Introduction

The phylogeography of European mammals has been extensively studied since the 1990s and a broad consensus has been reached regarding the importance of the Mediterranean peninsulas (Iberia, Italy and the Balkans) as refugia during the Last Glacial Maximum (LGM) (Taberlet et al., 1998; Hewitt, 1999). However, many of these findings were based on limited data sets in terms of both the number of samples and molecular markers used, ultimately leading to incomplete postglacial recolonisation scenarios. Increased sampling in particular has led to the identification of important refugia present further north in Europe, such as the Carpathian Basin (Kotlík et al., 2006; McDevitt et al., 2012). Therefore, it is clear that continent-wide sampling—including adequate sampling of putative refugia—is vital for accurate inference of the phylogeographic history of European species. Furthermore, over recent years phylogeographic studies have included data from multiple markers in order to gain new insights into complex colonisation histories (McDevitt et al., 2011) and processes over different time scales (Avise, 2004).

In addition to broad sampling and the use of multiple markers, a number of methodological innovations can also contribute to a better understanding of the factors influencing the broad-scale genetic structure of a species. Recent advances in approximate Bayesian computation (ABC), for example, have paved the way for robust comparison of phylogeographic/colonisation scenarios (Beaumont, 2010). In addition, likelihood-based methods paired with Monte Carlo sampling are becoming increasingly popular tools to estimate past demographic parameters (Girod et al., 2011 and references therein), as standard summary statistic-based bottleneck tests have low power at the typical sample sizes of phylogeographic studies and their results can depend on the choice of mutation model parameters (Peery et al., 2012). The full Bayesian model developed by Beaumont (1999) and Storz and Beaumont (2002), for example, infers posterior probability distributions of population parameters using information from the full microsatellite allelic distribution in a coalescent-based framework under a stepwise mutation model. The method by Storz and Beaumont (2002), implemented in the program MSVAR v.1.3, estimates the posterior distributions of three demographic parameters—the log of the current, log(N0), and ancestral, log(N1), effective population sizes and the log of time since the demographic change, log(ta)—assuming an exponential or linear decrease in a panmictic, isolated population.

The phylogeography of the European badger (Meles meles) is an example where extensive sampling, multiple markers and up-to-date analytical techniques might help to resolve outstanding issues. It is now well established that the European badger is one of the four Eurasian badger species (Del Cerro et al., 2010 and references therein). The broad-scale genetic structure of the European species is of interest, as it may be the result of historic restriction to and expansion from glacial refugia and/or recent impact by extensive anthropogenic interference (Pope et al., 2006). Indeed, badger densities declined in some European countries owing to the large-scale gassing of setts during the rabies control campaigns of the 1960s and 1970s (Griffith and Thomas, 1997). Consequently, a number of studies have attempted to describe the genetic variation of the species in a European context (Marmi et al., 2006; Pope et al., 2006; O'Meara et al., 2012). However, they were essentially based on the same data set of samples from western and north–central Europe, with little to no sampling from eastern Europe and Italy, making it difficult to draw robust conclusions about the location of refugia, patterns of postglacial expansion and recent demography.

Although subfossil evidence suggests that the badger was present in Iberia, Italy, the Carpathians and the Balkans during the LGM (Pazonyi, 2004; Sommer and Benecke, 2004, Marmi et al. (2006) inferred the presence of only one European mitochondrial DNA (mtDNA) clade. However, despite the absence of samples from southeastern Europe, the haplotype network led these authors to suggest recolonisation of the continent from multiple refugia. Working with microsatellites, Pope et al. (2006) interpreted the presence of a broad-scale isolation-by-distance signal between western European populations to result from an expansion out of a single glacial refugium. These authors also found a higher genetic diversity in badgers from western central Europe, compared with populations from the southwest and the north but could not ascribe this pattern with confidence to either historic process or recent demographic declines. Although in principle, MSVAR allows recent and ancient population declines to be distinguished (see, for example, Goossens et al., 2006; Heller et al., 2008), the simulation study by Girod et al. (2011) suggested that both the accuracy and precision of the estimated demographic parameters were greatest for old and severe events. As suggested by Peery et al. (2012), it would be advisable to estimate a method’s power to detect a bottleneck—or assess the likelihood of a false bottleneck signal—given realistic demographic scenarios and the available number of samples and genetic markers.

The overall objective of the current work was to gain a thorough understanding of the factors influencing the broad-scale genetic structure of the European badger. Based on continent-wide sampling, we specifically aimed to infer postglacial colonisation routes of the species in Europe, analysing mtDNA sequence data and microsatellite loci with traditional phylogeographic analyses coupled with ABC. Specifically, we wanted to test whether badgers recolonised Europe out of a single or from multiple refugia. Furthermore, we aimed to use microsatellite data to investigate the demography of European populations. Specifically, we wanted to test whether the reduced genetic diversity in peripheral badger populations was due to either historic or recent processes. Finally, we performed limited simulations to test the power of the MSVAR method to allow robust inference of past demographic events in the specific context of this study.

