DISENTANGLING GENETIC AND PRENATAL SOURCES OF FAMILIAL RESEMBLANCE ACROSS ONTOGENY IN A WILD PASSERINE

Authors


Abstract

Cross-fostering experiments are widely used by quantitative geneticists to study genetics and by behavioral ecologists to study the effects of prenatal investment. Generally, the effects of genes and prenatal investment are confounded and the interpretation given to such experiments is largely dependent on the interests of the researcher. Using a large-scale well-controlled experiment on a wild population of blue tits (Cyanistes caeruleus), we are able to partition variation in body mass across ontogeny into the effects of genes and the effects of between-clutch variation in egg characteristics. We show that although egg effects are important early in ontogeny they quickly dissipate, suggesting that the genetic interpretation of cross-fostering experiments may be preferable for many types of trait. However, the heritability of body mass is smaller than has previously been reported. Our results suggest that this is due to a combination of controlling postnatal environmental effects more carefully and accounting for viability selection operating early in ontogeny.

To understand the evolutionary process when individuals interact, it is important to ascertain which aspects of an individual's phenotype are under the control of others (Grafen 1988). An important class of interactions are those between parents and their offspring, and in particular the effect of parentally determined egg/seed characteristics on offspring phenotype (Roff 1992). In animals, yolk deposition usually occurs prefertilization and in oviparous species the embryo is usually at a very early stage of development at the point of oviposition (Blackburn 1999). Consequently, any effect of the egg on offspring phenotype is thought to be mainly, if not wholly, under the control of the parents. Relationships between egg size and offspring phenotype are well documented (Fox and Czesak 2000; Krist 2011), and studies on the effects of egg composition are starting to accumulate (Groothuis and Schwabl 2008). However, obtaining a holistic measure of egg effects (i.e., how much variation in offspring phenotype is caused by variation in egg characteristics generally) is challenging and few, if any, reliable estimates exist.

Holistic measures can be obtained by comparing the phenotypes of individuals that belong to a common group and estimating the proportion of variation in phenotype that can be explained by the grouping. Groupings could be defined at any level, such as eggs within mothers or eggs within clutches. However, egg effects are likely to be confounded with other sources of shared similarity among group members, such as those caused by a common natal environment or shared genes. Experimental or statistical tools are required to control for these additional effects. Cross-fostering experiments in which maternal siblings are distributed over one or more postoviposition environments have been used to assess the importance of between-mother egg effects (e.g., Hinde et al. 2010). However, such studies assume that the main sources of similarity between the phenotypes of siblings (or offspring and their parents) raised apart are a consequence of the size and composition of the eggs they hatched from. In reality, both additive and nonadditive genetic effects are also likely to contribute to the similarities of relatives, and indeed similar experimental designs have been used by quantitative geneticists under the assumption that pre-crossfostering common environmental effects, including egg effects, are of minor importance (Rutledge et al. 1972; Riska et al. 1985). Despite researchers from both fields being keenly aware of the assumptions that they make, and often discussing them at length, there is little empirical work that directly tests these assumptions. When variation in relatedness within groups and/or between groups exist then it is possible to disentangle the effects of genes from between-mother or between-clutch egg effects. Using cross-fostered data on collared flycatchers (Ficedula albicollis) in conjunction with multigenerational pedigrees, Kruuk and Hadfield (2007) were able to exploit variation in between group relatedness to distinguish genetic effects from the effects of the pre-crossfostering environment on tarsus length and body condition. Although the effect of the pre-crossfostering environment was found to be substantial, the degree to which this was the result of between-clutch variation in egg effects was hard to ascertain, because egg effects remain largely confounded with nonadditive genetic effects and also, because chicks rather than eggs were cross-fostered, the effects of incubation behavior (Nord and Nilsson 2011) and early post-hatching care.

Here, we present the results of a large-scale reciprocal cross-fostering experiment in a wild population of passerine bird, the Blue tit (Cyanistes caeruleus). In contrast to the numerous other studies employing this experimental design, we cross-foster eggs on the day they were laid rather than chicks some days after they hatch (see Krist and Munclinger 2011, also). As such, we remove the confounding effects of incubation behavior and early posthatching care. We then exploit pedigree relationships generated through extra-pair mating, polygamy, and divorce to separate the effects of shared genes from between-clutch egg effects. We show that such egg effects are an important source of variation early in ontogeny, which can be largely captured through measures of egg mass. However, later in ontogeny the similarity of siblings raised apart is almost entirely due to the effects of genes transmitted from their parents. Nevertheless, our heritability estimates are considerably lower than many other published studies. We argue that this is because our experimental design minimizes the artifactual effect of generating mixed broods that differ in age as well as genes, and that by obtaining data throughout ontogeny, we adequately control for the effect of ongoing selection in removing considerable environmental variation.

Methods

Work was carried out in the springs of 2010 and 2011 on a nestbox population of blue tits (C. caeruleus) in predominantly broad-leaved woodland on the Dalmeny Estate, Edinburgh, United Kingdom. The 225 26-mm-hole woodcrete nestboxes were distributed over two sites, with 180 on Craigie Hill (grid reference NT 156 766) and 45 along the Almond River (NT 179 758) with approximately 30 m separating adjacent boxes. The experiment described below is part of a larger study initiated in 2009 into patterns of inheritance when relatives interact, particularly the genetics of sibling competition. Data files and analysis scripts are available from the Dryad data repository (doi:10.5061/dryad.2j77t).

From 4th April, nestboxes were checked every four days if the nestbox contained <15 mm of nest material, every two days if there was >15 mm of nest material and the nest was not lined, and everyday once the nest was lined. In most cases a single egg was found during daily checks, but in four (2010) and seven (2011) cases, a nonlined nest was found to contain two eggs on a subsequent visit, and four late nests were found (two in each year) at a stage when four (2010) or three (2011) eggs had been laid. In 2011, temperature data loggers (Thermochron iButtons 1921G) were placed in the cup of all nests when the first egg was found. Nests in which the first egg was laid on the same day were randomly allocated into triads where possible and the eggs redistributed among members of the triad such that the egg from nest A was moved to nest B, the egg from nest B was moved to nest C, and the egg from nest C moved back to nest A. On days with only two nests, the eggs were reciprocally swapped, and on those days where the number of nests was greater than three but not divisible by three, we employed a round-robin design identical to that for triads but with additional nests in the chain. Nests were checked daily after the first egg(s) were found and new eggs were weighed (2011 only) and labeled with a nontoxic marker until either (1) a female was found incubating for the second day or (2) a female was found incubating for the first time but on the previous visit the eggs had been warm and no new egg had been laid. On alternate days, these new eggs either remained in their original nest or were cross-fostered in the same manner as the first egg, resulting in an alternating sequence of eggs laid in different nests being incubated together. In the case of triads we end up with mixed clutches of math formula, math formula and math formula where the first letter is the nest in which the eggs are reared. Cross-fostering ceased when one member of the triad satisfied condition (1) or (2), but the remaining nests were still visited and eggs marked. A schematic of the experimental design is given in Figure 1. Twelve eggs from 10 nests (2010) and three eggs from three nests (2011) were laid after nest checking had ceased, and we assume they were laid consecutively on the following days. On days when a triad was due for cross-fostering but one or more of the nests had a laying pause, eggs were cross-fostered when egg laying resumed and the pattern of crossing and noncrossing alternated as before. In 2010, a total of 108 of 114 nests were subject to cross-fostering with nine reciprocal crosses, 18 triad crosses, and nine round-robin crosses involving four nests. In 2011 a total of 151 of 163 nests were subject to cross-fostering with a single reciprocal cross, 39 triad crosses, and eight round-robin crosses involving four nests.

