Juniperus communis L. (Cupressaceae) is a diploid, dioecious, wind-pollinated, woody shrub or small tree. The distribution range of common juniper basically covers the entire Northern Hemisphere . It is the most widespread conifer taxon worldwide and known to have a broad ecological amplitude. Female individuals bear cones that ripen fully in the autumn of the second or third year of development and contain 1 - 3, rarely 4 seeds. Seeds are mainly dispersed by birds, especially thrushes (Turdus spp.) and common juniper does not produce a long-term persistent seed bank . Lifespan is estimated to be about 100 years, although exceptionally individuals reach over 200 years . Genets with clonal shoots, however, may readily exceed this age. Common juniper habitats have been accorded a legal protection status in Europe (EU Habitat Directive, Annex I, code 5130) and despite its wide distribution, the species is on the Red List in several European countries (e.g. the Netherlands , the UK  and Belgium [51, 52]). For more information about the species, we refer to Thomas et al. .
In 2005, we sampled 13 natural common juniper populations along a north-south transect from north-Germany and northern Netherlands to north France, and along an east-west transect from northwest France to northwest Germany (Figure 1). Populations located along the transect were selected for sampling when they contained at least 100 individuals and were presumed natural. A population was presumed natural if there was evidence based on historical topographic maps or if personal communications with local people and nature conservationists revealed that the population was at the site for many centuries. When no populations with at least 100 individuals were present, presumed natural populations with less individuals were included. The sampled populations included three populations in northwest France, three populations in Belgium, four in the Netherlands and three populations in northwest Germany. The populations occurred on heathlands or calcareous grasslands, both strongly fragmented habitats in the study area [7, 8]. For each population, the census population size was estimated based on the point-centred quarter method (PCQ), which is a plotless sampling method to estimate the population density . Therefore, in each population, one to three random sampling points were laid out depending on the size of the population. In each of the four quarters around the sampling points, distances were measured to a maximum of four trees closest to this point; one for each of four height classes (when available): < 1 m, 1 - 2 m, 2 - 3 m, > 3 m. These height classes broadly reflect the following development classes of the tree: seedlings, young plants, mature plants and old plants . Fresh needles were collected from the measured trees in each quarter. The needles were dried with silica gel in zip-locked bags until analysis. Next to this, a random sample of ripe cones was collected from the female shrubs at each sampling point, resulting in 6 to 27 plants per population. Table 1 provides information about the populations sampled.
In 2008, we sampled the three natural (defined as above) populations in Flanders that contain more than 100 individuals: the populations Heiderbos, Kattevennen and Hesselberg. The populations occurred on heathlands, a habitat that has become highly fragmented in Flanders . Moreover, we sampled 29 other locations in Flanders with one to 38 relict individuals. The three natural Flemish populations were analysed for genetic diversity statistics while the spatial genetic structure and clonal structure was inferred from Flemish samples collected on all the 32 locations. All sampling locations were located in the east of Flanders (northeast Belgium). They were selected from an earlier full inventory of Juniperus communis in Flanders . Again, individuals from different age classes were sampled. The height of the shrubs was recorded as mentioned above. Shrubs were sampled at random at each location since census population sizes often were too small to use the PCQ-method. Information about the sampling sites and their location is given in Table 1 and Figure 1, respectively.
Per shrub, 10 ripe cones were opened, the number of seeds was counted and filled seeds were exposed to 1% 2,3,5 triphenyltetrazolium chloride (TTC) solution in order to determine the viability of the embryos. Initially colourless, TTC is converted to formazan-red in the presence of living tissue (see e.g. Miller  for more details on the method).
DNA extraction and molecular genotyping
Genomic DNA was extracted from 20 mg of dried needles using the Dneasy Plant Miniprep Kit (Qiagen, Helden, Germany), according to manufacturer's instructions and followed by an additional treatment with 0.4 μg RNAse (Fermentas) at room temperature for 2 min. DNA concentrations were estimated and standardised against known concentrations of λDNA (Fermentas) on 1.5% agarose gels.
