• We are sorry, but NCBI web applications do not support your browser and may not function properly. More information
Logo of geneticsGeneticsCurrent IssueInformation for AuthorsEditorial BoardSubscribeSubmit a Manuscript
Genetics. Apr 2007; 175(4): 1895–1910.
PMCID: PMC1855112

Prediction of Breeding Values and Selection Responses With Genetic Heterogeneity of Environmental Variance

Abstract

There is empirical evidence that genotypes differ not only in mean, but also in environmental variance of the traits they affect. Genetic heterogeneity of environmental variance may indicate genetic differences in environmental sensitivity. The aim of this study was to develop a general framework for prediction of breeding values and selection responses in mean and environmental variance with genetic heterogeneity of environmental variance. Both means and environmental variances were treated as heritable traits. Breeding values and selection responses were predicted with little bias using linear, quadratic, and cubic regression on individual phenotype or using linear regression on the mean and within-family variance of a group of relatives. A measure of heritability was proposed for environmental variance to standardize results in the literature and to facilitate comparisons to “conventional” traits. Genetic heterogeneity of environmental variance can be considered as a trait with a low heritability. Although a large amount of information is necessary to accurately estimate breeding values for environmental variance, response in environmental variance can be substantial, even with mass selection. The methods developed allow use of the well-known selection index framework to evaluate breeding strategies and effects of natural selection that simultaneously change the mean and the variance.

THE standard genetic model in quantitative genetics is that phenotype P is the sum of genotype G and environment E: equation M1 (Falconer and Mackay 1996). The phenotypic variance can be written as equation M2, assuming no covariance between G and E. This model allows for genetic differences in mean (G), with a genetic variance equation M3. For different genotypes, environmental variances (equation M4) are assumed to be constant. On the basis of analysis of field data and laboratory (selection) experiments, there is, however, some empirical evidence that genotypes differ in equation M5.

Several studies have been carried out to quantify genetic differences in environmental variance in field data. SanCristobal-Gaudy et al. (2001), Sorensen and Waagepetersen (2003), and Ros et al. (2004) explicitly modeled genetic differences in environmental variance and found substantial genetic variance in environmental variance for litter size in sheep, litter size in pigs, and body weight in snails, respectively. Van Vleck (1968) and Clay et al. (1979), in analysis of milk yield in dairy cattle, and Rowe et al. (2006), in analysis of body weight in broiler chickens, found large differences between sires in phenotypic variance within progeny groups. In these studies, it was not possible to distinguish whether these differences were due to heterogeneity of environmental variance, genetic variance, or both.

Several selection experiments have been carried out to investigate whether phenotypic variance can be changed by selection. Phenotypic variance changed in some selection experiments with Drosophila melanogaster and Tribolium castaneum (Rendel et al. 1966; Kaufman et al. 1977; Cardin and Minvielle 1986), while it did not in an experiment with mice (Falconer and Robertson 1956). In these experiments, it was not always clear whether the response in variance was due to a change in environmental variance, genetic variance, or both.

Mackay and Lyman (2005) derived 300 isofemale lines of Drosophila and computed the coefficient of variation (CV) for environmental variance within each homozygous line, effectively a clone, and within crosses of each line with another inbred line. They found highly significant genetic variance in the CV and in environmental variance between lines. Homozygotes had higher environmental variance, in agreement with findings of Robertson and Reeve (1952). This study is probably the cleanest known example of showing genetic variance in environmental variance because the design allowed repetition of genotypes.