Figure 1.

Schematic of the experimental design for a stylized triad. The color of the eggs (white, gray, or black) designate their nest-of-origin and their number designates their laying order. Each row represents the sequence of eggs in the nest-of-rearing. All eggs are switched individually on the day they are laid.

In addition, 6% of eggs (74 of 1122 in 2010 and 70 of 1456 in 2011) were part of an additional experiment looking at laying sequence effects on chick development (Hadfield et al. 2013). Briefly, single eggs from some nests were moved into nests other than their usual destination. These additional crosses consisted of eggs laid first replacing late laid eggs and eggs laid late replacing first laid eggs together with the associated controls: eggs laid first replacing other first laid eggs and eggs laid late in the sequence replacing other late laid eggs. Chicks hatching from these eggs were omitted from the analyses.

DATA COLLECTION

Nests were checked daily for hatching from 11 days after the end of clutch completion. In 2010, nine of 114 (7.9%) nests were abandoned prior to hatching and three eggs were accidentally damaged, leaving 1044 eggs from 106 broods. In 2011, 20 of 163 (12.3%) nests were abandoned prior to hatching and eight eggs were accidentally damaged leaving 1334 eggs from 143 broods. In all but three cases (one in 2010, two in 2011), no chicks had hatched on the first visit. On the first day on which chicks hatched (Day 0), all chicks were weighed and individually marked by clipping down tufts on the head and a rear toenail. The identity of unhatched eggs was recorded and the nest revisited the following day (Day 1) at approximately the same time. Those chicks that had hatched in the preceding 24 h were marked and all chicks weighed. The nest was revisited approximately 48 h later (Day 3) and the procedure repeated and all chicks bled from the medial metatarsal vein under home office license. No egg hatched after Day 3, and eggs unhatched by Day 6 were removed for genotyping. All chicks were weighed on Day 6, 9, 12, and on Day 15 when tarsus length, wing length, and head-bill length were also taken. All morphological measurements were made by JDH and repeated measurements were taken on a random chick in 181 broods so that measurement error could be quantified. On Day 10 and onwards, parents were trapped, bled from the ulna vein, weighed and measured for the morphological traits. In total, 12,359 weights were taken from 2042 chicks with the number measured per day depending on patterns of hatching and survival. After Day 25, all nestboxes were visited and the identity of any unfledged dead chicks recorded.

Molecular methods

When multiple eggs hatch between nest visits it is not possible to assign chicks to the egg they hatched from, and consequently their nest-of-origins are unknown. To obtain this information, we genotyped all chicks, adult birds, and unhatched eggs for which blood or tissue samples were available at seven highly polymorphic microsatellite markers so that we could assign parentage.

DNA from both blood and tissue samples was extracted using DNeasy Blood and Tissue kit and the relevant protocol (Qiagen, Hilden, Germany). Individuals were genotyped using microsatellites previously isolated from a range of passerines and identified as polymorphic within blue tits (Olano-Marin et al. 2010): Pca7, Pca3 (Dawson et al. 2000); Pk12 (Tanner et al unpublished); Pocc1 (Bensch et al. 1997); PmaD22 (Saladin et al. 2003); Titgata02 and Titgata68 (Wang et al. 2005).

These were multiplexed within the following 10μl PCR; 24mM (NH4)2SO4, 0.2 mM of each dNTP, 3.5mM MgCl2 0.5 units Taq DNA polymerase, 1 μl of template DNA, and a primer mix containing 0.07μM of each Pca7, 0.15μM each Pca3, 015μM of each Titgata02, 0.12μM each Titgata68, 0.12μM each Pk12, 0.2μM of each Pocc1, 0.244μM of each PmaD22. Polymerase chain reaction (PCR) conditions were as follows: an initial denaturation step of 94°C for 1 min; then eight cycles of 94°C for 20 sec, a touchdown of 1° each cycle from 60°C to 52°C for 30 sec, 72°C for 20 sec; followed by 25 cycles of 94°C for 30 sec, 52°C for 30 sec, and 72°C for 30 sec with a final extension of 72°C for 30 min. Fragment size was resolved on an ABI 3730xl relative to 500 LIZ Size standard (Applied Biosystems) and analyzed using GeneMapper software version 4 (Applied Biosystems). In addition, we determined sex by amplifying the sex-linked markers P2 and P8 using standard protocols (Griffiths et al. 1998).

STATISTICAL ANALYSES

Parentage analysis

Genetic and nongenetic data were used to reconstruct the pedigree of sampled individuals using the Bayesian Markov chain Monte Carlo (MCMC) approach, MasterBayes (Hadfield et al. 2006). Maternity was restricted to the two possible females under the assumption that mixed-maternity clutches did not exist: for example, the potential mothers of chicks reared in nest A were restricted to those attending nests A or C. Paternity was not restricted, but we estimated the relative odds that a male attending a nest was the father of chicks whose mother was also attending that nest. In addition, we estimated the exponential decay of the odds ratio as a function of the distance between the nest attended by a male and the nest at which the mother of the chicks attended. Uniform improper priors were used for all phenotypic parameters. In cases where neither parent or only a single parent were caught at a nestbox, the model included a dummy parent with missing genotypes for the missing parent(s). This may be unjustifiable if the absent parent had been sampled at another nest (e.g., a replacement nest of a deserting female, or the primary nest of a polygynous male) or if the female at a nest had no “social” mate. The number of unsampled males were also estimated using methods outlined in Koch et al. (2008) and Waser and Hadfield (2011), and these represent the number in addition to the dummy males. In the absence of egg dumping, unsampled females are effectively dealt with by introducing dummy mothers. Approximations for dealing with genotyping error and missing genotypes were employed because exact solutions mix poorly with large sibships. The parameters of Wang's (2004) model of genotyping error were estimated from repeat samples with allelic dropout and stochastic error rates of 0.023 (0.020–0.025) and 0.008 (0.006–0.009), respectively.

