Kause Section of Ecology, Department of Biology, and Kevo Subarctic Research Institute, University of Turku, FIN-20014 Turku, Finland. Tel: +358 2 3335717; fax: +358 2 3336550; e-mail: Antti.Kause@utu.fi
Rapid larval growth in insects may be selected for by rapid ephemeral phenological changes in food resources modifying the structure of phenotypic and genetic (co)variation in and among individual traits. We studied the relative effects of three processes which can modify expression of additive genetic and nongenetic variation in traits. First, natural selection tends to erode genetic variation in fitness-related traits. Second, there may be high variance even in traits closely coupled with fitness, if these traits are themselves products of variable lower level traits. Third, traits may be canalized by developmental processes which reduce phenotypic variation. Moreover, we investigated the phenotypic and genetic role played by the underlying traits in attaining simultaneously both large size and short development time.
We measured phenotypic and genetic (co)variation in several pre- and post-ingestive foraging traits, growth, development rate, development time and size, together forming a hierarchical network of traits, in the larvae of a flush feeding geometrid, Epirrita autumnata. Rapid larval growth rate and high pupal mass are closely related to fitness in E. autumnata. Traits closely associated with larval growth displayed low levels of additive genetic variation, indicating that genetic variability may have been exhausted by selection for rapid growth. The body size of E. autumnata, in spite of its close correlation with fitness, exhibited a significant additive genetic variation, possiblye because caterpillar size is the outcome of many underlying heritable traits. The low level traits in the hierarchical net, number (indicating larval movements) and size of feeding bouts in leaves, relative consumption rate and efficiency of conversion of ingested food, displayed high levels of residual variation. High residual variation in consumption and physiological ability to handle leaf material resulted from their flexibility which reduced variation in growth rate, i.e. growth rate was canalized. We did not detect a trade-off between development time and final size. On the contrary, large pupal masses were attained by short larval periods, and this relationship was strongly genetically determined, suggesting that both developmental time and final size are expressions of the same developmental process (vigorous growth) and the same genes (or linkage disequilibrium).
In many cases, life-history traits display lower heritability values than morphological, behavioural and physiological traits. This has been suggested to be the result of strong selection on closely fitness-related life-history traits (Gustafsson, 1986; Mousseau & Roff, 1987; Roff & Mousseau, 1987). On the other hand, life-history traits may be constructed hierarchically from underlying morphological, physiological and behavioural characters, in which case genetic and environmental components of life-history traits are indirectly made up by the underlying characters (Price & Schluter, 1991; Houle, 1992; Falconer & Mackay, 1996). Therefore, it has been suggested that both genetic and environmental variation can accumulate step-by-step from many sequential basal traits to higher level traits. Since heritability is the proportion of phenotypic variance attributable to genetic effects, low heritabilities in life-history traits, which are typically at the top of the trait hierarchies, may result from high environmental variation accumulated from lower level traits (Price & Schluter, 1991). Accordingly, the use of heritability as a measure of genetic variation has been criticized, and after reanalysing previously reviewed studies Houle (1992) concluded that life-history traits generally show high levels of both additive genetic and nongenetic variation, possibly because they are affected by many heritable traits and a great number of environmental events.
Accumulation of variation in trait networks is not an additive process. There are developmental mechanisms which reduce both environmental and genetic variation in traits during development of an individual, i.e. the traits are canalized (Waddington, 1942; Atchley, 1984; Scharloo, 1991; Stearns & Kawecki, 1994; Stearns et al., 1995). This further complicates the interpretation of observed patterns in variation among different traits. Accordingly, a major task in evolutionary research is to detect and explain the degree of variation in traits which occupy different hierarchical positions, and which experience different selection pressures.
Analyses of life-history strategies are commonly based on two assumptions: that selection maximizes the fitness of an organism, and that there exist trade-offs which limit the set of possible life-history combinations (Roff, 1992; Stearns, 1992). Indeed, genetic trade-offs, i.e. negative genetic correlations, are more common among life-history traits compared to more weakly selected traits (Roff, 1996). Behavioural or physiological traits may be genetically correlated with life-history traits, and certain environmental variables may in practice allow only positive correlations; hence evolutionary changes in an upper level trait may be constrained by both the genetic properties of lower level traits and by the environment. Thus the simultaneous study of underlying traits, which respond directly to environmental variability, and life-history traits, may lead to a better understanding of the constraints and evolvabilities of life-history strategies than the study of life-history traits alone.