Materials and methods

Laboratory work

We collected muscle, ear or hair samples from a total of 675 badgers from 21 different European countries (Figure 1). Following geographic criteria, samples from five countries were further subdivided into different sampling locations, giving rise to a total of 30 predefined populations (Supplementary Table S1). Altogether, 177 samples (originating from Finland, Great Britain, Luxembourg, Norway, the Republic of Ireland and Zurich) had already been microsatellite genotyped for the study by Pope et al. (2006). Tissues samples from road-killed or legally hunted individuals were stored in absolute ethanol until extraction. DNA was extracted from tissue samples using an ammonium acetate-based salting-out procedure (Miller et al., 1988) and from hair samples (3 localities, 43 samples) using a Chelex protocol (Walsh et al., 1991). We used primers MelCR1 and MelCR6 (Marmi et al., 2006) to amplify a 594-base pair (bp) fragment of the 5′-end of the control region of 327 individuals (Supplementary Table S1, Supporting Information, Supplementary Appendix S1). Samples were also genotyped using 18 microsatellite markers in three multiplex reactions using the QIAGEN Multiplex kit (QIAGEN, Hilden, Germany; see also Pope et al., 2006). Further information on the multiplex composition and the PCR conditions are given in the Supporting Information, Supplementary Appendix S2. The genetic profiles of all samples consisted of at least 12 loci.

Figure 1
figure 1

Sample distribution and location of the genetic subpopulations inferred using the BAPS algorithm. The size of the pie charts indicates the number of samples collected from a locality, whereas the pattern of the pie chart indicates the identity of the genetic clusters.

Full size image

Data analysis: mtDNA

The number of polymorphic sites, as well as nucleotide and haplotype diversity within Europe overall, were calculated using DnaSP v.5 (Librado and Rosaz, 2009). Expected heterozygosity and haplotype richness were calculated for populations containing at least seven individuals in CONTRIB v.1.02 (Petit et al., 1998). Furthermore, our own data set was supplemented with 82 European individuals from two previous phylogeographic studies of Eurasian badgers (Marmi et al., 2006; Tashima et al., 2011) and a phylogenetic network was constructed using the software NETWORK v.4.6 (www.fluxus-engineering.com) with a median-joining algorithm based on maximum parsimony (Bandelt et al., 1999). Previous studies using multiple markers have demonstrated that European badgers are a monophyletic group (Marmi et al., 2006; Tashima et al., 2011), and under the circumstances of closely related sequences, simulation studies have demonstrated that this method provides reliable estimates of the true genealogy (Cassens et al., 2005). Because of missing data, all the sequences were truncated to 467 bp.

Past demographic expansion in European badgers as a whole was tested using three methods implemented in ARLEQUIN v. 3.5.1.2 (Excoffier and Lischer, 2010): Fu’s (1997) FS statistic, Tajima’s (1989) D statistic and mismatch distributions of pairwise nucleotide differences. Mismatch distributions were calculated and compared with expected values for an expanding population (Rogers and Harpending, 1992) by testing for goodness-of-fit statistics based on the sum of square deviations for a model of sudden expansion. Tests for goodness of fit for all three methods were generated using parametric bootstrapping with 10 000 replicates.

Data analysis: microsatellite loci

We estimated observed (Ho) and unbiased expected (Heu) heterozygosities (Nei, 1978) for the 25 predefined populations with 10 or more sampled individuals (Supplementary Table S1) using GENETIX 4.05.2 (Belkhir, 2004, unpublished software). We tested for the significance of heterozygote deficiency or excess in these 25 populations using the Markov chain method in GENEPOP 3.4 (Raymond and Rousset, 1995), with 10 000 dememorisation steps, 100 batches and 10 000 subsequent iterations. The populations were tested for linkage disequilibria among loci using an exact test based on a Markov chain method as implemented in GENEPOP 3.4. The false discovery rate technique was used to eliminate false assignment of significance by chance (Verhoeven et al., 2005).

We used the STRUCTURE v2.3.1 (Pritchard et al., 2000) and GENELAND v.4.0.3 (Guillot et al., 2005) Bayesian clustering algorithms to analyse the population genetic structure of European badgers. To estimate the number of subpopulations (K), 10 independent runs of K=1–15 were carried out with 106 Markov chain Monte Carlo (MCMC) iterations after a burn-in period of 105 iterations. Further details on the exact parameters used for running the programs can be found in the Supporting Information, Supplementary Appendix S3. We also used BAPS v5.4 (Corander et al., 2004) to cluster the data at group level, as it partitions the sampling units into populations with nonidentical allele frequencies (Corander et al., 2004). The program was run 10 times for each K=15–25, using the 30 sampling locations (Supplementary Table S1) as predefined groups. We used GENETIX v.4.05.2 to perform a factorial correspondence analysis to visualise the genetic distance between the 30 predefined badger populations.