In livestock and plant breeding, uniformity of end product is an important topic. In meat type animals, for instance, uniformity has economic benefits because excessive variability in carcass weight or conformation is penalized by slaughterhouses. Hohenboken (1985) reviewed the potential of mating systems (crossing, inbreeding) and breeding schemes to change variability. To evaluate breeding strategies, methods to predict responses to selection for uniformity are necessary. SanCristobal-Gaudy et al. (1998) derived prediction equations and SanCristobal-Gaudy et al. (2001) evaluated different selection indices using Monte Carlo simulation when the aim was to select for an optimum phenotype and thereby decrease the variance around the optimum (canalizing selection). Sorensen and Waagepetersen (2003) evaluated response to selection using an index including the mean and the variance of multiple records of an individual and Ros et al. (2004) discussed the use of a restricted index aiming at decreasing the environmental variance, while maintaining the mean. Hill and Zhang (2004) derived simple equations to predict response to directional mass selection with genetic heterogeneity of environmental variance. In general, these prediction equations can be used only in special cases. A general framework to predict responses in mean and variance is lacking.

The objective of the present study was to develop a general framework for prediction of breeding values and responses to selection with genetic heterogeneity of environmental variance. Responses to selection were predicted for different forms of selection based on a single phenotype, as well as selection on a mean or variance of a group of relatives. Furthermore, a measure of heritability of environmental variance was developed, enabling a direct comparison between selection to change the environmental variance of a trait and the well established framework of selection to change its mean.

DERIVATION AND EVALUATION OF EXPRESSIONS

In this section, the model incorporating genetic heterogeneity of environmental variance is defined and the framework for prediction is explained. Prediction of breeding values and selection responses based on a single phenotype and a group of relatives are then considered, in each case using Monte Carlo simulation to investigate the relationships between true breeding values and phenotypic information. Using these observations, multiple-regression equations are derived for one generation of selection and their goodness of fit is evaluated by simulation. Finally, a measure of heritability for environmental variance is proposed.

Genetic model and framework for prediction

The classical model, in the absence of dominance and epistasis, equation M6 (Falconer and Mackay 1996), is extended to include an additive genetic effect for the environmental variance,

equation M7
(1)

(Hill and Zhang 2004), where μ and equation M8 are, respectively, the mean trait value and the mean environmental variance of the population, equation M9 and equation M10 are, respectively, the breeding value for the mean and environmental variance, and χ is a standard normal deviate for the environmental effect. It is assumed that equation M11 and equation M12 are the sum of the effects at an infinite number of loci each with small additive effects and follow a multivariate normal distribution equation M13, where A is the additive genetic relationship matrix,

equation M14

equation M15 is the additive genetic variance in equation M16, equation M17 is the additive genetic variance in equation M18, equation M19, and equation M20 is the additive genetic correlation between equation M21 and equation M22. The term χ is normally distributed equation M23 and is scaled by equation M24 to obtain the environmental effect. The notation is listed in Table 1.

TABLE 1
Notation used

The genetic model in Equation 1 does not allow for random environmental effects on the magnitude of the environmental variance, because without repeated measurements on each individual these cannot be separated from the usual random environmental effects. With repeated measurements on each individual, these environmental effects on environmental variance become equivalent to permanent environmental effects (e.g., SanCristobal-Gaudy et al. 1998; Sorensen and Waagepetersen 2003).

To predict the breeding values equation M25 and equation M26 and selection responses equation M27 and equation M28, multiple regression was used. Selection index theory is essentially an application of multiple regression (Hazel 1943). Multiple regression gives the best linear prediction (BLP), which is equal to best linear unbiased prediction (BLUP) when fixed effects are known without error (Henderson 1984). When variables are multivariate normally distributed, regressions are linear and homoscedastic (Lynch and Walsh 1998). Although the distribution of P deviates slightly from normality with genetic heterogeneity of environmental variance (equation M29), P, equation M30, and equation M31 follow an approximately multivariate normal distribution for values of equation M32 observed in the literature (e.g., SanCristobal-Gaudy et al. 2001; Sorensen and Waagepetersen 2003; Ros et al. 2004; Rowe et al. 2006), justifying the use of multiple regression.