In this paper we present the results of a quantitative genetic analysis of several pre- and post-ingestive foraging traits, growth, development time and final size, together forming a hierarchical network of traits, in autumnal moth caterpillars, Epirrita autumnata (Borkhausen) (Lepidoptera: Geometridae). By estimating heritabilities and by dividing phenotypic variances into coefficients of additive genetic and residual variation, our aim was to elucidate the relative effects of natural selection, the position-specific accumulation of variation in the trait hierarchy and developmental canalization on additive genetic and residual variation. We also describe the roles of phenotypic and genetic factors for simultaneously attaining both short larval period and large size by behavioural and physiological traits related to food utilization.
The study was conducted at the station of Kevo Subarctic Research Institute in northern Finland (69° 45′ N, 27° 01′ E). The stock of animals used were the offspring of individuals used in experiments during the previous years. These originated from the local natural population and had never been reared on artificial diets; the genetic variability of the experimental strain was maintained by annually crossing wild-caught moths with individuals from the rearings. To obtain paternal half-sib families for the estimation of genetic parameters, each of 14 haphazardly selected males was mated with two females. The eggs were overwintered in an underground cellar and brought to the laboratory in the spring. Before the experiment, the larvae were individually reared in 48-mL plastic vials outdoors on fresh mountain birch leaves from trees other than those used in the experiment. The leaves were renewed every 3–4 days. The positions of the vials were randomized among vial frames. To synchronize larval development, larvae which had started to moult to the fourth instar were kept at a temperature of +1 °C. By this treatment we were able to halt the moulting process, apparently without any detrimental effects on these subarctic larvae.
The 26 mountain birch trees used in the experimental rearings were selected from the slope of the hill near the station. The large number of trees produced a magnitude of phenotypic variation in autumnal moth size similar to that observed under natural conditions. In the current study, the coefficient of phenotypic variation in pupal masses was 0.16. The average coefficients in femur length of pheromone-trapped adult males and in pupal masses achieved by larvae grown in mesh bags in natural birch stands were 0.06 and 0.23, respectively (both calculated for 17 sites from the data in Kaitaniemi et al., 1997). Accordingly, laboratory conditions did not reduce natural variation, which may increase calculated heritabilities and weaken differences between genetic and phenotypic correlations (see Willis et al., 1991; Weigensberg & Roff, 1996).
The growth experiment in the laboratory was started with larvae at the beginning of the fourth instar on leaves from our experimental trees. We used 26 larvae from each of the 28 full-sib families; each individual larva from a full-sib family was reared on leaves from one of the 26 experimental trees in a 100-mL plastic vial. Consequently, each larvae in a full-sib family received leaves from a different tree. We offered leaves attached to intact short shoots, carrying on average three leaves, haphazardly picked from the canopy. After the larvae had moulted to the fourth instar, they were weighed to the nearest 0.1 mg, allowed to feed on short shoot leaves from the experimental tree for 24 h at +12 °C, and reweighed. After the experiment, the experimental leaves were collected and pressed. The leaf areas fed on and the numbers and sizes of meals were analysed using an image analyser system (MCID, M4, Imaging Research Inc., Brock University, Ontario, Canada) connected to a computer. An additional sample of five leaves per tree was collected, weighed fresh, pressed and dried at +60 °C and reweighed; leaf areas were also measured, in order to transform areas fed into leaf masses fed. The numbers of leaves given to the larvae were excessive to consumption; the caterpillars on average consumed only 10 percentage of the total available leaf mass (23.7 mg of the total 243 ± 60.2 (SD) (mg)).
After completing the 24-h growth bioassay, we continued rearing the larvae individually up to pupation, each on leaves excised from its respective experimental tree. Consequently, each individual larva was reared on leaves from one individual tree throughout the experiment, as is likely to happen in the nature during late larval development. The leaves were replaced every 3 days, and old leaf remains and faeces were removed. The larvae stop feeding when they start to prepare for pupation, which can be easily recognized from the stubby form of the larvae. The larvae were checked twice a day to detect the cessation of feeding. They were allowed to pupate individually in moist moss; the pupae were weighed and sexed 2 weeks after pupation.
