Quantitative genetics of migration syndromes: a study of two barn swallow populations

Authors


Céline Teplitsky, Département Ecologie et Gestion de la Biodiversité UMR 7204, Case Postale 51, 55 rue Buffon, F-75005 Paris, France.
Tel.: +33 1 40 79 34 43; fax: +33 1 40 79 38 35; e-mail: teplitsky@mnhn.fr

Abstract

Migration is a complex trait although little is known about genetic correlations between traits involved in such migration syndromes. To assess the migratory responses to climate change, we need information on genetic constraints on evolutionary potential of arrival dates in migratory birds. Using two long-term data sets on barn swallows Hirundo rustica (from Spain and Denmark), we show for the first time in wild populations that spring arrival dates are phenotypically and genetically correlated with morphological and life history traits. In the Danish population, length of outermost tail feathers and wing length were negatively genetically correlated with arrival date. In the Spanish population, we found a negative genetic correlation between arrival date and time elapsed between arrival date and laying date, constraining response to selection that favours both early arrival and shorter delays. This results in a decreased rate of adaptation, not because of constraints on arrival date, but constraints on delay before breeding, that is, a trait that can be equally important in the context of climate change.

Introduction

Recent climatic alterations already have a measurable impact on the distribution, ecology and behaviour of animals (Root et al., 2003; Walther et al., 2005; Parmesan, 2006; Møller et al., 2010). Organisms that cannot adapt to these rapid alterations of their environment are declining and likely to go extinct (Parmesan, 2006). Along with poleward shifts in distribution, changes in phenology such as earlier breeding dates (McCarty, 2001; Stenseth et al., 2002; Walther et al., 2002; Root et al., 2003; Brommer & Møller, 2010; Dunn & Winkler, 2010) or earlier arrival dates for migratory birds (Jonzén et al., 2006; Gienapp et al., 2007; Lehikoinen & Sparks, 2010) are some of the most striking examples of rapid changes due to climate change. Such changes are crucial for population persistence, for example, because species that did not show a phenological response to climate change and did not advance their spring migration are declining (Møller et al., 2008).

The mechanisms underlying changes in migration dates are still unclear, although rates of change are not incompatible with microevolution (Gienapp et al., 2007), despite there being no clear-cut cases of micro-evolutionary change (Sheldon, 2010). There is strong selection for early arrival in migratory birds because of the reproductive advantages achieved by early arriving individuals (review in Newton, 2008). However, we still know very little about the evolutionary potential of phenology in terms of arrival dates and more generally in terms of correlations between phenological traits (Clements et al., 2011) or between phenological and other ecologically important traits (Dingle, 2006; van Noordwijk et al., 2006). The evolutionary potential depends not only on the amount of genetic variation in single traits but also on genetic correlations between traits. As evolution will proceed along the direction with most genetic variance, genetic correlations may affect the pace and even direction of response to selection (Lynch & Walsh, 1998). Depending on the direction of selection and the sign of correlations, genetic correlations could either hamper or accelerate responses to selection (Falconer & Mackay, 1996; Arnold et al., 2001).

Response to selection may particularly depend on genetic correlations in the case of migration dates: migration is a complex trait that integrates many different characters, including behaviour, physiology, morphology and life histories into a single syndrome (Dingle, 2006; Pulido, 2007a). So far, only genetic correlations between migratory traits (e.g. amount, intensity and onset of autumn migration, narrow-sense syndrome sensuDingle, 2006) have been shown in the blackcap Sylvia atricapilla (Pulido & Berthold, 1998). Those studies suggest potential constraints on response to selection notably because of genetic correlations between timing of egg laying and timing of autumn migration (Coppack et al., 2001) and between onset of autumn migration and timing of juvenile moult (Pulido & Coppack, 2004).

Broad-sense migratory syndromes include not only traits such as morphology that directly affects movement abilities but also life history traits. As migration implies selection on morphology for long distance flight (van Noordwijk et al., 2006), migrant birds generally have longer and more pointed wings, shorter tails and smaller body size than residents (Leisler & Winkler, 2003). Within populations, phenotypic correlations have been established between arrival dates and wing shape (Bowlin, 2007; Choi et al., 2010) and arrival dates and individual condition (Hatch et al., 2010). In terms of life history traits, migration affects the life cycle as spring arrival dates will determine laying date to a great extent (e.g. Both & Visser, 2001). As laying date is correlated with clutch size (e.g. Sheldon et al., 2003), arrival date may also affect the number of eggs laid in one season either in terms of clutch size or time constraints on the number of broods (Dingle, 2006; Møller, 2007b).

Clements et al. (2011) showed in a population of red deer Cervus elaphus that phenotypic correlations between phenological traits had an environmental rather than a genetic basis, emphasizing that phenotypic correlations do not provide estimates of genetic correlations. Direct evidence for genetic correlations between migratory traits and morphological or life history traits is extremely scarce so we do not know which traits are integrated into the syndrome and which traits can evolve independently (Pulido & Berthold, 2003; Pulido, 2007a). Only Møller (2001) showed both phenotypic and genetic correlations between morphology and arrival date.

A recent review (Kirkpatrick, 2009) reported that G matrices, the matrix of genetic variance–covariance, tended to be ill conditioned; that is, genetic variance tends to be distributed along a restricted number of dimensions that could decrease the evolutionary potential. However, theory also predicts that action of selection should lead to alignment of G matrices with selection (Arnold, 1992; Arnold et al., 2001; Jones et al., 2003) so that evolutionary constraints could be reduced. This would be in line with results reported by Agrawal & Stinchcombe (2009), who found no overall significant effects of genetic correlations on rate of adaptation. In this context, we pose the question whether a migration syndrome represents an effective constraint on response to selection of arrival date.

In this study, we used long-term monitoring data from a Danish and a Spanish population of the barn swallow Hirundo rustica, an insectivorous, long-distant migratory bird breeding in temperate and subtropical regions of the Northern Hemisphere and wintering in subtropical and tropical areas. The two study populations winter in West Africa and Southern Africa, respectively (Ambrosini et al., 2009). Our aims were to determine the evolutionary potential for arrival date and assess the potential constraints imposed by genetic correlations on evolutionary responses. Tightly linked to the later question is the question of the stability of G matrices, as such constraints are important only if G is relatively stable. There is still debate about G matrices stability (Eroukhmanoff, 2009), and comparing the Spanish and Danish populations is a small contribution to this issue. Moreover, although we cannot test formally for specific hypotheses as the impact of a latitudinal gradient or migration distance, comparing the two populations may help interpreting results.

