Irene Keller, Queen Mary, School of Biological Sciences, Mile End Road, London E1 4NS, UK. Tel.: ++44 20 7882 7528; fax: ++44 20 8983 0973; e-mail: firstname.lastname@example.org
We assess the impact of habitat fragmentation on the effective size (Ne) of local populations of the flightless ground beetle Carabus violaceus in a small (<25 ha) and a large (>80 ha) forest fragment separated by a highway. Ne was estimated based on the temporal variation of allele frequencies at 13 microsatellite loci using two different methods. In the smaller fragment, Ne estimates ranged between 59 and a few hundred, whereas values between 190 and positive infinity were estimated for the larger fragment. In both samples, we detected a signal of population decline, which was stronger in the small fragment. The estimated time of onset of this Ne reduction was consistent with the hypothesis that recent road constructions have divided a continuous population into several isolated subpopulations. In the small fragment, Ne of the local population may be so small that its long-term persistence is endangered.
The effective size of a population (Ne) is a fundamental concept in evolutionary biology and conservation. It is defined as the size of an ideal population subject to the same rate of random genetic change as the studied population (Wright, 1931) and is generally smaller than the actual number of observed animals (Frankham, 1995). The significance of Ne stems from its central role in determining the rate of increase in inbreeding and loss of neutral genetic variability through random drift. Furthermore, at low Ne, natural selection becomes less effective compared with drift, which can result in the random loss of favourable alleles and the fixation of slightly deleterious mutations (Hedrick, 2000; Frankham et al., 2002). Several studies have shown that these genetic processes can indeed reduce the fitness and adaptive potential of small populations (Crnokrak & Roff, 1999; Hedrick & Kalinowski, 2000; Amos & Balmford, 2001; Hedrick, 2001; Keller & Waller, 2002; Reed & Frankham, 2003) and may, along with demographic and environmental factors, affect their long-term survival (Primack, 2002).
Today, the fragmentation and isolation of natural habitats is one of the main threats to the persistence of many animal and plant species. Roads, for example, may represent significant barriers to dispersal for species with limited mobility and can reduce or prevent the exchange of individuals between different populations (e.g. Baur & Baur, 1990; Bhattacharya et al., 2003). Wingless ground beetles are very reluctant to cross paved roads (Mader, 1984), and this isolation is strong enough to lead to significant genetic differentiation between populations within a few decades (Keller & Largiadèr, 2003a).
If a formerly continuous populations is divided into several isolated fragments, this will inevitably reduce Ne within the individual fragments (e.g. Frankham et al., 2002). Thus, Ne estimates may be particularly valuable in the context of habitat fragmentation to identify small populations with an increased risk of extinction. Fortunately, such estimates have become easier with the advent of molecular techniques, as it is now possible to assess current Ne from genetic measures without the need for long-term demographic data (Waples, 1989, 2002). Recent methods also allow the detection and dating of Ne changes in the history of a population based on highly polymorphic markers (Cornuet & Luikart, 1996; Beaumont, 1999; Storz & Beaumont, 2002). These analyses may be useful to assess if a population has indeed experienced a size contraction as a consequence of fragmentation or if it has remained unaffected by potential barriers to dispersal. This is of particular importance because we often do not have any information on the value of Ne or other parameters before fragmentation.
An earlier study on the ground beetle Carabus violaceus showed that genetic variability in samples from a small and isolated forest fragment was significantly reduced (Keller & Largiadèr, 2003a). This indicated high rates of random genetic drift, thought to be due to a small effective population size in this fragment. Here, we present a more detailed quantification of the impact of fragmentation due to roads on the effective size of beetle populations. We used temporally spaced samples to estimate the current effective size of two populations in differently sized forest fragments. We then tried to estimate the timing and the magnitude of a potential size reduction associated with the construction of roads in the area.