For each larva, we recorded a total of nine traits. From the measurements of the 24-h growth bioassay, we calculated three pre-ingestive traits (number and size of meals, as indicated by holes or notches in leaves, and relative consumption rate, RCR), one post-ingestive trait (efficiency of conversion of ingested food to larval biomass, ECI), relative growth rate (RGR) and larval mass at the beginning of the experiment. In addition, development rate as well as length of larval period and pupal mass were recorded. All traits are based on fresh masses.
The number of meals describes distribution of feeding sites and movements of larvae. RCR, ECI, RGR and development rate were calculated using analyses of covariance to avoid ratio-based variables (Waldbauer, 1968; Scriber & Slansky, 1981). Ratio-based food utilization indices are not mathematically independent from each other and their distributions are usually unknown, reducing their usefulness in quantitative genetics (Raubenheimer & Simpson, 1992; Roff, 1992; Falconer & Mackay, 1996). By using ANCOVA (procedure GLM, SAS Institute, 1990), interdependence of variables can be reduced (e.g. Raubenheimer & Simpson, 1992), although probably not totally removed, and traits can be presented as a value of a dependent variable. In each ANCOVA analysis, variation caused by the covariate(s) was removed from the dependent variable (Table 1). The residuals of the models, with a mean of zero, are deviations of individual trait values from a predicted population mean, statistically independent of the covariates. To make calculation of coefficients of variation possible, the population mean of a given dependent variable was further added to the residuals. Consequently, we calculated new individual values which were corrected from the systematic variation due to heterogeneity of our sample. These values were used in further analyses. It is worth mentioning that our nutritional indices describe food utilization similar to the traditional ratio-based indices shown in Table 1, and that conclusions based on the two methods did not differ. For instance, we investigated whether there is genetic variation in relative mass gain during 24 h feeding (RGR), not whether there is variation in final mass after 24 h feeding. However, our method reduced the phenotypic correlations between the covariate(s) and the corresponding dependent variable to zero. Our measure of relative growth rate covers the exponential phase of growth within the fourth instar thus representing a strict physiological measure of growth, while development rate consists of episodes of growth and moulting (Ayres & MacLean, 1987a) within the last two instars and therefore describes more generally the larva's development and use of resources to gain biomass. Development time and length of larval period here refer to the duration of the last two instars. Statistical analyses were performed with individuals for whom all the above traits had been recorded.
Table 1. Dependent variables and covariates used in analysis of covariance to calculate food utilization and growth traits. Corresponding traditional food utilization indices (Waldbauer, 1968) are also shown.
Fitness connections of the traits
Fitness connections of the studied traits are well known. Rapid growth and development are strongly selected because delay in development and consequent asynchrony with the declining leaf quality of maturing mountain birch leaves, the main host plant, leads to high mortality and small size among Epirrita-larvae (Ayres & MacLean, 1987b; Haukioja et al., 1988; Tammaru, 1998; for other species see references in Scriber & Slansky, 1981; Ayres & MacLean, 1987a). In addition, especially during population peaks, rapid growth is favoured by selection, since the lack of mountain birch leaves may to a great extent prevent long larval growth periods (Bylund, 1995). Consequently, E. autumnata is thought to be a growth rate maximizer (Haukioja, 1981; Ayres & MacLean, 1987b), and its growth rate is one of the highest among lepidopteran larvae (Slansky & Scriber, 1982). Selection for rapid development in a time-limited environment is also expected in accordance with life-history theory (Roff, 1992 ; Stearns, 1992). Adult size is also strongly selected, since large adult Epirrita-female size correlates linearly with high potential and realized fecundity, also in field conditions. Large males find mates more quickly and are able to fertilize more eggs compared to smaller males (Haukioja & Neuvonen, 1985; Tammaru et al.,1996a,b; reviewed in Honek, 1993; but see Leather, 1988). It is not known how outbreaks in population density change the selection pressures, but it is possible that selection gradients of size and length of larval period differ considerably among different phases of population densities. Effective physiological utilization of ingested food and fast relative consumption rate construct rapid growth (Ayres & MacLean, 1987b; reviewed in Scriber & Slansky, 1981; Slansky, 1993), and hence they may be vital for Epirrita. Larval movements may be weakly selected, since the fitness consequences appear to be realized indirectly via diet selection within heterogeneous tree canopies (Niemelä & Haukioja, 1982), and this selection pressure may vary in time and space.