We used program ADZE v1.0 (Szpiech et al., 2008) to calculate the allelic richness (AR) of the 25 predefined populations with a sample size of N⩾10, excluding genetic markers that had been typed at <50% of all individuals in at least one population (loci Mel110 and Mel112). The estimates of allelic richness were based on a sample of 16 individuals in all 25 populations. We used regression analysis in R 2.13.0 (R Development Core Team, 2011) to test for the effect of both longitude and latitude on diversity measures based on microsatellites (Heu and AR) and mitochondrial sequence data (haplotype richness and haplotype diversity). The analysis was performed for all 25 predefined populations, but also, in order to avoid issues of spatial autocorrelation, using only one population—the one with the largest sample size—per BAPS-defined genetic clusters (see Results). Geographic coordinates for each predefined population were obtained by averaging coordinates of individuals.

Phylogeographic reconstruction

The ABC was implemented in DIYABC v1.0.4.46 (Cornuet et al., 2008) to further investigate the dynamics of the recolonisation process of badgers in Europe. This software can produce estimates of the relative likelihood of alternative phylogeographic scenarios in a coalescent framework. Here we used a relatively simplified situation, with three ‘lineages’ of interest: Iberia, Scandinavia and the Balkans. We aimed to clarify the phylogeographic origin of the Scandinavian populations (see Results).

The ABC analysis was performed with the microsatellite data only, as preliminary analyses with a combined mtDNA and microsatellite data set did not yield satisfactory support (admixture between groups was one of the scenarios to be tested). The microsatellite data set consisted of 74 individuals from Spain and Portugal (Iberia), 51 individuals from Norway and Sweden (Scandinavia) and 73 individuals from Serbia, Croatia, Hungary and Bulgaria (southeast Europe). After preliminary analyses, effective population sizes were allowed to vary between 10 and 10 000 for Scandinavia and Iberia, as well as between 10 and 20 000 for the Balkans. In scenario 1, the Scandinavian lineage split from the Iberian lineage at t1 (13 000 years to 1 year before present (BP)) to coincide with the emergence of various landbridges connecting mainland Europe to Scandinavia (Björck, 1995), the first appearance of badger fossils in Scandinavia (Sommer and Benecke, 2004) and to allow this event to have occurred any time since then (Supplementary Figure S1). In scenario 2, the Scandinavian lineage split from the Balkan lineage at t1. In scenario 3, each of the Iberian, Scandinavian and Balkan ‘lineages’ coalesced at t2 (LGM; 19 000–26 000 BP) and had independent histories since. Finally, scenario 4 has the Scandinavian lineage forming as a result of admixture between the Iberian and Balkan lineages at t1 (Supplementary Figure S1).

Simulated data sets were created by requesting a total of 21 summary statistics, including the number of alleles and heterozygosity (per population and per pairs of populations), Garza and Williamson’s M-ratio (per population) and both FST and δμ2 pairwise divergence statistics. One million simulated data sets per scenario were used to produce posterior distributions. Each scenario was considered equally probable and reliability of scenarios was visualised through principal component analysis, whereas posterior probabilities of scenarios were compared by means of logistic regression, using the closest 1% of simulated data sets to the observed data (Cornuet et al., 2008).

Demographic reconstruction

We used the method by Storz and Beaumont (2002), implemented in the program MSVAR 1.3, to infer past population dynamics of the badger populations. MSVAR 1.3 assumes that a stable population of size N1 started to either increase or decrease a time of ta ago to a current population size N0. The loci are assumed to be evolving according to a strict single-step mutation model with a mutation rate μ. Prior distributions for the parameters are assumed to be log-normal. The means and s.d. of these prior log-normal distributions are themselves drawn from prior (or hyperprior) distributions. Hyperpriors for the means were specified by normal distributions with a mean of α and a s.d. of σ. Hyperpriors for the standard deviations were assumed to be zero-truncated normal distributions with a mean of β and a s.d. of τ (see Supplementary Table S2).

We applied MSVAR to the 19 predefined populations with a sample size of N⩾20. Two loci (Mel115 and Mel14) had repeat length variations that were not a consistent multiple of two or four and were therefore excluded from the MSVAR analyses. We modelled an exponential change in population size as well as a generation time of 1 so that the estimate of log(ta) should be indicative of the number of generations since the change in population size. Initially, 12 independent runs were performed using different random seeds, starting values, priors and run lengths (Supplementary Table S2). We discarded the first 10% of each MCMC chain to avoid biases due to starting conditions.

