The study species
The collared flycatcher is a small, socially monogamous, insectivorous hole-nesting passerine bird, which inhabits deciduous woods of central Europe and the Baltic islands of Gotland and Öland. Adults overwinter in central Africa, and arrive at breeding grounds in late April–early May, where they establish territories and lay a single clutch of, on average, six eggs. Eggs hatch after about 12 days of incubation, and nestlings are fed (mainly with caterpillar larvae) by both parents until fledging, which usually occurs 2 weeks after hatching. By the time of fledging, the nestling tarsus has attained its final size (Alatalo et al., 1990; Merilä, 1997), and the fledglings have gained large fat reserves and weigh more than the adults (Merilä, 1996). After reaching independence, fledglings undergo a partial moult in the breeding grounds, and start migration to winter areas. Surviving birds return to breed to their natal area, and the first breeding usually takes place in the year after birth. More information on the life history of the species is given in Gustafsson (1989).
Data were collected between 1980 and 1999 as part of the long-term study of the nest-box breeding population on the Swedish island of Gotland in the Baltic Sea. In each year, breeding was monitored in 6–18 different woodlands (the number has increased with time) from early in the breeding season until the last young had fledged. When 11–12 days old, the nestlings were measured for tarsus length with digital callipers (to nearest 0.1 mm) and weighed with a Pesola spring balance to nearest 0.1 g. A condition index was estimated as the residual from a linear regression (see Pärt, 1990 for analysis of linearity) of body mass on tarsus length, so that one unit of relative mass corresponds to a deviation of 1 g from the allometrically expected mass. We note that this procedure may not remove all ‘body size’ dependent variation in ‘condition’ as tarsus length (or any other measure of body size for that matter) is not likely to capture all aspects of body size variation, and hence, part of the measured variation may still reflect individual size differences. At the time of measurement, or a few days earlier, the parents of all offspring were captured and identified (from individually numbered aluminium rings), and all nestlings were provided with individually numbered aluminium rings to enable identification if they were recaptured. See Kruuk et al. (2001) for a full analysis of the genetics of, and selection on, the skeletal component of body size, tarsus length.
Pedigrees were determined by assuming that the parents attending a given nest were the true parents of all offspring in the nest. This assumption is almost certainly correct in the case of the females as no egg dumping is known to occur (Sheldon & Ellegren, 1999), but some error is involved in the determination of paternal pedigrees because of the occurrence of extra-pair paternity. About 15% of nestlings in this population are known to result from extra-pair copulations (Sheldon & Ellegren, 1999), and consequently the paternal pedigrees are likely to be wrong in a corresponding proportion of cases. In the animal model approach adopted in this study (see below), these erroneous paternities could lead to a deflated estimate of additive genetic variance (e.g. Geldermann et al., 1986). However, previous studies of this population have shown that the effect of extra-pair paternity on estimates of heritability are minimal (cf. Meriläet al., 1998). Another potential problem with our analyses involves the pooling of data from both sexes in the same analyses; this was necessary as fledglings cannot be sexed from appearance, and data on sex determined by molecular methods are only available for a small proportion of the population. However, there is no evidence to suggest that males and females differ in fledgling condition in the collared flycatcher (Sheldon et al., 1998), or that selection acts differently on male and female condition (Meriläet al., 1997). Hence, combining data from both sexes for the analyses should not affect the results.
Some of the nestlings in the data were subject to brood size manipulations and cross-fostering experiments (see: Gustafsson & Sutherland, 1988; Merilä, 1997, for details). As brood size manipulations are known to affect fledgling growth and condition (Merilä, 1996, 1997), the average effects of the experiments were accounted for by including brood size manipulation as a fixed effect in the models (see below). For our main analyses, we restricted the data to nonfostered offspring. However, as explained below, offspring from broods involved in cross-fostering experiments were used in part of the analyses to evaluate the importance of pre-manipulation common environment, post-manipulation common environment, and dominance effects on offspring condition. The total number of breeding attempts monitored over the study period was 7575 (28 759 nestlings), but for the reasons detailed below, only 3836 breeding attempts (17 717 nestlings), were included in the analyses of nonfostered offspring. First, some of the nestlings were not measured for tarsus length (for example, all individuals ringed in 1982), and hence were excluded from the analyses as their body condition was not estimable. Secondly, some of the nests were subject to partial clutch size manipulations, in which eggs rather than nestlings were transferred between nests, with the result that it was not possible to distinguish between native and fostered fledglings in a nest. Thirdly, some nestlings were the result of pairings between collared and pied flycatchers F. hypoleuca: these breeding attempts, and any involving known first generation hybrid parents, were excluded from the data.