To address these aims, we first estimated phenotypic and genetic correlations for arrival date at the breeding grounds and morphology to assess the extent to which traits are integrated into a migration syndrome. We investigated broad-sense migration syndromes by estimating genetic correlations between (i) arrival date and proximate traits (wing length, tail length, length of the short central tail feathers and body weight) in both sexes, and (ii) arrival date and life history traits (time elapsed between arrival and laying dates, clutch size) in females. We then estimated multivariate phenotypic selection, as selection will act simultaneously and not independently on these traits. Using the breeder’s equation, we compared predicted responses to selection and rates of adaptation in the presence and absence of genetic correlations. The advantage of the breeder’s equation is that it allows understanding the part of genetic correlations between traits that slow or accelerate response to selection. However, the drawback is the use of a phenotypic selection gradient, so if selection is acting on the environmental and not on its genetic part, the breeder’s equation will overestimate response to selection (Alatalo et al., 1990; Morrissey et al., 2010). Hence, we also evaluated predicted responses to selection based on the genetic covariance between traits and fitness, as both methods provide complementary answers.

Materials and methods

Study species and data collection

Barn swallows are small (approximately 20 g), monogamous, semicolonial passerine birds that feed on insects caught on the wing. Sexual size dimorphism is slight with the exception of the outermost feathers of the forked tail. We used data obtained from the long-term study of two barn swallow populations in Spain (1991–2007) and Denmark (1985–2009). Data were available for spring arrival date, first and second laying date, first and second clutch size, annual reproductive success (total number of fledglings produced during one breeding season) and morphological measurements (wing length, body weight, shorter central tail feathers (hereafter ‘central tail’) and outermost tail feathers (hereafter ‘outer tail’). To remove the intrinsic correlation between arrival and laying date, we instead used arrival dates and the delay between arrival and laying dates.

Barn swallows were captured at least weekly from the arrival in early spring until the end of the breeding season using mist nets at all entrances to the barns with breeding birds. Capture–mark–recapture analyses have revealed that capture probability of adults exceeds 98% in Spain and Denmark (Møller & Szép, 2002).

The pedigree for each population was based on social relations, assuming that the two adults attending a nest are the genetic parents. In the Spanish population, 18% of offspring are extra-pair offspring (Møller et al., 2003), whereas 28% of offspring are extra-pair offspring in the Danish population (Møller & Tegelström, 1997). Males with long tails sire almost all the offspring in their own nests and many offspring in neighbouring nests, whereas short-tailed males sire few offspring in their own nests and few offspring elsewhere, as demonstrated by observations and a field experiment (Møller & Tegelström, 1997; Saino et al., 1997; Møller et al., 1998b, 2003). Females are almost always the mothers of offspring in their nests (Møller & Tegelström, 1997; Møller et al., 2003). In total, the pedigree of the Danish population contained 489 individuals for four generations (137 individuals with both parents known, 131 fathers, 146 mothers and a maximum family size of 4). In the Spanish population, the pedigree contained 1407 individuals for five generations (159 individuals with both parents known, 123 fathers and 115 mothers and a maximum family size of 7).

Animal model

We estimated the G matrix using multivariate animal models (Wilson et al., 2010). Random effects included additive genetic effects (Va) and permanent environment effects to account for repeated measurements of the same individual (Vpe). Males and females were not analysed separately, so models contained sex as a fixed effect because of protandry (Møller, 2007a) and size differences between sexes. Age was included due to traits differing among age classes, to be able to include young animals in the analysis. These fixed effects were included when significant in preliminary analyses. To avoid that our conclusions were based on traits with larger means dragging the entire pattern, we standardized the traits prior to analysis by dividing the trait by its standard deviation as is classically done (Hadfield et al., 2007; Agrawal & Stinchcombe, 2009).

Hence, the models we used were as follows:

– for proximate traits,

image(m1)

– for females life history traits,

image(m2)

with stDate, stWing, stWeight, stOuterTail, stTail, stDelay and stClutchSize being the standardized traits, animal the individual effect linked to the pedigree to assess additive genetic variance, and ID the identity not linked to the pedigree to assess permanent environment effects, i.e. effects linked to the individual that are not due to additive genetic effects, and year, the year of reproduction. For the analysis of proximate traits, we assessed whether including maternal effects would change the results. Although power is very low, it seems that the estimates of additive genetic (co) variances did not differ in models with and without maternal effects. Hence, as including maternal effect reduces sample size considerably, we only present results from models without maternal effect. For life history traits, sample size was too small to include maternal effect. However, a couple of studies showed low impact of maternal effects on laying date and clutch size (Sheldon et al., 2003; e.g. Charmantier et al., 2006; Garant et al., 2008), as can be expected for traits expressed in adults (Husby et al., 2011), so we believe that not including maternal effect should not introduce a strong bias in our study.

From this, the variance decomposition model will be as follows:

image(m3)

where the matrices of variance–covariance are the following: P, phenotypic; G, additive genetic; PE, permanent environment; and R, the residual. Note that in Results, phenotypic correlations were corrected for year effects, as the P matrix was calculated as G + PE + R, excluding year effects.

As arrival dates can constrain the ability to lay a second clutch, we also ran for females a model with five traits: adding delay between first and second clutch (stDelay2) and second clutch size (stClutchSize2) to the traits already in equation 1b.

image(m4)

However, those traits showed no correlations with arrival date, so we kept the simpler model, but results for additive genetic covariances and subsequent analyses from model 1c can be found in Appendix S1.

All pedigrees were pruned so they contained only informative individuals (Morrissey & Wilson, 2010), and the relevant statistics are presented in Table 1.