Material and methods
The ground beetle C. violaceus L. (Coleoptera, Carabidae) has a Palaearctic distribution and is common throughout most of Europe. Our study focused on the subspecies C. v. violaceus, which is found in various types of forests. The species is nocturnal, 22–35 mm in length and flightless (Marggi, 1992; Wachmann et al., 1995). The animals hatch in late summer, hibernate as larvae, and enter their first reproductive period in the following summer. Age determination based on mandible wear (Houston, 1981; Wallin, 1989) showed that most sampled animals were in their first reproductive period and an average generation time of 1 year seemed to be likely for our study area. We assessed the accuracy of this method in a small pilot study on 67 individuals (I. Keller, unpublished) by comparing the results with an unambiguous age determination based on the state of the gonads (Krehan, 1970).
Starting in 2000, three temporal samples were obtained at 1-year intervals from two locations in a mixed forest near Bern. The two sampling sites were at a distance of c. 250 m from each other but on opposite sides of a highway (Fig. 1). Samples L (large)-2000 to L-2002 were from a large forest fragment (location 2c in Keller & Largiadèr, 2003a) with an area of more than 80 ha. Samples S (small)-2000 to S-2002 were collected in a fragment of c. 25 ha, which had been cut off from the large fragment by the construction of a two-lane highway in 1969 (location 3b in Keller & Largiadèr, 2003a). The two sampling sites were at an equal distance from the edge of the forest and located in areas with similar vegetation. In the following, the abbreviations L (large) and S (small) will also be used to refer to the entire forest fragments.
At each site, 16 dry pitfall traps were installed in a 15 × 15 m grid during several months in summer and early fall. The traps consisted of plastic cups with a diameter of 11.5 cm and were emptied twice a week. The captured animals were taken to the laboratory, where they were anaesthetized with CO2 and one middle tarsus was cut off and stored in absolute ethanol. All animals were released back into the forest on the same day. A pilot study had shown that this treatment did not lead to increased mortality; none of 10 animals subjected randomly to either no treatment or to the amputation of one tarsus died after 1 month in captivity. At each location, we sampled at least 49 individuals in two of the 3 years.
The density of C. violaceus was assessed at study site L using a capture-recapture method (see e.g. Southwood & Henderson, 2000). A total of 256 plastic cups with a diameter of 6.5 cm were buried level with the soil surface in a grid of 16 × 16 traps at regular intervals of 4 m. These dry pitfall traps were closed with plastic Petri dishes when not in use. During the eight trapping sessions carried out in warm and dry weather between May and August 2001, the traps were opened for two nights. All beetles were individually marked with a code drilled into the surface of their elytrae (Mühlenberg, 1993) and then released where they had been caught. The population density was estimated with the program jolly version 01/24/91 (Hines, 1988) based on the model by Jolly (1965) for open populations. The result was used as a guideline for choosing reasonable ranges for Ne estimation and appropriate priors in the Bayesian analysis to detect changes in population size below.
Microsatellite isolation and typing
The collected tarsi were ground with plastic pestles, and the DNA was extracted with a Chelex resin protocol (Estoup et al., 1996). All individuals were typed at 13 polymorphic microsatellite loci, five of which were described in Keller & Largiadèr (2002). For these five markers, the forward primers were labelled with the following fluorescent dyes (ABI): CVI05136CMPG with PETTM, CVI08071CMPG with 6-FAMTM, CVI09106CMPG with NEDTM, CVI09194CMPG with NEDTM and CVI10036CMPG with VICTM. A further locus, 828A, had originally been developed for C. solieri (Rasplus et al., 2001). The forward primer was labelled with 6-FAMTM, and the reverse primer was modified by adding the sequence GTTTCTT (pig-tail) at the 5′ end to reduce the presence of 1 bp stutter bands (Brownstein et al., 1996). Of the remaining seven loci, four were isolated from the library described in Keller & Largiadèr (2002) and three from a partial genomic library of C. violaceus constructed by Primm srl (Milan, Italy). For the latter, DNA from one individual from our study population was ligated into pGEM11z(f+) and amplified in Escherichia coli. The remaining laboratory procedures used for the isolation and the characterization of these seven loci are described in Keller & Largiadèr (2003b). The primer sequences and general characteristics of these seven polymorphic loci are listed in Table 1. Two odd-sized alleles were present at locus CVI02156CMPG, which were slightly (<1 bp) but consistently larger than alleles 206 and 212, respectively. Cloning and sequencing of the polymerase chain reaction (PCR) products revealed considerable differences in the sequences of these alleles as compared with regularly sized alleles, which varied only in the number of repeats in the microsatellite. Alleles deviating from the regular size pattern by half a repeat unit were present at loci CVI05136CMPG, CVI05216CMPG and CVI08083CMPG.
