Altitude shapes the environmental drivers of large‐scale variation in abundance of a widespread mammal species

Abstract Aim Habitat quality and heterogeneity directly influence the distribution and abundance of organisms at different spatial scales. Determining the main environmental factors driving the variation in species abundance is crucial to understand the underlying ecological processes, and this is especially important for widely distributed species living in contrasting environments. However, the responses to environmental variation are usually described at relatively small spatial scales. Here, we studied the variation in abundance of a widely distributed mustelid, the European badger (Meles meles), across France. Location The whole metropolitan France. Methods We used (a) direct detections of 9,439 dead and living badgers, from 2006 to 2009, to estimate badger relative abundance in 703 small agricultural regions of metropolitan France and (b) a Bayesian modeling approach to identify the main environmental determinants influencing badger abundance. Results Despite a continuous distribution of badger in France, we found large variation in badger abundance between regions, explained by environmental factors. Among a set of 13 environmental variables, we demonstrated that badger abundance in lowlands (<400 m a.s.l.) was mostly driven by biotic factors such as potential food resources (earthworm abundance and fruit crops) and forest fragmentation. Conversely, in mountainous areas, abiotic factors (i.e., soil texture and climate) drove the variation in badger relative abundance. Main conclusions These results underline the importance of mapping the abundance of wildlife species based on environmental suitability and highlight the complexity of drivers influencing species abundance at such large spatial scales. Altitude shaped the environmental drivers (biotic vs. abiotic) that most influenced relative abundance of a widespread species. In the case of badger, such abundance maps are crucial to identify critical areas for species management as this mustelid is a main wild vector of bovine tuberculosis in several countries.


