Inferring local adaptation from QSTFST comparisons: neutral genetic and quantitative trait variation in European populations of great snipe


Stein Are Sæther, Netherlands Institute of Ecology (NIOO-KNAW), Centre for Terrestrial Ecology, PO Box 40, 6666 ZG Heteren, The Netherlands.
Tel.: +31 26 4791111; fax: +31 26 4723227;


We applied a phenotypic QST (PST) vs. FST approach to study spatial variation in selection among great snipe (Gallinago media) populations in two regions of northern Europe. Morphological divergence between regions was high despite low differentiation in selectively neutral genetic markers, whereas populations within regions showed very little neutral divergence and trait differentiation. QST > FST was robust against altering assumptions about the additive genetic proportions of variance components. The homogenizing effect of gene flow (or a short time available for neutral divergence) has apparently been effectively counterbalanced by differential natural selection, although one trait showed some evidence of being under uniform stabilizing selection. Neutral markers can hence be misleading for identifying evolutionary significant units, and adopting the PSTFST approach might therefore be valuable when common garden experiments is not an option. We discuss the statistical difficulties of documenting uniform selection as opposed to divergent selection, and the need for estimating measurement error. Instead of only comparing overall QST and FST values, we advocate the use of partial matrix permutation tests to analyse pairwise QST differences among populations, while statistically controlling for neutral differentiation.


It is a long-standing debate in evolutionary biology whether isolation is sufficient, necessary or only helpful for populations to diverge, and whether natural selection can generate divergence in the face of gene flow (Mayr, 1963; Endler, 1977). An increasing number of studies from a diverse range of taxa are highlighting morphological divergence among populations in the absence of differences at neutral genetic polymorphisms, suggesting that local adaptations can evolve over ecological timescales and/or in the absence of population isolation (e.g. Karhu et al., 1996; Kinnison & Hendry, 2001; Piertney et al., 2001; Wilding et al., 2001; Koskinen et al., 2002; Irwin et al., 2005). Spatial variation in selection regimes might therefore be indicated by geographic variation in functional genes, morphology and/or life-history traits, which may prove independent of spatial patterns of neutral diversity (Ekblom et al., 2007). However, if the extent of geographic differentiation is similar to that of selectively neutral genes, random genetic drift is sufficient for explaining the pattern without invoking differential selection in different areas. Therefore, it is important to have a baseline of neutral genetic structure as a null model when making inferences about possible local adaptations in, e.g. quantitative traits.

An analysis of intraspecific geographic variation in selectively neutral genetic markers has provided insight into many aspects of population biology, including gene flow between populations, historical demographic events, colonization history and phylogeography, and the identification of conservation units (Avise, 1994, 2000). A growing number of studies are describing substantial levels of geographic structure in bird species, brought about by limited gene flow caused by processes, such as natal philopatry and territoriality (Avise & Ball, 1991; Piertney et al., 1998, 1999). Thus, it cannot be taken for granted that any species will show low levels of geographic structure simply because it is, or has recently been, continuously distributed.

Recently, a number of studies have compared FST and QST estimates for the same populations to make inferences about local adaptation (Merilä & Crnokrak, 2001; McKay & Latta, 2002). FST (and related statistics) measures the extent of population structuring of genetic variation, and QST (Spitze, 1993) is a precisely analogous measure of differentiation in quantitative genetic traits. This implies that when FST = QST there is no evidence for geographically varying natural selection, whereas if QST > FST drift-migration balance cannot explain the entire pattern, and if QST < FST there is evidence for uniform natural selection across populations (Rogers, 1986; Lande, 1992; Whitlock, 1999; Merilä & Crnokrak, 2001; McKay & Latta, 2002). Hendry (2002) draws attention to some potential problems with the assumption that QST and FST should be equal under neutrality.

QST is the among-population proportion of the total additive genetic variance, and not phenotypic variance, of a quantitative trait. It is therefore ideally measured in a randomized ‘common garden’ experimental design to exclude the effect on the trait of environmental differences between populations (with appropriate design to partition the within-population variance, and preferably also to remove nonadditive effects). However, for many organisms it is not possible or practically feasible to conduct laboratory rearing experiments, but phenotypic as well as molecular data may be available from several populations. Fortunately, using realistic assumptions about the additive genetic components of variation within and among populations in lieu of proper quantitative genetic data, it may still be possible to say something about the extent of geographical differentiation in quantitative traits when compared with what is expected from neutral geographic differentiation (Barrowclough, 1980; Rogers & Harpending, 1983; Prout & Barker, 1993; Spitze, 1993; Podolsky & Holtsford, 1995; Kremer et al., 1997; Merilä, 1997; Smith et al., 1997; Storz, 2002; Saint-Laurent et al., 2003; Bernatchez, 2004; Roseman, 2004; Østbye et al., 2005), provided that a sensitivity analysis is undertaken of the assumptions. We may call this the PSTFST approach (P symbolizing ‘phenotypic’-QST, or ‘pseudo’-QST if one so desires). Often, general patterns are very similar when comparing FST with QST based either on phenotypic or genetic variance (Lynch et al., 1999; Schluter, 2000). One improvement to the phenotypic estimation of QST might be to derive conservative (minimum) estimates of QST, by measuring the repeatability of suitable traits (e.g. annually re-grown traits) to obtain maximum estimates of the genetic within-population variance component. Note that both the PST approach and common garden QST investigations are sensitive to deviations from purely additive gene action. Epistasis can potentially mask some effects of divergent selection (Whitlock, 1999), and nonadditive neutral gene action can sometimes result in patterns falsely suggesting selection, i.e. deviations from FST (López-Fanjul et al., 2003). However, these problems are probably most relevant to traits, such as life history, rather than to morphological traits that typically show substantial additive genetic variance (Crnokrak & Roff, 1995; DeRose & Roff, 1999; Merilä & Crnokrak, 2001; López-Fanjul et al., 2003).