The average posterior probability that the most likely female was the true mother was 0.994 and was virtually identical for the two years. The average posterior probability that the most likely male was the true male was 0.973 in 2010 and 0.99 in 2011 for chicks who had math formula posterior probability of being fathered by a sampled male. For genotyped eggs, the mean posterior probabilities for paternity were lower despite maternity being known: 0.953 (2010) and 0.985 (2011). The log odds ratio of paternity dropped by 0.032 (0.027–0.036) per meter as a function of the distance between a potential father and the offspring, but at distances of zero the effect was stronger (1.528 [1.302–1.908]), indicating that the social father is more likely to be the true father even after accounting for the effect of distance. The number of unsampled males (in addition to dummy males) was estimated to be 3.797 (2.145–6.729).

If a set of chicks reared together but belonging to two (or more) sibships have no genetic parents sampled, it is not possible to infer maternal sibship membership using parentage analysis alone. However, the round robin design means that these sibships have members in other nests which could be used to infer maternal sibships using sibship reconstruction. For example, in a triad cross, we end up with mixed clutches of math formula, math formula and math formula. If the genetic parents of A and C have not been sampled, then the clutch math formula cannot be partitioned into A and C sibships. However, any maternal sibships identified across two nests are unique in a triad design: for example, members of a sibship spanning nest A (math formula) and B (math formula) must belong to sibship A. Moreover, if the genetic parents of B have been sampled, then the clutches math formula and math formula can be partitioned using parentage analysis and members of sibship A and C identified. This information can then be used to partition math formula using sibship reconstruction. In addition, sibship reconstruction can also be used to detect paternal full/half sibships for which the father is unsampled. To use this additional information, we used rcolony (Jones and Wang 2010) an R (R Development Core Team 2012) interface to COLONY2 (Wang 2004; Wang and Santure 2009) to reconstruct sibships for offspring whose mother could not be assigned and/or whose posterior probability of being fathered by a sampled male was less than 0.5. Paternal sibships were restricted to individuals whose mothers (if known) were in nestboxes less than 340 m apart: twice the distance at which the odds ratio of extra pair paternity dropped to 0.05 from the MasterBayes analysis.

For all but one chick COLONY2 was able to assign individuals with unknown mothers into maternal sibships. Because we knew the number of eggs in each nest that were cross-fostered, additional information was available (e.g., if a nest contained five noncrossed eggs and three crossed eggs, then assigning four eggs to each of the two possible females should have a probability of zero). This information could not be used by MasterBayes or COLONY2, but was used as an external check on the assignment process. The proportion of cross-fostered eggs to noncross-fostered eggs was correct in all but two cases, both in 2011. One of these appeared to be the result of a mixed-maternity clutch for which there was supporting behavioral evidence. The 38 chicks that had been weighed at least once but died before an adequate blood sample could be taken (usually before Day 3), could be assigned maternity based on the preceding logic.

MasterBayes assigned 2.7% (2010) and 1% (2011) of chicks to an unsampled male, and 15.7% (2010) and 21% (2011) of chicks to a dummy male, reflecting the higher (male) desertion rates in 2011. To assign individuals with unsampled/dummy fathers into paternal sibships, we adopted a heuristic two stage strategy. First, we took each maternal sibship and partitioned it into full-sibships if disjoint groups could be found where all cross-group pairwise full-sib probabilities were less than some threshold math formula. These groups of full-sibs were then merged into paternal half-sibships spanning nests if the average of the cross-group pairwise half-sib probabilities was greater than some threshold math formula. The motivation behind the heuristic algorithm is that we expect most offspring within nests to be full-sibs rather than half-sibs given the mating patterns of blue tits (Kempenaers et al. 1997). We chose math formula on the following logic. If the probability that two individuals are full-sibs based on their genotypes is math formula and the probability based on prior information about the breeding system is math formula, then the total probability is math formula. If we only want to accept assignments with probability greater than math formula 0.95 and math formula 0.842 (the proportion of chicks with known social father who are within pair), then the threshold math formula 0.781, assuming that nests without an attendant male have similar patterns of paternity. Conditional on the full-sibship assignments, we derive the prior probability that two individuals in different nests have the same father as math formula 0.06, which is one over the average number of nests within a 170 m radius (32.664) multiplied by the proportion of sampled males that gained extra-pair paternity (0.507). Setting math formula 0.95, we obtain math formula 0.997. Many unjustifiable assumptions underlie these calculations, but we note that the results are largely insensitive to our choice of threshold. Even setting math formula such that math formula 0.158 and math formula 0.94 resulted in only two changes: two full-sibs were merged into a single full-sibship with their nine nest mates, and two singletons from different nests were merged into a paternal half-sibship. The 38 ungenotyped chicks were assigned to their social father. Patterns of paternity assignment are summarized in Table 1 with the aim of showing where the information is coming from for the quantitative genetic analyses. In addition, 38 of 80 males and 42 of 98 females in 2010 returned to breed in 2011, and of the 21 breeding pairs in 2010 where both individuals survived to 2011, 13 divorced. Within years, 26 of the 184 nests in which a male was caught were attended by a male that had been caught at two nests (i.e., a known polygamous male).

Table 1. Number of within-pair and extra-pair blue tit chicks when the social father was present at the nest or absent. The subscripts U and K designate genetic/social fathers of unknown (U) and known (K) identity. Nests at which the social father was absent are designated the male with the greatest share of paternity as their social father. The 38 ungenotyped chicks are excluded form the table
 Within-PairExtra-PairKExtra-PairU
Present1311106108
AbsentK1851842
AbsentU2313427