Table 1.   Pedigrees statistics after pruning of noninformative individuals.
 MorphologyLife history traits
Danish populationSpanish populationDanish populationSpanish population
Number of individuals4871407191709
Number of maternities1451854038
Number of paternities162166609
Pedigree maximum depth3424
Relatedness    
 0.02501600
 0.05021801
 0.12576906
 0.25044192615
 0.5003194014647
Pairs with known relatedness3726965269

Estimating selection

To assess the directional selection, we used the classic equation by Lande & Arnold (1983): selection gradients were estimated by regressing fitness (total annual offspring fledging success) against traits measured that same year. Similarly, nonlinear selection was estimated using quadratic regression including cross-products between traits. For each population, selection was estimated over all available years, averaging over yearly fluctuations (Charmantier et al., 2004). Because of nonindependence of data within 1 year and also because some individuals were included several times in the data set, we included year and ID as random effects in the model.

One bias in these analyses is that fledging success does not take extra-pair paternity into account. Although this can be an important issue as extra-pair paternity is high, we only have data on extra-pair paternity for 1 year in the Spanish population (1994) and 2 years for the Danish population (1988–1989), so we present results based on all data while not taking extra-pair paternity into account (but see Appendix S2 for selection analyses with the subsets including information on extra-pair paternity).

To obtain estimates consistent with G matrices (Hansen & Houle, 2008), traits were corrected for age and sex before analysis when those effects were significant, and selection was estimated by standardizing the data (σ = 1, μ = 0).

Predicting responses to selection

Responses to selection were predicted using two methods. First, we used the multivariate breeder’s equation R = Gβ, where G is the G matrix and β the phenotypic directional selection gradient. Second, in case selection was acting on the environmental and not genetic part of the traits, we used the Price–Robertson identity that predicts response to selection based on the genetic covariance between trait and fitness and then compared the two expected responses.

For that purpose, we estimated the genetic correlation between life history traits and fitness (fledging success) using a bivariate animal model. We used a Poisson distribution for fledging success and a normal distribution for other traits, and none of them was standardized for this analysis. As advocated by Morrissey et al. (2010), we then compared posterior distributions for environmental and additive genetic covariances to assess whether selection is acting more on the genetic or the environmental part of the trait. Note that we found no support for additive genetic variance in fledging success in the Danish population, so we restrained those analyses to the Spanish population. In the Danish population, absence of detectable genetic variance in fledging success can be due to a real absence of genetic variance in fitness, or a lack of power. As estimating genetic variance for fitness is notoriously difficult (Burt, 2000; Morrissey et al., 2010), we cannot choose between these alternatives.

Estimating constraints

Geometry of G

The shape of G will determine greatly how much G is a constraint on response to selection. A round matrix will allow response in all phenotypic directions whereas a more cigar-shaped (ill-conditioned) matrix will determine greatly along which direction the population is able to respond. To assess the shape of G, we used two estimates: (i) evenness of the eigenvalues of G (E(G)) as used by Agrawal & Stinchcombe (2009) with rounder matrices having more even eigenvalues, so E(G) ranges from 0 for ill-conditioned matrices to one for matrices where genetic variation is evenly distributed in all directions.

image(m5)

where inline image, λi is the ith eigenvalue of G and n is the number of traits. (ii) The index defined by Kirkpatrick (2009) measuring the effective number of dimensions where genetic variance is distributed:

image(m6)

When all the genetic variance is accounted for by the first dimension, then nD = 1, whereas if the variance is evenly distributed, nD is equal to the number of dimensions (five for proximate traits, three for life history traits).

Impact on the rate of evolution

We used the metric R defined by Agrawal & Stinchcombe (2009) to assess the impact of genetic correlations on the rate of evolution. This metric is the ratio between the rate of evolution in the presence of genetic correlations relatively to what it would have been without these correlations. It is defined as:

image(m7)

with

image(m8)

where inline image is the rate of change in fitness of the mean phenotype; inline image, the vector of average traits; β, the vector of selection gradient; and γ, the matrix of nonlinear selection. In equation (6), ΔWg is the rate of adaptation with genetic correlations whereas ΔWo is the rate of adaptation when all the covariances between traits are set to 0. The ratio R is then compared to 1, with a ratio superior to one implying increased rate of adaptation in the presence of genetic correlations, whereas a ratio lower than one implying that genetic correlations slow down adaptation.

Comparisons between populations

To compare G matrices, we used Krzanowski’s (1979, 2000) method that compares different subspaces of the matrices and gives an overall index of similarity between subspaces as well as an angle between those subspaces (see Blows et al., 2004; Hadfield et al., 2007 for details on application to genetic data). Briefly, this method compares a set of n eigenvectors (for each covariance matrix), to estimate (i) the angles between the most closely aligned n-dimensions subspaces, and (ii) an index S of overall similarity between subspaces (Krzanowski, 1979). If S = 0, then the subspace have nothing in common; if S = n, then subspaces are identical. The number of eigenvectors (n) should always be lower or equal to half of the total number of dimensions; otherwise, common axes will always be recovered (Krzanowski, 2000; Hadfield et al., 2007). Here, when comparing the G and/or P matrices for proximate traits, the total number of dimensions is the number of traits (5), so n = 2 at most, implying that we compared the populations in terms of vectors (n = 1) or plans (n = 2). When comparing matrices for life history traits, we had three traits and were then only able to compare vectors. When comparing vectors, we only used the first dimension, i.e. gmax.

Estimation method

Both animal models and selection analyses were run using Bayesian methods with the software MCMCglmm (Hadfield, 2010). The advantage of this method is that the results obtained are not point estimates, but the whole posterior distributions of fixed and random parameters. Here, we used for each analysis 1 200 000 iterations, with a burn-in phase of 200 000 and thinning of 1000, unless mentioned otherwise (because of autocorrelation issues). Hence, for each estimate we had a sample of 1000 values. We assessed several priors for each analysis: (i) uninformative improper prior (V = diag(n)*0.01, nu = n-3, 5 500 000 iterations, burnin of 500 000, thinning of 5000), (ii) Flat prior (V = diag(n), nu = 1.002), (iii) a parameter expanded prior (12 000 000 iterations, burnin of 2 000 000, thinning of 10 000) and (iv) a slightly informative prior (V = diag(n)*Vp/r, nu = n), where Vp is the phenotypic variance, n the number of traits and r the number of random factors. Flat priors and slightly informative priors gave the exact same results for both sets of traits in the two populations. In turn, although we obtained similar results with parameter expanded and improper priors for proximate traits in the Danish population, the choice of prior affected the outcome of the other analyses. Parameter expanded priors tended to more often result in null estimates and improper priors led either to (i) a crash of the model because of ill-conditioned matrices in the case of life history traits or (ii) to extremely high autocorrelation between iterations for proximate traits in the Spanish population even for long running time (at least up to a thinning of 50 000 iterations), indicating strong convergence problems for this model. In the following, we chose to present results from the model with slightly informative prior as it has a direct biological interpretation: the prior specification implies that (i) variance is distributed evenly across the random terms, and (ii) traits are independent (Hadfield, 2010). Posterior mode for additive genetic (co) variances, as well as comparison with estimates from models using other priors, can be found in Appendix S3.