The QUERCUS-program (Shaw & Shaw, 1994), applying the restricted maximum likelihood method (Shaw, 1987), was used for the quantitative genetic analyses. This method is especially appropriate since it can accommodate unbalanced designs such as ours. The program estimates additive, dominance and environmental components from phenotypic variances and covariances of known relatives, and tests their significance. After square root-transformation of size of meals, trait values did not deviate from the normal distribution. The only exception was number of meals whose results should be viewed with caution.
Phenotypic and genetic correlations were first calculated for each sex separately. We tested in two ways whether the correlations differed between sexes (according to Willis et al., 1991; Roff, 1996). First, we regressed the correlations of males on the correlations of females and tested whether the slopes differ from unity (reduced major axis-method of type-II regression, Clarke, 1980). Second, the mean absolute differences between pairs of correlations were calculated to assess how close, on average, are the two sets of correlations. The slopes of regressions did not deviate from unity (phenotypic correlations: b = 1.00 ± 0.04 (SE), n = 32; intercept: −0.04 ± 0.01; genetic correlations: b = 0.98 ± 0.11, n = 35; intercept: −0.06 ± 0.04), and the mean absolute differences were small (phenotypic: 0.07 and genetic correlations: 0.23). The analyses indicate that the correlations of both sexes revealed similar patterns, although phenotypic correlations of females were slightly stronger compared with males. Also heritabilities and the coefficients of variation revealed similar patterns for both sexes. Therefore, we analysed sexes simultaneously in the same analyses, to increase statistical power.
Variation in the traits
Univariate analyses, with sex as a fixed effect, were performed for each trait to estimate additive genetic and phenotypic variances. We tested whether the additive genetic component was significantly different from zero by calculating twice the difference between the likelihood of a full model and a model in which the tested additive component was set to zero. The test statistics were compared to the χ2 distribution with degrees of freedom being the number of tested components constrained (one in this case) to zero (Shaw & Shaw, 1994).
Two measures of additive genetic variation were computed. Narrow-sense heritability was calculated as the ratio of additive genetic variance to phenotypic variance (Falconer & Mackay, 1996). The standard error of heritability was computed after Becker (1984). The heritability was considered nonzero when the additive genetic variance component differed significantly from zero. Heritability, together with the selection differential, determines the absolute short-term evolutionary change of a trait. The additive genetic coefficient of variation, on the other hand, is regarded as a measure of relative evolvability to compare variation in traits measured in ratio scale within and between species (see Zar, 1984; Charlesworth, 1987; Houle, 1992; Lynch & Walsh, 1998). Additive genetic and residual coefficients of variation were calculated as
respectively, where VA is additive genetic and VP phenotypic variation and x¯ is the mean of a given trait.
Covariation between traits
Bivariate analyses, with sex as a fixed effect, were performed to estimate narrow-sense genetic and phenotypic correlations. We tested whether a given covariance between two traits was significantly nonzero by doubling the difference between the likelihoods of a full model and a model in which respective covariances were restricted to zero. Degrees of freedom again were the number of tested covariance components set to zero. All correlations between two traits, X and Y, were calculated as
where V is the variance and COV the covariance, either additive genetic or phenotypic (co)variance, between the two traits (Falconer & Mackay, 1996). The correlation was considered nonzero when the corresponding covariance differed significantly from zero. We followed the practice of Shaw & Platenkamp (1993) and calculated narrow-sense genetic correlations between traits whose VA were positive. In the case of the covariance between number and size of meals, we were unable to calculate the genetic correlation because the bivariate analysis yielded negative additive genetic variance estimates despite positive estimates in the univariate test. Given the low heritabilities of certain traits, estimates of genetic correlations among these traits should be viewed with caution.
If the estimation of (co)variances was restricted by negative dominance variance components, negative values were constrained to zero, after which the restricted model was regarded as equivalent to the full model. The hypotheses (V = 0 or COV = 0) were then tested against this model with corresponding degrees of freedom.
Difference between phenotypic and genetic correlations was first tested by regressing (reduced major axis-method of type-II regression, Clarke, 1980) genotypic correlations on phenotypic ones (Roff, 1996) and, second, by calculating the average of the absolute differences between pairs of correlations (Willis et al., 1991).