Growth

We treated weight as a Gaussian response in a linear mixed model fitted using ASReml-R (Butler et al. 2007), an R interface to ASReml (Gilmour et al. 2002). As fixed effects, we had age (as a factor), year, clutch size, hatch date (the date on which the first chick in the nest hatched), time of day (in units of 24 h), sex, and hatching interval (a three-level factor indicating whether the chick had hatched by Day 0, Day 1, or Day 3). Interactions between all variables and age (day as continuous) were fitted to capture any age-dependent changes in their effects. All continuous predictors except day were mean centered to facilitate interpretation. As random effects, we had a nest-of-rearing effect, a nest-of-origin effect, a genetic effect, and a residual or permanent environment effect. The genetic effects were assumed to have a correlation structure defined by the pedigree-derived additive genetic relationship matrix (i.e., an animal model: Henderson 1950). Age-specific covariance (7 x 7) matrices were fitted for each type of effect, although in some cases they were structured by a lower dimensional parameterization rather than the full unstructured 28-parameter model. A range of lower dimensional variance structures were used (see Table S1) and results are presented from a model with the variance structures that gave a low Akaike Information Criterion (AIC). Given the high within-clutch repeatability of egg weight (0.739 [0.704–0.792]: see also Christians (2002)), we also ran additional models to assess how much of the nest-of-origin variation could be explained by variation in egg weight across clutches. The explanatory power of egg weight could be either due to a causal effect of egg weight on body weight, a noncausal relationship mediated by correlations between egg weight and other egg characteristics, or a noncausal relationship mediated by a genetic correlation between egg characteristics and body weight. When the effect is causal, we expect within- and between-clutch regressions of egg weight on body weight to be equivalent. When the relationship is driven by a genetic correlation, we expect that the within-clutch regression is zero under the assumption that egg weight is determined by the breeding value of the individual that laid the egg rather than the individual that hatched from the egg. When the relationship is mediated by correlations between egg weight and some unmeasured egg characteristics, then the difference between the within- and between-clutch regression is not predictable without knowing how the suite of egg characteristics covary within and between clutches. We compared the within- and between- clutch regressions to assess whether there were any deviations from parity. Ideally, we would fit the expected egg weight of a clutch and the deviation of the individual's egg weight from this expectation, but unfortunately not all chicks can be uniquely assigned to eggs. Instead of the deviation of an individual's egg weight, we therefore take the deviation of the mean egg weight of chicks within the same nest-of-origin/nest-of-rearing combination that hatched in the same hatching interval. Because eggs were only weighed in 2011, we could only fit these effects for the 2011 data and so we fitted separate nest-of-origin covariance matrices for each year.

To test whether the decomposition of body mass variation at the end of ontogeny was similar to that for other morphological traits, we also fitted equivalent models to tarsus length, head-bill length, and wing length. However, these traits were only measured at one point during ontogeny (Day 15), so, we fitted univariate models to each trait by essentially dropping all terms involving day. Measurement error was estimated by fitting bird identity as a random effect and removed from the denominator in the heritability estimates (Merila and Sheldon 2001).

The significance of groups of fixed effects was assessed using conditional Wald tests, and individual effects were tested using a t-test with conservative degrees of freedom (the number of nests: 249). Significance of random terms were assessed using likelihood ratio tests, and standard errors on functions of variance components (e.g., heritabilities, correlations, covariances from autoregressive models) were obtained via the Delta method (Lynch and Walsh 1998).

Survival

Survival was modeled as a series of binary variables in a Bayesian Generalised Linear Mixed Models (GLMMs), the parameters of which were drawn from their posterior distribution using MCMC techniques implemented in the R (R Development Core Team 2012) package MCMCglmm (Hadfield 2010). Each individual has a history of eight events corresponding to Days 0, 1, 3, 6, 9, 12, 15, and 25, where the event is either alive (1) died since previous visit (0) or neither (−). For example, a chick hatching on Day 1, and found dead on Day 12 gets an event history of math formula, where the − events are essentially treated as missing. Because our analysis does not include prehatching mortality the first nonmissing event is one with probability one and so, we omit it from the analysis to give event histories of the form math formula.

As fixed effects, we had day as a categorical predictor to account for age specific changes in survival probability. Hatching interval, hatch date, clutch size, year, and weight were fitted as main effects and also interacted with day as a continuous predictor to capture any age-specific changes in their effects. All continuous predictors expect day were mean centered (either globally for time-invariant predictors, or by day for time-varying predictors [i.e., weight]) to facilitate interpretation. The event at time t is the outcome of a continuous process operating between time math formula and t. Because body weight also continuously changes over time, an exact analysis would require us to jointly model survival and weight in continuous time (Hadfield 2008), but this is challenging. Following earlier authors (Lynch and Arnold 1988), we assume that the probability of surviving from math formula to t depends only on the weight at time math formula.

Nest-of-rearing, nest-of-origin, individual, and nest-of-rearing/day were fitted as random terms. The latter term captures any daily variation in survival probability for a specific nest, and was fitted in an attempt to capture the effects of brood desertion. Residual variances were fixed at one because they are not identifiable from the data. Parameter expanded priors were used for the remaining variance components to give scaled F-distributions with numerator and denominator degrees of freedom set to one and a scale parameter of 1000. We used a prior correlation matrix for the fixed effects that would result in independent priors if the predictors had been subject to Gelman's (2008) scaling and centering. The correlation matrix was scaled to a covariance matrix by math formula where math formula is the posterior mean of the sum of the variance components from an initial analysis. Conditional on math formula, this induces a marginal prior for the fixed effects which is close to being flat on the (0,1) interval on the probability scale and can alleviate the consequences of extreme category problems. Nevertheless, extreme probabilities were predicted and so we truncated the latent variables at absolute values of 25 to avoid under/overflow. Estimates with and without truncation were very similar. Chain length, thinning interval, and burn-in period were such that 2000 samples were collected from the posterior with minimal autocorrelation. The significance of pairs of coefficients corresponding to a main effect and an interaction with day were tested by calculating a Wald statistic from the joint posterior distribution and evaluating with a χ2 test.

Results

GROWTH

The model of age-specific weights supported significant effects of all predictors either as main effects or as interactions (Table 2). The main effects are the effects at Day 0 and the interaction with day models the linear change of the effect with age. Although hatch date and clutch size had little effect on hatching weights, interactions with age existed such that by Day 15 the addition of an egg or the postponement of hatching by a day reduced weight substantially (0.081 g/egg and 0.034 g/day). Hatching interval had a strong effect throughout ontogeny and even at Day 15 chicks that hatched in the intervals 0–1 and 1–3 weighed 0.455 and 1.184 g less than chicks that hatched on Day 0. Males and females had equivalent weights at hatching, but by Day 15 males were 0.456 g larger than females.