| INTRODUC TI ON
Distribution and abundance of species are determined by the interaction of ecological processes (such as movement, reproduction, or social interactions) and environmental factors (such as resource availability and habitat configuration; Brown, 1984).
Understanding how a species' abundance is likely to vary over space and time helps to plan sound biological management strategies and to identify areas requiring conservation/management attention (Kaiser, 1997). Spatial and temporal environmental heterogeneity can determine the population trends of wild populations, either a population decrease or extinction, or the ability of some species to become overabundant. Species interactions and environmental factors are usually the most important factors limiting distribution and abundance (Boyce et al., 2016;Hoffmann & Blows, 1994), but may act at different temporal and spatial scales, environmental factors being preponderant at large scales (Byrne, Fogarty, O'Keeffe, & Newman, 2015).
Abundance can be highly variable within a species distribution range in response to environmental gradients (Morrison et al., 2006).
Areas with high environmental suitability tend to support larger populations (Weber, Stevens, Diniz-Filho, & Grelle, 2017) because of increased favorability of local conditions (e.g., climatic conditions or available food resources), in particular when ecological requirements are met (Hutchinson, 1957). For example, in an assemblage of freshwater fish, species abundance was associated with species' environmental preferences in the river (e.g., stream size and water clarity; Taylor, Winston, & Matthews, 1993). Landscape characteristics and trophic resource availability can also influence spatial variation in abundance in other taxa such as carnivorous mammals (Pita, Mira, Moreira, Morgado, & Beja, 2009), but to fully understand the habitat driving forces linked to animal abundance, we need better knowledge on how qualitative and quantitative aspects of environmental factors impact the spatial variation in abundance (Morris, 1987).
Identifying patterns of covariation between abundance and environmental factors at the macroscale, that is, over all or at least a large part of the distribution area, requires both a very large amount of occurrence data that allows the estimation of species abundance, and information about environmental predictors distributed over the same large area. However, data of spatial variation in abundance for a species within its distribution range, or at least over a large area far beyond the population scale, are difficult to obtain. Consequently, predictions of spatial abundance patterns over large areas remain scarce for diverse taxa (Sagarin, Gaines, & Gaylord, 2006). Most field and statistical methods rely on complex designs to estimate local absolute abundance (e.g., distance sampling approaches or mark-recapture methods) and are inapplicable or too expensive to be considered for abundance estimation in large spatial scale studies. The relative abundance estimates allow us to study the species over broad spatial scales, under the assumption that the population index is proportional to the population density, and the probability of "detection" for animals in the survey is known (Pollock et al., 2002).
Among mammals, the badger Meles meles appears to be a good model species to investigate how environmental factors shape the pattern of variation in abundance over a large spatial scale due to its widespread distribution (Johnson, Jetz, & Macdonald, 2002;Newton-Cross, White, & Harris, 2007). This nocturnal medium-sized carnivore, distributed throughout temperate Eurasia, can occupy a large range of biomes, such as woodlands, forests, and arid or mountainous landscapes to a lesser extent (Griffiths & Thomas, 1993).
Badger social structure and population abundance are thought to be driven mainly by sett or trophic resource dispersion/availability (Johnson et al., 2002;Kowalczyk, Zalewski, Jedrzejewska, & Jedrzejewski, 2003), which are both driven by landscape patterns (Hammond, McGrath, & Martin, 2001). Characterized by an omnivorous diet, the species can use a wide range of food items including plants (e.g., fruits) and animals such as earthworms or insects (e.g., beetles) depending on the temporal variation of resources availability (Cleary, Corner, O'Keeffe, & Marples, 2011;Kruuk, 1978).
The badger is therefore a generalist or opportunist feeder (Roper, 1994). Several environmental determinants have been proposed to explain variation in badger abundance, using predicted sett abundance in large-scale studies, among which food resource availability, landscape patterns, and climate were the most important drivers (Acevedo et al., 2014;Etherington, Ward, Smith, Pietravalle, & Wilson, 2009;Reid, Etherington, Wilson, Montgomery, & Mcdonald, 2012). A previous study showed that habitat fragmentation due to forest loss affected badger sett density in Spain and led to population isolation (Virgós, 2001). However, using the density of badger setts as an abundance proxy at large scale might be irrelevant due to social group size variation in this species, within seasons (Revilla & Palomares, 2002), or with trophic resources availability (Da Silva, Woodroffe, & Macdonald, 1993). Badger absolute abundance and densities are not known in France, but a recent study has estimated the badger relative abundance (i.e., the actual abundance multiplied by an unknown constant) in each small agricultural region (SAR) of France (Calenge et al., 2015). How these spatial variation in relative abundance over SARs correlate with environmental variation remains unknown.
In the present study, we aim at determining the environmental factors driving spatial variation in badger relative abundance at the scale of all metropolitan France. We first estimated relative badger abundance per SAR over France during 4 years, that is, 2006-2009, using a Bayesian modeling approach expanding previous work of Calenge et al. (2015). Second, we investigated the main ecological drivers of badger relative abundance variation among a set of relevant environmental variables. Although badger is present throughout France, we expected local abundance to be linked to the large variation within the country in terms of agricultural practices, human pressure, and forest cover, and leading to quantitative and qualitative habitat variation between regions. Based on the scientific literature, we predicted that relative badger abundance would be positively correlated with the availability of different potential food resources, such as earthworm abundance (described as the most frequent item in the badger's diet; Kruuk, 1978;Neal & Cheeseman, 1996), which varies widely between spatial units in France (Rutgers et al., 2016). We also expected higher relative badger abundance in woody lowland environments, as this climatic zone offers both the highest level of food resources (Mysłajek et al., 2016) and suitability for sett installation, with favorable mosaic landscapes (Cresswell, Harris, Bunce, & Jefferies, 1989;Kruuk, 1989) including small agricultural fields (i.e., pastures) and forest patches.

| Badger data collection
Direct observations of badgers (M. meles), that is, presence data, were collected by wildlife protection officers from the French National Hunting and Wildlife Agency (i.e., ONCFS). This organization implements the "small carnivorous logbook program" (SCSL), a program conducted over the whole of metropolitan France, and which asks wildlife protection officers to report all dead and living animals belonging to small carnivorous species (in particular Mustelidae) that they randomly meet during their fieldwork (Calenge et al., 2015). Detections of dead or living animals were recorded mainly along roads, when the officers were driving by car (e.g., police patrol missions). Although this program began in 2001 and is still running, we restricted our dataset to 4 years of detections, between 2006 and 2009. Because we needed to make reasonable the assumption of constant relative abundance over the study period, we had to define a shorter study period. Moreover, due to several logistics constraints in recent years (smaller number of officers, smaller budget for gasoline, etc.), the program has collected a much smaller volume of data in recent years. The years TA B L E 1 Environmental variables (n = 13) calculated for each spatial unit (i.e., small agricultural region) in metropolitan France