Table 1. Characterization of seven Carabus violaceus microsatellite loci based on one sample of 49 individuals. The sequences have EMBL accession numbers AJ549834–AJ549840. NA, number of alleles; HO, observed heterozygosity; HE, expected heterozygosity.
All 13 loci were amplified simultaneously with a QIAGEN (Basel, Switzerland) Multiplex PCR kit according to the recommendations of the manufacturers, using 2 μL of template DNA (Chelex extraction), an annealing temperature of 60 °C and 30 PCR cycles. A quantity of 5 μL of a 1 : 10 dilution of the PCR reaction were added to a mix of 19 μL Hi-Di Formamide (ABI, Rotkreuz, Switzerland) and 1 μL LIZTM size standard, and the denatured fragments were resolved on an automated DNA sequencer (ABI 3100).
Genetic variation within populations
Each locus in each sample was tested for departure from Hardy-Weinberg equilibrium with arlequin version 2.000 (Schneider et al., 2000) based on the approach of Guo & Thompson (1992). The tests were conducted with 1 000 000 steps in the Markov chain and 5000 dememorization steps. Bonferroni correction for multiple testing was applied within each sample separately. The program fstat version 22.214.171.124 (Goudet, 2001) was used to test for deviations from genotypic equilibrium between all pairs of loci in all samples.
Detection of potential immigrants
The estimation of effective population size and the detection of population growth or decline rely on the assumption that the samples were obtained from a closed population. Therefore, we used an assignment test implemented in geneclass version 1.0.02 (Cornuet et al., 1999) to identify individuals whose genotypes had low probabilities of belonging to the population where they had been caught, i.e. potential immigrants. The analysis was carried out for each temporal sample separately, and loci deviating from Hardy-Weinberg equilibrium (see below) were omitted. For each individual, the probability of its genotype to occur in its supposed population of origin was calculated based on the approach of Rannala & Mountain (1997). In a next step, we used the exclusion method described in Cornuet et al. (1999) to detect potential immigrants. With the simulation option in geneclass, 10 000 genotypes each were generated based on the allele frequencies in the two samples S and L, respectively. The probability of each simulated genotype in the respective sample was then calculated to obtain a distribution of these values for animals of known origin. An individual was considered a potential immigrant and omitted from the following analyses if its probability in a population was lower than the values of 99% of the simulated genotypes.
Note that this procedure will only detect recent immigrants and provide no information on gene flow further in the past. It serves the purpose of avoiding a bias that such individuals could introduce into our estimation of Ne. Recent immigrants from genetically distinct populations will cause more pronounced allele frequency changes between different generations, which can lead to an underestimation of Ne.
Estimation of effective population size
The variance effective population size (Ne) was estimated with both an F-statistic-based (Waples, 1989) and a likelihood-based method (Anderson et al., 2000). Simulations had shown that the likelihood estimator was more accurate and had a smaller variance than the F-statistic-based approach when genetic drift was strong (Williamson & Slatkin, 1999). However, the performance of the estimator was not assessed for weak drift. In such situations, the F-statistic-based estimator could potentially give more reliable confidence intervals (CIs) (Berthier et al., 2002). As a minimum sample size of 50 individuals is recommended especially for F-statistic-based Ne estimation (Waples, 1989), the analyses were based on the two largest samples from each population (L-2001 and L-2002; S-2000 and S-2001). Additionally, the maximum-likelihood estimation was repeated with all three temporal samples from each location.