In this study, we analyse geographical variation in microsatellites and morphological traits among distributional regions and among populations within these regions of the great snipe (Gallinago media), a migrating lekking shorebird. In western Europe (Scandinavia), the great snipe is a scarce inhabitant of earthworm-rich mountain fens around the tree line (Kålås et al., 1997a), whereas in eastern Europe, the great snipe is patchily distributed predominately in lowland meadows subject to annual flooding (flood plains), eastwards of the Yenisey (Fig. 1; Gromadzka et al., 1985; Tomkovich, 1992; Kuresoo & Leibak, 1994). Hence, habitat differences may predict different selection regimes in the two disjunct regions, whereas differences among populations within the regions are more likely to be affected by drift and migration alone.

Figure 1.

 Sampling locations for great snipe populations included in this study. Shaded areas roughly indicate the breeding distribution, but breeding populations occur very fragmented within these areas (especially in the westernmost part of the eastern region).

Previously, these two distributional regions were more closely, if not completely, connected. Up until the mid-19th century, great snipe were breeding over large parts of the lowlands also of western Europe (Germany, Denmark, southern part of the Scandinavian Peninsula and Finland), occupying a similar habitat as is still present in eastern Europe. The great snipe is now extinct in lowland western Europe, mainly because of the extensive man-made transformation of suitable habitat for agricultural purposes. The remaining western population (in the Scandinavian mountains) is estimated to be in the range of 10 000–30 000 males at present (Gjershaug et al., 1994). Although there is little information about the size of the eastern population, it is clear that it has recently declined (Gromadzka et al., 1985; Panchenko, 1985; Tomkovich, 1992; Kuresoo & Leibak, 1994). The species is currently classified as ‘Near Threatened’ at a global level (BirdLife International, 2000).

Here, we first analyse the patterns of genetic structure derived from microsatellite DNA. Second, we compare variation in morphological traits (body size measures and a secondary sexual trait) among great snipe populations, to the variation expected under neutrality. If gene flow is limited, we may expect adaptation to local conditions if selection is sufficiently strong. Such local adaptation may, on the other hand, be swamped by extensive migration between populations. If the geographical structure of quantitative traits were more pronounced than the geographical structure of neutral genetic markers, differential natural selection among populations, rather than drift alone, would have to be invoked to explain differences between populations.

Our approach was to compare FST with a pseudo-QST measure derived from estimates of within- and between-population variance components of phenotypic traits. This approach is largely similar to those of Merilä (1997), Storz (2002) and Saint-Laurent et al. (2003). As quantitative trait data were purely phenotypic, assumptions about the additive genetic components of variance had to be made to be able to directly compare the magnitude of QST and FST estimates, and we performed sensitivity analyses of these assumptions. For traits re-grown annually, we measured in one population the between-year repeatability to obtain a maximum estimate of the genetic proportion of the within-population variance component of the trait (Falconer & Mackay, 1996; Lynch & Walsh, 1998; but see Dohm, 2002 and Discussion) to derive a conservative (minimum) estimate of QST. We took various steps to ensure that measurement error and measurer bias did not inflate our estimates.

Materials and methods

Collection of samples and morphological measurements

We caught great snipe with mist nets on leks distributed from Nordland, Norway, in the north to Biebrza, Poland in the south (Table 1, Fig. 1). Morphological measures of tarsus (true tibio tarsus length), total head (bill plus head), bill (to end of skin), bill to nostrils, wing length (maximum flattened) and tail white (the length of white on the outermost tail feather) were done according to Höglund et al. (1990b). A few 1-year-old birds had not yet moulted their juvenile tail feathers, and measurements of these feathers were excluded. All measurements were taken with digital callipers to the nearest 0.1 mm except wing length, which was measured to the nearest 1.0 mm using rulers with a riveted right angle stop. We used mean values of traits measured more than once for the same individual, and mean values of bilateral traits (tarsus, wing length and tail white), avoiding pseudoreplication and lowering the influence of measurement error.

Table 1.   Sampling locations for great snipe included in this study.
PopulationLocalityCountryGeographic locationnYear
  1. Within each locality, birds were sampled at 1–10 leks. Sample sizes for microsatellite DNA are indicated (samples were larger for morphological traits). In the NT population, samples were pooled from the three locations.

NTNamsskoganNorway64°53′N 13°12′E131995
 RøyrvikNorway64°55′N 13°28′E 1995
 HattfjelldalNorway65°40′N 13°48′E 1996
RoRørosNorway62°43′N 11°30′E301996
GaGåvåliaNorway62°17′N 09°36′E221993–1996
RiRindalNorway63°09′N 09°17′E121995–1996
VaValdresNorway60°53′N 09°01′E181996
PoBiebrzaPoland53°31′N 22°38′E121994
EEKärevere/TartuEstonia58°25′N 26°31′E471996, 1998
EWRoudeEstonia58°43′N 23°50′E161998