We performed path analyses to illustrate complex correlation table and trait network in a simple descriptive form (for introduction see Mitchell, 1993). Using path analysis, correlations among traits can be divided into direct effects and effects via other traits. Direct effects, i.e. path coefficients, are standardized partial regression coefficients which describe the effect of a given variable on another while keeping all other traits statistically constant. The CALIS procedure and its RAM statement were used to carry out the path analyses (SAS Institute, 1990). The maximum likelihood option was used to calculate the χ2 goodness-of-fit of the models, and high P values indicate no significant deviation between a path model and an observed correlation structure.
Calculated path analyses were exploratory, but knowledge, or assumptions, about the relationships among the traits is required to determine the directions of the path coefficients. The following causality was assumed. When a larva starts to disperse its feeding sites, the movement of the larva interrupts the current meal, leading to decreased meal size. Both the number and the size of the meals may affect the relative consumption rate. ECI may function as a feedback mechanism partially controlling the consumption of leaf material. The relative growth rate is mainly a product of RCR and ECI (Scriber & Slansky, 1981; Ayres & MacLean, 1987b). Development rate, i.e. relative mass gain during the two last instars, is assumed to be affected by relative growth rate, which both may influence length of larval period and pupal mass (see Tammaru, 1998). Development rate, larval period and pupal mass were assumed to be influenced by the lower level traits via relative growth rate. Consequently, number of meals (indicating larval movements) and meal size form the base level, RCR and ECI the second lowest, growth and development rates the next level and length of larval period and pupal mass the highest level in the hierarchical net (Fig. 1).
Our aim was to find path diagrams which did not deviate significantly from the underlying correlation structure. We first included all traits into one large path model, but all possible path diagrams deviated significantly from the observed correlations. Hence, we divided the traits into two temporally and functionally homogeneous groups. The first network included traits measured during late ontogeny: development rate, length of larval period, pupal mass, and to link the two networks together, relative growth rate during the fourth instar. The model was nonsignificant and there was no need to remove any paths, but it is worth mentioning that the path from development rate to larval period was not included since the corresponding phenotypic correlation was zero. The second network included five functionally related traits, i.e. number and size of meals, relative consumption rate, ECI and relative growth rate, all measured during the 24-h bioassay at the beginning of the fourth instar to examine how relative growth rate is determined by the underlying traits. To achieve a nonsignificant model, a weak and nonsignificant path between meal size and ECI was removed.
In both path analyses, exclusion of any additional path would have led to a significantly different model. To compare two nested models, the difference between their goodness-of-fit χ2 values was computed, which is distributed as a χ2 with degrees of freedom equal to the difference in degrees of freedom between the two models (Mitchell, 1993). The residuals of the models were found not to deviate from the normal distribution. Multicollinearity was tested by constructing the path diagrams using multiple regressions (REG procedure and its STB and VIF options). The largest variance inflation factors were less than 3, indicating lack of multicollinearity. Unanalysed causes, i.e. influences on each dependent variable that are unexplained by the model, were calculated as √(1 − R2), where R2 denotes the total explained variance in the variable. Before the path analyses were performed, the effect of sex was removed from each larval trait using an ANOVA (procedure GLM) in which sex was regarded as an independent effect, and the residuals of the models were used in the analyses.
Genetic and residual variation
Pupal and initial larval masses were the only traits to show significant heritabilities, while low heritabilities were found in traits closely related to larval growth, i.e. RCR, ECI, RGR, development rate and length of larval period (Table 2). This pattern was slightly modified for the coefficients of additive genetic variation: number of meals exhibited the highest variation, while meal size, pupal and initial larval masses showed moderate variation, and other traits the lowest. In general, residual variation differed marginally significantly among hierarchical levels (one-way parametric ANOVA after loge-transformation of CVR values, F3,4 = 4.83, P = 0.08, R2 = 0.78) being lowest at the two topmost levels (not transformed values: x¯=11.10±2.71 (SE) and x¯= 10.75±2.34, the top and second highest level, respectively), moderate at the second lowest (x¯=26.46±4.41) and highest at the base level (x¯=46.47±29.96). Of the individual traits, relative growth rate and length of larval period showed the lowest residual variation (Table 2). To ensure that the high levels of variation in number and size of meals were not the results of correlations with initial larval size or the differences in duration of feeding time, the original trait values were corrected with initial size and time in the analysis of covariance. The coefficients of variation were recalculated from the residuals of the models similar to the food utilization indices (Table 1). The recalculated coefficients of variation were of the same magnitude as the original ones (number of meals: CVA = 19.59, CVR = 67.30 and size of meals: CVA = 4.48, CVR = 25.23), although CVA of meal size was slightly reduced.