With the F-statistic-based approach, Ne was estimated from the standardized variance (F) of the allele frequencies between the two generations. F was computed using the equation proposed by Nei & Tajima (1981), as this was found to be slightly more accurate for loci with rare alleles such as microsatellites (Waples, 1989). For locus j
where Kj was the number of alleles at that locus and xi and yi the frequencies of allele i in the two temporal samples. The average standardized variance across loci was computed as the weighted mean (Tajima & Nei, 1984):
From this value, an estimate of Ne was obtained following plan II (sampling before reproduction without replacement) as (eqn 11 in Waples, 1989):
where t = 1 was the number of generations between samples and S the sample sizes at times 0 and t, respectively. As the sample size varied between loci due to missing data, S was calculated as the harmonic mean of the sample sizes Sj at locus j weighted by the number of independent alleles (Kj − 1) as (Sachs, 1973):
Our sampling had actually followed plan I of Waples (1989) as the animals were released back into their populations. Under this scheme, the denominator in eqn 11 of Waples (1989) changes to , with N being the census population size. However, the difference between the two equations becomes very small if N > 2Ne, which is likely to be true in most populations (Frankham, 1995). 95% CIs for were computed with eqn 16 from Waples (1989) as:
with (total number of independent alleles) and the critical values of the chi-square distribution with n degrees of freedom. These two bounds for the Fc estimate were then used to calculate the 95% CI for Ne by applying eqn 11 from Waples (1989).
The likelihood-based estimator of Ne was computed with the program mcleeps version 1.0 (Anderson et al., 2000). At site L, the log-likelihood for Ne was estimated for Ne values between 50 and 8000 in steps of 50 and using 50 000 Monte Carlo replicates at each step. The upper value of Ne = 8000 was chosen based on an extrapolation of the estimated population density of roughly 100 animals/ha (see below) to the total area of the fragment. In a second step, the regions containing the mode and the limits of the 95% CI were analysed in more detail by taking 50 000 replicates at intervals of 10. For population S, the procedure was the same with log-likelihood estimations for Ne values between 50 and 2500 in steps of 50 followed by a more detailed analysis of the relevant regions in steps of 10. The 95% CIs were defined by the two points with log-likelihood values two units less than the observed maximum (Sokal & Rohlf, 1995; see also Anderson et al., 2000).
Detection of population growth or decline
The program msvar version 0.4.1b (Beaumont, 1999) was used to detect potential population reductions in the two forest fragments. This software assesses the most probable values for certain demographic or genealogical parameters given our data. More precisely, we obtain a posterior probability distribution for these parameters, which depends on our initial knowledge of the parameters (the prior distribution) and on how this previous information is changed by the data (e.g. Beaumont & Bruford, 1999). The data determine the posterior distribution through the so-called likelihood (the probability of the data given the parameters), which is computed based on coalescent theory in msvar. The program assesses the posterior probability distributions for the three parameters r, tf and θ. The first parameter r = N0/Nt is the ratio of the current over the ancestral effective population size, where N0 and Nt are defined as the number of haploid chromosome sets in the respective populations (i.e. twice the number of individuals). tf = ta/N0 is the time of onset of the population change expressed in units of the present effective population size. Finally, the parameter θ = 2N0μ is a function of the mutation rate μ and the present population size.