Table 2. Summary of fixed effects from a model of age-specific weights in blue tits. The column Pr(>W) is the P-value from a Wald test that jointly tests the main effect and the interaction with day
CoefficientEstimateSEPr(>|Z|)Pr(>W)
Intercept0.9760.012<0.001 
Day 10.4600.009<0.001 
Day 31.8060.028<0.001 
Day 64.6930.064<0.001 
Day 97.6220.094<0.001 
Day 129.1880.116<0.001 
Day 159.6950.130<0.001 
Time0.2310.051<0.001 
Day:Time−0.0690.030.010<0.001
Sex (M)0.0060.0060.133 
Day:Sex (M)0.0300.002<0.001<0.001
Year (2011)−0.0740.016<0.001 
Day:Year (2011)−0.0290.0110.004<0.001
Clutch size0.0010.0040.396 
Day:Clutch size−0.0050.0030.0150.071
Hatch date0.0020.0010.090 
Day:Hatch date−0.0020.0010.0030.003
Hatch day (0–1)−0.1760.007<0.001 
Day:Hatch day (0–1)−0.0190.003<0.001<0.001
Hatch day (1–3)−0.6380.039<0.001 
Day:Hatch day (1–3)−0.0360.007<0.001<0.001

For the nest-of-rearing and residual effects, a fully unstructured matrix gave the lowest AIC and was found to be a significantly better fit than any structured model. For the nest-of-origin and genetic effects, which are smaller in magnitude and for which there is less information, it was not possible to fit some of the high-parameter structured matrices or a fully unstructured matrix. Of the structured matrices that could be fitted, the eight-parameter first order autoregressive (AR1: nest-of-origin effects) and constant correlation (CC: genetic effects) models with heterogenous variances gave the lowest AIC. Because the differences in estimates and AIC were small between an AR1 and CC model for the genetic effects, and the fact that the AR1 structure also provided one of the best fits for the nest-of-rearing and residual effects where all structured and unstructured matrices could be fitted (see Table S1), we chose to impose an AR1 structure to the genetic effects also. Random regression using orthogonal polynomials, the method of choice in many quantitative genetic studies (Kirkpatrick et al. 1990), performed poorly compared to the more widely used parametric techniques from time-series analysis and spatial statistics (Pletcher and Geyer 1999).

The proportion of variance explained by nest-of-rearing increased over ontogeny from 20.9 ± 3.1% on Day 0 to 77.8 ± 2.2% by Day 15, whereas the proportion explained by nest-of-origin was substantially less and declined over ontogeny from 13.1 ± 2.9% to 0 ± 0%. The proportion explained by the residual variance although much larger in magnitude followed a similar pattern, declining from 59.3 ± 3.8% to 14 ± 1.9%. The heritability was low but relatively stable across ontogeny, being 6.7 ± 3.6% at Day 0 and 8.2 ± 2.2% at Day 15. The full results are presented in Table 3. If nest-of-rearing and nest-of-origin effects are repeatable across the same mother in different years then our estimates of the genetic effects may be inflated. We had little power to separate these maternal effects into repeatable and nonrepeatable effects, but in the Supporting Informations (Tables S2 and S3), we present the results of models with these terms included and show that the results remain qualitatively and quantitatively very similar.

Table 3. Correlations and (co)variances for different components of blue tit weight (in grams) across different ages. Correlations are in the upper triangle. Standard errors were obtained using the delta method as implemented in the msm package (Jackson 2011). Means and variances are calculated on the raw data
ComponentDayDay 0Day 1Day 3Day 6Day 9Day 12Day 15h2
Nest-of-rearing         
 00.006 ± 0.0010.773 ± 0.0430.623 ± 0.0600.397 ± 0.0810.234 ± 0.0900.122 ± 0.0950.022 ± 0.0980.209 ± 0.031
 10.007 ± 0.0010.013 ± 0.0250.830 ± 0.0290.647 ± 0.0530.475 ± 0.0690.309 ± 0.0810.155 ± 0.0880.241 ± 0.031
 30.014 ± 0.0020.029 ± 0.0600.092 ± 0.0140.859 ± 0.0220.671 ± 0.0450.469 ± 0.0640.282 ± 0.0760.357 ± 0.032
 60.022 ± 0.0030.056 ± 0.0080.194 ± 0.0350.558 ± 0.0840.900 ± 0.0140.706 ± 0.0380.497 ± 0.0580.561 ± 0.030
 90.020 ± 0.0040.060 ± 0.0130.225 ± 0.0890.745 ± 0.1361.227 ± 0.0040.879 ± 0.0170.693 ± 0.0400.685 ± 0.025
 120.012 ± 0.0120.046 ± 0.0330.186 ± 0.1410.690 ± 0.1791.273 ± 0.0121.710 ± 0.0330.879 ± 0.0180.757 ± 0.022
 150.002 ± 0.0060.024 ± 0.0830.114 ± 0.1810.494 ± 0.2011.021 ± 0.0061.529 ± 0.0831.768 ± 0.1810.778 ± 0.022
Nest-of-origin         
 00.004 ± 0.0010.982 ± 0.0120.964 ± 0.0230.947 ± 0.0330.930 ± 0.0440.913 ± 0.0540.896 ± 0.0630.131 ± 0.029
 10.005 ± 0.0010.006 ± 0.0010.982 ± 0.0120.964 ± 0.0230.947 ± 0.0330.930 ± 0.0440.913 ± 0.0540.116 ± 0.026
 30.009 ± 0.0020.012 ± 0.0030.023 ± 0.0060.982 ± 0.0120.964 ± 0.0230.947 ± 0.0330.930 ± 0.0440.089 ± 0.022
 60.010 ± 0.0030.013 ± 0.0040.026 ± 0.0080.030 ± 0.0120.982 ± 0.0120.964 ± 0.0230.947 ± 0.0330.031 ± 0.012
 90.009 ± 0.0030.012 ± 0.0040.023 ± 0.0080.027 ± 0.0110.025 ± 0.0110.982 ± 0.0120.964 ± 0.0230.014 ± 0.006
 120.003 ± 0.0020.005 ± 0.0020.009 ± 0.0050.010 ± 0.0060.010 ± 0.0060.004 ± 0.0030.982 ± 0.0120.002 ± 0.001
 150.000 ± 0.0000.000 ± 0.0000.000 ± 0.0000.000 ± 0.0000.000 ± 0.0000.000 ± 0.0000.000 ± 0.0000.000 ± 0.000
Genetic         
 00.002 ± 0.0010.999 ± 0.0000.998 ± 0.0000.997 ± 0.0000.996 ± 0.0000.995 ± 0.0000.994 ± 0.0000.067 ± 0.036
 10.003 ± 0.0010.004 ± 0.0020.999 ± 0.0000.998 ± 0.0000.997 ± 0.0000.996 ± 0.0000.995 ± 0.0000.078 ± 0.038
 30.006 ± 0.0030.010 ± 0.0040.023 ± 0.0090.999 ± 0.0000.998 ± 0.0000.997 ± 0.0000.996 ± 0.0000.088 ± 0.036
 60.013 ± 0.0050.019 ± 0.0080.045 ± 0.0160.088 ± 0.0290.999 ± 0.0000.998 ± 0.0000.997 ± 0.0000.088 ± 0.029
 90.016 ± 0.0060.025 ± 0.0090.058 ± 0.0190.114 ± 0.0340.148 ± 0.0420.999 ± 0.0000.998 ± 0.0000.082 ± 0.024
 120.018 ± 0.0060.027 ± 0.0090.063 ± 0.0180.124 ± 0.0330.161 ± 0.0410.176 ± 0.0440.999 ± 0.0000.078 ± 0.020
 150.018 ± 0.0060.028 ± 0.0080.065 ± 0.0160.128 ± 0.0290.166 ± 0.0380.182 ± 0.0430.187 ± 0.0470.082 ± 0.022
Residual         
 00.016 ± 0.0010.870 ± 0.0100.821 ± 0.0140.752 ± 0.0200.589 ± 0.0310.351 ± 0.0450.006 ± 0.0590.593 ± 0.038
 10.020 ± 0.0010.031 ± 0.0120.880 ± 0.0080.811 ± 0.0140.643 ± 0.0260.387 ± 0.042−0.007 ± 0.0590.564 ± 0.037
 30.036 ± 0.0020.053 ± 0.0220.119 ± 0.0060.877 ± 0.0090.716 ± 0.0220.419 ± 0.041−0.036 ± 0.0620.466 ± 0.035
 60.054 ± 0.0020.080 ± 0.0050.171 ± 0.0130.318 ± 0.0210.849 ± 0.0120.584 ± 0.0320.104 ± 0.0600.320 ± 0.029
 90.047 ± 0.0030.070 ± 0.0070.155 ± 0.0230.299 ± 0.0260.390 ± 0.0030.771 ± 0.0200.353 ± 0.0520.218 ± 0.023
 120.027 ± 0.0070.041 ± 0.0140.088 ± 0.0290.200 ± 0.0300.292 ± 0.0070.368 ± 0.0140.676 ± 0.0300.163 ± 0.019
 150.000 ± 0.004−0.001 ± 0.025−0.007 ± 0.0320.033 ± 0.0330.124 ± 0.0040.231 ± 0.0250.317 ± 0.0320.140 ± 0.019
Mean         
  0.9971.3502.6505.5738.52310.24810.944 
Variance         
  0.0260.0740.3541.1021.8271.7531.226 