Multiple regression with selection on a single phenotype

Monte Carlo simulation:

Monte Carlo simulation was used to investigate the relationships between equation M33 and equation M34 with P, with the objective to decide which order of fit would be required for accurate prediction and then to evaluate the fit of predictions on the basis of multiple regressions. Fifty replicates with one phenotypic observation on each of 500,000 unrelated animals in each replicate were generated according to the genetic model in Equation 1, assuming equation M35. The breeding values equation M36 and equation M37 and the environmental effect χ were randomly drawn from equation M38 and scaled by their corresponding standard deviations. When the genetic correlation between equation M39 and equation M40 was nonzero, equation M41 was sampled given the expected value based on equation M42 with variance equation M43. Expected breeding values were calculated as the mean equation M44 and equation M45 within successive intervals of 0.01 units of P (equation M46) and averaged over replicates. Expected selection responses to directional mass selection were calculated as the mean equation M47 and equation M48 of all selected animals having equation M49 and averaged over replicates, where x is the truncation point. The selected proportion was assumed to be the same in both sexes.

Breeding value estimation:

Figure 1, A and B, presents, respectively, the expectation of equation M50 given P and the expectation of equation M51 given P2 when equation M52, obtained by simulation. These show that the relationship between equation M53 and P is almost linear and that the relationship between equation M54 and P2 is also almost linear (quadratic in P). Therefore, equation M55 roughly predicts equation M56 and regression on equation M57 roughly predicts equation M58. As a consequence of genetic heterogeneity of environmental variance, the distribution of P is slightly leptokurtic and is slightly skewed when equation M59. By fitting curves to the simulation results, it was found that regression on equation M60 explained most of the residual nonlinearity and skewness when equation M61. Moments of P of higher order did not improve the fit and were therefore not considered in the rest of this study.

Figure 1.
Expected equation M369 (A) and equation M370 (B) as a function of P and P2, respectively (equation M371; equation M372; equation M373; equation M374; equation M375).

On the basis of this curve fitting, breeding values were predicted using multiple regression on the first through third order of P,

equation M62
(2)

where

equation M63
equation M64
equation M65

and

equation M66

Elements of P and G were derived using the higher-order moments of the normal distribution (e.g., Stuart and Ord 1994) and standard variance–covariance rules (see appendix a). Elements in these matrices were verified with Monte Carlo simulation.

Evaluation of predictions:

Predictions of Equation 2 were close to the expectations obtained from Monte Carlo simulation when equation M67, as could be expected from the almost linear relationships shown in Figure 1, A and B (equation M68). For equation M69, Figure 2A shows that equation M70 is approximately linear in P, with a slope close to equation M71, for P within two standard deviations (SD) of its mean, but becomes curvilinear for extreme P. The predicted equation M72 using the full model with multiple regressions on P, P2, and P3 fitted well to the expectations obtained from Monte Carlo simulation (equation M73) and, in contrast to multiple regressions on only P and P2, also explained the nonlinearity in the extremes.

Figure 2.
Expected (MC) and predicted breeding values equation M376 (A) and equation M377 (B) based on a single phenotype as a function of phenotype using the full model with multiple regression on P, P2, and P3 (MR3) or the reduced model with multiple regression on P and P2 (MR2) (equation M378; equation M379; ...

Figure 2B shows that the relationship between equation M74 and P is highly curvilinear for equation M75, with higher equation M76 for more extreme P. As for equation M77, the predicted equation M78 using the full model with multiple regressions on P, P2, and P3 fitted well to the expectation from Monte Carlo simulation (equation M79). The use of only P and P2 was adequate only within 2 SD of the mean, but was biased for extreme P.

Response to mass selection:

Response to selection (equation M80) is predicted as equation M81, where b is the regression coefficient of the breeding value on the selection criterion, and S is the selection differential in units of the selection criterion (e.g., phenotype) (Falconer and Mackay 1996). With homogenous environmental variance and directional mass selection, equation M82 and equation M83, where i is the selection intensity, and equation M84 is the breeders' equation (Falconer and Mackay 1996; Lynch and Walsh 1998). With genetic heterogeneity of environmental variance, directional mass selection leads to responses in mean and variance (Hill and Zhang 2004). To predict this, Equation 2 can be rewritten as equation M85, giving

equation M86
(3)

where equation M87, equation M88, and equation M89 are the respective means for the selected animals.