Depth
Mean index of depth class of an obstacle to roots (measured in 4 categories: 1 = no obstacle to roots between 0 and 80 cm, 2 = obstacle to roots between 60 and 80 cm depth, 3 = obstacle to roots between 40 and 60 cm depth, 4 = obstacle to roots between 0 and 40 cm depth) 2006-2009 corresponded to the period with the largest number of detections by the data volume.
The spatial unit of interest in our study is the small agricultural region, which is delimited by the intersections between "large agricultural regions" (n = 429 in metropolitan France) and administrative departments (n = 90). Large agricultural regions correspond to a group of neighboring administrative units (i.e., municipalities) characterized by similar landscapes, agricultural practices, and activities (for further details see www.agres te.agric ulture.gouv.fr/defin ition s/zonages). These intersections define 703 homogeneous SARs in metropolitan France (mean = 766.7 km 2 ; SE = 722.2 km 2 ) and for which variation in relative abundance was studied.

| Environmental variables
Each SAR was characterized by 13 environmental variables that were grouped into five categories describing landscape, terrain features, potential food resources, climate, and the urbanization level ( given in Appendix S1. Each selected continuous variable was standardized (common mean equals to 0, and common standard deviation equals to 1) to provide comparable effect sizes in the modeling procedure.
Landscape variables were mainly related to forest cover rep- to roots in the first 40 cm; Depth). All three related to the ability to dig and build a sett. Sloped areas are attractive as they aid excavation, and the depth of the soil should allow for sufficient digging (Thornton, 1988). Most relevant potential food resources attractive to badgers were also recorded, that is, the percentage cover of (a) . We also used (d) the median predicted abundance of earthworms in topsoil (Earthworm) from Rutgers et al. (2016), as a proxy of the major potential food resource for badgers (Cleary et al., 2011;Mouches, 1981). To describe the climate of each SAR, we used 24 temperature and precipitation maps of average monthly temperature and rainfall. We applied a principal component analysis using the ade4 package (Dray & Dufour, 2007)  SARs into two groups: (a) nonmountainous SARs with an average elevation lower than 400 m (n = 531) and (b) mountainous SARs with an average elevation higher than 400 m (n = 172). We set the limit of 400 m based on a principal component analysis on all variables, including altitude (further details in Appendix S2). We then modeled the relationship between the predictors and badger abundance in these two groups separately.