Table 2. Means, variance components,† heritabilities and their standard errors, and coefficients of variation for food utilization, growth and size measures. The total number of individuals used was 664 from 14 half-sib families.
Phenotypic and genetic covariation
In general, phenotypic and genetic correlations showed convergence, as reflected by a high regression coefficient between them. However, phenotypic correlations were significantly lower than genetic ones, as indicated by a slope (± SE) of a type-II regression significantly higher than one (b = 1.90 ± 0.26, T = 4.78, P < 0.0001, d.f. = 25.4) and intercept significantly different from zero (a = 0.21 ± 0.08, T = 2.63, P < 0.02), and the large absolute difference between the correlations (0.47).
Neither phenotypic nor genetic trade-offs were observed between pupal mass and length of larval period. Rather, high pupal masses were attained after short larval periods (Table 3, Fig. 2A). Rapid development rate showed a strong and significant phenotypic correlation with pupal mass, and rapid relative growth rate had significant phenotypic correlations with both high pupal masses and short larval periods. Fast development rate had a direct effect on pupal mass, as indicated by a strong path among the traits. Instead, the impact of relative growth rate on pupal mass was an indirect one via improved rate of longer term mass gain, i.e. development rate (Fig. 2A). Genetic correlations between rapid development rate and both large pupae and short larval periods were high and significant. There was also a tendency for rapid relative growth rate to be genetically connected with high pupal mass; the correlation was strong but nonsignificant (Table 3).
Table 3. Narrow-sense genetic (above diagonal) and phenotypic correlations (below diagonal) for foraging, growth and size traits. The total number of individuals used was 664 from 14 half-sib families. Genetic correlations larger than 1 are reported as 1.
Size differences between genotypes (progeny means of 14 fathers) increased from the beginning to the end of the experiment, as suggested by the correlation between initial larval mass and absolute larval mass gain during the experiment (r = 0.60, P = 0.02, n = 14). Also, initially large larvae converted leaf biomass to larval biomass more efficiently than initially small larvae, as shown by the significant phenotypic and genetic correlations between initial size and ECI (Table 3).
Phenotypic correlations and path analysis revealed a link from high relative consumption rate and efficient ECI to rapid relative growth rate (Table 3, Fig. 2B). Genetic correlations between the variables showed similar trends. Large meals showed strong and significant phenotypic correlations with fast relative consumption and growth rates. Furthermore, decreased larval movement, as indicated by the low number of meals, correlated at the phenotypic level with large meals and rapid relative growth rate, and weakly, but significantly, with rapid relative consumption rate (Table 3). Path analysis showed that limited dispersion of feeding bouts directly increased meal sizes leading to high relative consumption and growth rates (Fig. 2B). The corresponding genetic relations among number and size of meals, RCR, and RGR were strong and marginally significant, the only exception being the weak correlation between number of meals and RGR (Table 3). The relationship between number of meals and ECI in the path analysis was illustrated as a correlation because of uncertainty about causality. Interestingly, rapid relative consumption rate was genetically linked with a high ECI, while at the phenotypic level the relationship was the reverse, due to the negative environmental correlation (re = −0.22, P < 0.001, d.f. = 1).
Genetic and residual variation
The two measures of E. autumnata size, initial larval and pupal mass, exhibited significant heritabilities (Table 2); on the other hand, none of the traits closely associated with larval growth displayed significant heritabilities. Heritabilities were further partitioned into coefficients of additive genetic, CVA, and residual variation, CVR, which both revealed low levels of variation in relative growth rate and high levels in size as well as in the two lowest level traits in the trait hierarchy, number and size of meals. Three nonexclusive explanations may account for the results. First, natural selection tends to erode additive genetic variation in traits closely associated with fitness, a conclusion derived from Fisher's fundamental theorem of natural selection (Fisher, 1930). Secondly, variation may occur in traits at the top of a hierarchical network, since they are subject to variation of the underlying traits (Price & Schluter, 1991; Houle, 1992). Third, fitness-related traits may be canalized, i.e. buffered against genetic and environmental perturbations, by developmental processes which reduce phenotypic variation (Waddington, 1942).