In a model including the 2011 egg weight covariates, the mean clutch egg weight was a significant predictor of mass throughout ontogeny (P<0.001): on Day 0 an increase in mean egg weight by 0.1 g increased mass by 0.061 ± 0.005 g and the change in the effect across days (0.01 ± 0.003 g/day) results in a predicted difference at Day 15 of 0.211 g. The regression on the within-clutch deviations was shallower with an increase in the egg weight deviation of 0.1 g causing an increase in mass of 0.027 ± 0.015 g on Day 0. The change in the effect across days was 0.01 ± 0.006 g/day, resulting in a predicted difference at Day 15 of 0.175 g. A joint test of whether the within- and between-clutch regression parameters (the main effect and the change with age) differed was not significant (P = 0.091), although power was low due to the low within-clutch variation in egg mass together with the limited ability to capitalize on this variation because many chicks could not be assigned to unique eggs.

When egg weights were not used as predictors, a model in which year specific nest-of-origin effects were fitted resulted in a minimal change in log-likelihood despite the addition of eight parameters (math formula = 1.741) and the proportion of variance explained by nest-of-origin was comparable in the two years (declining from 11.9 ± 3.6% to 0 ± 0% in 2010 and from 13 ± 3.5% to 0.2 ± 0.3% in 2011). However, when fitting egg weight effects (in 2011 only), the change in log-likelihood was greater (math formula = 10.989) and year specific nest-of-origin effects were supported using a likelihood ratio test (P = 0.005). In this model, nest-of-origin effects declined from 15.6 ± 4.1% to 0 ± 0% in 2010 and from 0.9 ± 1% to 0 ± 0% in 2011.

Under a missing at random process, the full multivariate model accounts for any reduction in variance caused by selection (Lynch and Arnold 1988; Hadfield 2008). Because there is a strong association between body mass and survival, either directly or through common predictors of mass and survival (see below), we may expect the effects of selection bias to be substantial if univariate analyses at the end of ontogeny were performed. Indeed a univariate analysis of Day 15 weights with the same fixed effects had a predicted phenotypic variance (after accounting for the fixed effects) of 1.166 compared to 2.273 from the multivariate analysis (Fig. 2). Moreover, the phenotypic variance was partitioned more equitably in the univariate model: nest-of-rearing, 69.3 ± 2.8%; nest-of-origin, 0 ± 0%; genetic, 16.5 ± 3.5%; residual, 14.1 ± 2.5%, indicating that the association between survival and nest-of-rearing effect was disproportionately large.

Figure 2.

Growth trajectories of blue tits (in grams) with the observed mean (thick solid line) ± 2 standard deviations (SD; thin solid lines) and the predicted mean (thick dashed line) ± 2 predicted SDs (thin dashed line). The predicted SDs include variation due to the fixed effects to make them more comparable to the raw standard deviations. The predicted trends take into account the selective mortality experienced by earlier age classes under the assumption of a missing at random (MAR) process.

Univariate models for the three morphometric traits measured on Day 15 showed that males hatching first in an early nest have longer tarsi, wings, and head-bill lengths, consistent with the effects found on weight (see Table 4). The models all gave estimates of zero for the nest-of-origin effects (see Table 5), but varied substantially with respect to the proportion of variance explained by nest-of-rearing (tarsus: 33.8 ± 3.8%; head-bill: 62.4 ± 3.5%; wing length: 73.6 ± 2.4%) and heritabilities (tarsus: 63.2 ± 9.3%; head-bill: 31.8 ± 7.3%; wing length: 7.4 ± 3.2%). Measurement error was excluded from the denominator of these proportions.