Using posterior distributions when performing subsequent calculations (evenness, number of dimensions, rate of adaptation etc.) allows estimates of confidence intervals around their estimates. Significance was inferred when 95% confidence intervals did not include zero. When the lower or upper estimate was very close to zero, we also checked for the number of times the estimate was positive or negative, thus giving the equivalent of a P value (Appendix S3). This was not applicable to variance estimates, so we compared Deviance information criterion (DIC) of univariate models including or not including additive genetic variances. DIC are equivalent to AIC, so lower DIC is better. If DIC is smaller than 5, there is little difference between models and for DIC larger than 10, the models with larger DIC is ruled out (Spiegelhalter et al., 2007; Barnett et al., 2010). As we believe that in outbred natural populations additive genetic variance is likely to differ from zero, we considered additive genetic variance to be significant if ΔDIC < 5, as analyses suffering from a lack of power is more likely than an absence of additive genetic variance.

Results

Phenotypic and genetic correlations

Proximate traits

There were weak-to-moderate phenotypic correlations between traits (Table 3a,b) in both populations: individuals that arrived earlier had longer wings and longer tail feathers, and there was no significant correlation with body weight. Only in the Spanish population did earlier birds have longer tails. In the Spanish population, heavier birds also had longer wings, and longer central and outer tail feathers (Table 3b). Morphological traits were also positively correlated in the Danish population, except for weight and tail measurements (Table 3a). There was additive genetic variance in all traits for both populations except for weight in the Danish population (Table 2). As traits were standardized to a variance of 1, estimates of Va are equivalent to traditional heritability measures. Heritability of arrival date was 32% in the Danish population, but only 11% in the Spanish population. Heritability of morphological traits was also moderate (range 0.15–0.50, Table 3). Almost all morphological traits were genetically positively correlated between each other in both populations. In turn, arrival date was in general not significantly genetically correlated with morphological traits, except for a strong negative correlation between arrival date and outer tail feathers and a significant negative genetic correlation with wing length in the Danish population (Table 3a), but not in the Spanish population (Table 3b). Genetic correlations were of the same sign as phenotypic correlations, and although some of the correlations were significant at the phenotypic but not genetic level, we did not detect any hidden genetic correlations.

Table 3.   Matrices of phenotypic (below diagonal) and genetic (above diagonal) correlations with their 95% confidence interval. On the diagonal are the additive genetic variances. Values in bold are significant estimates.
 Arrival dateWing lengthWeightOuter tail lengthCentral tail length
a – Danish population
 Arrival date0.316 [0.18; 0.481]0.312 [−0.6;0.007]−0.135 [−0.513; 0.208]0.525 [−0.751;0.261]−0.289 [−0.594; 0.186]
 Wing length0.164 [−0.262;0.08]  0.476 [0.278; 0.712]  0.489 [0.069; 0.719]0.515 [0.165; 0.681]0.601 [0.196; 0.779]
 Weight0.02 [−0.07; 0.111]0.161 [0.095; 0.266]  0.283 [0.167; 0.508]  0.243 [−0.114; 0.544]0.299 [−0.176; 0.573]
 Outer tail length0.292 [−0.391;0.238]0.365 [0.278; 0.441]0.014 [−0.073; 0.11]  0.157 [0.094; 0.223]0.373 [−0.003; 0.659]
 Central tail length−0.025 [−0.146; 0.053]0.305 [0.207; 0.383]0.088 [−0.01; 0.175]0.132 [0.053; 0.239]  0.238 [0.129; 0.468]
b – Spanish population
 Arrival date0.115 [0.075; 0.179]−0.165 [−0.375; 0.205]−0.11 [−0.385; 0.168]−0.189 [−0.407; 0.124]−0.091 [−0.376; 0.192]
 Wing length0.051 [−0.111;0.017]  0.361 [0.226; 0.489]0.512 [0.235; 0.688]0.625 [0.436; 0.771]0.688 [0.54; 0.819]
 Weight−0.006 [−0.049; 0.04]0.204 [0.161; 0.249]  0.194 [0.123; 0.266]0.345 [0.153; 0.594]0.401 [0.105; 0.575]
 Outer tail length0.105 [−0.165;0.072]0.4 [0.345; 0.43]0.14 [0.086; 0.182]  0.191 [0.13; 0.251]0.524 [0.353; 0.706]
 Central tail length0.052 [−0.098;0.013]0.322 [0.286; 0.368]0.143 [0.098; 0.183]0.227 [0.195; 0.285]  0.275 [0.193; 0.378]
Table 2.   Deviance information criterion (DIC) for univariate models including individual variance only (VID) or decomposing this variance into its additive genetic (VA) and permanent environmental (VPE) components. The prior used for all these models was slightly informative with V = 0.25 and nu = 1 for each variance parameter. In bold, traits for which ΔDIC < 0, i.e. the model including additive genetic variance has a smaller DIC. In italic, models where ΔDIC < 5, i.e. both models are equivalent.
ModelDanish populationSpanish population
VA + VPEVIDΔDICVA + VPEVIDΔDIC
Males and females
 Arrival date1245.6161250.6004.9845752.9795744.1808.799
 Wing length235.392221.27114.1224686.2664680.8135.453
 Weight794.849820.27325.4245721.2455694.35926.886
 Outer tail length574.496569.7614.7353234.9043220.44414.460
 Central tail length686.163690.0013.8385797.0845762.56034.524
 Fledging success2329.0202343.40114.3815687.8185689.4521.634
Females
 Arrival date379.937383.0103.0732869.2722864.1565.116
 Delay between arrival date 1st laying191.939187.7884.1511596.7211592.9323.789
 First clutch size387.602387.6350.0331785.5191783.6441.875
 Fledging success969.489981.33111.8422936.3712932.6323.739