High pupal mass and rapid relative growth rate are the two best depictors of fitness in our data set (Haukioja, 1981; Haukioja & Neuvonen, 1985; Ayres & MacLean, 1987b; Haukioja et al., 1988; Tammaru et al.,1996a,b), and the results showed that growth rate was genetically less variable than size. Populations of a species with cyclical outbreaks in population density are unlikely to be at an evolutionary equilibrium, and selection gradients of the traits may vary. Nevertheless, selection for rapid growth seems to be considerably stabile in time and space. Although speculative, it is possible that this constant selection has eroded genetic variation in relative growth rate (RGR), as well as in consumption rate (RCR), and efficiency of conversion of leaf mass to larval biomass (ECI), in accordance with Fisher's (1930) statement. As expected (Scriber & Slansky, 1981; Ayres & MacLean, 1987b; Slansky, 1993), in our data rapid relative growth rate was achieved by fast consumption rate and efficient conversion of leaf mass to larval biomass, and the genetic correlations between these traits were positive (Table 3, Fig. 2B), indicating no constraints on correlated evolutionary change. However, the lack of genetic variation in food utilization and growth constrains microevolution for more rapid mass gain. The coefficients of additive genetic variation were among the highest in number and size of meals, which are several causal steps down from relative growth rate (Figs 1 and 2B), and their fitness consequences are probably environment-dependent. Thus, the impact of selection on erosion of genetic variation may be weak. Assumed among-trait differences in the strength of selection are, however, indicative.
Price and Schluter: hierarchical network of traits
The second explanation for the observed distribution of additive genetic variation among traits results from their hierarchical structure. Traits at the upper levels of the causal network are affected by lower level traits, which themselves are heritable. Accordingly, levels of additive genetic variation in life-history traits can be high, regardless of whether they are strongly selected for or not (Price & Schluter, 1991; Houle, 1992). This trend is partly evident in the coefficients of additive genetic variation found in our data. Relative growth rate displayed the lowest CVA, while the upper level traits, development rate, length of larval period and pupal mass showed higher levels of CVA. Development rate and larval period, both of which consist of two episodes of growth and moulting in our experiment, as well as larval size, which is affected by many underlying traits, may have exhibited high genetic variation owing to their high position in the hierarchical network.
Waddington: canalization of traits
Developmental processes may reduce phenotypic variation of traits (Waddington, 1942; Atchley, 1984; Scharloo, 1991; Stearns & Kawecki, 1994; Stearns et al., 1995) and hence canalization may prevent accumulation of environmental and genetic variation in trait networks. In addition, Price & Schluter's (1991) idea can be reversed: low level traits, out of the direct scrutiny of selection, may vary as long as their combined effects lead to a fit value in an upper level trait, and thus to a fit phenotype (e.g. Bader & Hall, 1960 ; Wagner & Misof, 1993 ; and references therein). Interestingly, the coefficients of additive genetic and residual variation were highest at the base level traits, number and size of meals, and moderate in efficiency of conversion of ingested food and relative consumption rate. Instead, relative growth rate, the product of ECI and RCR, and length of larval period showed lowest levels of additive genetic and residual variation, which indicates that these traits may be buffered against variation more efficiently than the other traits. This would be possible by modifying the lower level traits in order to achieve a more optimal value in the upper level trait.
Our data revealed a mechanism that reduced residual variation in relative growth rate. Larvae of E. autumnata counterbalanced the reduced efficiency to physiologically utilize ingested food by increasing consumption rate, as indicated by the negative environmental correlation between ECI and RCR. This compensatory consumption is a common phenomenon among insect herbivores (reviewed in Slansky, 1993). The growth of E. autumnata is, however, weakly canalized against extensive differences in quality and quantity of food, as shown by the experimental manipulations of diets by Neuvonen & Haukioja (1985) and Tammaru (1998). Compensatory mechanisms may occur at the genotypic level too (e.g. Atchley, 1984), but a more detailed genetic analysis would be needed to separate the impact of genetic canalization and natural selection on genetic variation (Stearns & Kawecki, 1994).
Phenotypic and genetic covariation
Neither phenotypic nor genetic trade-offs were found between development time and pupal mass (Table 3, Fig. 2A). Instead, large size was attained by a short larval period, and size differences among E. autumnata genotypes increased from the beginning to the end of the experiment. This might be an ecological necessity: owing to the rapid seasonal deterioration of leaf quality, a prolonged development time would prevent high mass gain. Correlation between short development time and large final size presumably resulted from both traits being phenotypically and genetically determined by larval mass gain. The high genetic correlations among these traits further suggest that the traits seem to be part of the same developmental process, and expressions of the same genes (or linkage disequilibrium). Furthermore, large body size is converted into high expected fecundity (Haukioja & Neuvonen, 1985; Tammaru et al.,1996a,b), and thus resources allocated to rapid growth and development are innately converted to larger size and higher fecundity.