Directional selection:

With directional selection by truncation, equation M90, where equation M91 for normally distributed observations, z is the height of the standardized normal at the truncation point x, and p is the selected proportion (Falconer and Mackay 1996; Lynch and Walsh 1998). equation M92 and equation M93 were calculated by integration assuming that P is normally distributed, which is approximately the case for observed values of equation M94 in the literature:

equation M95
(4a)
equation M96
(4b)

The predicted response to directional mass selection is thus

equation M97
(5)

The term equation M98 is similar to the term equation M99 derived by Hill and Zhang (2004), who calculated the probability of selection by using a Taylor series approximation, where the factor equation M100 appears here in the B-matrix, assuming that equation M101 and equation M102 is small. Equation 5 can be rewritten using the regression coefficients in B and ignoring the terms involving P3, which were not included by Hill and Zhang (2004),

equation M103
(6)
equation M104
(7)

where equation M105. When equation M106 and equation M107 are close to zero, Equations 6 and 7 approach equation M108 and equation M109, which are Equations 10 and 11 of Hill and Zhang (2004). Differences arise when equation M110 is substantially different from zero, as the covariance between P and P2 was ignored by Hill and Zhang.

Stabilizing and disruptive selection:

Stabilizing and disruptive selection can be considered as selecting the animals with low or high P2, respectively (Falconer and Mackay 1996). Assuming that selection is by truncation, selection differentials equation M111 for P2 can be obtained from Equation 4a and selection differentials for P and P3 are zero for these types of selection when P is normally distributed. With stabilizing selection by truncation, animals only in the middle of the distribution are selected, giving a selection differential

equation M112
(8a)

where equation M113 and equation M114 are, respectively, the selection intensity and truncation point corresponding to equation M115, the proportion of animals culled on one side of the distribution. With disruptive selection by truncation, the extreme animals in both tails of the distribution are selected, giving a selection differential

equation M116
(8b)

where equation M117, the proportion of animals selected on one side of the distribution. The standardized selection differentials equation M118 for directional, stabilizing, and disruptive selection are in Table 2.

TABLE 2
Standardized selection differentials of equation M433 for directional, stabilizing, and disruptive selection by truncation on a normal distribution corrected for the expectation of P2 (= 1) for different selected proportions (p)

Evaluation of predictions with directional mass selection:

The predictions of Equation 5 [multiple regressions (MR) 3] and Equations 6 and 7 (MR2) and those of the Hill–Zhang (HZ) model (Hill and Zhang 2004) are compared in Table 3 with the observed responses obtained from Monte Carlo (MC) simulation, for different values of equation M119 and selected proportions. Multiple regressions on P, P2, and P3 (MR3) predicted the responses well, with prediction errors <5% when the selected proportion was at least 5% (prediction error relative to equation M120 with equation M121, equation M122, equation M123). Prediction errors were on average smaller for MR3 than for MR2, although occasionally larger. They were also on average slightly smaller for MR2 than for the Hill–Zhang model, especially with equation M124, because in the latter the covariance between P and P2 was not accounted for in calculation of the regression coefficients. Prediction errors using MR were mainly due to poor prediction of selection differentials because of deviations from normality and thus increased with decreasing selected proportion as the tails of the distribution were most affected (results not shown). When equation M125 increased to 0.10 or 0.15, which reflects the upper range of estimates in the literature (e.g., SanCristobal-Gaudy et al. 2001; Ros et al. 2004), prediction errors increased up to 10–20% (prediction error relative to equation M126 with equation M127, equation M128, equation M129), especially with selected proportion of 1% (results not shown). Increasing equation M130 increases deviations from normality in P, but it seems that the multiple-regression framework is robust against these relatively small deviations from normality, except when the selected proportion is very small. It can be concluded that MR3 is the preferred method for predicting responses in equation M131 and equation M132 with directional mass selection, having prediction errors <5% when at least 5% are selected.