Female life history traits

Female life history traits were correlated with some extent at the phenotypic level. In both populations, we found a negative phenotypic correlation between arrival date and clutch size: birds that arrive early have larger clutches (Table 4). Only in the Spanish population did we detect a negative correlation between arrival date and delay before laying, suggesting that females arriving earlier also take more time before they lay their first egg. When investigating genetic correlations between life history traits in females, we found no genetic correlation between any of the three traits (arrival date, delay before laying and clutch size) in the Danish population (Table 4a). However, in the Spanish population we found a strong negative genetic correlation behind the phenotypic correlation between arrival date and the delay before laying (rg = −0.78 [−0.87; −0.43], Table 4b).

Table 4.   Matrices of phenotypic (below diagonal) and genetic (above diagonal) correlations. On the diagonal are the additive genetic variances. Values in bold are significant estimates.
 Arrival dateDelay before breedingClutch size
a – Danish population
 Arrival date  0.496 [0.183; 0.779]  0.358 [−0.357; 0.713]−0.486 [−0.8; 0.204]
 Delay before breeding−0.003 [−0.147; 0.152]  0.278 [0.086; 0.547]0.028 [−0.64; 0.534]
 Clutch size0.227 [−0.366;0.077]−0.102 [−0.23; 0.065]  0.278 [0.086; 0.547]
b – Spanish population
 Arrival date  0.119 [0.049; 0.214]0.784 [−0.87;0.434]−0.102 [−0.555; 0.354]
 Delay before breeding0.771 [−0.829;0.65]  0.099 [0.036; 0.177]−0.017 [−0.45; 0.464]
 Clutch size  −0.06 [−0.199; 0.072]−0.044 [−0.195; 0.104]  0.086 [0.046; 0.176]

Selection patterns

Proximate traits

There was strong directional selection for earlier arrival in both populations, but no directional selection on morphological traits (Table 5a,b). There was no strong evidence for nonlinear selection on individual traits (Appendix S4). When using canonical analysis as prescribed by Blows & Brooks (2003), and Phillips & Arnold (1989), several eigenvalues of the γ matrices were significantly different from 0 (Appendix S4), although they were very small and of opposite sign, implying the absence of stabilizing selection (rather a saddle and unstable equilibrium) and also very little curvature (Phillips & Arnold, 1989).

Table 5.   Selection gradients on arrival date and four morphological traits and predicted responses to selection when this response is calculated with a G matrix including estimates of genetic covariance’s, or when those covariances are fixed to zero. Lower and upper estimates are based on the 95% confidence interval. Estimates in bold are statistically significant.
 Directional selectionPredicted response to selection
With estimated genetic covariancesWith genetic correlations fixed to zero
EstimateLower estimateUpper estimatePEstimateLower estimateUpper estimateEstimateLower estimateUpper estimate
a – Danish population
 ID0.0780.0570.100       
 Season0.0000.0000.002       
 Residual0.0410.0270.056       
 Intercept1.0541.0001.114< 0.001      
 Age0.075−0.0340.1900.188      
 Age2−0.033−0.0720.0070.088      
 Sex−0.086−0.155−0.0140.022      
 Arrival date0.0920.1310.062< 0.0010.0260.0490.0120.0270.0480.013
 Wing length−0.034−0.0700.0060.1000.010−0.0180.026−0.012−0.0370.005
 Weight0.014−0.0190.0500.4380.004−0.0100.0230.003−0.0080.016
 Outer tail length0.009−0.0250.0460.6340.0120.0010.0210.000−0.0040.007
 Central tail length0.031−0.0040.0690.1120.011−0.0040.0280.006−0.0050.020
b – Spanish population
 ID0.0020.0000.015       
 Season0.0710.0170.146       
 Residual0.1670.1470.183       
 Intercept0.6510.4870.858< 0.001      
 Age0.3180.2370.409< 0.001      
 Age2−0.042−0.058−0.026< 0.001      
 Sex          
 Arrival date0.1160.1430.090< 0.0010.0140.0230.0080.0130.0220.008
 Wing length−0.001−0.0300.0240.930−0.001−0.0120.012−0.002−0.0100.010
 Weight0.013−0.0120.0400.3260.004−0.0040.0110.003−0.0030.008
 Outer tail length0.004−0.0240.0290.8060.002−0.0050.0100.002−0.0040.006
 Central tail length−0.015−0.0400.0120.238−0.001−0.0110.009−0.004−0.0120.003

Female life history traits

When investigating directional selection on female life history traits, in both populations we found directional selection for earlier arrival date (Denmark: βAD = −0.08 [−0.13; −0.04]; Spain: βAD = −0.23 [−0.29; −0.17]) and larger clutch size (Denmark: βCS = 0.10 [0.07; 0.14]; Spain: βCS = 0.12 [0.08; 0.15]). Only in the Spanish population did we find selection for shorter delay before laying (Denmark: βDelay = −0.03 [−0.07; 0.01]; Spain: βDelay = −0.15 [−0.21; −0.10]).

As for proximate traits, we found little evidence for nonlinear selection (Appendix S4), except for a significant positive arrival date by clutch size interaction, suggesting that selection on arrival date is stronger for birds laying smaller first clutches. When using canonical analysis, in the Spanish population we found one eigenvalue significantly different from 0 (λ1 = 0.04 [0.01; 0.16]), suggesting concave selection, but with little curvature as this value was small. In the Danish population, two eigenvalues differed from 0 (λ1 = 0.09 [0.02; 0.13]; λ2 = −0.09 [−0.13; −0.03]), and they were of opposite signs suggesting unstable equilibrium. No stabilizing selection was detected in either population.