Table 4. Summary of fixed effects from univariate mixed models of blue tit tarsus length, head-bill length, and wing length (all in mm)
CoefficientTarsusHead-BillWing
Intercept16.845 ± 0.058, P<0.00122.361 ± 0.068, P<0.00145.111 ± 0.376, P<0.001
Time−0.059 ± 0.199, P = 0.383−0.096 ± 0.257, P = 0.354−0.052 ± 1.356, P = 0.485
Sex (M)0.505 ± 0.023, P<0.0010.341 ± 0.021, P<0.0010.891 ± 0.102, P<0.001
Year (2011)−0.326 ± 0.078, P<0.001−0.518 ± 0.102, P<0.001−1.673 ± 0.579, P = 0.002
Clutch size0.007 ± 0.018, P = 0.3450.011 ± 0.024, P = 0.3240.068 ± 0.134, P = 0.306
Hatch date−0.023 ± 0.006, P<0.001−0.046 ± 0.008, P<0.001−0.197 ± 0.046, P<0.001
Hatch day (0–1)−0.061 ± 0.026, P = 0.010−0.185 ± 0.024, P<0.001−1.610 ± 0.116, P<0.001
Hatch day (1–3)−0.370 ± 0.056, P<0.001−0.666 ± 0.051, P<0.001−5.477 ± 0.249, P<0.001
Table 5. Summary of different variance components for blue tit tarsus length, head-bill length, and wing length (all in mm). Standard errors for the nest-of-origin variance components were obtained from models where the variances were not constrained to be nonnegative. Means and variances are calculated on the raw data
ComponentTarsusHead-BillWing
Nest-of-rearing0.104 ± 0.0150.222 ± 0.0267.862 ± 0.865
Nest-of-origin0.000 ± 0.0100.000 ± 0.0090.000 ± 0.095
Genetic0.194 ± 0.0290.113 ± 0.0200.787 ± 0.255
Residual0.009 ± 0.0160.021 ± 0.0132.039 ± 0.200
Measurement error0.030 ± 0.0030.036 ± 0.0040.326 ± 0.037
Mean16.95722.24744.233
Variance0.3950.45512.245

SURVIVAL

All pairs of fixed effects (the main effect and interaction with day) were all significant except clutch size (P = 0.497). The effects were all in the predicted direction with light chicks hatching last within a large late clutch suffering the greatest mortality (Table 6). The intraclass correlation in survival (after conditioning on the fixed effects) for nest-of-rearing/day was high 0.625 (0.483–0.697) and is likely to mainly capture the effects of brood desertion. The general nest-of-rearing intraclass correlation was moderate 0.190 (0.095–0.329), indicating that between nest variation in mortality also existed even after accounting for brood desertion. The nest-of-origin intraclass correlation was very small 0.000 (0.000–0.032) as was the permanent evironment intraclass correlation 0.000 (0.000–0.007).

Table 6. Summary of fixed effects from a model of survival in blue tits. Mean is the posterior mean, l-95% and u-95% are the lower and upper 95% credible intervals, and pMCMC is twice the posterior probability that the estimate is negative or positive (whichever probability is smallest). The column Pr(>W) is the P-value from a Wald test that jointly tests the main effect and the interaction with day
CoefficientMeanl-95%u-95%pMCMCPr(>W)
Intercept12.43910.75713.978<0.001 
Day 3−1.418−2.9770.2650.080 
Day 6−1.795−3.370−0.2670.026 
Day 90.175−1.5561.8130.807 
Day 12−0.689−2.5060.9130.440 
Day 15−1.814−3.5100.0320.046 
Year 2011−1.691−3.3420.1830.059 
Day:year (2011)−0.197−0.320−0.0700.003<0.001
Clutch size−0.156−0.5740.2880.467 
Day:clutch size−0.003−0.0330.0250.8670.497
Hatch date−0.219−0.357−0.0700.002 
Day:hatch date−0.004−0.0150.0050.446<0.001
Hatch day (0–1)−0.390−1.2430.4410.370 
Day:hatch day (0–1)−0.019−0.0740.0350.4870.005
Hatch day (1–3)−0.125−2.1651.6230.917 
Day:hatch day (1–3)−0.065−0.1860.0500.3080.219
Weight2.1561.5422.806<0.001 
Day:weight−0.001−0.0360.0370.957<0.001

Discussion

To our knowledge this is the first study that has been able to distinguish the effects of genes, eggs, and postlaying parental effects on phenotypic expression. All of these effects are shown to play a role in determining body mass at some point during ontogeny. In many studies these effects are confounded, and yet researchers tend to ascribe the effects to the single source that most interests them. For example, both Kölliker et al. (2000) and Hinde et al. (2010) performed similar cross-fostering experiments on passerine birds and showed a positive correlation between parental provisioning and offspring begging. However, the former study interpreted the correlation as being predominantly genetic, whereas the latter study interpreted the correlation as being predominantly due to prenatal environmental effects. The fact that neither experiment could distinguish between these hypotheses was acknowledged in both cases and indirect evidence supporting the alternative hypotheses was presented. Our results provide direct evidence that for some types of trait both processes may contribute, but for traits expressed later in ontogeny genetic effects may be more important. Whether this generalization will hold across different types of traits remains to be tested.

In this study, we were able to obtain holistic measures of the magnitude of egg effects without needing to understand the mechanisms by which variation in these egg effects are generated. Studying the effects of specific egg characteristics, such as size or maternal androgen concentrations, offer another route for understanding how between-clutch variation in egg effects can be generated. Here, we show that these effects can be largely captured by measures of egg mass, and note that the rapidly diminishing importance of egg effects over ontogeny is consistent with the literature on the effects of egg mass on offspring phenotype (Williams 1994; Christians 2002; Krist 2011). However, as stated by Krist and Remes (2004), direct genetic correlations between egg mass and offspring traits may generate patterns that although consistent with between-clutch egg effects are actually due to the direct effect of genes on offspring phenotype. Because the within-clutch relationship between egg mass and offspring phenotype cannot be genetic (Krist and Remes 2004), we compared it to the between-clutch relationship to gauge whether the latter may be biased by genetic effects. Although there was no compelling evidence for bias, the power of this test was relatively low.

In a review of egg-size effects in birds, Krist (2011) finds that effect sizes on various offspring traits are roughly comparable between observational studies and cross-fostering studies in which any correlation between parental investment pre- and postoviposition is broken. Although this result suggests that such correlations may be weak, Krist (2011) also demonstrated the generality of early results (Amundsen and Stokland 1990; Reid and Boersma 1990) showing that offspring traits are positively correlated with the egg size of their foster parents. Together these observations provide support for an earlier suggestion (Krist and Remes 2004) that parents may modulate the degree of parental care they give depending on the egg size from which their foster-offspring hatch (see also Russell et al. 2007). If such a process is operating in our system, then we can underestimate the variation in egg effects if parents reduce the amount of care to chicks from large eggs. However, if any modulation of parental care is determined by the mean egg size of the raised clutch, then the reciprocal cross-fostering design used here would be superior to the complete cross-fostering design more commonly used. The design used by Krist and Munclinger (2011) whereby a foster clutch contains eggs from many (>2) original clutches would be even less biased under these circumstances because the correlation between an individual's egg-size and the mean egg-size of the raised clutch would be even smaller. However, such a design has limited power for exploring the genetics of sibling competition (Bijma 2010)—the main motivation behind our experiment.

