Ecology rather than people restrict gene flow in Okavango‐Kalahari lions
Abstract
Reduced gene exchange between animal populations may be an indicator of the effects of anthropogenic fragmentation or it may reflect natural gradients in the landscape that can also result in population fragmentation. It can be difficult, therefore, to disentangle the role of local ecology from anthropogenic factors, creating a risk of attributing a lack of gene flow as being due to human activities, leading to ill‐informed management decisions. Here, we test the ecological and anthropogenic factors driving population differentiation and show how the relative influence of such effects can be identified. Using Bayesian clustering and a causal modelling approach, we combine genetic and remote sensing data to disentangle the confounding influences of ecological and anthropogenic fragmentation. We investigate a region where such confusion may arise; in and around the Okavango Delta in northern Botswana. Specifically, we used 20 microsatellites to investigate the genetic structuring of African lions Panthera leo occupying a landscape dominated by two very different environments, the wetland Okavango and the surrounding Kalahari Desert. We find that differences in ecology, rather than anthropogenic barriers, are driving genetic differences in the population and that despite their ability to disperse long distances these lion populations are differentiated into two distinct genetic groups, one inhabiting the wetland Okavango Delta and the other one inhabiting the surrounding dryland Kalahari, divided by an apparently unobstructed boundary. The genetic structure observed could easily have been misinterpreted as a response to anthropogenic disturbance reducing gene flow. This reinforces the need to consider non‐anthropogenic hypotheses, such as ecological differences between habitats, when assessing possible mechanisms of gene flow and their implications for population management. As anthropogenic pressure increases in this region, we recommend conservation managers consider the Okavango population as a separate conservation unit, but also recognize the importance of maintaining the current structural landscape connectivity.
Introduction
The exchange of genes between populations is essential to maintain genetic diversity and reduce the risk of inbreeding depression (Keller & Waller, 2002; Björklund, 2003; Spielman, 2004). From an evolutionary perspective, however, the opposing force of local adaptation to contrasting environments may result in naturally reduced gene exchange between populations (Aitken & Whitlock, 2013; Arnegard et al., 2014). Given the emphasis in recent literature for the prevalence of isolation by ecology (IBE – “The correlation between neutral genetic differentiation and environmental or phenotypic divergence among populations”), as opposed to isolation by distance (IBD – “The correlation between neutral genetic differentiation and geographic distance; Shafer & Wolf, 2013; Bradburd, Ralph & Coop, 2013; Sexton, Hangartner & Hoffmann, 2014), it is reasonable to assume that even relatively subtle transition zones in an environment may lead to reduced gene flow (Doebeli & Dieckmann, 2003). These implications for gene flow may have ramifications for the management of species at a landscape scale. While habitat fragmentation has been widely identified as a threat to wild populations (Fischer & Lindenmayer, 2007) the risk of identifying a population as being anthropogenically fragmented, when it is in fact naturally fragmented by local ecology, could easily result in ill‐informed management decisions or unnecessarily prevent local land use (Stockwell, Hendry & Kinnison, 2003). In landscapes unaffected by human activities with no obvious geographic barriers to dispersal, evidence of reduced gene exchange strongly supports an argument for natural fragmentation and possible local adaptation (Rueness et al., 2003). However, when humans actively change environments, it can become much more difficult to disentangle the role of local ecology from that of anthropogenic habitat fragmentation (Ewers & Didham, 2006), and in turn the relative need for conservation intervention.
A number of studies have reported reduced gene flow and genetic isolation among felids and other carnivores resulting from urbanization and other types of anthropogenic fragmentation; these include tigers Panthera tigris Linnaeus (Kenney et al., 2014), Atlantic forest jaguars Panthera onca Linnaeus (Haag, Santos & Sana, 2010), mountain lions (Puma concolor Linnaeus Ernest et al., 2014; Riley et al., 2014) and bobcats Lynx rufus Schreber (Lee et al., 2012). Similarly, loss of genetic diversity has been correlated with a reduction in populations and human population expansion in both ocelots Leopardus pardalis Linnaeus (Janecka et al., 2014) and African wild dogs Lyacon pictus Temminck (Marsden et al., 2012). In many of these studies anthropogenic drivers are clearly the primary isolating factor, however, few studies explicitly test for alternative, non‐anthropogenic hypotheses. One recent exception is a comparison of dispersal and gene flow in pine marten Martes martes Linnaeus (Ruiz‐González et al., 2014), which uses resistance surfaces to depict the unit cost to an individual of moving through a landscape, allowing an assessment of which elements of the landscape determine rates of gene flow between regions.