All birds were measured by JAK except in Gåvålia (JAK 86%, PFI 8%, SAS 6%), Nord-Trøndelag/Nordland (PFI 100%) and Poland (SAS 100%). The noise introduced by differences between these observers was generally small, as inferred from a sample of birds measured independently by several persons in the same year (Table 2). Measurement error was low (generally 1–4%) and repeatability high both within and among observers (Table 2), apart from wing length, which was excluded from geographical analyses. However, slight systematic differences between observers could still potentially bias QST estimates, in particular as two populations were not measured by JAK. These problems were overcome first, by the exclusion of wing length in geographical analyses, second by adjusting the tarsus measures of PFI to those of JAK by subtracting the mean difference of pairwise measurements between PFI and JAK (0.44 mm), and third by using a principal component measure of bill measures with high repeatability that did not show systematic difference between measurers. The first axis of a principal component analysis (conducted separately for males and females) of total head, bill and bill to nostrils captured 91.36% of the variation in these traits, which were all highly correlated. As great snipe males are much more likely to be caught on leks than are females, sample size for females was small in most populations, and females were hence excluded from most morphological analyses because there are systematic size differences between the sexes (Höglund et al., 1990b).

Table 2.   Repeatabilities and measurement errors of morphological traits of great snipe males measured on different dates in the same year in the Gåvålia population, either by the same person (JAK; within-observer repeatability), or by two or three persons (JAK, PF and SAS; between-observer repeatability) (all values of P < 0.001).
TraitR*Fd.f.ME† (%)
  1. *Repeatability.

  2. †Measurement error.

  3. ‡First principal component, accounting for 95.27% of the variation in total head, bill and bill to nostrils in this sample.

Within observer
 Total head0.96150.2146,473.91
 Bill to nostrils0.96353.0746,473.71
 Mean tarsus0.96352.5946,473.77
 Mean wing0.82110.2045,4617.87
 Mean tail white0.990197.3547,481.01
Between observers
 Total head0.96356.0839,453.71
 Bill to nostrils0.96966.9940,463.11
 Mean tarsus0.88216.8740,4611.83
 Mean wing0.8069.7536,4119.40
 Mean tail white0.96458.3742,483.55

As QST and FST are attributes specific to the populations and cohorts compared, finding evidence for spatial structure might alternatively be an artefact of temporal variation if populations are not sampled at the same time. Our genetical data were collected over a short time span (Table 1). Morphological data were collected over a longer time span in one of the populations (GA, 1986–2003), rendering temporal effects more problematic. General linear mixed models revealed very low effects of year sampled in this population (data not shown), as well as of age (except for wing length), justifying pooling data from different years and ages. Age was estimated at first capture as either 1-year old or older using feather wear (Sæther et al., 1994). Limiting the analyses to the same years as for genetical data or to measures of more than 1-year-old males had only minor effects on estimates of variance components (overall QST values changed in the third or fourth decimal) and did not change any conclusions (data not shown). We therefore chose to include all years, and use mean values of individuals irrespective of age. We chose not to present detailed analyses of geographical variation in mean wing length as this trait showed large measurement error within season (both within and between observers), as well as age-related variation (and we did not have accurate age estimates for all birds in all populations to remove this effect).

Microsatellite genotyping

Five hypervariable tetranucleotide microsatellites were isolated using an enrichment protocol similar to that of Piertney & Dallas (1997) and Piertney et al. (1998). Individuals were genotyped at these loci (SNIPE B2, 3, B5, 12, 20; primers described in.The 10-μL PCR mixture contained approximately 10 ng of DNA, 1 μL of 10× buffer without MgCl2 (MBI Fermentas, Ontario, Canada), 1 μL of 25 mm MgCl2, 1 μL of 2.5 mm dNTPs, 0.5 μL of each of 10 μm forward and reverse primer, 0.25 units of Taq-polymerase (MBI Fermentas) and 5 μL of ddH2O. PCRs were performed on a GeneAmp PCR System 9600 (Perkin Elmer, Waltham, MA, USA) using the following conditions: initial denaturation 94 °C for 3 min; c cycles of denaturation at 94 °C for 30 s, annealing at a °C for 30 s, extension at 72 °C for 40 s (c = number of cycles, a = annealing temperature); then a final extension step at 72 °C for 2 min. For locus SNIPE B2, c = 30 and a = 51; SNIPE 3 and SNIPE 20, c = 32, a = 52; SNIPE B5, c = 31, a = 51; SNIPE 12, c = 30, a = 56.

Aliquots (∼3.5 μL) of the PCR products were separated on denaturing 6% polyacrylamide gels (Sambrook et al., 1989). PCR products with shorter fragment sizes (140–200 bp; loci B2, 3 and 12) were run on the gels for at least 1 h, whereas PCR products with longer fragment sizes (300–350 bp, loci B5 and 20) were run for at least 1.5 h. After electrophoresis, the PCR products were visualized by silver staining (Sambrook et al., 1989). Individuals were assigned genotypes by comparison with a standard set of samples of known allele size. The microsatellite sequences obtained in this study are deposited at GenBank under the accession numbers, AY363298AY363302.

Data analyses

Microsatellite DNA

None of the females was heterozygous for the microsatellite loci, SNIPE 3 and SNIPE 12. These are therefore probably located at the Z chromosome and data from females for these loci were excluded from the following analyses. Each locus in each population was tested for deviations from Hardy–Weinberg equilibrium, and the probability of deviations from Hardy–Weinberg equilibrium for all loci combined in each population was calculated according to Fisher's method for combining probabilities (Sokal & Rohlf, 1981). The presence of linkage disequilibrium was also tested for each pair of loci in each population. These tests were done using Genepop on the Web (; Raymond & Rousset, 1995). Each locus was also tested for the proportion of multistep mutations vs. single-step mutations with the program MISAT (Nielsen, 1997).