| Modeling framework
We used a Bayesian variable selection approach to identify the major environmental determinants influencing the relative abundance of badgers. We developed a hierarchical modeling framework extending the modeling approach described in Calenge et al. (2015), to allow both the estimation of the relative abundance of the badger and the identification of variables affecting this relative abundance (for each elevation group of SARs).
The first level of this model was an observation model de- status k (k = 1 or 2 for dead and living animals, respectively). N ijk is assumed to follow a Poisson distribution: And we supposed as the expectation of this distribution the following model: where S j is the surface area of the jth SAR, A ij is the true density of the species i in the SAR j, E jk is the sampling effort in SAR j for animals in status k, and finally, P ik is the detection probability of the species i in the status k. The parameters in the above model are not all identifiable, but Calenge et al. (2015) showed that all the parameters of the following reparametrized model are identifiable: where s j is the log surface of the jth SAR, a ij the log-relative abundance of species i in each SAR j, e jk the log-relative sampling effort for, respectively, dead or alive animals, and p ik a complex function of the detectability of the species i with status k. The equivalence between the two models can be demonstrated with the following: This model can be fitted under the assumption that the log-relative sampling effort is known for dead animals. We assumed that this effort was proportional to the known number of kilometers Vj travelled in the jth SAR by officers, that is, e j1 = log Vj + β (see Calenge et al., 2015 for further details on this model and properties of the estimated parameters). Note that this reparametrized model allows in particular the estimation of the desired log-relative abundance a ij .
We extended this framework for the particular case of the badger, which is the focus species here. For this species only, we designed a second level in our hierarchical modeling framework: We modeled the relative abundance a bj of the badger in the SAR j as a linear combination of the environmental variables of interest. We used the Bayesian approach for variable selection set by Kuo and Mallick (1998), in order to select the variables that are most likely to explain the relative abundance of badgers among the 13 independent environmental predictors X r (r = 1,…,R; R = 13). Thus, we supposed that the log-abundance a bj of the badger follows a normal distribution of parameters: We modeled the expectation of this distribution with the following: where a 0 is the intercept of the model, and X jr is the value of the variable X r (r = 1…R) for the SAR j. The influence of the environmental variable X r on the mean log-relative abundance of the badger is supposed to be the result of the combination of two parameters. First, the parameter γ r is a binary coefficient indicating the presence (γ r = 1) or absence (γ r = 0) of the environmental variable X r in the model. In a Bayesian context, the value of this parameter is supposed to be the realization of a Bernoulli variable with posterior probability p r that the variable X r is in the model. Second, the coefficient β r is the classical regression coefficient associated with variable X r , can take any real value, and determines the importance of the rth variable on the mean log-relative abundance of the badger when this variable belongs to the model, as in a classical regression model. This approach consists in separating the presence of a variable in a model from its importance and then estimating the probability of the presence of each variable in the model from the data, as suggested by Kuo and Mallick (1998 chains for an initial period of 1,000 cycles (burn-in period) and then collected information for the next 300,000 iterations with a thinning of 100. We implemented the MCMC simulations with JAGS software (Plummer, 2018) operating in R software.
From our analyses, we could identify those variables with the largest influence on the relative abundance of the badger and calculate the posterior probability p r that each variable X r belongs to the best model (γ r = 1). We also identify the best models predicting the relative abundance of the badger and calculated the posterior probability p (γ 1 , γ 2 ,…γ r ), for each possible combination of the coefficients (γ 1 , γ 2 ,…γ r ). We checked the mixing properties of the MCMC by verifying that the posterior probabilities estimated for the coefficients γ r were identical across the three chains.

| Index of badger relative abundance
For each class of SARs (mountainous and nonmountainous), the procedure of variable selection above allowed us to identify the variables affecting the badger relative abundance a bj (we retained the predictors with a posterior probability p r > .5). We therefore fit a final model of μ j as a linear combination of these variables (γ r set to 1 for retained variables and 0 for other variables). For each class of SARs, the best models of μ j allowed the estimation of the coefficient β r of each selected variable. Badger relative abundance a bj for each SARs was therefore estimated by a bj ~ N (μ j , τ) and rescaled to represent relative abundance as an index, from 0 to 1. showing high variability between SARs (Figure 1).

| Nonmountainous SARs group
The relative abundance of badger in nonmountainous SARs (<400 m) was influenced by four environmental variables among the 13, each one being characterized by a posterior probability of being in the true model equal to one ( Table 2). The most important variables were related to food availability, such as percentage cover of fruits, pastures, and abundance of earthworms in the topsoil fraction of SARs (Table 2). We found a positive relation between these variables and the relative abundance of badger, showing the importance of resource availability in increasing badger abundance. Edge density had a strong negative influence on the relative abundance of badgers, suggesting that the forest cover should not be too fragmented to support high abundance at this scale.
Among the 2 13 possible combinations of variables, the best model predicting the relative abundance of the badger in low elevations included these four variables. This model was the most often visited by the three MCMC chains, with a posterior model probability p (γ 1 , γ 2 ,…γ r ) of .440 (Table 3). The second half of selected models included the effect of other variables characterized by low posterior probabilities, which were not considered here.

