting grounds in the CV but may reduce their hunting activity in these areas if coyotes are present to avoid competitive interactions.

In the MD, kit foxes were consistently more likely to be detected in areas of higher human disturbance, even after accounting for coyote presence. Kit foxes are known to use anthropogenic food (Iossa et al., 2010), so their elevated activity in more disturbed areas could indicate that they use these areas to search for food. However, kit fox occupancy was higher in human-disturbed areas in 2017 and lower in high disturbance areas during the drought in 2016. This remained true following two-species occupancy modeling. As an arid-adapted species, the kit fox may outcompete other species in natural landscapes when there is a resource shortage, reducing their need to occupy areas of high human disturbance. However, kit foxes are known to inhabit urban areas and to benefit physiologically in urban areas in times of drought (Cypher & Frost, 1999). This could suggest that kit fox avoidance of higher disturbance sites in 2016 was in response to coyotes' increased activity in those areas during the drought. This aligns with a study from Utah where kit foxes foraged in resource-scarce habitats as a form of spatial isolation from coyotes (Kozlowski et al., 2012).

Raccoon detection probability in the CV was significantly higher in areas of lower human disturbance during the drought. This might indicate that raccoons had to increase foraging activity during the drought to meet their resource needs, especially in less disturbed areas where natural resources were likely limited. Raccoons eat a diversity of food but are selective when food is plentiful and diversify when resources are scarce (Lotze & Anderson, 1979), which could explain their higher detection rates in more natural areas during the drought. Conversely, raccoons in human-disturbed areas, where anthropogenic resources were readily available, may not have had to adjust their foraging activity, leading to lower relative detection rates. The intense winter precipitation of 2016–2017 increased green vegetation cover (Potter, 2019), likely stimulated prey population growth (Wilson, 2018), and therefore reduced raccoons' obligate foraging activity in natural areas, lowering their detection probability.

Management implications

Mesopredators in California are facing major changes to their ecosystems, including drought and the expansion of human disturbance. The risk of drought is increasing (Diffenbaugh et al., 2015), and future projections predict the replacement of greater than 4.8 million hectares of wild and agricultural lands with exurban development (Mann et al., 2014). Our research indicates that mesopredators respond differently to human disturbance and dominant predators during times of drought, and that human modified areas may serve as drought refugia in California.

Species in the CV could have been buffered to some of the more extreme drought effects by the managed landscape and water systems, which ensure productive agriculture (Lund et al., 2018). While agricultural lands are not typically considered and managed as wildlife habitat, future conservation might need to redefine these areas as important for the persistence of wildlife in the region, especially during drought. However, it is also important to recognize that increased use of habitat near humans can be problematic. Our research revealed that coyotes in the MD likely increased their activity in human-disturbed areas during the drought, probably in search of anthropogenic resources. While this may have benefited coyotes, it may also have increased the probability of human–coyote conflict. Indeed, an analysis of wildlife incidents reported to the California Department of Fish and Wildlife (CDFW) found that the odds of human–mesopredator conflict increased 1%–2% for every 25 mm decrease in precipitation (Bowman et al., 2020). In response to the drought, CDFW established the “Human Wildlife Conflicts Program” to improve wildlife incident responses (California Natural Resources Agency, 2021).

Consistent tracking of human–wildlife conflict and drought conditions will be necessary for understanding and managing for climate-driven increases in human–wildlife incidents in the future (Abrahms, 2021). Similarly, the continued large-scale camera trap monitoring of terrestrial vertebrate species in California will provide more robust data about mesopredator occurrence and co-occurrence as precipitation patterns fluctuate over time. These data, paired with small mammal abundance studies, or second cameras set to maximize small mammal detections (Seidlitz et al., 2021), would be valuable in examining the hypothesized influence of prey availability on mesopredator activity in times of drought and drought recovery.

ACKNOWLEDGMENTS

The data and funding for this project were provided by the California Department of Fish and Wildlife. We thank Karen Miner for her oversight of the project and Scott Newton and the many technicians who collected the data, as well as the numerous partners and landowners who made this large-scale monitoring project a reality.

CONFLICT OF INTEREST

The authors declare no conflict of interest.

DATA AVAILABILITY STATEMENT

Data (Parren, 2021) are available from Dryad: https://doi.org/10.5061/dryad.6t1g1jx08.