To check for evidence of recent bottlenecks we used the program Bottleneck 1.2.02 (Cornuet & Luikart, 1996). We chose to use a Wilcoxon test under the assumptions that all loci fit the stepwise mutation model, or that all loci fit a two-phased mutation model with the proportion of multistep mutations found by the program MISAT.

The genetic structuring of populations was examined by a hierarchical analysis of molecular variance (amova; Excoffier et al., 1992) computed with Arlequin 1.1 (Schneider et al., 1997). Variance was partitioned between eastern (Estonian and Polish populations) and western (Norwegian) populations, between populations nested within these two groups, and among individuals within populations. Pairwise population differentiation was calculated based on FST (Weir & Cockerham, 1984) using Genetix 4.05 (Belkhir et al., 1996). We also estimated RST (Slatkin, 1995), and did analyses using both FST and RST to ensure that conclusions did not depend on the choice of differentiation statistic. RST is a measure of genetic differentiation based on the stepwise mutation model, and is often more appropriate for microsatellites as differentiation might be underestimated by FST if mutations create allelic homoplasy and mutation rate is high relative to the migration rate. However, if mutation rates are low relative to migration rates FST can be expected to provide more accurate estimates of genetic differentiation than RST (Slatkin, 1995). As an estimator of RST we used Goodman's unbiased ρ. RST was calculated using RstCalc 2.2 (Goodman, 1997) after standardizing allele sizes to a global mean of zero and unit standard deviation, and after averaging variance components over loci. P-values of global RST estimates over all populations or over regions were obtained by permutation tests, and approximate 95% confidence intervals by the range of the central 95% of 1000 bootstrap estimates.

Note that differentiation estimates below zero are most likely because of sampling variation (and not because alleles from different populations actually are more similar to each other than the alleles within the same population). The best estimate for negative values would therefore be zero. We did not adjust negative pairwise estimates to zero, as this would have created a bias when sampling variation of positive estimates is not similarly adjusted.

Isolation by distance was tested with Mantel tests assuming a linear relationship between pairwise values of FST/(1 − FST) and the natural logarithm of geographic distances (km) between all population pairs (Rousset, 1997). Geographic distances were calculated following the Earth's curvature, using the GeoDistances module in R 4.0 (Casgrain & Legendre, 2001).

Quantitative traits

Repeatability [varbetween/(varbetween+varerror)], measurement error (1 − repeatability) and pairwise QST values were estimated using the VARCOMP procedure in SPSS 11, applying the anova (type III sum of squares) approach. Maximum likelihood-based estimates were very similar (data not shown). Repeatability as a maximum estimate of heritability is also reported corrected for the separately estimated within-observer measurement error (Lynch & Walsh, 1998) as


QST was estimated as


where g is the assumed additive genetic proportion of differences between populations, h2 (narrow-sense heritability) is the assumed additive genetic proportion of differences between individuals within populations, varpopulation is the observed between-population variance component and varerror is the observed within-population variance component. A sensitivity analysis was performed, simulating different values of g and h2 (including the corrected between-year repeatability of tail white), but the pairwise QST values used are those obtained assuming g = 1 and h2 = 0.5, unless otherwise stated. The advantage of using these particular assumptions is that significance testing of the estimate (or rather the between-population component) can then be conducted by standard methods of analysis of variance.

Nested analyses, partitioning the variation in morphology among regions and among populations within regions, were performed using procedure GLM in SPSS 11 and associated variance components estimated with procedure VARCOMP. Simple and partial matrix permutation tests were performed using R 4.0 (Casgrain & Legendre, 2001). Statistical analysis of whether pairwise QST values involving populations in different regions were larger than those within regions were conducted by calculating the standardized Mantel statistic (rM) between a distance matrix A of pairwise QST values and a matrix B of kind of comparison (within or between regions). Matrix permutation on A 10 000 times was then used to obtain a randomization P-value of the null hypothesis of no difference. To test if among-region QST values were larger than within-region values when controlling for neutral genetic variation, a distance matrix C of pairwise FST or RST values was also constructed and the partial rM (Smouse et al., 1986) computed between A and B while controlling for C. A randomization P-value was obtained by matrix permutation of A, holding B and C constant. Matrix permutation of B instead of A while controlling for C gave the same conclusions (results not shown).


Microsatellite DNA

The number of alleles at the five loci ranged from two to 19 (SNIPE B2 = 13 alleles; SNIPE 3 = 10 alleles; SNIPE B5 = 19 alleles; SNIPE 12 = 10 alleles; SNIPE 20 = 2 alleles). The mean number of alleles per locus was between 5.6 and 9.0 in the populations.In the western Estonian population locus SNIPE 20 was monomorphic, whereas all other populations were polymorphic for all loci. We found statistically significant deviations (heterozygote deficiencies) from Hardy–Weinberg equilibrium for one or more loci in five of the eight examined populations (data not shown). After combining probabilities for all loci in each population, four populations (eastern Estonia, Rindal, Gåvålia and Røros) showed evidence of deviations from Hardy–Weinberg equilibrium (P < 0.05). However, none of these remain significant after Bonferroni adjustment of the α level for the number of populations. There was no consistent linkage disequilibrium between pairs of loci across populations.

The maximum likelihood tests for proportion of multistep mutations (pmm) vs. single-step mutations indicated a very low proportion of multistep mutations (pmm ≤ 0.01 for all loci), indicating that the use of R-statistics is appropriate (Nielsen, 1997). There was no evidence of a recent bottleneck in any population (data not shown) although the power of these tests is low with only five loci. The observed heterozygosity did not deviate from what could be expected under a strict stepwise mutation model, or a two-phased mutation model assuming that 5% of the variation in allele size is attributable to an infinite allele model, and 95% to a stepwise mutation model.