We used a mixed-model analysis of variance based on an ‘animal model’ and restricted maximum likelihood (REML) estimation procedure to quantify the different causal components of variance for nestling condition (Groeneveld & Kovac, 1990; Groeneveld, 1995; Knott et al., 1995; Lynch & Walsh, 1998). Although mixed models have been widely used in animal breeding science, they have rarely, and only very recently, been implemented in the analysis of data from natural populations (but see Réale et al., 1999; Kruuk et al., 2000, 2001; Milner et al., 2000). The animal model expresses the phenotype of each individual i as a sum of fixed and random effects, with the latter comprising a component of additive genetic value and other random effects such as a maternal effect value:
where yi is the individual’s phenotype, μ is the population mean, bi,j are fixed effects, ai is the individual’s additive genetic value, uij are other random effects and ei is a random residual value. In matrix form, for measurements on many individuals, this gives:
where y is the vector of phenotypic values, β and u are column vectors of fixed and random effects, respectively, X and Z are the corresponding incidence matrices and e is a vector of residual values (Knott et al., 1995; Lynch & Walsh, 1998). These methods are more flexible and make less assumptions about the data than the conventional models used in estimating quantitative genetic parameters (Shaw, 1987; Lynch & Walsh, 1998). In particular, they have the advantages of allowing for highly unbalanced data sets and the inclusion of fixed effects; also, by exploiting all relationships between individuals in a pedigree, they are considerably more powerful than the conventional models used in estimating quantitative genetic parameters (Sorenson & Kennedy, 1986; Knott et al., 1995; Lynch & Walsh, 1998). The animal model provides estimates of components of variance in a base population that are unbiased by any effects of finite population size, selection or inbreeding in subsequent generations (Thompson, 1973; Sorenson & Kennedy, 1984; van der Werf & Boer, 1990). Because information in any pedigree rarely dates back to a true base population, an assumption concerning the base population is usually made, namely that the first generation of animals with data form the base population; the subsequent analysis will then estimate the components of variance in this generation. If the first generation does consist of selected individuals, the resulting estimates will still apply to that generation, but not to an unselected (‘true’) base population (van der Werf & Boer, 1990). Here, we estimated components of variance, including the additive genetic variance and hence the heritability, using REML-VCE (version 3.2, Groeneveld, 1995), and single trait animal model best linear unbiased predictor (BLUP) estimates of breeding values were obtained using the software package PEST (Groeneveld & Kovac, 1990; Groeneveld et al., 1992).
In our basic model, the total phenotypic variance (VP) in offspring condition was partitioned into its causal components: VP=VA + VEy + VEa + VEn + VR, where VA is the additive genetic variance, VEy, VEa and VEn are the environmental components of variance attributable to year, area and box of rearing (common nest environment) effects, respectively, and VR is the residual variance. Narrow sense heritabilities were calculated from (h2=VA/VP) (Falconer & Mackay, 1996), and the environmental effect variance quantified as the summed effects of different environmental sources of variance and residual variance to total phenotypic variance (E=VEy + VEa + VEn + VR). Year, area and common nest environment were included as random effects in the model because these sources of variation are relevant for estimating heritability in the population: for example, as parents and offspring are necessarily measured in different years, large year-to-year variation would reduce the similarity between parents and offspring and hence the heritability and potential response to selection. However, inclusion of year and area effects as fixed terms did not lead to any significant increases in heritability (h2) estimates. The VCE program returns standard errors for the estimates of variance components and heritability, from which significance was assessed by t-tests. Significance levels for variance components were assessed from z-scores, calculated as the ratio of the estimate to its standard error, and tested against an asymptotic (large sample) standard normal distribution.
Note that a component of variance due to box of rearing will include variance due to both common environment and maternal (or paternal) effects. As most of the parents in our data occur only once, or a few times, and almost never together in more than 1 year, the separation of maternal and common environment effects is difficult. However, we attempted to evaluate the relative importance of maternal vs. common environment effects by splitting the box of rearing term into components because of paternal and maternal identity. In the presence of strong maternal effects, we expected the random effect of maternal identity to capture more of the box of origin variance than the random effect of paternal identity. As a further test of this, we ran a model that included both the box of rearing, as well as maternal and paternal identities as random effects: if maternal or paternal effects were present, part of the variance in nestling condition should be accounted for by maternal and/or paternal identities beyond the effect accounted for by the random effect of box of rearing.