AFLP analysis was performed on the northwestern European and the Flemish samples according to Vos et al.  and Van Der Merwe et al.  with following modifications. Restriction and ligation were performed in a single step, e.g. 200 ng of genomic DNA was restriction digested using the enzyme combination PstI (Fermentas) /MseI (Fermentas) and ligated to the PstI and MseI adaptors. Primer combinations used for the generation of fingerprints were PstI-ACT + MseI-ACA, PstI-ACT + MseI-ACC, PstI-AGT + MseI-ACC and PstI-AGT + MseI-ACA. Fragment separation and detection took place on a NEN IR2 DNA analyzer (Li-Cor Biosciences) using 36 cm denaturing gels with 6.5% polyacrylamide. IRDye size standards (50 to 700 bp) were included for sizing of the fragments. Fragments within the size range of 75 bp to 677 bp were scored with Saga Generation 2 (Li-Cor Biosciences) as present or absent. Prior to data analysis, monomorphic loci were discarded. Due to the long time period between the analysis of the northwestern European samples and of the Flemish samples, different PCR-machines and fabrication batches of products were used. Hence, following Coart et al. , the AFLP-data of the northwestern European samples and of the Flemish samples were processed separately. The number of individuals typed with AFLP markers is given in Table 1.
AFLP error rate and reproducibility
Reproducibility was evaluated on the dataset obtained from the individuals sampled on the Flemish scale using intra- and intergel replicates. 23 samples (12%, according to the recommendations of Bonin et al. ) were chosen randomly and analysed twice independently starting from the same DNA-extraction. Samples with a profile that was doubtful, for example profiles showing low band intensities, were discarded. We estimated the error rate at the allele level as described by Bonin et al.  based on the binary matrix obtained for the replicate samples. This error rate is effectively the average Euclidean distance (= 1 - Simple Matching similarity index ) between replicate pairs. The error rate was first used to eliminate unreliable markers (markers difficult to score or unstable markers) and to clean up the binary data matrix . Secondly, we recalculated the error rate based on the replicated samples for the final markerset. In order to evaluate this error rate in accordance to the goal of the study, we performed a UPGMA-cluster-analysis based on the Simple Matching similarity index calculated from the binary matrix of the replicated samples using the programme TREECON . We calculated the number of replicate pairs that were correctly assigned (i.e., as 'sister' to one another) in the cluster analysis (e.g. see ). We also calculated the mean pairwise inter-individual genetic distance based on the Simple Matching index for all the genotypes from the Flemish dataset and from the northwestern European dataset, and compared this with the mean intra-individual genetic distance (= equivalent to the error term). Furthermore, we implemented the simulation procedure in the programme AFLPOP  to investigate the power of the data for the assignment test.
Habitat fragmentation and seed-mediated dispersal
To estimate the degree of habitat fragmentation in the study area, we mapped the natural habitats suitable for Juniperus communis with the programme ArcView version 3.1. (ESRI) and based on CORINE Land Cover 2006 vector data (CLC06) (version 13 - 02/2010, http://www.eea.europa.eu/). CLC06 classifies the European land cover into 44 categories derived from Landsat and SPOT satellite images at a 1:100,000 scale and with a minimum mapping unit of 25 ha . The following land cover types from CLC06 were considered as suitable habitats for Juniperus communis: natural grasslands (class 3.2.1), moors and heathlands (class 3.2.2), sclerophyllous vegetation (class 3.2.3) and transitional woodland-shrub (class 3.2.4). The percentage of suitable habitat was calculated for northwestern Europe within an area of 1.4E+7 ha, and within a 30 km radius buffer zone surrounding each sampled population.
Seed-mediated dispersal events were estimated by individual-based population assignment tests using the computer program AFLPOP 1.1 . Because we are aware that we did not sample all the potential source populations, our aim was not to allocate individuals to some of the sampled populations, but to estimate migration rates by identifying immigrants for the populations sampled on the northwestern European scale. First, the likelihood was computed that an individual genotype (G) may be found in each of the candidate populations based on their respective dominant AFLP band frequencies. G is then assigned to the population showing the highest likelihood for G . Given that pollen flow might result in ambiguous assignments when low levels of stringency are used , allocation tests were conducted setting the minimum log-likelihood difference (MLD) to 1 and 2. At these MLD = 1, MLD = 2 stringency levels, an assignment to a population is made when the probability of the given assignment is ten or 100 times more likely than the next most probable assignment, respectively. Other settings in the program were: replace zero frequencies by (1/(sample size+1)) and calculate a p-value for each individual's log-likelihood by creating empirical distributions from 1000 randomly generated genotypes based on the presence frequencies of each population. When the p-values for an individual were below a certain warning threshold (< 0.001 in our case) for all candidate populations, it was concluded that the individual did not originate from any of the sampled populations.