An analysis of molecular variance confirmed that great snipe are weakly structured into one western and one eastern group. There seems to be no variation among populations within these groups, and most of the variance is accounted for within populations (Table 3). Pairwise RST and FST estimates indicate low, but significant, population differentiation between the eastern Estonian population and most Norwegian populations (Table 4). There was a strong correlation between pairwise FST and RST estimates (Mantel test, rM = 0.859, P < 0.001).

Table 3.   Hierarchical analysis of molecular variance (amova) for eight great snipe populations categorized into two regions (Norway and Estonia/Poland) (a) based on weighted average R over five microsatellite loci, (b) based on microsatellite allele frequencies.
Source of variationd.f.Sum of squaresVariance componentsPercentage of variationP-valueRST
 Among regions151.72 0.263 2.410.0530.024
 Among populations  within regions651.81−0.055−0.50.46−0.005
 Within populations3323559.0510.7297.51  
Source of variationd.f.Sum of squaresVariance componentsPercentage of variationP-valueFST
 Among regions12.6550.0171.720.150.017
 Among populations within regions60.73−0.023−2.330.99−0.023
 Within populations332335.691.01100.6  
Table 4.   Population pairwise RST estimates from microsatellite variation (above diagonal) and pairwise estimates of Weir–Cockerham FST (below diagonal).
 Eastern regionWestern region
  1. Significant values are highlighted in bold, and comparisons involving populations in different regions are highlighted in italics. Population abbreviations as in Table 1.

Po 0.096−0.0360.0290.1620.0820.0680.058
EE0.055 0.0950.0860.0320.1040.0920.084
EW−0.0160.052 0.0920.1920.1310.1090.102
Ri0.0270.0490.033 0.052−0.0050.0030.009
Ga0.0900.0110.1040.040 0.0630.0730.107
NT0.0230.0260.0320.0000.019 −0.0310.003
Ro0.0090.0310.0240.0050.021−0.018 −0.002

The global estimates of divergence over all populations were (±bootstrap 95% confidence limits): RST = 0.059 (±0.041, P < 0.001) and FST = 0.026 (±0.026, P = 0.007). The estimates over regions were: RST = 0.051 (±0.037, P < 0.001) and FST = 0.018 (±0.022, P = 0.012). The higher estimates of RST than of FST may indicate that stepwise-like mutations rather than drift alone have contributed to the differentiation.

Isolation by distance

We found some indication of a linear isolation-by-distance pattern when comparing population pairwise genetic and geographic distances. The pattern using Goodman's unbiased ρ as an estimator of RST (Mantel r = 0.356, P = 0.048) is shown in Fig. 2a. The Weir–Cockerham FST estimator indicated a somewhat weaker pattern (Mantel r = 0.224, P = 0.17; Fig. 2b). This isolation-by-distance effect, although small, may explain the small differentiation between regions, but it is hard to analyse whether there is an additional effect of region as there was no overlap in geographic distance.

Figure 2.

 Neutral genetic divergence increases somewhat with geographic distance, as shown in these isolation-by-distance plots using either RST (a) or FST (b) against the natural logarithm of distance in kilometres along the curvature of the Earth. Comparisons of populations in different regions are indicated by solid symbols, and comparisons of populations within the same region are indicated by open symbols (diamonds in the east and squares in the west). Also shown is the least squares linear regression.

Quantitative traits

Annually re-grown traits (tail white and wing length) showed moderate to high repeatability between years, indicating substantial heritability of these traits (Table 5). After correcting for measurement error, repeatability estimates were of similar magnitude (∼0.8) for both traits, although we found somewhat lower values for tail white in females. Limiting the analysis to adults (to remove potential noise introduced by age-related variation) only marginally increased repeatability of tail white, but increased the repeatability of wing length to 0.95 (Table 5).

Table 5.   Between-year repeatabilities of annually re-grown traits measured on the same individuals in two or more years irrespective of age, and in two or more years as adults only.
  1. *Repeatability.

  2. †Repeatability adjusted for within-observer measurement error, 1.009% for tail white and 17.868% for wing length.

  3. All P < 0.001.

 Tail white0.8230.83213.53382, 647
 Wing length0.6490.7906.01296, 510
 Tail white0.7200.7286.8393, 119
 Wing length0.6700.8165.5993, 119
(Adult males)
 Tail white0.8290.83713.78266, 438
 Wing length0.7760.94510.22194, 323
(Adult females)
 Tail white0.6890.6965.9867, 85
 Wing length0.7290.8887.6166, 84

Populations nested within region had only little influence on trait values, but different traits showed striking variation in the degree of differentiation among regions.Tarsus length and amount of white on tail showed very strong divergence among regions,whereas a composite measure of bill length (PC1) showed only very weak (but significant) differentiation. The divergence in tail white and tarsus appeared to be entirely independent of each other. Although there was a slight overall correlation between the two measures (r = 0.102, P < 0.001, n = 1502 males) this was an artefact of both traits differing among regions: within region, there was no correlation (Fig. 3).

Figure 3.

 Both tarsus length and tail white is larger in the eastern (black diamonds) than in the western region (open grey circles). The overall correlation between the two traits is an artefact of these differences. Within regions there is no relationship (least squares regression lines and 95% confidence limits shown, east r2 = 0.001, n = 88; west r2 < 0.001, n = 1416).