If we had assumed that egg effects were absent, like most analyses in studies with a quantitative genetic focus, then estimates of the additive genetic variance would be biased by twice the egg effect variance (assuming individuals from the same clutch are full siblings) to give heritabilities declining from 0.329 (Day 0) to 0.082 (Day 15) across ontogeny. However, heritability estimates from other cross-fostering studies in blue tits are often considerably larger than these expectations (Kunz and Ekman (2000): 0.80 (Day 4) to 0.35 (Day 15), Hadfield (2005): 0.782 (Day 1) to 0.295 (Day 15), and Nilsson et al. (2009): 0.084 (Day 14)). In most reciprocal cross-fostering experiments, chicks rather than eggs are usually swapped, and swapping occurs between nests in which some chicks had hatched the previous day (reviewed in Merila and Sheldon 2001). Our data suggest that the difference between these results and our expectations, at least early in ontogeny, is due to the existence of posthatching but pre-cross-fostering effects in chick cross-fostering experiments. Our data show that the nest-of-rearing variation in mass on Day 1 (i.e., the day on which chick cross-fostering is usually done) would contribute 0.021 to the covariance between chicks from the same nest-of-origin resulting in heritability estimates of 0.708 on Day 1, roughly in line with that observed. This estimate is based on a model where hatching interval was not controlled for, and so includes variation due to different levels of parental care prior to Day 1 and also the effects caused by age differences (on the order of hours) between broods that hatch on the same day. Protocols that match the nests between which chicks are exchanged by mean chick mass (Brinkhof et al. 1999) or the proportion of chicks hatching on Day 0 (Hadfield and Owens 2006) should reduce these biases.

Age differences between the two half-broods may also exist when eggs are cross-fostered if eggs from the same clutch have more similar hatching times after controlling for incubation behavior. Between-clutch variation in hatching rate would then generate greater variance in egg effects than would be observed in unmanipulated nests. In a companion paper (Hadfield et al. 2013), we find evidence that there is indeed between-clutch (and within-clutch) differences in intrinsic hatching rate. The magnitude of these differences is such that the expected absolute difference in hatching time between two half broods raised together is equal to 3.439 (2.450–4.663) h. By controlling for hatching interval in the analysis, we hoped to reduce the upward bias caused by age differences, but we acknowledge that because the exact time of hatching is not known for most chicks then such biases are likely to persist. Although not ideal, we note that in chick cross-fostering designs, nests are usually checked every 24 h, and so if hatching is uniform across this interval then the average absolute difference in hatching time between two broods is expected to be 8 h, generating an even larger bias. Although the bias due to age differences may dissipate later in ontogeny, it is possible that under resource limitation these differences become magnified as the smaller size of the younger chicks puts them at a competitive disadvantage that further reduces their relative size.

Using observational data, Charmantier et al. (2004) estimated heritabilities of 0.267, 0.349, and 0.638 for body mass (Day 14/15) in three populations of blue tit; estimates which are more comparable to the results from chick cross-fostering than those reported here. These estimates cannot be biased upwards by experimental artifacts that generate age differences between chicks raised together. However, in such observational data, information regarding heritability comes from comparing the phenotypes of relatives that have not been raised in the same brood, such as siblings born in different years, or parents and their offspring. Sources of similarity between these types of relative not due to the direct effect of genes have been shown to be common (van der Jeugd and McCleery 2002; Kruuk and Hadfield 2007; Stopher et al. 2012), including inherited aspects of the environment provided by parents (Räsänen and Kruuk 2007) even those associated with eggs (Tschirren et al. 2009). Cross-fostering not only reduces these sources of similarity (because [some] relatives are distributed across different environments), but also provides greater power for identifying what aspects of the environment may cause familial resemblance. Whether these unaccounted for environmental sources of similarity can fully explain the discrepancies between studies, and if so, what these sources are remain to be explored. In this study, nest-of-rearing effects have been treated as environmental in origin, and no attempt was made to decompose them into nongenetic effects and genetic effects that arise through the action of heritable traits expressed in parents and/or siblings. In the future, we hope to obtain sufficient cross-fostering replicates to address this difficult question.

Body mass was found to be under strong positive directional selection across ontogeny with odds ratios of 2.185 (1.698–2.673) and 2.163 (1.936–2.366) per gram on Day 0 and Day 15, respectively. Moreover, the consistent negative effects of increased clutch size, hatch date, and hatching interval on both body mass and survival imply that the association, causal or not, between body mass and survival is even greater. Because most studies do not collect information throughout ontogeny, and if they do the data are not always analyzed in a multivariate framework, then most studies are expected to generate downwardly biased estimates of the variance components due to selection bias (Lynch and Arnold 1988; Im et al. 1989; Hadfield 2008). In this study, there was nearly a twofold difference in the predicted phenotypic variance at Day 15 between a univariate and a multivariate model, indicating that selection had already eliminated a great deal of variation by the end of the fledgling period. In addition, the univariate analysis gave smaller estimates of the environmental variance components (particularly nest-of-rearing) and consequently estimates of the heritability that were twofold higher than the multivariate analysis. This is consistent with the large nest-of-rearing effect on survival and suggests that unaccounted for sources of covariation between mortality and body mass exist at the nest-of-rearing level.

In conclusion, our results underline the importance of exercising caution when interpreting the causes of familial resemblance both in experimental systems but particularly in observational studies where such biases are harder to gauge. In addition, we urge researchers to collect relevant phenotypic data throughout early development given that the distribution of weights toward the end of the nestling phase had already been dramatically modified by phenotypic selection in this study. When looking at how variation in egg characteristics translate into selected phenotypic differences between offspring, studies that only focus on traits expressed toward the end of development are in real danger of being too late.

Associate Editor: A. Charmantier

ACKNOWLEDGMENTS

The authors thank S. Farrell, A. Leeuw, and K. Stopher for helping in collecting data, B. Hansson, M. Ljungqvist, and J. Pemberton for advice on the genotyping protocol, and E. Cunningham, P. Korsten, L. Ross, P. Smiseth, T. Uller, A. Charmantier, and two anonymous reviewers for comments and useful discussions. This work was funded by Natural Environment Research Council and Royal Society fellowships to JDH and supported by Lord Rosebery and Dalmeny estate.

Ancillary