Here, we test the ecological and anthropogenic factors driving population differentiation in one such carnivore, the African lion Panthera leo Linnaeus. Lions are large, charismatic and highly mobile predators, displaying a broad habitat tolerance with populations persisting in virtually all habitats across Africa, with the exception of the rainforests and the interior of the Sahara Desert (West & Packer, 2013). Lions are now believed to survive in <17% of the area they occupied 150 years ago (Ray et al., 2005; Barnett et al., 2009, 2014). This recent and rapid decline (Riggio et al., 2013) has meant that there is a strong interest in the most effective management strategies to ensure the species long‐term persistence in the face of increased habitat encroachment. Despite numerous long‐term studies on these predators (Schaller, 1972; Radloff & Toit, 2004; Packer et al., 2005), there are still large areas of lion habitat that are poorly known in terms of lion ecology, such as Okavango Delta (but see, Hemson et al., 2009; Cozzi et al., 2013; Winterbach et al., 2014). This leaves many broader‐scale factors largely unanalysed and the intrinsic effects of varying environmental factors in need of further research (Celesia, 2010). The result is a lack of information regarding the impact of environmental vs. anthropogenic factors on population structure.
The Okavango Delta landscape is a multi‐island, wetland wildlife management area of approximately 21 000 km2 with dramatic seasonal and episodic flooding. It is surrounded by Kalahari desert which, following rains, provides highly nutritious seasonal grazing for migrating herbivores, but throughout much of the rest of the year has a far sparser density of potential prey for lions (Bartlam‐Brooks, Bonyongo & Harris, 2011; Bartlam‐Brooks et al., 2013). Only four lion conservation units are estimated to hold more than 2000 lions and the Okavango region forms a part of one of these (Morandin et al., 2014). Development, settlement and agriculture bisect the Okavango/Kalahari region with the Okavango Delta in the north, the Central Kalahari Game Reserve (CKGR) to the south and human settlement in the middle. A number of veterinary cordon fences, designed to prevent the movement of wild ungulates and associated zoonotic diseases, also divide the area (Fig. 1; Mbaiwa & Mbaiwa, 2006; Hemson et al., 2009; Winterbach et al., 2013). Both fencing and human settlement may restrict lion movement through the intervening habitat (Kesch, Bauer & Loveridge, 2015; Oriol‐cotterill et al., 2015), resulting in reduced gene flow, and has led to the assumption that the population is likely to have been similarly fragmented. In contrast, the East of the Okavango has little or no substantial human presence and thus there is clear potential for lion movement between this and the surrounding dry Kalahari habitat. A recent commentary based on mitochondrial DNA has suggested that the lions of the Okavango wetland are characterized by a different habitat specialization to those found in dryer savannah habitat (Moore et al., 2015).

To disentangle the role of human and non‐human impacts on the lion population, we first examine the population for the presence of landscape‐level genetic structuring. We then use this landscape to test the competing hypotheses that any genetic structure observed is primarily driven by impediments to gene flow caused by human influences upon the landscape, non‐human ecological factors or a combination of the two. Finally, we discuss the implication for regional conservation management of the species.
Materials and methods
Study sites
Our study centres on the lion population of the Okavango region of northern Botswana (23°06'–13'E, 19°30'–32'S). This highly heterogeneous wetland landscape (Ringrose, Vanderpost & Matheson, 2003) incorporates the Moremi Game Reserve, as well as the surrounding regions of the Chobe National Park, Central Kalahari Game Reserve, Makgadikgadi Pans National Park and Nxai Pans National Park (Fig. 1). The broader study area encompasses a region of approximately 160 000 km2 but is generally of much drier Kalahari sandvelt, mopane woodland and expansive saltpans.
Sampling
Blood and tissue samples were collected from a total of 149 wild lions from across the study region from 2010 until 2013; representing approximately 10% of the local population (Bauer et al., 2015). Samples were collected from seven broad sampling locations (Fig. 1) identified as areas that well represented the diversity of the study area, encompassed as broad an area of gene flow as possible, were known or suspected to contain lions and for which access could be granted. Sampling represented multiple prides across each area. Furthermore, the locations of the seven areas were targeted so individual pathways between and within them bisected as many potential barriers to gene flow as possible. Samples included 113 fresh tissue samples collected using a remote biopsy dart delivery system (Karesh, Smith & Frazier‐Taylor, 1987), blood (n = 23) and dry tissue from trophy hunted animals (n = 13). Trophy hunted samples originate from outside the sampling areas.