TABLE 3
Response to directional mass selection in AmAm) and AvAv) for different values rA and selected proportions comparing predictions [as prediction errors (predicted minus observed)] with observed responses obtained from Monte Carlo ...

Multiple regression with selection based on a group of relatives

Monte Carlo simulation:

In animal breeding sires are often selected on performance of their half-sib progeny. Monte Carlo simulation was used to investigate relationships between the equation M133 and equation M134 of the sires and statistics on phenotypes of their progeny and then to evaluate the fit of predictions on the basis of multiple regressions. Fifty replicates were generated of 500,000 unrelated sires each with 10 or 100 half-sib progeny or of 50,000 unrelated sires each with 1000 or 10,000 half-sib progeny. Data were simulated according to the genetic model in Equation 1. The breeding values of sires (equation M135 and equation M136) and unrelated random-mated dams (equation M137 and equation M138) were randomly sampled with variance equation M139 or equation M140, respectively. For each progeny, the Mendelian sampling terms equation M141 and equation M142 were randomly sampled with variance equation M143 and equation M144, respectively, to give breeding values for each progeny (equation M145 and equation M146):

equation M147

When the genetic correlation between equation M148 and equation M149 was nonzero, breeding values equation M150 and Mendelian sampling terms equation M151 were sampled given their expected value based on equation M152 or equation M153 and with variance equation M154 or equation M155, respectively. For each progeny, the environmental effect χ was randomly sampled and scaled with its standard deviation.

Expected breeding values of sires were calculated as the mean equation M156 and equation M157 within successive intervals of 0.01 SD units of progeny mean equation M158 or log-transformed within-family variance [equation M159] and averaged over replicates. Expected genetic selection differentials of directional selection on equation M160 were calculated as the mean equation M161 and equation M162 of all selected sires with equation M163 and averaged over replicates.

Breeding value estimation:

When there is an observation on only a single phenotype, there is no independent information available on the mean and variance of the genotype, although P and P2 provide point estimates. When phenotypes of a group of relatives each having the same relationship to an individual are available (e.g., progeny), statistics such as equation M164, equation M165, and equation M166 can be used to predict its equation M167 and equation M168. Here equation M169 is the mean phenotype and equation M170 is the mean P2 of the relatives, equation M171 is the within-family variance, and n is the number of relatives within the group. With large n, equation M172 becomes the main predictor of equation M173, but otherwise equation M174 contains additional information because animals with a high equation M175 have a higher probability of having a very high or low equation M176. This is similar to directional mass selection and the term equation M177 therefore plays an equivalent role to P2. Although the Monte Carlo simulation was based on sires with half-sib progeny, the prediction of breeding values generalizes to any group of relatives with the same relationship. The multiple-regression equation can be represented as

equation M178
(9)

where equation M179 is the additive genetic relationship between relatives within the family,

equation M180
equation M181

and equation M182 is the relationship of relatives to individual j. Elements in the P- and G-matrices, derived in appendix a, were verified with Monte Carlo simulation for the case of sires with half-sib progeny.

Log-transformation of var W:

In multiple regression linearity is assumed and is typically satisfied if the explanatory variables are normally distributed. Because variances of normally distributed variates are equation M183- distributed, the distribution of equation M184 is not normal. As the number of relatives increases, equation M185 approaches a normal distribution, but the convergence is slow (Stuart and Ord 1994). The relationship between equation M186 and equation M187 is therefore nonlinear if there are a finite number of relatives (see Figure 3B), and also the sampling variance of equation M188 increases with its mean. A logarithmic transformation of equation M189 seems a logical choice to reduce both the nonnormality of equation M190 and the positive relationship between the mean and its sampling variance. When using equation M191 instead of equation M192, the elements in the P- and G-matrices involving equation M193 were transformed using a first-order Taylor series approximation (Lynch and Walsh 1998). The matrices equation M194 and equation M195 involving equation M196 were calculated, respectively, as equation M197 and equation M198, where equation M199 with equation M200 based on a first-order Taylor series approximation. The quantity equation M201 was calculated using a second-order Taylor series approximation (Lynch and Walsh 1998): equation M202, replacing equation M203 in Equation 9.