Prior to the allocation test, we assessed the power of our dataset for accurate assignment of the real genotypes with the population assignment simulator of AFLPOP 1.1 . The simulator generated 1000 random genotypes based on the observed allele frequencies in each sampled population. Those 1000 simulated genotypes were then blindly reassigned to their most probable population. The simulation process was repeated 10 times to check for the consistency of the results. Because of geographical affinity and small population size we pooled the samples from populations Kootwijkerzand and Loenen, and also the samples from the populations Meenser Heide and Weinberg. The latter populations were also pooled because of statistically non-significant ΦPT-pairwise values. This reduces the risk of misassignment due to similar allele frequencies between population pairs.
Population genetic structure was analysed based on AFLP data on both spatial scales. Total genetic diversity was partitioned among and within populations by carrying out a hierarchical analysis of molecular variance (AMOVA) on Euclidian pairwise genetic distances . The ΦPT analog for FST  was calculated based on Euclidian genetic distances, and its significance was determined using the Monte Carlo procedure (999 permutations). Based on these Euclidian pairwise genetic distances a principal coordinates analysis (PCoA) was performed. These analyses were carried out using GENALEX 6.2 . To further identify possible spatial patterns of genetic diversity, the software BAPS 5.3  was used to identify clusters of genetically similar populations using a Bayesian approach. A population mixture analysis was performed for the maximum number of clusters (K) ranging from K = 1 up to K = 15. We ran the cluster analysis ten times in order to test the reproducibility of the results. In order to identify a significant isolation-by-distance effect , a Mantel test was performed on pairwise genetic distances and geographic distances. At the local (Flemish) scale, we investigated the existence of a fine scale spatial genetic structure. We plotted and regressed average pairwise kinship coefficient of relatedness for dominant markers  against geographical distances with the software SPAGeDi 1.3 .
As common juniper can reproduce clonally, we first checked whether the dataset contained similar ramets of the same clone. This was done with AFLPdat  by setting the maximum number of differences among identical individuals to 15 bands. The latter was estimated from 23 replicated samples. Further population genetic analyses were restricted to the individuals derived from sexual reproduction (i.e. genets). Genetic diversity statistics were calculated based on AFLP data. We calculated AFLP fragment frequencies with AFLPsurv 1.0  to estimate allele frequencies for each population. This was based on a Bayesian approach with a non-uniform prior distribution of allele frequencies following Zhivotovsky , assuming either no, or some deviation (FIS = 0.1) from Hardy-Weinberg genotypic proportions according to the outcrossing nature of the species. However, the results based on the different FIS
-values were very similar and therefore, only those based on FIS = 0 are presented. Allele frequencies were then used to calculate Nei's gene diversity (Hj) and the percentage of polymorphic loci at the 5% level corrected for the sample bias (PPL). Furthermore, band richness corrected for the sample bias (Br) was computed on the AFLP data with AFLPDIV (first described in ). This measure of genetic diversity represents the number of phenotypes expected at each locus (i.e. each scored AFLP fragment) and can be interpreted as an analogue of allelic richness .
The level of inbreeding (FIS) was estimated from AFLP fragment frequencies using FAFLPcalc . FAFLPcalc uses AFLP frequencies to estimate band frequencies that are used to simulate data with a range of inbreeding coefficients. This approach assumes that half of the individuals in a population are outbred, and that inbred individuals will be more homozygous (exhibit more null phenotypes). Scoring errors and high levels of non-independence between bands can lead to poor results, which is why we compare calculated FIS values only among our sample populations and within a sampling year and not to those from other studies.
In order to identify whether habitat fragmentation resulted in a decrease of genetic diversity in younger (height < 1 m) compared to older individuals (height > 1 m) of the common juniper populations, values of ΦPT were calculated between the four different development classes by AMOVA based on AFLP data for each population separately. Also, values of ΦPT and genetic diversity statistics (PPL, Br, Hj) were calculated for pooled population samples per height class within each spatial scale. Average inbreeding coefficients (FIS) were compared between young plants (height < 1 m) and older individuals (> 1 m height) of the common juniper populations sampled on the northwestern European scale by a t-test. Finally, we used Spearman rank correlations to identify possible relationships between AFLP-based genetic diversity measures (PPL, Br, Hj and FIS) and population characteristics (population size, population density, % filled seeds and % viable seeds).