Overall constraints due to genetic correlations

Proximate traits

When considering arrival date and the four morphological traits, both matrices were quite even: for a maximum value of 1, evenness of G was at least 0.7 in both populations for a maximum of 1 in the absence of genetic correlations (Denmark: EG = 0.82 [0.68; 0.87]; Spain: EG = 0.706 [0.605; 0.825]). Surprisingly, for a maximum number of dimensions of 5, the effective number of dimensions was around 1.6 in both populations (Denmark: nDim = 1.83 [1.52; 2.41]; Spain: nDim = 1.58 [CI = 1.44; 1.93]), suggesting higher overall genetic constraints than estimated by evenness. When estimating number of dimensions fixing genetic correlations to 0, dimension numbers increased significantly to almost three (Denmark: nDim = 3.09 [2.37; 3.82], Spain: nDim = 3.17 [2.64; 3.92]), but remained significantly lower than 5, implying that the number of dimensions is reduced not only because of genetic correlations but also because of a lack of additive genetic variance in some directions.

Female life history traits

For the Danish population, in spite of the absence of significant genetic correlations, we found that the number of dimensions was lower than 3 (nDim = 1.52 [1.23; 2.13]), and evenness lower than 1 (EG = 0.82 [0.64; 0.98]).

For life history traits, in females of the Spanish population, we found rather similar results, with high evenness (EG = 0.78 [0.60; 0.92]) and a number of dimensions of 1.47 [1.24; 1.99] for a maximum of 3. Fixing genetic correlations to 0 tended to increase the number of dimensions (nDim = 2.32 [1.91; 2.88], difference = 0.89 [0.10; 1.25]), suggesting constraints are also due to distribution of variance among traits.

Responses to selection and rate of adaptation

Proximate traits

Because of directional selection acting on arrival date only, we would not expect any difference between the expected response (Rg = Gβ) and the response when genetic covariances have been fixed to zero (Ro = Goβ), as correlative responses alone are pretty weak (Table 5a,b). If anything, we would predict longer outermost tail feathers in the Danish population (Table 5a), resulting from the negative genetic correlation between length of outermost tail feathers and arrival date. Accordingly, we found the rate of adaptation being similar to what is expected in the absence of genetic correlations with ratio of rates of adaptation with and without genetic correlations not differing significantly from 1 (Denmark: 0.91 [0.41; 1.26], Spain 0.93 [0.69; 1.35]).

Female life history traits

We only investigated potential constraints on response to selection in the Spanish population as neither genetic correlations between traits nor additive genetic variance for fledging success were detected in the Danish population.

In the Spanish population, based on the breeder’s equation, we would predict an increase in clutch size (RCS = 0.009 [0.003; 0.03]) and an earlier arrival date (RAD = −0.016 [−0.04; −0.006]), but no response in delay between arrival and laying date (RDelay = 0 [−0.009; 0.015]). This is in line with expectations from genetic covariances between traits and fledging success. Genetic correlations between fledging success and arrival date (−0.218 [−0.399; 0.077], P = 0.106) and fledging success and clutch size (0.282 [−0.076; 0.414], P = 0.085) were almost significant, whereas the genetic correlation between fledging success and delay between arrival and laying date were far from significant (0.067 [−0.197; 0.314], P = 0.299). For the three traits, the sign of genetic covariances was the same as for environmental (residual) covariances, although environmental covariances were much stronger (Fig. 1).

Figure 1.

 Comparison of the posterior distributions for genetic (dashed line) and environmental (residual, full line) covariances for three life history traits (a) Arrival date, (b) Delay between arrival and laying date, and (c) Clutch size.

Absence of response in delay before breeding may be due to the negative genetic correlation with arrival date, as in the absence of genetic correlation, we would predict a significantly shorter delay (RDelay = −0.016 [−0.03; −0.006]). In this scenario, constraints due to genetic correlation between arrival date and delay before breeding resulted in an expected rate of adaptation divided by two (relative rate: 0.48 [0.23; 0.86]). Response of arrival date would be roughly similar in the absence of genetic correlations (RAD = −0.025 [−0.05; −0.013]).

Populations comparisons

Proximate traits

We found that matrices of genetic correlations between proximal traits were very similar in their first dimension (Table 6, Fig. 2a). This first dimension includes variance in morphological traits and arrival dates (Appendix S5). The similarity along gmax was almost perfect (S = 0.9 for a maximum of 1), and the angle was significant but rather small (19.8° [9.6; 97.5]). However, they were extremely dissimilar along their second axis (S = 0.004 [0; 0.8], see Table 6 for the comparison of planes), with an angle of 73° [24.85°, 90°] as the second axis for the Spanish population mostly correlated with weight while including mostly variance from arrival date in the Danish population.

Table 6.   Comparison between the two G matrices for proximal and life history traits using Krzanowski’s method. S is index of similarity varying from 0 to n, n being the number of dimensions included in the analysis. For n = 1, we compare the major axes of the matrix, whereas with n = 2 we compare the plans defined by the 2 first axes.
ComparisonnS (Index of similarity)Angle
12
 10.90 [0.66; 0.99]19.83 [9.60; 37.53] 
Proximate traits20.99 [0.84; 1.76]13.62 [1.90; 24.40]72.87 [27.13; 89.91]
Life history traits10.01 [0; 0.79]61.89 [27.4; 89.84] 
Figure 2.

 Three-dimensional representation of G matrices. The grid ellipsoid represents GDenmark, the genetic matrix of the Danish population, whereas the full ellipsoid represents GSpain, the genetic matrix of the Spanish population. (a) Proximate traits. The subspace is defined by the first three eigenvectors of GSpain. Hence, the three axes represent 84% of genetic variance for the Spanish population, but only 76% of genetic variance for the Danish population. The first axis is similar among populations and contains variance for all traits. In contrast, the second axis for the Spanish population mostly correlates with weight while including mostly variance from arrival date in the Danish population. See Appendix S5 for more details on the weight of each variable on each eigenvector. (b) Female life history traits.

Female life history traits

For female life history traits, matrices were very dissimilar with respect to all axes (for gmax: S = 0.005 [0, 0.79], Angle = 61.9 [27.4; 89.84], Fig. 2b).