We used information from our field observations and from the literature to choose appropriate prior distributions for each of these three parameters. Within these intervals, which were likely to include the true values of the parameters, all values were assumed to be equally probable, i.e. we used uniform prior distributions. Based on the values estimated for census population sizes (see below), the following ranges were expected for present and ancestral Ne at site S: N0 = 20–5000 haploid chromosome sets, Nt = 200–200 000 haploid chromosome sets. The mutation rate was estimated to be between 10−2 and 10−7 (Jarne & Lagoda, 1996), and the period of population change ta between 1 and 1000 generations. From these values, the most extreme combinations of parameters were chosen to calculate θ, r and tf, the results were log10 transformed and rounded outward to the nearest integer on a log10 scale. This led to the following uniform priors for the three parameters: log10θ [−6, 2], log10r [−4, 2] and log10tf [−4, 2]. For site L, the expected ranges for the parameters were N0 = 200–100 000 and Nt = 200–200 000 haploid chromosome sets, μ = 10−2–10−7 and ta = 1–1000 generations, which resulted in the following priors: log10θ [−5, 4], log10r [−3, 3] and log10tf [−5, 1]. It is clear that the choice of these prior distributions is to a certain extent arbitrary. However, it is important to note that this was taken into account by using very broad limits for the priors, which reduced their impact on the results.
This analysis to detect historical population size changes was based on samples S-2001 and L-2001. The four loci with odd-sized alleles (CVI02156CMPG, CVI05136CMPG, CVI05216CMPG, CVI08083CMPG) were omitted because the method assumes a pure stepwise mutation model for the analysed microsatellites. Further, we assumed that the population size change from Nt to N0 had been exponential. msvar relies on Markov chain Monte Carlo (MCMC) simulation to explore the posterior distributions of the parameters. At each step in the chain, the algorithm draws a random sample from the parameter space delimited by the priors. This sampling is performed in such a manner that the equilibrium distribution of the MCMC is the posterior distribution we are interested in. Each simulation run consisted of 1010 steps with a thinning interval of 100 000 resulting in 100 000 samples from the posterior distributions. The MCMC scale parameter, which determined the size of the changes used for updating θ, r and tf, was set at 0.5. We ran four independent chains for sample S and three for sample L with different starting points for the three parameters. The first 25% of the samples in each run were treated as burn-in and omitted in the subsequent analyses. To check for convergence of the MCMC, the marginal posterior distributions for log r and log tf were plotted for each chain separately with the program locfit version 08/03/00 (Loader, 1996) available as a package for R (Ihaka & Gentleman, 1996) and compared visually. Additionally, we computed the Brooks, Gelman and Rubin convergence diagnostic (Gelman & Rubin, 1992; Brooks & Gelman, 1998) with the baysian output analysis program version 1.0 available from http://www.public-health.uiowa.edu/boa/. We further determined the modes of the posterior distributions and the 2.5 and 97.5% quantiles.
A total of 61 males and 49 females of C. violaceus were caught and individually marked. Of these, only nine individuals were recaptured, which made the estimation of population density very difficult. The program jolly yielded an estimate of 88 individuals/ha, but no CI could be calculated due to the low recapture rate.
Genetic variation within populations
In the 2001 and 2002 samples from site L, a significant deviation from Hardy-Weinberg equilibrium was detected at locus CVI09106CMPG. As this locus does not regularly show homozygote excess in other samples, the presence of a frequent null allele is unlikely. In the sample collected in 2000, a significant deviation from Hardy-Weinberg equilibrium was observed at locus CVI02233CMPG. In sample S-2001, significant homozygote excess was detected at locus CVI02156CMPG.
Of the 468 pairwise tests for genotypic disequilibrium, 41 were significant at the 0.05 level and seven at the 0.01 level. As these significant tests were generally observed between different pairs of loci in the six samples, they were likely to be the result of type I errors or random deviations in finite populations and not of physical linkage between loci.
Detection of potential immigrants
In samples L-2000, L-2001 and L-2002, the assignment test revealed six, six and five genotypes, respectively, with probabilities of occurring in population L, which were lower than 99% of the respective probabilities inferred from 10 000 simulated genotypes. As these individuals could have been potential immigrants, they were excluded from the following analyses, which left 32 genotyped animals in sample L-2000 and 54 in each of the other two samples. Two potential immigrants were detected in sample S-2000, three in S-2001 and one in S-2002. The final sizes of the three samples from S were 47, 56 and 27 animals, respectively.