DNA extraction and genotyping
Genomic DNA was extracted from each sample using approximately 25 mg of tissue or 100 μL of raw blood using DNeasy® Blood and Tissue kits (Qiagen) according to the manufacturer's instructions. Twenty microsatellite loci previously identified and amplified in the domestic cat Felis catus (FCA1, FCA6, FCA8, FCA31, FCA45, FCA69, FCA75, FCA77, FCA96, FCA97, F115, FCA126, FCA129, FCA133, FCA193, FCA205, FCA224; Menotti‐Raymond, David & Lyons, 1999) and subsequently successfully used in lions (Driscoll, 1992; Spong & Creel, 2001; Driscoll et al., 2002; Dubach et al., 2013; Lyke, Dubach & Briggs, 2013) were used in this study. These loci were selected based on previous high polymorphism in lions. To rule out genotyping errors each sample was amplified and genotyped at least three times independently, following suggestions by Bonin et al. (2004). Two DNA samples of previously identified genotypes were included on each PCR as positive controls and to ensure consistency of allele size scoring across runs. Amplification methods and multiplex combinations are presented in Appendix S1. The allele sizes and genotypes were scored in genemapper v4.1 (Applied Biosystems). Each sample and locus was checked for null alleles and possible scoring errors using microchecker 2.2.3 (Van Oosterhout et al. 2004).
Analysis of genetic diversity and structure
Genetic diversity was measured by the number of alleles per locus (A), allelic richness (AR), the inbreeding coefficient (FIS), observed heterozygosity (HO) and expected heterozygosity (HE) using, fstat 2.9.3.2 (Goudet, 2001).
A Bayesian maximum likelihood approach was performed in the software structure 2.2.3 (Pritchard, Stephens & Donnelly, 2000) to infer the number of populations (K) derived from the individual microsatellite genotype data. The programme was run without information on the sampling locations, allowing for admixture and correlated allele frequencies among populations. K was tested from 1 to 10, with 10 independent runs each, 500 000 iterations and a burn in of 50 000 iterations. The optimal value of K was selected using the ∆K approach (Evanno, Regnaut & Goudet, 2005) implemented in structure harvester 0.6.94 (Earl & von Holdt, 2012). Because of limitations with the Evanno method (Pritchard et al., 2000), it is unable to detect if K = 1. Therefore, to confirm that a result of K = 2 is not a false positive, we checked that individual membership of the clusters was not distributed randomly across the landscape with no true genetic structure. To detect the presence of nested genetic structuring, particularly important when K = 2, each identified cluster was re‐run in structure (Evanno et al., 2005).
Genetic differentiation and gene flow
The differences in allele frequencies between the population clusters identified in structure were assessed using pairwise Jost's DEST (Jost, 2008), calculated in the software smogd 1.2.5 (Crawford, 2010). Because of its fundamental theoretical importance and its usefulness for comparative purposes, we also report FST (Neigel, 2002). Calculations of FST were performed in fstat 2.9.3.2, with significance assessed using 10 000 permutations. Genetic diversity for each cluster (FIS, HO, HE, AR) was also measured in the same manner as the population‐wide assessment. The number of alleles (NA) and private alleles (PA) in each cluster was calculated using convert 1.31 (Glaubitz, 2004). The importance of demographic factors over mutations was assessed by testing the correlation of GST and Nei's marker diversity parameter (HS) using the software CoDiDi 1.0 (Wang, 2015).
To detect gene flow between patches, a Baysean MCMC (Markov chain Monte Carlo) algorithm was used in BayesAss 3.0 (Wilson & Rannala 2003). This estimates the percentage of individuals from each population found in all others and infers levels of recent population movement between each and every sampling groups and thus determines routes of significant gene flow.
Construction of landscape resistance models
We constructed a suite of resistance surfaces, representing ecological and anthropogenic factors, to be incorporated into a reciprocal causal modelling framework. Ecological factors include wetland topography, natural prey availability and the ecological contrast between the environments of the Okavango wetland and the Kalahari‐Linyanti‐Makgadikgadi dryland complex. Anthropogenic factors examined were the presence of veterinary cordon fences, and human settlement and associated livestock grazing. In addition, we controlled for the influence of geographic distance between individuals.
In the Okavango Delta region the most dominant feature structuring the landscape is water. Previous research (Cozzi et al., 2013) has shown a negative relationship between the water level of the Okavango and a carnivore's ability to cross its channels. To create a map of water extent across the Okavango, where much of the region is inaccessible, we used a Normalised Differential Water Index (NDWI) derived from the Moderate Resolution Imaging Spectrometer (MODIS) on board NASA's Terra satellite. The seasonal and episodic fluctuations of the Okavango Delta's floodwaters result in vastly contrasting wet and dry periods (Andersson et al., 2003). Acknowledging this and knowing a small number of dispersal events can have a strong influence on gene flow (Lowe & Allendorf, 2010), as well as our own observations of swimming lions, we did not want to overestimate the strength of water as a barrier to dispersal. Therefore, satellite data from a ‘low flood’ period (April/May 2005) were chosen for calculating water extent. The resulting NDWI landscape layer assigns water features with a highly positive value, and vegetation and soil with zero or negative values (Xu, 2006). However, due to the nature of the Okavango's wetland landscape, many areas that may appear from satellite‐derived data as vegetation are in fact floating vegetation or waterlogged semi‐permanent swampland. For this reason a cut‐off value of ≥‐0.43 was used to delineate water areas based upon ground‐truthing against known water extent. The resulting map provided a binary landscape of areas containing cells with water and cells without. To test whether resistance is non‐linear the layer was transformed to produce multiple layers conforming to linear, exponential and cubic increases.