Convergence among chains was tested with the Gelman and Rubin (1992) statistic using the CODA library (Plummer et al., 2006) of the R package. A point estimate of <1.1 is normally taken as an indicator of good convergence (Gelman et al., 2004), with a value of <1.2 sometimes being used as a guideline for approximate convergence (see, for example, Brooks and Gelman, 1998). As MSVAR has been shown to reach convergence with difficulty, especially with recent and severe bottlenecks (Girod et al., 2011), we also considered chains that only converged approximately as giving rise to informative point estimates. If the 12 independent chains did not converge, we ran 2 further chains for a total of 1010 steps (Supplementary Table S2) and, in the case of successful convergence, limited the inference to the 3 longer chains. Because of the computational complexity of MSVAR, running a larger number of longer chains was impractical (see also Girod et al., 2011). Independent runs were pooled into one data set to produce larger samples of the posterior distribution. The marginal posterior distributions of the model parameters were estimated using the R library LOCFIT (Loader, 1999). Point estimates of log(N0), log(N1) and log(ta) were obtained from the mode of their marginal posterior distribution. The 90% highest probability density intervals were obtained with the CODA package.

We performed a limited number of simulations to assess the power of the method by Storz and Beaumont (2002) to correctly infer the past demography of the different badger populations, given the sample sizes of the empirical data sets and the number of genetic markers used. We used the program DIYABC to simulate nine different scenarios of population decline. We fixed the number of simulated microsatellites to 16 (evolving according to a strict stepwise mutation model) and the ancestral effective population size to N1=5000, but varied the current effective population size N0 (50, 150, 1000) and the generation time since the decline ta (25, 100, 500). For each combination of these parameters, we generated five genetic data sets consisting of 20 individuals and five data sets of 50 individuals. We analysed these simulated data using the same methodology as used for the empirical data. If approximate convergence was not achieved for the log(N0) estimate, we simulated and analysed a new data set to ensure a balanced sampling design for the statistical analysis. We performed three-way analysis of variance in R to test for the influence of simulated sample size, current effective population sizes and times since decline on the accuracy of the MSVAR point estimates for log(N0) and log(N1). The difference between the simulated and estimated effective population size was log-transformed to improve normality.

Results

Altogether, 49 haplotypes were observed in the control region data set, with 28 of these newly described in this study (GenBank accession nos.: KJ161328–KJ161355). A total of 27 polymorphic nucleotide sites were found, of which 16 were parsimony informative. Overall nucleotide and haplotype diversities were equal to 0.00591 (s.d.±0.00018) and 0.894 (s.d.±0.01), respectively. The median-joining network did not reveal clear genetic structuring within European badgers (Figure 2), with closely related haplotypes differing only by a single bp (Figure 2). A star-like pattern was found with core haplotypes, meles1–3, distributed throughout eastern and central Europe, the Balkans and Britain and a newly identified haplotype mm20 found in the Balkans, Poland and Estonia. However, a group of haplotypes—mm1–4, mm6–14 and meles15–20—were largely restricted to Iberia (with a few occurrences in Western Europe). Furthermore, haplotypes meles12–14 were almost exclusively Scandinavian (Sweden and Norway). Most Irish individuals also had these ‘Scandinavian’ haplotypes (and mm4). These ‘Scandinavian’ haplotypes formed a third group in the median-joining network that was more closely related to the Iberian than the eastern haplotype groups. ‘Scandinavian’ haplotypes also occurred in Finland, Estonia and central Russia. The entire European data set conformed to a model of demographic expansion (sum of square deviations: 0.0067; P=0.42) and was also consistent with a model of spatial expansion (sum of square deviations: 0.0079; P=0.21). Fu’s (1997) FS value was significant (−26.08331; P<0.001) but Tajima’s (1989) D was not (−0.88218; P=0.19). The mismatch distribution revealed that there was evidence of a bimodal distribution of pairwise differences (Supplementary Figure S2).

Figure 2
figure 2

Median-joining network (a) and geographic distribution of 49 badger mitochondrial control region haplotypes (b). The same colours represent the same groups of haplotypes in both figures. Missing haplotypes are indicated by a small black square in the network. Horizontal bars represent mutational steps when greater than one. The sizes of the symbols in the network and the map are representative of haplotype frequency and sample sizes, respectively.

Full size image

Although the microsatellite Heu of the 25 predefined populations with N⩾10 was almost always larger than Ho (Supplementary Table S1), no locus systematically deviated from Hardy–Weinberg equilibrium after correcting for multiple tests (Supplementary Table S3). Three different pairs of loci were in linkage disequilibrium in three predefined populations after correcting for multiple tests. All loci were therefore included in the analyses. The results of the three clustering methods did not converge on one optimal solution. In the case of STRUCTURE, K=10 was the highest K at which the log-likelihood values converged reasonably well (Supplementary Figure S3), whereas GENELAND inferred the presence of only nine genetic populations. The badgers in Iberia, Ireland, Britain, Denmark and Scandinavia always formed separate clusters in both analyses (Supplementary Figure S4). GENELAND also inferred the presence of a separate Scottish population. However, the composition of the clusters in the rest of mainland Europe differed between STRUCTURE runs and between both methods (Supplementary Figure S4). The BAPS algorithm, run by clustering 30 predefined groups, inferred the presence of 14 clusters in the optimal partition and appeared to resolve the incompatibilities between the different STRUCTURE and GENELAND runs (compare Figure 1 and Supplementary Figure S4). It also identified Iberia, Ireland, Denmark and Scandinavia as being genetically distinct. Furthermore, it split the mainland data set into a further seven geographically coherent clusters and classified the three sampling locations in Britain as being genetically different.