Estimation of effective population size
The effective population size (Ne) at site L was difficult to estimate with both methods due to the weak drift observed between the temporal samples. The F-statistic-based estimator yielded a value of positive infinity for the mean and limits of the 95% CI of 144–positive infinity (Table 2). Using the two largest samples L-2001 and L-2002, the likelihood-based estimator gave concordant results with a basically flat log-likelihood distribution for Ne values larger than c. 250 (data not shown). The largest log-likelihood was obtained for Ne = 7750, although not too much importance should be attached to this result, as the log-likelihoods were very similar across a wide range of Ne values. The lower limit of the 95% CI was estimated as Ne = 160 (Table 2). The upper limit was not contained in the interval of Ne values [50, 8000] for which log-likelihoods were evaluated. When all three temporal samples were included, tighter estimates were obtained with a mean of 190 and a CI of 90–850 (Table 2).
Table 2. Effective size of local C. violaceus populations in a large (L) and a small (S) forest fragment. For the likelihood-based estimator, the mode (m) and the 95% confidence interval (CI) are given. The analysis was performed twice using the two largest samples (2 samples) and all three temporal samples available per location (3 samples), respectively. For the F-statistic-based estimator, m is the mean and CI the 95% CI.
*Very similar log-likelihood values for a wide range of Ne values.
Likelihood (2 samples)
Likelihood (3 samples)
Likelihood (2 samples)
Likelihood (3 samples)
At site S, the drift signal was considerably stronger. The F-statistic-based estimator gave a mean Ne of 59 and a 95% CI of 24–499 (Table 2). The point estimate obtained with the likelihood-based approach using the two largest samples was a bit higher with a maximum log-likelihood value for Ne = 140 (Table 2). The lower limit of the 95% CI was estimated as Ne = 50. Again, the upper limit was outside the interval of Ne values [50, 2500] for which the log-likelihoods were evaluated. The results did not become clearer when all three temporal samples were included in the analysis. In this case, the log-likelihood distribution was flat for Ne values larger than c. 150 (data not shown). The maximum log-likelihood was observed for Ne = 530 but it is important to note that the log-likelihoods were very similar across a wide range of Ne values. The lower limit of the 95% CI was at Ne = 120, while the upper limit was not contained in the interval between 50 and 2500.
Detection of population growth or decline
The Brooks, Gelman and Rubin convergence diagnostic was close to 1 for log r and log tf at both sites, indicating convergence of the individual chains. In both samples, we detected a clear signal of population decline (Fig. 2), as none of the intervals between the 2.5%- and 97.5%-quantiles included positive values for log r. For sample L, the modes of the posterior distributions for log r were between −1.52 and −1.43 indicating a c. 30-fold reduction of Ne (Fig. 2a). In sample S, the size contraction may have been even more severe. The modes of the posterior distributions for log r were between −2.03 and −1.85, which suggested a 100-fold reduction of Ne (Fig. 2b). In sample L, the average mode of the posterior distributions for log tf was −0.33 and the average 2.5%- and 97.5%-quantiles −0.66 and 0.07, respectively. In sample S, the mode was −0.14 and the two quantiles −0.4 and 0.11, respectively. The marginal posterior distributions of log tf obtained in the different MCMC runs are plotted in Fig. 3. In order to obtain an estimate for the onset of the population reduction ta = tf × N0 in the two fragments, we used the point estimates for tf, and approximated the present effective population size N0 with the 95% confidence limits of our Ne estimates (see Table 2). We used the lowest of the three estimates of the lower 95% confidence limit available for each study site. This analysis suggested a start of the population reduction between 84 and 795 years ago in fragment L. For population S, the size contraction was estimated to have started at least 35 and not more than 723 years ago.