Discussion

Migration syndromes

To our knowledge, this is the first study presenting results on integration between traits involved in a migration syndrome at both phenotypic and genetic levels in wild populations. We found support for a migratory syndrome in terms of correlations between arrival dates and both morphological and life history traits.

Considering proximal traits such as morphological traits that can directly influence flight performance, at the phenotypic level arrival date was negatively correlated with wing and tail length. This is in line with previous results showing that birds with more elongated feathers arrive earlier (wing shape: Bowlin, 2007; Choi et al., 2010; tail length: Møller, 2001). However, considering the functional link between flight performance and wing length, the correlation between arrival date and wing length could be seen as surprisingly small: It is as low as −0.05 in the Spanish population, suggesting that traits are integrated, but not so strongly that flexibility is impossible. Migration is not only about flight performance but also about timing and matching environmental conditions. In the Spanish population, it has been shown that ecological conditions during migration (NDVI in North Africa) determine arrival date to a great extent (Balbontín et al., 2009a).

A negative correlation between outer tail feather length and arrival date may be surprising from an aerodynamic point of view given that tail length in males is often longer than the aerodynamic optimum, especially in the Danish population (Thomas, 1993). Hence, we should expect that longer outermost tail feathers should impair migration and hence delay arrival. This is opposite to what we found, most likely because birds with longer outer tail feathers probably arrive earlier as they are also birds in better condition as evident for example from their lower parasite load (Møller et al., 2004).

Considering life history traits, we found evidence for negative phenotypic correlations between arrival date and clutch size in both populations, and also a negative correlation between delay before breeding and arrival date in the Spanish population only. Negative phenotypic correlations between phenology (laying date) and clutch size are also common (e.g. Winkler et al., 2002; Both & Visser, 2005), but estimates of correlations between phenological traits, especially arrival date and other life history traits are still lacking. The negative correlation between arrival date and delay before breeding in the Spanish population may appear surprising: as selection for earlier breeding is in general very strong (Visser et al., 1998; Sheldon et al., 2003; Charmantier et al., 2006; Teplitsky et al., 2010; this study), why would earlier arriving bird wait longer before breeding in this population and not in the Danish population? The annual cycle of the Spanish and the Danish populations differs considerably (Ambrosini et al., 2009). The Danish population arrives on spring migration in April–June and produces one–two clutches in May–August. In contrast, the Spanish population arrives at the breeding grounds in February–March and produces two–three clutches in March–July. Hence, Danish barn swallows start laying their first clutch 2–4 weeks after arrival, whereas Spanish barn swallows start laying several weeks later, apparently because conditions for survival are better in Spain than in West Africa during spring (Balbontín et al., 2009b), but insufficient for early start of reproduction, leading to different optimal waiting times before breeding.

So far, genetic correlations were tested for autumn migration syndrome sensu stricto only. Genetic correlations between amount, intensity and onset of migration have been shown (Pulido & Berthold, 1998) as well as correlations between migration timing and post-juvenile moult (Pulido & Coppack, 2004). Here, we report the first estimates of genetic correlations for the broad-sense migration syndrome. In contrast to predictions from Pulido (2007a) that ‘morphological and life history traits do not tend to covary genetically with migratory behaviour, or do so only loosely’, many phenotypic correlations were detected at the genetic level. We found a well-defined broad-sense genetic migration syndrome for both morphology (Danish population) and life history traits (Spanish population).

For phenotypic correlations that were not detected at the genetic level, it is possible that these correlations are indeed due to environmental covariances. For example, Clements et al. (2011) found in an island population of red deer that most of the phenotypic correlations between phenological traits were of environmental origin and did not arise from genetic correlations. However, lack of power and extra-pair paternity may also contribute to our results. Lack of power is an important issue in studies of genetic correlations (Roff, 1997; van Noordwijk et al., 2006), and our data sets were relatively small compared to other data sets used in many of the other animal model studies. Although power issues may not be a crucial issue in the Spanish population where we managed to detect some significant genetic correlations on a data set restricted to females only, power may be a greater problem in the Danish population. Although we managed to detect some significant genetic correlations, some of the point estimates were really high, but still not significant: e.g. the correlation between clutch size and arrival date of −0.39 was not significant.

Extra-pair paternity is also likely to bias our results to some extent. Extra-pair paternity is common in the barn swallow, and a 2-year study of paternity using multilocus DNA fingerprinting demonstrated that 33% of 63 broods and 28% of 261 offspring were sired by extra-pair males (Møller & Tegelström, 1997). Extra-pair paternity can lead to underestimation of genetic (co-) variances. For example, Charmantier & Réale (2005) showed that for 20% of extra pair paternities, heritabilities could be underestimated by at least 20%. In any case, genetic variances and covariances could be higher than what we found, but no lower.

Evolutionary potential

The evolutionary potential was estimated not only by heritability of traits but also by evenness of G matrices and their number of dimensions. We found heritabilities of arrival date to be smaller in the Spanish than in the Danish population (0.15 vs. 0.31). Heritability of arrival dates in the Danish population is of a similar magnitude to what has been found for traits related to timing of spring migration (mean estimate 0.43, Pulido, 2007b). However, estimates of heritability are extremely population and model dependent (choice of fixed and random effects, Wilson, 2008), so comparing such estimates can only give a very rough idea of the similarity in evolutionary potential.

Overall evolutionary potential was similar in both populations for both types of traits. When considering evenness, evolutionary potential was rather high: ca. 0.7 for a maximum of 1, and it seems that genetic correlations have a strong impact on rate of adaptation when evenness is much lower than 1 (Agrawal & Stinchcombe, 2009). However, the number of estimated dimensions where additive genetic variance was available was greatly reduced compared to its maximum (proximate traits: 1.6 dimensions for a maximum of 5, life history traits: 1.5 for a maximum of 3). Comparing the two estimates suggests that the number of dimensions is more sensitive to the lack of additive genetic variance in some directions, probably because evenness is calculated on a log scale. As estimates of constraints due to genetic correlations are increasing (e.g. Hansen & Houle, 2008; Agrawal & Stinchcombe, 2009; Kirkpatrick, 2009; Walsh & Blows, 2009), more comparisons would be needed to evaluate the causes of similarities among estimates.