A factorial correspondence analysis (Figure 3) illustrated a strong differentiation of the badgers in Ireland, Britain and Scandinavia. Scandinavian badgers were—in contrast to the results of the mtDNA analysis—closely related to populations from eastern Europe and British/Irish badgers and to western–central Europe. The DIYABC analysis supported scenario 4 as the most probable, demonstrating that the Scandinavian population resulted from admixture between the Iberian and Balkan ‘lineages’ (Figure 4). This event was estimated to have occurred at 1760 BP (95% confidence interval: 525–6390 BP). The principal component analysis demonstrating the reliability of the chosen scenarios parameters and the posterior distributions of model parameters for the most likely scenario are provided in Supplementary Figure S5 and Supplementary Table S4, respectively.

Figure 3
figure 3

Factorial correspondence analysis of badgers from different predefined European populations. The analysis was based on 18 microsatellite loci. The percentage of the total variation explained by each of the two axes is given.

Full size image
Figure 4
figure 4

Results from the ABC analysis: graph of linear regressions showing posterior probabilities of the scenarios on the Y axis and the number of simulations used to calculate it (1% of total simulations) on the X axis. The plot for the best-supported scenario 4 is in grey graphic and represented on the right.

Full size image

The microsatellite-based genetic diversity measures (AR and Heu) showed a significant decline in variability from east to west and from south to north (Supplementary Table S5), irrespective of whether all 25 predefined populations were used or only one per BAPS-defined genetic cluster. These relationships mainly resulted from badgers in Iberia, Ireland, Britain, Denmark and Scandinavia having a reduced genetic diversity compared with the remaining populations in mainland Europe (Supplementary Figure S6). In the case of mtDNA, there was evidence of a decline in haplotype richness (and in haplotype diversity using all populations) from south to north (Supplementary Table S4), with badgers in Britain, Ireland and Scandinavia in particular being less variable than the southern populations (Supplementary Figure S7).

When trying to estimate values for log(N0), log(N1) and log(ta) using the method by Storz and Beaumont (2002), the 12 initial MCMC chains for the three demographic parameters converged in 7 populations at the <1.1 level and in a further 7 populations the chains reached approximate convergence at least (Supplementary Table S6). In the case of 5 populations, the 12 initial chains relating to the log(N0) and log(ta) estimates did not converge. When these data were analysed using the three longer chains (a total of 1010 steps), the analyses for Serbia did not converge and were not considered further.

The MSVAR-based point estimates of log(N1) were all relatively similar and varied between 3.7 and 4.1 (which does correspond to a range of ∼5000–12 500 individuals though; Figure 5). The point estimates of log(N0) were particularly low (⩽1.9 or ⩽79 individuals) for Belgium, Northern Ireland, Norway, Scotland and Sweden. In contrast, Croatia had the highest estimate of log(N0)=3.2 (∼1580 individuals), followed by Barcelona (2.9; 800 individuals) and Northern Italy (2.8; 630 individuals). All other populations had point estimates of 2.0⩽log(N0)⩽2.7 (≅200–500 individuals). There was a significant positive correlation between point estimates of log(N0) and log(ta) (rs=0.72; P<0.001). The five populations with estimates of log(N0)⩽1.9 were also estimated—with one exception—to have suffered the most recent declines (2.1⩽log(ta) ⩽2.4; Figure 5). However, the point estimates suggested that no decline occurred before 100 generations ago. On the opposite end of the spectrum, the populations in Luxembourg and Barcelona were estimated not to have declined before 2500 generations (log(ta)⩾3.4) ago. It should be emphasised that all estimates of log(N0) and log(ta) had wide 90% credible intervals.

Figure 5
figure 5

Quantification of populations size changes using the MSVAR method by Storz and Beaumont (2002), consisting of (a) estimates of the log of present (N0) and past (N1) effective population sizes and (b) log of time since the decline in population size. The point estimates represent the mode of the highest posterior density interval, and the error bars and grey bars represent the corresponding 90% credible intervals.