Our results are in agreement with the hypothesis that habitat fragmentation due to major roads may have severe detrimental effects on populations of the flightless ground beetle C. violaceus. A dramatic 100-fold reduction of Ne was detected in the small forest fragment S (Fig. 2b). A similar demographic history was likely for the population in the larger fragment L, although the size reduction may have been slightly less severe (Fig. 2a). Our estimates of the time of onset of these contractions are consistent with the hypothesis that the population decline occurred in the recent past. Although the present analysis clearly cannot pinpoint the start of the population change, it can still provide a lower bound which agrees well with known historical disturbances of the habitat. Thus, our results are consistent with the hypothesis that road constructions have divided a formerly much larger population into several isolated subunits. Of course, it cannot be excluded that other factors such as, e.g. changes in forest management could be responsible for the observed population decline (Niemelä, 1997). The recent population contraction has led to a low effective size in the small fragment S (Table 2). It is remarkable how well the estimated magnitude of the population declines in the two fragments seems to correlate with the respective forest fragment sizes. Indeed, the estimated population decline is about three times stronger in fragment S, whose area is approximately three times smaller than that of the large fragment.
The estimation of Ne in population L was difficult as is generally the case in large populations where genetic drift is weak and allele frequency changes between temporal samples are mainly due to sampling errors (Waples, 2002). The tightest estimates could be obtained from the maximum-likelihood analysis based on all three temporal samples, which gave a 95% CI of 90–850. This result indicates that Ne is unlikely to be critically small in L although it may be lower than might be expected given the large size of the fragment. However, it has to be kept in mind that what we estimated was the effective size of a local population whose distribution did not necessarily cover the entire fragment. Indeed, previous results showed that the beetles within one fragment did not constitute a single random-mating population, i.e. some weak but significant genetic structure was observed (pairwise FST values between different sites within fragments 0.0045–0.0211; Keller & Largiadèr, 2003a). Such a population subdivision within fragments could, for example, be due to habitat heterogeneity leading to a patchy distribution of favourable areas, or to weak isolation by distance within fragments. Nonetheless, it seems reasonable to assume that individual subpopulations occupy a relatively large area considering the mobility of large ground beetles, which may be capable of movements across more than 200 ha in an unfragmented habitat (den Boer, 1990b). Indeed, we detected a clear connection between fragment size and Ne, which would not be expected if the area occupied by individual subpopulations were considerably smaller than the fragments.
The effective size of population S was found to be lower than that of population L, with a mean of 59–140 (Table 2). The maximum-likelihood estimation based on all three samples gave a mean of Ne = 530. However, the log-likelihood values were very similar across a wide range of Ne values. This is in contrast to the analysis based on the two largest samples only, where the log-likelihood curve showed a distinct peak. The finding that the inclusion of an additional temporal sample did not lead to tighter Ne estimates may be due to the very limited size of sample S-2002, which consisted of only 27 individuals.
The two local populations from fragments S and L both appear to receive immigrants from other populations. Such individuals immigrating from genetically differentiated populations can introduce a bias into the estimation of effective population sizes (Waples, 1989), although this effect was found to be rather small provided that the sampling intervals were relatively short (Wang & Whitlock, 2003). To deal with this problem, we removed beetles with deviating genotypes from the analyses (see Material and methods for details). As expected based on theoretical considerations, this resulted in more conservative Ne estimates (i.e. slightly larger point estimates and somewhat wider CIs) than the same analysis based on all sampled individuals(e.g. F-statistic-based estimate for site S using all individuals: mean = 55; 95% CI = 24–290). As there is a certain amount of genetic structure within each fragment (see above) animals with deviating genotype probabilities may originate from other parts of the same fragment and need not necessarily be immigrants from outside. Further, some of these individuals might not even be true immigrants at all but could, by chance, have genotypes which are rather untypical for their population of origin.