As the number of dimensions containing genetic variance is reduced, our results suggest potential constraints on response to selection depending on the direction of selection. For proximate traits, we did not detect any change in evolutionary rates due to genetic correlations, the reason being that we did not detect selection on associated morphological traits. This lack of selection, notably on the length of outer tail feathers, which has been repeatedly shown to be under positive sexual selection (Møller et al., 1998a, 2006), notably through experiments (Møller, 1988; Møller & de Lope, 1994; Saino et al., 1997), most likely arises from the lack of data on extra-pair paternity. As selection tends to be positive on the length outer tail feathers, there should not be a constraint on response to selection arising from correlation with these traits, when considering fecundity selection. However, as survival selection on outer feathers is negative in males (Møller & Szép, 2002), predictions may differ depending on the selection estimate used, and also when considering males and females separately, which was not possible here due to small sample size. To properly predict responses and constraints on response to selection, we would need to add survival selection or even use more integrative measures of fitness such as lifetime reproductive success or pt(i) (Coulson et al., 2006).

When looking at correlations with morphological traits only, we conclude that genetic correlations impose little constraints on rates of adaptation. In contrast, when looking at life history traits, we found stronger constraints in the Spanish population, with an evolutionary rate divided by two, because of the negative correlation between arrival date and delay before breeding, while both are under negative directional selection. In terms of trait responses, decreased rate of adaptation is due to the absence of response in delay before breeding, whereas the predicted response in arrival date does not seem to be affected by the genetic correlation. This is in full agreement with responses predicted from the genetic correlations between traits and fitness. Hence, breeder’s equation and Price–Robertson identity predict qualitatively similar responses to selection. However, although using the Price–Robertson identity can predict evolutionary responses, it does not allow explaining why this response should or should not occur. Here, information from the breeder’s equation suggest that the predicted absence of evolutionary response in delay before laying could be either due to the genetic correlation with arrival date rather than to selection acting only on the environmental part of the trait. In the context of climate change, inability to reduce the delay before laying could be as important as being unable to arrive earlier, both in terms of mismatch in timing of reproduction and in terms of ability to benefit from a longer breeding season: in barn swallows, earlier first breeding allows longer intervals between first and second clutch, which is associated with higher reproductive success (Møller, 2007b).

Based on same estimates of evenness, we found rather different constraints on rate of adaptation. These results emphasize once more the need for multitrait studies when trying to understand and predict responses to selection. We found more constraints when looking at life history traits. This is in line with prediction from Walsh & Blows (2009): as life history traits in general are subject to stronger directional selection, in a multivariate context genetic constraints should be stronger because of genetic correlations, even in the presence of additive genetic variance.

Population differences

Although we are aware that conclusions from comparison of two populations have to be made with extreme caution, we assessed population similarities in their G matrices and related these results to previous studies on barn swallows. Overall, matrices of genetic correlations between arrival date and morphological traits were quite similar whereas they were extremely different when considering life history traits. This suggest that stability of G matrices actually depends on traits considered and lend support to Arnold et al.’s (2001) hypothesis that they should be least stable for life history traits. Further investigation is still needed, and more empirical data required to truly understand the stability and the factor at the origin of stability/divergence of G matrices and how these matrices in turn constrain divergence between populations (Hohenlohe & Arnold, 2008; Eroukhmanoff, 2009).

Although matrices of correlations of arrival date with morphology were very similar along gmax, but were different along their minor axes, previous studies have found very strong differentiation of G matrices between the two populations studied here and an Italian population. Roff et al. (2004) in an analysis of females only found that the matrices had not one single axis in common, and hence that the Spanish and Danish population had totally different G matrices. Divergence between results can be explained by four differences between the studies: (i) in the study by Roff et al. (2004), more morphological traits were included such as beak shape, and arrival date was not included; (ii) traits were not standardized, which influences estimates of variance–covariances and hence the shape of the G matrix and subsequent comparisons (Hansen & Houle, 2008); (iii) the methods used to estimates G matrix similarities were very different; and (iv) Roff et al. (2004) only analysed females. In their study, Roff et al. (2004) used a common principal component analysis (Phillips, 1998). Using the same software on our variance–covariance matrices, we also found no evidence for matrix similarities. However, although the first axis is not exactly the same, the estimate by Krzanowski’s method is more precise and show that this angle is only of 27°. The advantage of the method we used is that we can actually estimate the amount of similarity/divergence between matrices. The discrepancy of results between the two methods is best explained by Steppan et al. (2002), who noted that if a principal component analysis finds that two matrices are unrelated, it ‘does not mean that there are no similarities, but rather that all of the tested null hypotheses fit less well than do the alternative hypothesis of no similarity’.

Recently, Santure et al. (2010)– while comparing Qst and Fst among barn swallows populations, including the Spanish and Danish populations analysed here – showed differentiation between morphological traits, suggesting population structure due to divergent selection among populations. The absence of divergence in morphological traits in this study along gmax tends to suggest that divergence among population went along some minor axis. This would deserve further investigation, but would be in line with results from Chenoweth & Blows (2008) and McGuigan et al. (2005) showing that when drift is accounted for, population differentiation actually occurred along axes of minor variation rather than along gmax.

Although we found evidence of a migration syndrome, this evidence seems to differ between populations. Further studies including more populations/species would be needed to assess the extent to which differences in life cycles contribute to such divergence, as opposed to drift. This should allow a better understanding of how selection shapes G matrices and in turn how they affect evolutionary trajectories.

Acknowledgments

We are very grateful to A. Husby and F. Jiguet for comments on an earlier version of this manuscript and to O. Gimenez, P. de Villemereuil and N. Dochtermann for help with Bayesian analyses. Thanks also to R. Lorrillière for help in R coding. Thanks to all the people who helped obtain field data, especially A. Barbosa, N. Cadée, J. Cuervo, L. Garamszegi, D. Gil, I. G. Hermosell, A. Marzal, F. Mateos, S. Merino, J. Moreno, C. Navarro and P. Ninni. CT was funded by ANR EPICE (ANR-08-JCJC-0041). The Spanish Ministry of Education and Science (CGL-2009-08976) supported this research.

Data deposited at Dryad: doi: 10.5061/dryad.41np7

Ancillary