Much of the Kalahari region consists of a relatively structurally homogeneous landscape, however, prey density fluctuates across the wider landscape, leading to fluctuations in the amount of resources available to predators and this is likely to be a limiting factor for dispersal (Winterbach et al., 2014). For example, the Okavango has much higher prey densities than the Kalahari. Over such vast areas it is prohibitively restrictive to obtain accurate wildlife densities; however, past studies have identified correlations between primary productivity calculated via remote sensing and herbivore density (Bommel et al., 2005; Pettorelli et al., 2009). Therefore, as a proxy for potential prey density we used methods described by Pettorelli (2013) to calculate yearly integrated normalized differential vegetation index (INDVI) across the entire study region from NASA's MODIS data. The resulting map provided a resistance surface of between 0 and 1, where higher values indicate higher primary productivity and thus potential prey density. Herbivore biomass and lion density (Celesia et al., 2010), as well as herbivore biomass and NDVI (Pettorelli et al., 2009), have been shown to have a linear relationship, therefore the functional shape was considered linear.
As livestock grazing is the dominant land use of the region and the one with highest levels of predator conflict, resulting in persecution of lions found in these areas by cattle ranchers, we accounted for human land use by using the distribution of populated areas and associated livestock grazing extracted from Winterbach et al. (2014) and Elliot et al. (2014). This allowed us to produce a binary resistance map where areas of grazing were coded to resist gene flow.
The cells within each of the three satellite‐derived maps (NDWI, INDVI and human land use) had a resolution of 1150 cells per square kilometre. Each layer was aggregated by a factor of 10 to reduce the variance around clusters of cells within each layer. In the case of the binary layers (e.g. water and livestock grazing) aggregating produced a resistance layer with values between 0 and 1, with a lower value for those regions with less consistent cover, and higher values for continuous cover. To account for our null hypothesis of isolation by distance, a homogenous resistance map was also created where all cells were equal to 1.
The layers were tested by multiplying with resistance values ranging from 2 to 100 (Appendix S2) and this value added to the homogeneous resistance map to enable testing of how changes to resistance affect the detectability of a genetic correlation, with preferential areas for dispersal assigned a value of 1. In the case of human land use, where it may be possible for lions to derive an unintended benefit from utilizing free‐ranging cattle as easy prey, values of <1 were tested to allowing the possibility of a reduced resistance across those cells. A pairwise cost–distance matrix between every possible pair of lions was then calculated for each of the above resistance surfaces using an eight neighbour connection scheme, in the gdistance package addition in r (R Core Team, 2014; van Etten, 2015).
The resistance enforced by veterinary cordon fencing utilized a resistance surface at the same resolution as the other surfaces. Due to the linear nature of this variable conferring a limited cumulative effect, larger resistance values were investigated, from 2 to 250 000. As fences needed to be converted from a linear vector to a cellular raster layer, the connection scheme was reduced to four neighbours to remove the risk of ‘free’ movement between diagonal cells.
To test how lion gene flow may be restricted by a barrier effect caused by ecological differences in habitat between the wetland and the dryland regions, we constructed a binary matrix. Any lion within 25 km of the Okavango water was considered a ‘wetland’ lion and all others were considered ‘dryland’. Pairwise differences in membership to either the wetland or the dryland were given a score of 1 representing a barrier between individuals, and similarities scored as 0 representing no barrier. Due to the likelihood that any barrier to gene flow between areas may be semi‐permeable, we created further resistance surfaces based on the above binary matrix. Wherever a difference was identified between individual pairs, a resistance cost, ranging from 2 to 250 000 (Appendix S2), was added as an additional to the simple IBD surface, in the same way as for the other landscape resistance costs.
Reciprocal causal modelling framework
We employed a reciprocal causal modelling approach (Shirk et al., 2010; Cushman et al., 2013; Castillo et al., 2014; Ruiz‐Gonzalez et al., 2015), a refinement of causal modelling first proposed by Cushman et al. (2006), to evaluate the null hypothesis of isolation by distance as well as the competing hypotheses that lion gene flow is restricted by (1) movement resistance due to wetland topography; (2) movement resistance due to likely prey density; (3) a resistance effect resulting from the wetland/dryland divide; (4) resistance effect from veterinary fence and (5) movement resistance due to human settlement and associated livestock grazing. Each of these was compared against pairwise Nei's genetic distance as the dependent variable, as calculated in GenAlEx 6.5 (Peakall & Smouse, 2012), using a multivariate optimization approach (Castillo et al., 2014).