From the discussion above, it becomes clear that a very accurate estimation of Ne in natural populations is difficult. However, it still seems reasonable to conclude that Ne is considerably smaller in local population S than in L, as was suspected from previous analyses, which had detected a reduction of neutral genetic variability in the small fragment (Keller & Largiadèr, 2003a). Some substructure may still be retained in the small fragment as indicated by the fact that site S was significantly differentiated from a second sampling location within the fragment (pairwise FST = 0.021; Keller & Largiadèr, 2003a). It is unsure if this will be sufficient to ensure the long-term survival of C. violaceus in fragment S, in view of the fact that the highway probably prevents significant immigration from other, larger fragments (Zangger, 1995; Keller & Largiadèr, 2003a), which would restore genetic variation or buffer population fluctuations (den Boer, 1990a; Ingvarsson, 2001).
In a previous study, we had detected a pairwise FST value (Weir & Cockerham, 1984) of 0.019 between study sites S and L (Keller & Largiadèr, 2003a). Taking into account the additional loci used in this study, we obtained a similar value of 0.020 with a 95% CI of 0.012–0.028. Using eqn 12 from Gaggiotti & Excoffier (2000), we assessed the correspondence between this observed value and the expectation for given (unequal) population sizes and divergence times. We assumed that the two populations had been separated by the construction of the highway c. 30 generations ago. To investigate the range of plausible values, the expected FST was calculated both with the minimum (i) and maximum (ii) Ne values estimated for the two populations. In case (i), Ne was 90 in population L and 24 in population S (Table 2). In case (ii), the effective sizes were assumed to be 850 for population L and 499 for population S (Table 2), which resulted in expected FST values of 0.21 for (i) and 0.022 for (ii). This indicates that our observed value of 0.020 is not entirely incompatible with a model of 30 years of complete isolation but still very much at the lower end of what would be expected in this particular setting. We suspect that this discrepancy can largely be explained by the fact that the population structure of C. violaceus may not correspond very well to the model used by Gaggiotti & Excoffier (2000), which assumed that an ancestral population had been split into two completely isolated populations, whose sizes add up to that of the ancestral population. Local populations S and L are unlikely to be completely isolated but rather appear to receive migrants from other demes, as indicated by the results of the assignment tests. The fact that, on average, pairwise FST values tended to be higher between samples from opposite sides of major roads than between samples from within the same forest fragment (Keller & Largiadèr, 2003a), indicates that this exchange of migrants could be disrupted by the roads and, today, largely restricted to individual fragments. In such a complex system, no precise expectations of pairwise FST values can presently be computed. However, these expected values would clearly be lower than in the case of completely isolated populations and might correspond better with our observed value.
If we combine all our results on population structure, genetic variability (see Keller & Largiadèr, 2003a), effective population size and population size changes, these are in good agreement with the hypothesis that roads are such strong barriers to dispersal in C. violaceus that they have led to the subdivision of a once continuous population into relatively isolated fragments. It is certainly possible that other factors may have contributed to the observed patterns. This is particularly true for the population decline, whose onset cannot be determined precisely. Although the estimated time frame coincides with road constructions in the area, we cannot exclude the possibility that the population decline was caused by other unknown historical events.
In agreement with our hypothesis that fragmentation due to roads may lead to critically small and isolated C.violaceus populations, den Boer (1990a,b) found that the distribution of population sizes in other flightless ground beetles was skewed towards larger values, probably because many small populations had disappeared. Thus, it may be all the more important to focus on the conservation of large, viable populations by preventing further fragmentation of continuous forests. Our results further emphasize that future efforts to improve the connectivity of fragmented habitats should consider the needs of invertebrates, for example by providing suitable microhabitats on wildlife overpasses (Zangger, 1995).
We thank M. von Fischer and the Burgher Community of Bern for the permission to collect samples, and C. Zurbrügg, E. Keller and M. Ulrich for help in the field. G. Rigoli and J.-C. Nicod provided assistance in the lab and P. Berthier help with computations. The authors also thank T. Giger, G. Hamilton, G. Heckel, W. Nentwig, S. Neuenschwander and G. Reckeweg for helpful discussions, as well as M. Beaumont and an anonymous reviewer for their comments on a previous version of the manuscript. This study was financed by the Swiss Federal Office for Education and Science in the framework of Cost 341.