As our data included a large number of full-sib families, we also evaluated the potential significance of dominance variance in condition. Dominance variance will increase the variance between full-sib groups (Falconer & Mackay, 1996), but it is usually not possible to separate this effect from nongenetic effects such as those caused by a common environment. Here, we make use of data from several cross-fostering experiments performed on the population in different years (Gustafsson & Sutherland, 1988; Gustafsson et al., 1995; Merilä, 1996) to test whether the variance between groups of full-sibs is greater than expected after accounting for that due to additive genetic and common nest environment effects. Cross-fostering experiments in this population involved the splitting and reciprocal transfer of broods so that full-sibs were reared apart in two different nest-boxes: one group in the nest-box with their true parents, the other in a foster parent nest-box (see, for example, Merilä, 1996; Merilä & Fry, 1998; Sheldon et al., 1998, for details). Each nest box therefore included nestlings from two full sibships. The animal model for these analyses included – in addition to the random effects of additive genetic, year and area effects, and a fixed effect of brood size manipulation – a random effect for box of rearing and a random effect identifying full-sibships (i.e. box of origin). The existence of any dominance variance for condition should be reflected in this final term, which estimates the variance between groups of full sibs after accounting for additive genetic and common environment variance. However, note that any early common environment effects acting prior to experimental manipulation (up to 2 days of age) will also contribute to this term. A priori, we expect early common environment effects to be minimal, for two reasons: in the pied flycatcher, there is no association between the size of eggs produced by a mother and the subsequent condition at fledging of the offspring (Potti, 1999), and in the current species there is no correlation between condition at 2 days of age and condition at 13 days (Merilä, 1996). Nevertheless, the estimate of the variance component due to nest of origin can only provide an upper limit on the potential contribution of dominance variance.
Viability selection on survival from fledging to adulthood was assessed from recapture data. During the study period, including 1982 when no tarsus length measurements were made, virtually all collared flycatchers breeding in study plots were captured and checked for aluminium rings. Individuals marked and measured as nestlings, and recaptured as breeding adults were classified as survivors (11.8% of nestlings in the data, with 35.6% of nests containing at least one survivor). All other individuals were assumed to be dead. Although some individuals assumed to be dead may have emigrated outside the study area, the natal dispersal distance has been previously shown to be uncorrelated with body condition index in this population (Pärt, 1990). Hence, the use of local recruitment rate as a criterion for survival is justified.
We quantified the effect of viability selection on variance components by repeating the quantitative genetic analyses on two different sets of data: (i) pre-selection, using data on all (nonfostered) nestlings in the database (with the restrictions detailed above), which estimates the different causal components of variance before most of the selection has taken place; (ii) post-selection, using a restricted data set consisting of only phenotypic measurements on those individuals that returned as adults to the breeding site; this approach is directly analogous to any animal model analysis of a trait expressed only in adults (e.g. Réale et al., 1999; Kruuk et al., 2000; Milner et al., 2000). As the animal model estimates the components of variance in a base population (Sorenson & Kennedy, 1984; Lynch & Walsh, 1998), any difference in the estimates from the two approaches will reflect the difference between the constitution of the total population and that of the subset that then survived to become breeding adults. It therefore provides an estimate of the effects of selection on the additive genetic variance similar to that adopted in standard selection experiments (e.g. Meyer & Hill, 1991), but allowing for the fact that, as a result of continuous immigration into the study area, different generations are distributed at different times throughout the pedigree. We also tested the prediction that, under truncation selection, selection should reduce the additive genetic variance of a trait by a factor of:
where VA′ is the additive genetic variance after selection, h2 is the narrow sense heritability and k is the factor by which the phenotypic variance is reduced (Falconer & Mackay, 1996, p. 202). Standardized selection differentials for body condition index were estimated using least squares regression of survival on standardized (zero mean, unit variance) trait values (Arnold & Wade, 1984a,b). Before analyses, the individual survival values (0=nonsurvivor, 1=survivor) were divided by the population mean rate of survivorship to give a measure of relative fitness (Arnold & Wade, 1984a,b). The use of standardized trait values and relative fitness in the analyses facilitates the comparison of results with those from other studies, and allows prediction of the selection response from standard quantitative genetic theory. Directional (S) and nonlinear (c) selection differentials were estimated using linear and second order polynomial regressions, respectively. Nonlinear selection differentials reveal the presence or absence of stabilizing (negative sign) or disruptive (positive sign) selection, and were estimated as detailed in Meriläet al. (1997). Although the selection differentials were estimated with least squares methods, their statistical significance was tested with logistic regression analyses (Mitchell-Olds & Shaw, 1987; Meriläet al., 1997). Analyses were performed for each of the study years separately, as well as for the data pooled over years. In each of the analyses, relative fitness was calculated over the entire data set used in the given analysis, but the condition index was standardized within the years. Selection analyses were performed with PROC GLM and PROC GENMOD available in SAS statistical package (SAS Institute Inc., 1996).