Each resistance surface was evaluated by assessing the correlation between individual pairwise genetic distance and pairwise resistance distance. Correlation was calculated using Mantel testing with 10 000 iterations implemented using the Ecodist package in r (Goslee & Urban, 2007).
To select the most appropriate resistance value for each landscape variable (Appendix S2), we identified the resistance surface providing the peak of support in partial Mantel r‐values while controlling for IBD (Shirk et al., 2010). When a plateau, not a peak, was identified the resistance value associated with the start of the plateau was chosen. Each of these resistance surfaces was then evaluated against genetic distance while accounting for each of the other resistance surfaces, and not just IBD. This causal modelling criteria was implemented in the same way as Castillo et al. (2014), so that;

Any resistance surfaces that had a positive and significant correlation with genetic distance was included in a multivariate model resistance surface. This involved summing the optimized resistance surfaces to be included. To account for interactions between variables, the parameter value of each variable was altered using the same value as for the univariate models (Appendix S2), while holding all others constant, until the peak of support for Mantel r‐values was identified (Shirk et al., 2010). Each variable was optimized in order of decreasing correlation identified by univariate optimization. For each additional variable, following optimization, the previously established variables were re‐optimized to ensure no change. The process was repeated until the peak of support for the model did not change and we had identified the best‐supported multivariate model.
To be accepted the multivariate model needed to pass the causal modelling criteria outlined above where the model with positive RS in every comparison represents the best candidate model (Cushman et al., 2013). In addition, partial Mantel testing of genetic distance against; i) the organizational model, while partialling out IBD, must be significant; ii) IBD, while partialling out the organizational model must be non‐significant.
Results
Genetic diversity
Genetic data were obtained from 149 individual lions. No genotyping error was detected by microchecker 2.2.3 (Van Oosterhout et al. 2004). The number of alleles per locus ranged from 5 to 12, with a mean of 7.3 per locus. The lion population sampled showed a similar mean observed (HO) and expected heterozygosity (HE) of 0.65 and 0.62 respectively (Appendix S3). Genetic diversity statistics for each sampling location are reported in Appendix S4. A slightly negative mean inbreeding coefficient (FIS) of −0.019 (−0.054 to 0.021 at the 99% confidence interval) indicated that there was no inbreeding across the population. One locus (Fca6) deviated significantly from Hardy–Weinberg equilibrium (P < 0.05), with CoDiDi (Wang, 2015) indicating population substructure and gene flow rather than mutation was likely to account for the observed deviations. Removing this locus did not alter the patterns of genetic structure or differentiation observed, however, results are reported for the reduced dataset.
Genetic structure
The results of structure analysis indicated that the lions are genetically partitioned into two distinct clusters across this landscape (K = 2, Appendix S5). In total 83% of individuals were assigned (Q > 0.80) to one of the two population clusters. The remainder were assigned based on their majority membership. This is visualized with a plot of percentage membership of each sample site to the two clusters identified (Fig. 2). Geographically, this broadly identified a wetland and a dryland cluster (Fig. 3). The wetland cluster is represented by the lions in the Okavango area, which are largely separate from the remaining dryland Kalahari lions. Sampling sites covered large heterogeneous areas and may include a landscape representing both the wetland and dryland types, meaning sampling areas should not be categorized as wetland or dryland, rather individual lions may belong to one of these two categories. Re‐running structure within the wetland cluster only identified weak hierarchical substructure overshadowed by the initial wetland–dryland observation. The structure result is further tested by forming the first resistance surface model within the causal modelling framework; a resistance effect resulting from the wetland/dryland divide restricting gene flow.


Genetic differentiation and gene flow
Pairwise DEST and FST between the wetland and dryland structure clusters was 0.074 and 0.058 (P = 0.01) respectively. The dryland group had far more private alleles (PA = 26) than the wetland group (PA = 7; Table 1). There is no evidence for significant inbreeding or outbreeding as measured by FIS in either cluster (Table 1).
| Pop. | N | F IS | H o | H E | NA | AR | PA |
|---|---|---|---|---|---|---|---|
| Wetland | 75 | −0.003 (−0.043 to 0.025) | 0.6286 (0.17) | 0.6307 (0.17) | 5.85 (1.8) | 5.8 (1.8) | 7 |
| Dryland | 74 | 0.023 (−0.023 to 0.053) | 0.6958 (0.12) | 0.6802 (0.12) | 6.8 (2.1) | 6.7 (2.1) | 26 |
- N, number of individuals; FIS, inbreeding coefficient; Ho, observed heterozygosity; He, expected heterozygosity; NA, mean number of alleles; AR, mean allele richness; PA, total number of private alleles. 95% confidence intervals or standard deviation in parenthesis when appropriate.