Figure 3.
Expected (MC) and predicted equation M383 as a function of the standardized mean phenotype of half-sib progeny (A) and expected and predicted equation M384 as a function of equation M385 (B) using either equation M386 (MR linear) or equation M387 (MR log) in multiple regression (equation M388; equation M389; equation M390; equation M391; equation M392; number of half-sib progeny/sire ...

Evaluation of predictions:

Predictions of Equation 9, which holds for any group of relatives with the same relationship, were evaluated for the case of sires with half-sib progeny. The expected and predicted values of equation M204 are shown as a function of the standardized equation M205 of 100 half-sib progeny in Figure 3A for equation M206. equation M207 is linear in standardized equation M208 with a slope of equation M209. The predicted equation M210 fitted well the expectation from Monte Carlo simulation (equation M211) and the predictions using equation M212 or equation M213 did not differ if equation M214.

Figure 3B shows that the relationship between equation M215 and equation M216 is curvilinear for the case of 100 half-sib progeny when equation M217. The predicted equation M218 using untransformed equation M219 (MR linear) overestimated equation M220 for extreme values of equation M221, whereas predicted equation M222 using log-transformed equation M223 (MR log) was curvilinear in equation M224 and fitted well the expectation from Monte Carlo simulation (equation M225).

As the bias in equation M226 was largest with extreme values of equation M227, the bias in equation M228 at 2 SD from the mean equation M229 predicted by multiple regression using either equation M230 or equation M231 was computed for different values of equation M232 and numbers of progeny per sire (Table 4). Multiple regression on equation M233 was seen to overestimate, but on equation M234 to underestimate, equation M235. The bias using equation M236 was negligible when the number of progeny was 100, but increased when the number of progeny was small (10) or large (10,000). The bias with equation M237 was negligible with 10,000 progeny, as could be expected from the slow convergence of a equation M238-distribution to a normal distribution, and was small with a few progeny. Note that a higher degree of symmetry between the expected equation M239 at −2 and 2 SD of the mean equation M240 corresponded with a smaller bias in equation M241 with equation M242. The value of log-transformation of equation M243 thus depends on the number of progeny per sire.

TABLE 4
The expectation of Av and bias in Âv at equation M441 using either var W or ln(var W) in multiple regression for different values of equation M442 and numbers of half-sib progeny per sire

Response to directional selection on family mean:

With the common procedure in livestock breeding to directionally select animals by truncation on the mean (equation M244) of relatives, e.g., progeny, information on the within-family variance (equation M245) is ignored. As for mass selection, if there is genetic heterogeneity of environmental variance, animals with a higher equation M246 would have a higher probability of selection when the selected proportion is <50%, diminishing as the number of relatives increases. If equation M247 and equation M248 are uncorrelated, the response in equation M249 is proportional to the selection differential equation M250,

equation M251
(10)

where

equation M252
(11)

There is therefore no selection pressure on equation M253 with an infinite number of relatives (Equation 11), as suggested by Hill and Zhang (2004) using a different argument.

Response in equation M254 and equation M255 with selection on equation M256 can be generalized as

equation M257
(12)

To compute (12), equation M258 was not included in the information vector x because selection is solely on equation M259. For infinitely many relatives, Equation 12 can be rewritten as equation M260, which is the corresponding standard breeders' equation, and equation M261, showing that response in equation M262 then becomes solely a correlated response to selection on equation M263.

Evaluation of predictions:

Table 5 shows predicted responses (Equation 12) in equation M264 and equation M265 when selecting on the mean of half-sib progeny (equation M266) in comparison to observed responses from Monte Carlo simulation. In general, these agreed well, although prediction errors were slightly higher when equation M267, especially with high selection intensity. As expected, the response in equation M268 increased with more progeny (higher accuracy of selection) and lower selected proportion (higher selection intensity). The response in equation M269 was small, becoming negligible with 100 progeny/sire when equation M270, but was, however, substantial when equation M271, basically as a correlated response to selection on the mean. Response in equation M272 increased nonlinearly with increasing selection intensity, similar to directional mass selection, but to a lesser extent. In conclusion, responses in equation M273 and equation M274 to selection on equation M275 can be predicted accurately using multiple regression.