Full size image

When analysing the simulated data with the MSVAR method, the majority of the chains estimating log(N0) only converged approximately; in 10 cases, the three longer chains needed to be run to achieve convergence and two data sets needed to be replaced because of nonconvergence (Supplementary Table S7). Although sample size did not affect accuracy, estimates of log(N0) were more accurate for older and severe events (Figure 6, Supplementary Figure S8 and Supplementary Table S8). Estimates of log(N0)<2.0 were obtained with 20 simulated data sets, all but one of which corresponded to a 100-fold simulated decline (5000 to 50 individuals; Figure 6). Point estimates of 2.0⩽log(N0)⩽2.7 were obtained for 34 simulated data sets, 33 of which contained at least two of every possible combination of N0 (50, 150) and ta (25, 100, 500). The 10 data sets with a simulated N0 of 150 and a decline 25 generations ago gave rise to point estimates of 2.6⩽log(N0)⩽3.5. The chains estimating N1 achieved good convergence (Supplementary Table S6). The strength of the bottleneck did affect the accuracy of the log(N1) point estimate (Supplementary Table S9). MSVAR tended to underestimate log(N1) for severe bottlenecks occurring more than 100 generations ago (the interaction was not significant though; Supplementary Figure S9 and Supplementary Table S9).

Figure 6
figure 6

Inference of the current effective population size, log(N0), of simulated data sets using MSVAR. The simulated data sets had different samples sizes (N), time since population decline (ta) and current effective population size (sim N0). Past effective population sizes were fixed to 5000 individuals. In each graph, the grey bars represent the 90% credible intervals of the point estimate represented by a cross. The dotted and the dashed lines indicate the simulated past and current effective population sizes, respectively.

Full size image

When using MSVAR to estimate the log(ta) of the simulated data, around a third of all runs did not achieve approximate convergence and/or had a bimodal posterior distribution and these were excluded from further analysis (Supplementary Figure S10 and Supplementary Table S7). This appeared to be particularly the case for less severe and more recent declines (Supplementary Figure S10). In the remaining results, the point estimates were always higher than the simulated values. Point estimates of time since decline varied between 1.9⩽log(ta)⩽3.2 and 2.0⩽ log(ta)⩽3.2 when analysing data sets with a simulated decline 25 and 100 generations ago, respectively. Point estimates of log(ta)⩾2.9 were obtained for data sets with a simulated ta of 500. In the case of the simulated data, there was also a significant positive correlation between point estimates of current effective population size and time since decline (rs=0.38; P<0.00).

Discussion

The results of our median-joining network were consistent with those presented by Marmi et al. (2006), revealing no clear mtDNA genetic structuring within European badgers. However, the geographic distribution of some groups of haplotypes in European badgers strongly suggested the presence of at least two separate source populations. Specifically, we had haplotypes with an Iberian distribution and another group of haplotypes spread throughout most of the continent, with secondary contact of the haplotypes from both ‘Iberian’ and ‘Eastern’ groups in western Europe. The evidence for a bimodal mismatch distribution (Patarnello et al., 2007) further supported a scenario of expansion out of at least two glacial refugia (Supplementary Figure S2). Although not converging on an optimal solution regarding the population genetic structure of European badgers, the different microsatellite-based analytical methods showed that the badgers in Iberia together with populations in Ireland, Britain, Denmark and Scandinavia were well differentiated from the remaining mainland populations.

The presence of many unique haplotypes in Iberia, coupled with extensive fossil data from the region (Sommer and Benecke, 2004), clearly points to the peninsula having been a glacial refugium for badgers. Fossil evidence also shows that badgers were present in the Carpathians and the Balkans during the LGM (Pazonyi, 2004; Sommer and Benecke, 2004). Southeastern and eastern Europe are also the only regions where all four main ‘Eastern’ haplotypes (meles1–3 and mm20) are found, suggesting that expansion started outwards from here. Neither the mtDNA nor the microsatellite data provided evidence for genetic divergence between the Carpathian and Balkan refugia, in contrast to results from smaller mammals with more limited dispersal ability (Kotlík et al., 2006; McDevitt et al., 2012). The greater dispersal ability possessed by larger mammals may have allowed connectivity between the Carpathian and Balkan refugia, as a similar pattern of distinct refugia but with apparently uninterrupted gene flow during the LGM was shown for another large mammal, red deer (Zachos and Hartl, 2011).

The origin of Scandinavian badgers is more uncertain and results from the two marker types suggested differing colonisation scenarios. The mtDNA haplotypes observed in Scandinavia were closely related to the Iberian haplotypes (Figure 2a). Considering that the first subfossil records from southern Sweden have been dated to ∼9000 BP (Sommer and Benecke, 2004), it appears most likely that the Iberian lineage reached Scandinavia first over a landbridge (Björck, 1995) after the Younger Dryas (because of the dominance of the closely related mtDNA haplotypes). Scandinavian haplotypes were also found as far eastwards as Russia, and this suggests an extensive expansion after this initial colonisation from Iberia. Based on the extensive distribution of these haplotypes and the recent nature of this recolonisation (after the Younger Dryas), an argument could be made for a third, unsampled refugium with badgers colonising Scandinavia from the east as opposed to Iberia. However, the close relationship between Iberian and Scandinavian haplotypes suggests this to be unlikely. The microsatellite-based ABC analysis provided most support for a colonisation of Scandinavia from both Iberian and southeastern refugia (Figure 4). This later admixture from the southeast (via Finland) occurred more recently (525–6390 BP). This pattern of admixture in Scandinavia has been previously observed in other mammalian species (Brunhoff et al., 2003; Ruiz-Gonzalez et al., 2013). The mtDNA haplotypes from the southeast of Europe do not occur in Scandinavia however (Figure 2b), but this may be because of a selective advantage of the haplotypes of the first wave of colonisers in the region (Ruiz-Pesini et al., 2004; McDevitt et al., 2012).