Migrant detection in BayesAss indicated a number of sampling groups with a small but statistically significant fraction of individuals more likely to have originated or be the offspring of lions from different sampling group to the one in which they were found (Table 2). In all cases, the direction of migration appears to be in an outward direction from the central Okavango Delta (Fig. 4). It is notable that there is still some dispersal from sampling groups on the periphery of the wetland region into dryland sampling groups (e.g. group IV to group V).
| Source | ||||||||
|---|---|---|---|---|---|---|---|---|
| I | II | III | IV | V | VI | VII | ||
| Destination | I | 0.8926 (0.0236) | 0.0076 (0.0075) | 0.0076 (0.0071) | 0.0073 (0.0071) | 0.0074 (0.0072) | 0.0071 (0.0069) | 0.0070 (0.0068) |
| II | 0.1668 (0.0517) | 0.6996 (0.0352) | 0.0163 (0.0154) | 0.0153 (0.0145) | 0.0147 (0.0138) | 0.0139 (0.0133) | 0.0147 (0.0144) | |
| III | 0.1660 (0.0469) | 0.0161 (0.0161) | 0.7016 (0.0260) | 0.0130 (0.0120) | 0.0131 (0.0125) | 0.0120 (0.0115) | 0.0119 (0.0117) | |
| IV | 0.0546 (0.0344) | 0.0164 (0.0160) | 0.0566 (0.0315) | 0.7472 (0.0330) | 0.0211 (0.0200) | 0.0160 (0.0165) | 0.0145 (0.0143) | |
| V | 0.0370 (0.0227) | 0.0121 (0.0114) | 0.0545 (0.0377) | 0.0627 (0.0361) | 0.7494 (0.0292) | 0.0177 (0.0162) | 0.0128 (0.0123) | |
| VI | 0.0148 (0.0115) | 0.0104 (0.0094) | 0.0098 (0.0094) | 0.0121 (0.0126) | 0.0553 (0.0215) | 0.8534 (0.0265) | 0.0083 (0.0080) | |
| VII | 0.0349 (0.0304) | 0.0186 (0.0179) | 0.0192 (0.0182) | 0.0372 (0.0311) | 0.0201 (0.0201) | 0.1059 (0.0459) | 0.6924 (0.0223) | |

Reciprocal causal modelling
After selecting resistance surfaces with peak support for Mantel r (Appendix S6) and controlling for isolation by distance, partial Mantel tests revealed just two resistance surfaces with a statistically significant partial Mantel r (P < 0.001) when controlling for all other variables (Table 3). These were a linear movement resistance due to wetland topography and a resistance effect from moving between the wetland dominated environment and the surrounding dryland regions. Fences also appear to be a statistically significant factor, but only when controlling for isolation by distance and the impact of human usage of the landscape. When controlling for all other variables, movement resistance due to wetland topography represented the best candidate model. Optimizing each resistance surface with in the multivariate model resulted in the optimal fence resistance value dropping to 0.0125. As this resistance acts on a maximum of two fenced cells between any lion pairing, it resulted in an undetectable change to any multivariate model which incorporated fences, and thus they were excluded. When the remaining two resistance surfaces were incorporated into a single multivariate resistance model, they performed no better than wetland topography alone, with only wetland topography having positive RS in every comparison (Table 4). Furthermore, when controlling for the resistance of moving between the wetland environment and the dryland regions the multivariate model became non‐significant.
| Landscape variable | Partialled out variable |
Peak partial Mantel r |
P |
|---|---|---|---|
| Water resistance | Isolation by distance | 0.13 | 0.0002 |
| Resistance value = 10 | Vegetation resistance | 0.11 | 0.0006 |
| Veterinary fence resistance barrier | 0.12 | 0.0010 | |
| Dryland/wetland resistance barrier | 0.19 | 0.0010 | |
| Human land use | 0.12 | 0.0004 | |
| Dryland/wetland resistance barrier | Isolation by distance | 0.12 | 0.0001 |
| Resistance value = 100 000 a Dryland/wetland resistance value is 3 orders of magnitude higher than the other values because non‐barrier values are cumulative. | Vegetation resistance | 0.11 | 0.0001 |
| Water resistance | 0.11 | 0.0001 | |
| Veterinary fence resistance barrier | 0.07 | 0.0005 | |
| Human land use | 0.09 | 0.0001 | |
| Vegetation resistance | Isolation by distance | 0.07 | 0.0411 |
| Resistance value = 50 | Water resistance | 0.01 | 0.3585 |
| Veterinary fence resistance barrier | 0.03 | 0.2332 | |
| Dryland/wetland resistance barrier | 0.05 | 0.0689 | |
| Human land use | 0.05 | 0.1113 | |
| Veterinary fence resistance barrier | Isolation by distance | 0.07 | 0.0011 |
| Resistance value = 0.75 | Vegetation resistance | 0.06 | 0.0409 |
| Water resistance | −0.06 | 0.0162 | |
| Dryland/wetland resistance barrier | −0.06 | 0.0143 | |
| Human land use | −0.07 | 0.0020 | |
| Human land use | Isolation by distance | 0.04 | 0.0924 |
| Resistance value = 0.0625 | Vegetation resistance | 0.01 | 0.4132 |
| Water resistance | 0.03 | 0.1590 | |
| Dryland/wetland resistance barrier | 0.04 | 0.0958 | |
| Veterinary fence resistance barrier | −0.003 | 0.5472 | |
| Isolation by distance (IBD) | Vegetation resistance | −0.02 | 0.7436 |
| Resistance value = N/A | Water resistance | 0.09 | 0.2429 |
| Veterinary fence resistance barrier | 0.05 | 0.0381 | |
| Dryland/wetland resistance barrier | 0.03 | 0.1702 | |
| Human land use | −0.01 | 0.5954 |
- a Dryland/wetland resistance value is 3 orders of magnitude higher than the other values because non‐barrier values are cumulative.