Pairwise QST estimates between populations show that comparisons involving populations in different regions often had large (and significant) QST values of tail white and tarsus, whereas comparisons within regions were small (often negative) and nonsignificant. No pairwise QST values for bill were significant. Matrix permutation tests confirmed that QST values were significantly larger between regions than within regions both for tail white (rM = 0.788, P = 0.018) and tarsus (rM = 0.736, P = 0.018), but not for bill (rM = 0.158, P = 0.218).

QSTFST comparisons

Partial matrix permutations of pairwise values showed that, even when controlling for the neutral genetic structure (or geographic distance), QST for both tail white and tarsus (but not bill) were higher for comparisons between populations in different regions than comparisons within regions (Table 6; Fig. 4). These results should be very robust to varying the assumptions of the proportion of additive genetic variance, because the relative difference between pairwise QST values in a matrix does not change much by varying the g and h2 parameters of that matrix. This was confirmed by partial matrix permutation tests using simulated QST matrices for a range of g and h2, and the observed RST and FST matrices (data not shown).

Table 6.   Partial matrix permutation tests of the relationship between pairwise differentiation in quantitative traits (A) and the kind of comparison (B, populations within or among regions), while controlling for pairwise differentiation in neutral genes or geographic distance (C).
Matrix AMatrix CrM (AB·C)*P-value
  1. *Partial Mantel statistic.

QST(tail white)FST0.7490.015
QST(tail white)RST0.7020.015
QST(tail white)ln(distance)0.6720.003
Figure 4.

 Population differentiation in quantitative traits (QST, assuming g = 1 and h2 = 0.5) in relation to neutral genetic differentiation (FST). For both tail white (a) and tarsus (b), comparisons among populations in different regions (solid symbols) are larger than expected from the neutral differentiation whereas this is not the case for comparisons within the regions (open symbols, diamonds in east and squares in west). A different pattern is found for bill length (PC1) where QST values among regions do not differ from within regions and are not larger than expected from neutral variation (c). Note the different scales on the QST axes. Dashed lines are expectations if QST = FST.

The overall divergence between regions in bill (QST = 0.01) was lower than, or similar to, the values of FST and RST (0.018 and 0.051, respectively, see above), whereas tail white and tarsus showed much stronger differentiation (QST = 0.568 and 0.416 respectively). The problematic wing length measure showed no indication of a divergence deviating from neutral expectations (QST = 0.043).

We recalculated QST values for different assumptions about heritability (h2, 0.25, 0.5, corrected between-year repeatability of tail white, and 1.0) and the magnitude of the additive genetic proportion of the between-population variance component (g, 0.05–1.0). This exercise showed that the conclusions are not sensitive to varying g and h2 even outside realistic parameter space (Fig. 5). Exceptionally small additive genetic proportion of the between-population variance have to be invoked to arrive at QST values comparable with neutral markers for tarsus (Fig. 5b) and in particular tail white (Fig. 5a). For the composite measure of bill length most simulations yielded lower QST values than expected from microsatellite differentiation, but these were often within the bootstrapped 95% confidence intervals of the neutral differentiation (Fig. 5c).

Figure 5.

QST sensitivity plots of varying the additive genetic proportion of between-population (g) and within-population (heritability) variance components. QST values are calculated treating the two regions as two populations. Estimates of neutral divergence is shown as horizontal lines ±95% bootstrap confidence limits (bold solid line for RST; dashed line for Weir–Cockerham FST). QST for (a) tail white and (b) tarsus is considerably larger than neutral genetic divergence for most parameter space, whereas QST for (c) bill (PC1) is similar or smaller.

For females, overall QST between regions (n = 513–523 in western region and 16 in eastern region) resembled male values (QST(tail white) = 0.545, QST(tarsus) = 0.454, QST(pc1 bill) = −0.007), but low sample size in most populations precluded calculation of pairwise values.


This study has highlighted that neutral genetic differentiation is not sufficient to explain geographic differentiation in some quantitative traits in great snipe, and suggests local adaptation to different habitats despite high gene flow. The two habitats coincide with distributional regions, but the trait divergence cannot be explained as an artefact of isolation by distance.


Any comparison of QST and FST to infer spatial variation in adaptations requires that these two measures are comparable and unbiased. Different ways of calculating subdivision for microsatellites (RST and Weir–Cockerham FST) gave similar results. Critical assumptions behind the QST estimates were investigated to ensure that conclusions did not depend on inflated values because of uncertainties about heritability and additive genetic proportion of differences between populations. Sensitivity analyses revealed that conclusions were very robust to variation in these parameters. It thus appears that a quantitative genetic common garden rearing scheme of great snipe to arrive at these conclusions would have been unnecessary. Moreover, measurement error was in general low and would not be expected to cause overestimation of QST values. The same person took most measures, but the deviating values of tarsus length in the NT population measured by another person suggest that our steps to eliminate observer bias were not fully adequate for this trait in this population (as indicated by the high within-region and low between-region pairwise QST values involving this population,.For the other traits, and for all other populations, there were no detectable systematic between-observer effects that could bias results. Excluding measures of tarsus in the NT population would have increased between-region QST estimates, but for the sake of being conservative we chose to keep it. It is important for any QST study to estimate measurement error (both within and among observers), and take steps to ensure that conclusions are not biased, e.g. because different persons measured traits in different populations. We urge all future studies of QST to include repeated independent measures of a subsample for this purpose.