The colonisation of Ireland provides another clear example of conflicting results obtained from mtDNA and microsatellite data. The question of how the island acquired its fauna and flora has long presented a problem (McDevitt et al., 2011). Using data from a shorter fragment of the control region and six microsatellites, O'Meara et al. (2012) recently concluded that badgers have colonised Ireland naturally, but failed to identify how and when this occurred. We show here that Irish badgers share the majority of their haplotypes with Scandinavia but are closely related to Britain at the microsatellite level (Figures 2a and 3). Natural colonisation from an Iberian source (as we have proposed for Scandinavia) is unlikely because of the absence of evidence for suitable landbridges post LGM (McDevitt et al., 2011) or appropriate fossil evidence before the presence of humans in Ireland (McCormick, 1999). The only securely dated badger fossil in Ireland dates to 1554 BP (P Woodman and M O’Dowd, personal communication) and other fossils are of a Bronze Age context (4000–1500 BP; McCormick, 1999). Therefore, human-mediated introduction(s), which has been reported for other Irish mammals (see, for example, McDevitt et al., 2011), is the most likely scenario and our genetic results support this. Early introductions from Britain (as supported by a number of mtDNA haplotypes shared between the islands and microsatellite data) would have been followed by later introductions during the Viking invasions from Scandinavia from ∼800 AD onwards (supported by the Scandinavian haplotypes found on the island). The Vikings were also responsible for the introduction of the house mouse (Mus musculus) onto Ireland (Searle et al. 2009) and badger bones were associated with Viking sites in Dublin from the tenth and eleventh centuries AD, possibly being a valuable food item at the time (McCormick, 1999).

Our results show that the populations on the European periphery—the western- and northern-most populations—have a reduced genetic diversity (at least at the microsatellite level) compared with the populations in central, eastern and southeastern Europe. As a result of successive founder events, species that expanded from a refugium may exhibit reduced genetic diversity in populations furthest from their origin (Hewitt, 1999), especially if the population in the refugium had a small effective population size. It is therefore likely that the reduced diversity in the peripheral populations results from historic processes. On the other hand, badgers might have undergone recent declines and genetic bottlenecks as a result of human-induced habitat modification (see, for example, van der Zee et al., 1992) and/or persecution in the context of disease management (Griffith and Thomas, 1997).

We aimed to use the MSVAR approach to assess the likelihood of recent demographic processes having contributed to a reduced diversity in some populations. Taken at face value, the estimates of log(N0) are very encouraging, suggesting that five populations have undergone a severe decline and that three populations located in glacial refugia (Barcelona, Italy and the Balkans) have the largest current effective population sizes, with the remaining populations having intermediate estimates. MSVAR also suggested, however, that no decline occurred before 100 generations ago. Because the characteristics of a bottleneck, such as its timing and severity, can influence the accuracy and precision of the MSVAR analysis (Girod et al., 2011), we performed our own simulations—based on realistic demographic scenarios and the number of samples and genetic markers available for the empirical analyses—to assess whether our empirical results were a realistic indication of past demographic events.

Similar to Girod et al. (2011), many chains—both with the empirical and the simulated data—only reached approximate convergence (Gelman and Rubin statistic <1.2), with a small number not converging at all. A less stringent convergence criterion thus appears necessary to obtain point estimates of the demographic parameters, despite the larger credible intervals this necessarily will entail. Results by Girod et al. (2011) suggested that scaled parameters (θ0≡4N0μ, θ1≡4N1μ add tf≡ta/(2N0); μ=mutation rate) provide more precise estimates of the natural parameters for contractions. However, most researchers are likely to find point estimates of the demographic parameters intuitively more appealing.