TABLE 5
Response to selection on mean of half-sib progeny (equation M454) in AmAm) and AvAv) for different values of rA, different numbers of progeny/sire, and different selected proportions comparing predictions (MR) [as prediction errors (predicted ...

Defining a measure of heritability for environmental variance at the phenotypic level

Heritability (equation M276) is a central parameter in quantitative genetics (Falconer and Mackay 1996; Lynch and Walsh 1998). For standardization of results of analysis of genetic heterogeneity of environmental heterogeneity in field data and for making comparisons to “conventional” traits easier, it would be helpful to define a measure of heritability (equation M277) for environmental variance at the phenotypic level. Heritability equals the regression coefficient of the breeding value A on the phenotype P. Here we propose a definition of equation M278, which equals the genetic variance in environmental variance as a proportion of the variance of P2. This definition is equal to the regression of equation M279 on P2, where equation M280 and equation M281, and equation M282 is therefore

equation M283
(13)

Alternatively, equation M284 could be defined at the level of environmental variance, which equals one in Equation 1. On the basis of single phenotypic records, the environmental variance of a genotype is, however, not estimable. The measure of heritability in Equation 13 is directly related to single squared phenotypic records and as such is the natural analogy of the classical heritability of the mean (equation M285), which can be used in prediction of response to mass selection when equation M286.

Under the assumption of equation M287 and making use of Equation 13, Equation 9 can be greatly simplified when selecting on information of a group of relatives, for example, half-sib progeny. When equation M288, equation M289 reduces to equation M290 because equation M291, so the multiple regression for equation M292 can be simplified by regressing solely on equation M293. The accuracy of equation M294 can then be derived as

equation M295
(14)

where equation M296 and equation M297 are columns of B and G corresponding to equation M298. The resulting expression is exactly the same as that for accuracy of equation M299 (Cameron 1997), except that h2 is replaced by equation M300. To investigate the effect of assuming equation M301, the accuracy of equation M302 predicted with Equation 14 was compared to that predicted using Equation 9 and Monte Carlo simulation when equation M303 (Table 6). In general, accuracies were slightly underestimated by Equation 14, increasingly so with greater equation M304 (equation M305), whereas the ones of Equation 9 were close to those from simulation. It seems that equation M306 can be used as a first approximation in standard prediction equations when equation M307, but predictions should be interpreted with caution.

TABLE 6
Realized (MC) and predicted accuracy of Âv for different numbers of half-sib progeny per sire and equation M462 using either the exact prediction (MR exact) or the approximate prediction (MR approx)

EXAMPLES OF CHANGING ENVIRONMENTAL VARIANCE BY SELECTION

In the previous section the focus was mainly on evaluating the goodness of fit of multiple-regression predictions with Monte Carlo simulation, but the results also show the effects of selection on environmental variance. For example, the response in equation M308 with directional mass selection increased nonlinearly with increasing selection intensity (Table 3), and environmental variance increased unless equation M309. The response in equation M310 was, however, negligible with directional selection on a half-sib progeny mean when equation M311 (Table 5), but was substantial when equation M312, due to a correlated response.

We now use the formulas (Equations 2, 3, and 9) to assess the effects of selection strategies aimed at changing the environmental variance, taking values of equation M313 between 0.01 and 0.10. These correspond to a low equation M314 but are large relative to equation M315, indicating a genetic coefficient of variation between 14 and 45%, higher than that for standard quantitative traits (Houle 1992). As expected, the accuracy of equation M316 increased with equation M317 (Table 7). The accuracy was low when using information only on own phenotype or a small number of progeny, but increased with number of relatives, especially with half-sib progeny, and when equation M318. With 1000 half-sib progeny, the accuracy was >0.90, unless equation M319.