Pairwise comparisons – avoiding some assumptions

By comparing pairwise QST values among populations either in the same or in different regions, and controlling for the neutral divergence among the same populations using partial matrix permutations, we could avoid relying on specific assumptions about g and h2 for our conclusions to hold. This is because – even if the absolute values of QST may, e.g. be overestimated – the difference in relative magnitude between pairwise comparisons either within or among the units of interest (regions, in our case) is less affected by these assumptions. We suggest that adopting this pairwise approach, together with a sensitivity analysis of varying g and h2 for the overall QSTFST comparison, allows for robust conclusions to be drawn using purely phenotypic data in lieu of common garden experiments. This highlights the advantage of sampling several subpopulations to do a more reliable analysis of local adaptation than only obtaining an overall estimate of QST.

Neutral differentiation

We found that great snipe populations are weakly structured across northern Europe. Microsatellite DNA markers detected a genetic division between western and eastern populations (Norwegian and Estonian/Polish samples respectively). This weak neutral genetic differentiation between the two regions might simply be an isolation-by-distance effect. Although it is possible that there is also an effect of region (see Fig. 2), it would be very hard to say, based on contemporary genetic variation, whether such an effect is because of the recent separation or because of more ancient restriction of gene flow between populations in different habitats. The recent fragmentation of the distribution – which is probably mostly because of habitat changes in the lowlands induced by humans during the 19th century (Kålås et al., 1997a) – has geographically separated the remaining lowland populations from the western mountain populations, but we cannot detect any genetic signature of this separation. In view of that, analyses of museum specimens from the now-extinct populations in lowland western Europe would be interesting.

Quantitative trait differentiation

Different traits showed different patterns of QST compared with FST. Birds from the eastern and western regions differed substantially more in both tail white and tarsus than expected from neutral genetic differentiation, but did not in bill length (see also Kålås et al., 1997b for morphological variation).

Divergent selection

Eastern birds had whiter tails than western birds. The amount of white in the tail has probably been subjected to sexual selection (Höglund et al., 1990a; Sæther et al., 2000). As birds in Poland and Estonia display at lower latitudes and earlier in the season (J.A. Kålås, S.A. Sæther, A. Kuresoo, L. Luigujoe, unpublished data), birds from these localities perform their displays under considerably darker light conditions. Hence, more extensive white in the eastern populations might be because of requirements of a more conspicuous signal there. It is thus possible that the difference in tail white between the regions represents a local adaptation to light conditions (in all populations males display during the night). Furthermore, it is more likely that western populations have evolved less white tails, rather than that eastern populations have become more white. This is because great snipe probably must have colonized Scandinavia from the south (-east), rather than vice versa, after the last glaciation. If so, our results suggest that it is the cost of maintaining extensive white in the western populations (because of, e.g. predation, see Höglund et al., 1992) – rather than the benefit of more white tails in the eastern populations – that has shifted the trade-off balance and is the ultimate cause of the differentiation, but further studies are needed to confirm this. Interestingly, also females had more white tails in the eastern populations.

Tarsus length also differed more than expected from neutral markers between regions, and showed little variation between populations within the regions. This could perhaps be because of the habitat differences between the western and eastern populations. In the east, great snipe occur largely in sites subject to annual flooding early in the breeding season (when males display at leks) and we may speculate that this has led to natural selection for longer legs than in the mountain populations. Unfortunately, we did not have samples from large parts of the very eastern distribution of great snipe. In particular, it would be interesting to compare the Scandinavian mountain populations with northern Russian ones that occur in similar habitat (Morozov, 1994). Our prediction is that those birds, despite being located further away, should have trait values more similar to Scandinavian than to Polish and Estonian birds.

Detecting uniform selection: limitations of the approach

The weak divergence among regions in pc1 (bill length) corresponded to a pattern expected from neutral differentiation, or was possibly lower. Optimal bill length in great snipe is likely to be influenced by the depth at which earthworms (their main food) occur, and the birds prefer habitat with a suitable balance between easier soil penetrability (wetter areas) and earthworms occurring closer to the surface (drier areas) (Løfaldli et al., 1992). Perhaps earthworms are sufficiently available at the same soil depth in the two regions to prevent divergence, or perhaps conditions at overwintering grounds in Africa are more important. We cannot exclude that there is in fact similar stabilizing selection on bill length in both regions. Given the low values of genetic population differentiation in this study, it would be very hard to statistically document QST < FST for these populations to convincingly show stabilizing selection on any trait across environments, although the overall region analyses suggest so (Fig. 5c). Given also the unknown magnitude of g and h2, and the difficulties involved in calculating standard errors for ratios, this illustrates one important limitation of the PST vs. FST method: by using this approach it would often be much harder to find evidence for uniform stabilizing selection than to find evidence for differential selection.

However, this shortcoming is also often shared by QSTFST comparisons involving common garden experiments. It is instructive to note that the few documented cases of QST < FST (Merilä & Crnokrak, 2001; McKay & Latta, 2002; Edmands & Harrison, 2003) often show large population differentiation in neutral genes. Also, as Hendry (2002) pointed out, when FST is approaching unity, it will be hard to show that QST is even larger. A related problem is that the maximum value of FST is in practice often less than unity (because of mutation), and that QST may be less constrained from reaching its maximum value under neutrality. Relative measures of between-population divergence, such as FST, is heavily affected by the within-population diversity and may therefore be poor measures of divergence for loci with high diversity such as microsatellites (Charlesworth, 1998; Hedrick, 1999). Any factor affecting the difference in within-population diversity, such as different degrees of inbreeding or demographic histories of bottlenecks, could therefore potentially result in different values of FST even if the absolute levels of divergence are similar (Charlesworth et al., 1997), and it seems likely that QST estimates will not be affected in a similar way. These problems must be traded against the straightforwardness of comparing dimensionless estimates of divergence at marker loci and quantitative traits.