Although the size of the empirical data sets varied between 20 and 37 individuals, our simulations suggested that the accuracy of the MSVAR point estimates was not influenced by sample size. However, the simulation results also suggested that it was very difficult to draw robust conclusions about the severity and timing of past bottlenecks in our empirical data. The estimates of time since decline appeared to be particularly unreliable, as the MSVAR point estimates of log(ta) were always higher (sometimes substantially so) than the simulated values. However, five empirical populations had an estimated current effective population size of log(N0)<2.0, with an estimated decline log(ta)⩽2.4 generations ago. Only five of the simulated data sets had similarly low estimates of both log(N0) and log(ta), and all of these had a simulated N0 of 50 individuals after experiencing a 100-fold decline 25 generations ago (Figure 6 and Supplementary Figure S10). In other words, the simulations provided support for a recent severe bottleneck of badger populations in Belgium, Northern Ireland, Norway, Scotland and Sweden.

However, drawing conclusions about the past demography of the remaining badger population seems problematic. Generally, our simulation results agreed with those by Girod et al. (2011), in that the estimates of the current effective population size tended to be more accurate for older and severe events. In other words, point estimates of log(N0) for bottlenecks of the same severity tended to be larger for more recent bottlenecks. For example, values in the range of 2.0⩽log(N0)⩽2.7 were obtained for all combinations of N0 (50, 150) and ta (25, 100, 500), and these data sets—especially those with a simulated decline 25 and 100 generations ago—also gave rise to estimates of log(ta) that were comparable to the ones obtained with the empirical data. Although it appeared encouraging that the populations in the 3 classical glacial refugia had the highest estimated values for log(N0), the 10 data sets with simulated N0=150 and ta= 25 gave rise to similarly high estimates of log(N0) and log(ta). In other words, the simulations suggested that for the majority of the empirical populations, the timing and severity of a possible bottleneck could not be established with certainty. Furthermore, given the convergence problems, the empirical estimates of the demographic parameters (with the possible exception of Luxembourg) had very large 90% credible intervals and were thus very imprecise as well.

During the twentieth century, badgers have declined dramatically in Belgium, partially as a result of the rabies eradication campaigns (Griffiths and Thomas, 1997): In 1992, only 28 badger setts were known from Flanders (where our samples were from). Although Scandinavian badgers are abundant and currently expanding their range, the population was restricted to the southern tip of the peninsula during the nineteenth century (Bevanger and Lindström, 1995). It is thus possible that the Scandinavian results were caused by a population expansion from small effective population size (MSVAR performs poorly when analysing data from expanding populations; Girod et al., 2011). Although badgers in Denmark have declined over the past 50 years and exhibit reduced genetic diversity, Pertoldi et al. (2005) have convincingly shown that they have not suffered from a genetic bottleneck during the past half-century. We therefore cannot exclude the possibility that the genetic diversity in Scandinavia and Denmark was historically low, even before possible recent human-mediated declines.

There are—to our knowledge—no records of recent badger declines in Northern Ireland or Scotland (Griffiths and Thomas, 1997; Reid et al., 2011). Although not directly comparable to this study, the estimates of allelic richness presented by both Pope et al. (2006) and O'Meara et al. (2012) suggested that most, if not all, British and Irish badger populations have similarly low genetic diversity than those analysed in this study. It is therefore likely that the reduced diversity observed both with microsatellites and mtDNA on the British Isles resulted from founder effects during natural colonisation, or—as probably in the case of Ireland—human-mediated introduction. It is possible that MSVAR proves the occurrence of recent unrecorded bottlenecks in Scotland as Pope et al. (2006) did not observe a decline in genetic diversity in Britain with latitude.

Given the general correlation of genetic diversity with latitude, and the support for southern refugia, it is likely that reduced genetic diversity in northern populations was contributed to by historic processes. Unfortunately, despite multiple markers and extensive sampling, any additional effects of recent processes remain unresolved. As also shown by others, the precision and the accuracy of the MSVAR method is influenced by the timing and severity of a bottleneck. Based on simulated data sets representing specific demographic scenarios and the available number of samples and genetic markers, we have established that there is insufficient power to differentiate between historic founder effects and recent declines with certainty in the context of this study. We therefore urge caution when trying to relate demographic declines inferred using MSVAR with particular historic or climatological events (see, for example, Goossens et al., 2006; Heller et al., 2008), especially if there is uncertainty about the study species’ generation time (which is required to estimate of the time since the decline). Other researchers should undertake similar testing if they wish to employ MSVAR in this way.

In summary, continent-wide sampling of badgers and analyses with multiple markers provided evidence for two glacial refugia (Iberia and southeast Europe) contributing to the genetic variation observed in badgers in Europe today. Because of lack of sampling in Russia, there remains a certain doubt, however, regarding the phylogeographic origin of badgers in the far east of Europe. The pattern of decline of genetic diversity with increasing latitude suggested that the reduced diversity in the peripheral populations probably resulted from a postglacial expansion processes. Because of methodological limitations, it was not possible to ascertain whether some of these peripheral populations also have undergone recent genetic bottlenecks. Therefore, despite our best efforts, additional questions remain to be answered about the large-scale genetic structure of M. meles in Europe.

Data archiving

Sample locations and microsatellite data: DRYAD entry doi:10.5061/dryad.5nm5g.