| Mountainous SARs group
Two variables among the 13 relevant environmental variables had the largest influence on relative badger abundance for mountainous SARs (elevation >400 m), with posterior probabilities calculated over all the MCMC iterations p r > .5 (Table 2). High badger abundance was mostly explained by a higher index of soil texture, which suggested a preference of the badger for fine soil texture. Continental climate also has a positive effect on abundance in mountainous regions. Unsurprisingly, low temperatures and heavy winter precipitation tend to limit badger abundance more in mountainous SARs than in nonmountainous SARs.
The best model predicting the relative abundance of badgers included soil texture and continental climate and was selected by the three MCMC chains, with a posterior model probability p (γ 1 , γ 2 , …γ r ) of .209 (Table 3). Less probable models included mostly the effects of environmental variables related to resource availability (i.e., Earthworm, Maize, and Pasture), with lower posterior model probabilities (p r < .30); they were not considered here.

| Badger relative abundance in France
Bayesian variable selection allowed the identification of four variables affecting the relative abundance of the badger in low elevation SARs and two for high elevation SARs. The final model of μ j was fitted as a linear combination of these variables (γ r set to 1) and considering the associated coefficient β r of each variable on badger relative abundance (estimates and 95% credible intervals in Table 2).
The estimated index of badger relative abundance a bj (rescaled between 0 and 1) confirmed the presence of the species throughout France, with locally varying abundance between SARs

| Nonmountainous SARs group (<400 m)
Among the relevant environmental variables used in our model, those related to potential food resources seemed to be the most important variables supporting higher badger abundance in nonmountainous SARs. Badgers have a broad omnivorous diet, although earthworms are often described as the preferred food resource (Kruuk & Parish, 1982). The results confirmed that landscapes at low elevations with a high abundance of earthworms tended to host higher badger abundance, as well as SARs showing a high proportion of permanent pastures, which are often described as foraging areas for badgers (Hammond et al., 2001;Reid et al., 2012). Permanent pastures are moreover often presented as areas with a high biomass of earthworms (Da Silva et al., 1993). The presence of badgers in lowland areas of France is also favored by arable lands supporting high trophic resources such as orchards and vineyards, which is consistent with previous studies on badgers (Barea-Azcón et al., 2010;Rosalino, Macdonald, & Santos-Reis, 2005). By contrast, the presence of maize fields did not increase (or decrease) the abundance of badgers at the large spatial scale of France, although badgers can use this food resource in their diet (Lanszki & Heltai, 2011;Moore et al., 1999), especially during late summer when food resources are more scarce.
F I G U R E 1 Number of detected badgers (Meles meles) in each small agricultural region (SAR) of France from 2006 to 2009 (the inset contains the legend). Bold black lines surround SARs above 400 m mean elevation, representing the limit between mountainous and nonmountainous SARs groups High badger abundance was also associated with less fragmented SARs. Highly fragmented landscapes (i.e., characterized by high edge density values) reduce the availability of suitable sett sites for badgers through loss of forest cover and increase isolation (Virgós, 2001). Badgers may be vulnerable to habitat loss and population fragmentation, mainly because of low dispersal rate (Woodroffe, Macdonald, & Da Silva, 1995). In the northern central part of France, high edge density seems to induce low relative abundance of badgers in these SARs (a bj < 0.3). However, in Brittany, despite high edge density values, SARs support high badger relative abundance (a bj > 0.6).
The presence of arable lands and a high percentage of pastures with high-predicted earthworm abundance seem to compensate for habitat fragmentation. Results suggest that the highest abundance of badgers in lowlands occurs in favorable connected landscapes, with a variety of trophic resources, suitable for this opportunistic feeder.

| Mountainous SARs group (>400 m)
The fine soil texture and continental climate (characterized by moderate precipitation trends, concentrated mostly in the warmer months) appeared as the main environmental drivers to increase badger abundance in this SARs group. These two variables seem linked to a gradient of elevation, from favorable lowlands composed of finer soil texture (allowing digging for sett installation; Obidziński, Pabjanek, & Mędrzycki, 2013)  low winter temperatures). As explained before, the SAR elevation variable was not included in our Bayesian model due to the high correlation between environmental variables and elevation (e.g., terrain ruggedness and edge density). Although SARs have been separated into two distinct groups (i.e., <400 and >400 m), the effect of elevation still seems important here and confirmed that elevation was the main driver of badger abundance for highest elevation SARs.
While biotic factors (such as earthworm biomass or the presence of fruit crops) drove badger abundance in lowland areas, lower variation in-and lower levels of-food supply in mountainous areas may explain why these biotic factors do not influence badger abundance variation at high elevations. Conversely, climate and soil structure played a preponderant role in shaping badger abundance above 400 m altitude, probably by imposing constraints on sett locations.