The QSTFST approach may therefore be most useful: (a) for providing indirect evidence of divergent, rather than uniform selection; and (b) in situations with low-to-moderate neutral subdivision (because of gene flow or relatively recent isolation), rather than for populations separated a very long time ago.

Repeatability as maximum heritability in PST analyses

Without additive genetic variance there can be no evolutionary response to selection on a trait, hence QST may remain low despite different selection ‘pressures’ being present. It may therefore be important to establish if a trait under study is heritable. If analysed carefully, repeatability of a trait suitable for such an analysis might be considered an upper limit on how much additive genetic variation is present for the trait within a population (Falconer & Mackay, 1996; Lynch & Walsh, 1998). It is well acknowledged that the true narrow-sense heritability may be substantially lower than this upper limit, but using repeatability as an estimate of heritability in the calculations of QST may provide conservative (low) estimates of QST (i.e. conservative in the context of showing QST > FST). However, it is important that the repeatability estimate is not downward biased for this reasoning to hold true. The worry is that repeated measurements of a trait might not be comparable, and that measurement error will deflate estimates. These two problems, and a solution, may be illustrated by our measures of mean wing length.

At first sight, wing length may appear to show low repeatability compared with tail white. However, after correcting for the substantial measurement error, wing length shows very similar between-year repeatability to tail white (Table 5). (The relatively large measurement error of wing length is probably not only because of low accuracy of measurements, but may also be an effect of wing feather wear during the breeding season, Sæther et al., 1994.) Moreover, unlike tail feathers, 1-year-old great snipe have not yet moulted their wing feathers (Sæther et al., 1994), whereas adult birds have fresh wings that are on average longer. Hence, after further restricting the analyses to known adult males (Table 5) the among-year repeatability estimate turns out to be extremely high (0.94), suggesting a potentially large genetic component of the wing length variation among adult individuals. This highlights the importance of using comparable measurements when calculating repeatability and to take measurement error into account. If not, repeatability might in fact potentially underestimate heritability (Widemo & Sæther, 1999; Dohm, 2002) and thus overestimate QST instead of providing a conservative estimate.

General patterns of QST vs. FST in natural populations

In our study, different quantitative traits showed very different patterns of divergence. There is no reason, in general, to expect traits in linkage equilibrium to show a correlation of QST values beyond what is expected from the neutral genetic differentiation. QST is therefore an attribute of the particular trait, and cannot be said to characterize the populations as such, unlike FST ideally would. The pattern observed of QST usually exceeding FST in published studies (Merilä & Crnokrak, 2001; McKay & Latta, 2002) is therefore heavily dependent on the particular traits that happen to have been studied. This pattern is likely to be affected by biases toward traits showing high QST because spatial variation might have motivated the study in the first place, and biases against publishing low QST values because such traits might be deemed uninteresting or because of the statistical difficulties of rejecting the null model when FST is low. Thus, it is not easy to say anything in general from studies comparing FST and QST about whether natural selection has a predominately diversifying or homogenizing effect on metapopulations, apart from the fact that both occur. A more fruitful approach might be to compare patterns emerging from different kinds of traits, such as those involved in premating isolation vs. other traits (e.g. Butlin & Tregenza, 1998). Our study indicates that both a sexual signal (tail white) and some, but not all, morphological traits show larger differentiation than expected from neutral loci.

Implications for conservation genetics

Our results may have practical implications for conservation biology. Although great snipe populations are very weakly differentiated at neutral loci, adaptive genetic differentiation (as measured by PST here and also in MHC divergence by Ekblom et al., 2007) makes it clear that the eastern and western regions might need to be treated as separate conservation units. Such units are often defined using divergence in neutral markers alone, but our results support the view that this approach risks failing to identify ecologically important genetic differences among populations (e.g. Karhu et al., 1996; Butlin & Tregenza, 1998; Hedrick, 1999; Crandall et al., 2000; Fraser & Bernatchez, 2001; Pearman, 2001; Reed & Frankham, 2001; McKay & Latta, 2002; Stockwell et al., 2003; Hansson & Richardson, 2005). It is not at all clear whether one can predict ecologically or evolutionary important differences among natural populations from neutral divergence, or indeed if there is a general correlation between FST and QST (Merilä & Crnokrak, 2001; Reed & Frankham, 2001; Crnokrak & Merilä, 2002; Latta & McKay, 2002; McKay & Latta, 2002; Ekblom et al., 2007). Ultimately, the maintenance of ecologically meaningful and adaptively significant genetic diversity should be the primary goal in conservation genetics, and not the maintenance of neutral variation. Although not a cure-all, adopting the PSTFST approach may open up avenues for putting the tools of neutral genetic variation into their proper organismal context in a whole new set of natural populations and species, which are otherwise unavailable to quantitative genetic analysis. Many organisms of conservation concern presumably fall into this category, and it may sometimes be important to go beyond neutral genetic variation because adaptive population divergence may have evolved in the face of gene flow.


Thanks to Sten Svartaas, Henrik Brøseth and Marek Borkowski for help with the fieldwork, Gunilla Olsson and Reija Dufva for assistance in the laboratory, and to Martin Carlsson, John Dallas, Johan Dannewitz and two reviewers for helpful comments. Financial support was received from the University of Trondheim and the Research Council of Norway (to SAS and PF), the Norwegian Institute for Nature Research (to JAK) and the Swedish Research Council (to JH).