Life history theory relies on an assumption that there are trade-offs which constrain possible life-history combinations (Roff, 1992; Stearns, 1992). However, it is not unexpected that large size is attained after short development time, since these traits can be constrained by a trade-off with other traits not studied here, for instance starvation resistance or survival (Scriber & Slansky, 1981). This has been demonstrated also by mathematical models (Charlesworth, 1990). Moreover, if even small among-genotype differences in resource acquisition exist, then genotypes acquiring more resources than others may be able to become large quickly in a long run, and no genetic trade-off is observed (van Noordwijk & de Jong, 1986; Houle, 1991). In fact, a minority of relationships among life-history traits show trade-offs (reviewed in Roff, 1996; Klingenberg & Spence, 1997).
Both phenotypic and genetic correlations as well as path analysis revealed that E. autumnata caterpillars with restricted movements fed large meals and had fast relative consumption rates, leading to fast growth rates (Table 3, Fig. 2B). Moreover, rapid growth resulted in rapid development rate and large size, reflecting benefits of limited larval movement. Indeed, the evolution of insect foraging should be considered in terms of the degree to which reproduction, both quality and quantity of offspring, and survival are maximized. The net nutrient gain, for example, is favoured only indirectly by selection favouring individuals which both survive and reproduce successfully (e.g. Lemon, 1991). Indeed, observations of caterpillar foraging have often shown that they do not maximize nutrient gain. They may move about considerably within the plant foliage and spend time resting or hiding (Leonard, 1970; Heinrich, 1979; Schultz,1982, 1983; Mauricio & Bowers, 1990; DeLoach, 1994), probably to avoid natural enemies (Heinrich, 1979; Heinrich & Collins, 1983) or plant allelochemicals (Edwards & Wratten, 1983), to gain physiological advantages (Casey, 1976), or to explore for high-quality leaves (Schultz,1982, 1983). Movements within and between leaves in autumnal moth larvae appear to be constrained by costs, but Schultz's (1982,1983) proposal that a qualitatively heterogeneous canopy causes insects to disperse leaf damage may well be true in the E. autumnata– birch system. Although E. autumnata larvae are usually found on branch tips (Kaitaniemi et al., 1997) and only minor movements seem to occur, they nevertheless favour small, young leaves within shoots (Niemelä & Haukioja, 1982), and these leaves differ in quality from the remaining leaves (Suomela et al., 1995).
We propose that the correlation structure of flush feeding autumnal moth caterpillars may be advantageous in a time-limited environment favouring rapid growth. Comparative data from other insect species feeding on mountain birch during early as well as mid and late summer, when leaf quality is more stable (Haukioja et al., 1978; Ayres & MacLean, 1987b; Hanhimäki et al., 1995; Nurmi et al., 1996), would provide insight for this hypothesis. In fact, natural selection may favour correlation structures as such (Bradshaw, 1986). For instance, populations which have evolved under different external environments have been shown to differ in their genetic correlation structures (Berven, 1987; Dingle et al., 1988).
Taken together, to elucidate patterns of (co)variation in a geometrid moth population, we used knowledge about the hierarchical relationships among larval traits. Our data showed reduced phenotypic and genetic variation in a strongly selected character like growth compared to high genetic variation in pupal mass, a closely fitness-related trait situated at the top of a trait network. We further showed a mechanism of how relative growth rate is canalized against environmental and random variation by intercorrelated responses in the underlying traits. We therefore stress that accumulation of environmental variation in trait networks can be retarded by developmental processes.
Virpi Lummaa, Toomas Tammaru, Kai Ruohomäki, Tero Klemola, Ulrika Candolin, Vesa Ruusila and two anonymous reviewers gave valuable comments on earlier versions of this paper. Kirsi Loponen, Elina Salmela and Lasse Iso-Iivari helped in the laboratory and field. The staff at the Kevo Subarctic Research Institute supported us in many ways. The Center for Biotechnology in Turku provided an image-analyser for leaf measurements. Ellen Valle kindly helped with translation. The study was financed by the Academy of Finland. We are grateful to all of them.