| Water resistance | Dryland/wetland resistance barrier | Multivariate model | |
|---|---|---|---|
| Water resistance | – | 0.07205735 | 0.06983185a Non‐significant P‐value for correlations between genetic distance and the model in the row partialling out the model in the column. |
| Dryland/wetland resistance barrier | −0.0720574 | – | −0.0993351 |
| Multivariate model | −0.0698319 | −0.02262906a Non‐significant P‐value for correlations between genetic distance and the model in the row partialling out the model in the column. | – |
- a Non‐significant P‐value for correlations between genetic distance and the model in the row partialling out the model in the column.
Discussion
Our study reveals that ecological rather than anthropogenic factors are restricting gene flow between lion populations, and this has divided the lions of northern Botswana into two genetically distinct clusters. This is initially suggested through structure analysis, and is then independently verified by incorporating this as a hypothesis into the organizational modelling framework. The population is separated into a dryland ‘Kalahari type’ and a wetland ‘Okavango type’, despite no obvious physical barrier to dispersal and gene flow, and is consistent with the observations and hypothesis of Moore et al. (2015). There is still some admixture between these regions, as confirmed by the migrant analysis which suggests limited gene flow persists in an outward direction with the Okavango Delta at the centre. The differentiation of the populations between these regions (FST) is similar to that found between many regional Lion Conservation Units (LCU's; Dubach et al., 2013) and other similarly structured carnivores (Carmichael et al., 2001; Mcrae et al., 2005; Rueness et al., 2014).
Our results also reveal that only a single organizational model is consistent with the pattern of genetic differences observed across the sampled population; the Okavango wetland topography acting to increase resistance to gene flow. The split between the wetland–dryland divide, hypothesized from the structure results, is also positively correlated and significant when controlling for all other variables, however, the wetland topography alone is a better model. Controlling for the other factors through partial Mantel tests indicates that isolation by distance, anthropogenic features including the veterinary fences or variations in primary productivity (INDVI) perform better than an isolation by ecology model. Furthermore, the indication that an invisible barrier to movement is correlated with reduced gene flow between the wetland and dryland areas, even when controlling for the resistance caused by water, supports the broad‐scale structure results of K = 2. The fact that resistance caused by the divide between the wetland and dryland regions does not improve upon the univariate wetland topography model, yet is positively correlated with gene flow, is likely due to the fact that the two variables are nested (Sheppard & McMaster, 2008), with the wetland topography variable not affecting lions outside the wetland region.
While fences appeared weakly correlated with gene flow in the univariate models, the very small resistance value they represent, which reduced further when incorporated into the multivariate model, rules them out as a significant factor contributing to the genetic structure observed. The fences largely follow the outline of the Okavango Delta and are thus likely to be correlated with other landscape variables; however, they do not themselves exert any greater influence than the Okavango topography or the wetland–dryland divide. The permeability of fences to lion movement has previously been documented in the region and may be due to the type and level of maintenance required for fences to be impermeable to lions (Kesch et al., 2015).