| SARs scale effect and model accuracy
Because the data were collected mostly along roads, the spatial distribution of detections is not independent of roads locations.
However, we do not think that roads locations resulted in a strong bias here. Indeed, on a large scale, the roads are very dense and well distributed throughout France. Moreover, the role of wildlife protection officers is to enforce the law uniformly over their territory (whatever the density of roads), resulting in a sampling effort mostly dependent on the distance travelled by the officers (i.e., our measure of effort) and not on the spatial distribution of the roads. On a small scale, it is important to note that our sampling units (i.e., SARs) are characterized by a homogeneous landscape, tending to limit such bias within a unit. Even though the density of a species is probably different along the road and in the surrounding habitat, the detectability parameters in our model take this difference into account provided that it is constant over SARs (e.g., density along the roads always 50% smaller than in the surrounding habitat).
We assumed that environmental variables, averaged over each SAR, were distributed uniformly over these spatial units in our Unfortunately, to date, most of the environmental data are not available on a yearly basis.

| CON CLUS ION
Different environmental factors drove the badger abundance depending on the elevation. While biotic factors acted in lowland parts of France, abiotic factors were associated with spatial variation in relative abundance in higher elevation areas. This calls for the use of spatial replications in habitat-resource-occurrence or abundance studies over heterogeneous landscapes. Mapping the relative abundance of wildlife species based on environment suitability can be used to develop conservation strategies for endangered populations (Gibson, Wilson, Cahill, & Hill, 2004) or to evaluate risks of a species to become overabundant, especially for species acting as wild reservoir hosts for infectious disease (Bessell, Orton, White, Hutchings, & Kao, 2012). The presence of high abundance of badgers in some parts of France (e.g., mean a bj = 0.72 in Dordogne) can, for example, help to identify areas of potential risk for zoonotic transmissions, such as bovine tuberculosis, which has recently been found in cattle and badgers in France (Payne et al., 2013).
Beyond environmental drivers, interspecific competition can be a major process shaping wildlife communities (Schoener, 1983).
However, interactions within mammal communities have only rarely been studied at large scale, and such an approach using relative abundance may allow us to study the coexistence of species. Interactions between mesocarnivores can influence competitive relationships (Tammeleht & Kuuspu, 2018) or abundance, as shown by the culling of European badgers for disease control in England, which was associated with increases in red fox Vulpes vulpes densities (Trewby et al., 2007). Detections of all mustelid species monitored by the SCSL program could thus allow studying the relationships between the relative abundance of these species in France and interactions within the mustelid community.

ACK N OWLED G M ENTS
We are grateful to all the ONCFS wildlife protection officers who have participated in data collection for the "Small Carnivorous Species Logbook" program in France since 2001. The French National Hunting and Wildlife Agency (i.e., ONCFS) funded this study; no grants were received.

CO N FLI C T O F I NTE R E S T
The authors declare that they have no conflict of interest.

AUTH O R CO NTR I B UTI O N S
SD and SR designed the study. MJ and CC conducted the statistical analyses. MJ drafted the manuscript. MJ, CC, LS, SR, and SD critically revised the manuscript. All authors approved final manuscript for submission.

PER M ITS
The fieldwork was made by 1,500 qualified officers working for the ONCFS, commissioned by the French Minister of Ecology, which allowed them to manipulate dead animals, according to French legislation. Data collection for the SCSL program was a part of their job, so their written or verbal consent for this study was not required. The director of the ONCFS approved the SCSL program.

DATA AVA I L A B I L I T Y S TAT E M E N T
The data that support the findings of this study (i.e., animal detections and kilometers travelled with cars) are available in a public accessible data repository: Dryad: https ://doi.org/10.5061/dryad.8pk0p 2nhz.