Dispersal by habitat generalists such as lions and those organisms with a high plasticity, and thus ability to persist within poor‐quality habitat, can be expected to be less influenced by the relative quality of habitat between patches and the distribution of landscape features (Zalewski et al., 2009) than by the ecology of the prospective destination patch. This is compounded in wide‐ranging animals with an ability to travel long distances with relatively low energetic constraints. Under such circumstances the ability of a disperser to settle and procreate will depend on the individuals degree of suitability or adaptation to a particular habitat (Clobert et al., 2009). This lack of importance for the quality of habitat through which highly mobile organisms are moving is highlighted by our causal models, which suggest that ecological differences are driving genetic differentiation. Similar finding have been reported in other highly mobile carnivores. For example, Rueness et al. (2003) identified a geographically invisible barrier to lynx gene flow coinciding with the ecological Continental and Atlantic regions of North America. Similarly, both Sacks et al. (2004) and Pilot et al. (2006) noted genetic subdivisions were associated with unobstructed boundaries between contiguous habitats, in coyotes Canis latrans and grey wolves Canis lupus. Furthermore, while we did not explicitly investigate selection, differences in the local environment seem likely to increase selective pressures and reinforce divergence. Such habitat‐specific adaptation has been observed previously even in the presence of low‐level immigration (Postma & van Noordwijk, 2005; Sacks et al., 2008).
We did not to find strong evidence that human habitation or other anthropogenic barriers have caused a significant barrier to gene flow in these lions. Despite numerous studies identifying either ecological or anthropogenic factors to be driving genetic differentiation in carnivores (Rueness et al., 2003; Sacks, Brown & Ernest, 2004; Mcrae et al., 2005; Pilot et al., 2006; Muñoz‐Fuentes et al., 2009; Haag et al., 2010; Reddy et al., 2012; Ernest et al., 2014; Dolrenry et al., 2014; McManus et al., 2014), until recently few had explicitly compared the effect of both together (but see Wasserman et al., 2010; Ruiz‐Gonzalez et al., 2015 for examples). Under the context in which we explored this phenomenon, it is clear that ecological factors have played a far more significant role in driving the current genetic landscape of lions than human interventions in the landscape. Despite this, it is important to acknowledge that while anthropogenic fragmentation does not appear to have driven the current pattern of genetic structure, increased development or persecution of lions could still reduce gene flow to a level where it become detrimental to population connectivity. Detecting temporal aspects of population structure and disentangling historic vs. more recent factors is a complex endeavour, with inherent biases and assumptions, and while the current pattern appears not to be associated with recent changes in the landscape, the importance of the anthropogenic landscape become even more important as human densities increase. Such an increase in human populations is likely to lead to an increase in human–wildlife conflict associated with large predators such as lions, thus human development may still become an important driver (Ewers & Didham, 2006).
The partitioning of the wetland lions into a separate genetic cluster, with no obvious restrictions to movement into the dryland areas, strongly suggests the need to consider lions from Okavango Delta as a somewhat isolated subpopulation. However, the gene flow that does exist might suggest that this population could act as a genetic resource for the wider regional lion population. The genetic division of Okavango lions from the wider landscape has considerable management ramifications given the risk of relatively small populations being disproportionately affected by human perturbation (Heller, Okello & Siegismund, 2010; Marsden et al., 2012) and merits further investigation into the Okavango as an ecosystem. As such, we would also suggest that for management purposes Okavango lions should be considered as a largely isolated single conservation unit, while still maintaining natural connectivity. In addition, the genetic divergence across this region, linked to habitat‐conforming genetic structure, suggests conservation measures that aim to protect the full range of genotypes will also preserve evolutionary potential within the species.
Acknowledgements
We gratefully thank the Botswana Department of Wildlife and National Parks for allowing this research (permit number EWT 8/36/4 XIII [35]). We thank Patricia Brekke and the anonymous reviewers whose comments greatly improved this manuscript. We thank Debbie Peak, Rob Jackson, Kyle Burger, Robyn Coetzee & Robert Riggs for contributing samples. We thank Map Ives, Kai Collins, Segametsi Monamorwa, Nic Prost, the management and camp staff of Wilderness Safaris Botswana and Chitabe, Crispin Sanderson, Grant Huskisson, Dane Hawk, Rick Nelson, Erik Verreynne, Alan Wilson, Anna Butterfield & Jaques Van deMerwe of Vision International, Wilton Raats, Chris Kruger of Machaba Camp, Dominik Bauer and Kristina Kesch for logistical and veterinary support. We thank Anton van Shalkwyk, Hanri Ehlers and Kanabo Conservation Link for invaluable support and funding; Wilderness Safari's for logistical support; PneuDart for the donation of a dart projector; Wilderness Wildlife Trust for financial support. SGD was supported by a BBSRC CASE studentship (BB/F017324/1). All relevant import and export permits were granted.
Authors' contributions
S.G.D. conceived the ideas; S.G.D. and G.M; S.G.D. and D.G. processed and analysed the data; C.C., V.S. and D.G. provided critical feedback and helped shape the manuscript; S.G.D. led the writing.
Open Research
Data availability statement
Sampling locations, microsatellite genotypes, distance matrices and resistance surface raster layers are archived at figshare doi: 10.6084/m9.figshare.11523114.