TABLE 7
Predicted accuracy of Âv based on a single phenotype or different numbers of full-sibs or half-sib progeny for different values of equation M475 and rA

Table 8 shows the predicted response in equation M320 for directional, stabilizing, and disruptive selection based on phenotype (Equation 3, selection differentials from Table 2, neglecting the terms involving P3) and for directional downward selection on equation M321 based on 100 half-sib progeny assuming equation M322 [calculated as equation M323, where equation M324). Predictions were close to observed responses in Monte Carlo simulation. For all selection strategies, responses in equation M325 increased with equation M326 due to a higher accuracy and a higher genetic variance in itself.

TABLE 8
Predicted response in AvAv) for directional (up/down), stabilizing, and disruptive selection based on phenotype and downward directional selection on equation M484 based on 100 half-sib progeny for different values of equation M485 and selected proportions

With directional and disruptive selection on phenotype, the predicted response in equation M327 was positive and environmental variance increased substantially and nonlinearly with selection intensity. Disruptive selection gave a slightly larger response because the selection intensity in each tail of the distribution of P was higher. With stabilizing selection on phenotype, the response in equation M328 was negative but small, even when the selection was intense because selection differentials remain small and were nearly constant (Table 2). With directional downward selection on equation M329 based on 100 half-sib progeny, response in equation M330 was negative and environmental variance decreased substantially, which in an agricultural context would imply increased uniformity of end product. Responses increased linearly with selection intensity and became large, especially with equation M331. When the best 5% of the sires are selected on equation M332 and dams are selected at random with equation M333, the environmental variance would be 0.554 in the next generation, which is only 79.1% of that in the current generation! In conclusion, a large number of progeny is necessary to predict equation M334 with high accuracy, but responses in equation M335 can be large relative to the environmental variance in the current generation.

DISCUSSION

A multiple-regression framework has been developed to predict breeding values and selection responses in mean and variance for mass selection and selection between families in the presence of genetic heterogeneity of environmental variance. The model of Hill and Zhang (2004) has been refined for directional mass selection and extended to stabilizing and disruptive selection based on phenotype and to between-family selection. The phenotypic variance increases nonlinearly with selection intensity under directional mass selection when equation M336. It increases even more with disruptive selection, but decreases only slightly with stabilizing selection, which is in agreement with results of Gavrilets and Hastings (1994) and Wagner et al. (1997). With selection on family mean, phenotypic variance is expected to change little unless equation M337, but with selection on within-family variance, response in phenotypic variance may be large providing equation M338, even though a large number of relatives is necessary to estimate equation M339 accurately.

Methodology:

Comparison of genetic models:

Different genetic models to account for genetic heterogeneity of environmental variance appear in the literature, basically either additive effects both at the level of the mean and at the level of the environmental variance (Hill and Zhang 2004; Zhang and Hill 2005; this study) or additive effects on the mean and an exponential model for the environmental variance (SanCristobal-Gaudy et al. 1998, 2001; Sorensen and Waagepetersen 2003; Ros et al. 2004). In the exponential model

equation M340
(15)

where equation M341 is the environmental variance when equation M342, and equation M343 is the individual's breeding value for environmental variance in the exponential model, such that environmental variances are multiplicative on the observed scale and additive on a log-scale. (Note that equation M344 in the notation of SanCristobal-Gaudy et al. 1998.) The distribution of true variances (not variance estimates) is unknown in practice and cannot help in guiding whether the additive model or the exponential model better reflects the real world. Clearly, each model has specific (dis)advantages. The exponential model has tractable properties so it is easier to use in data analysis; for example, the environmental variance can never become negative, whereas in the additive model the term equation M345 is defined only when equation M346. The additive model, however, fits nicely in quantitative genetic theory, leading to better properties for deterministic predictions of selection response. A disadvantage of the exponential model is that the average environmental variance in the population is equation M347, so there is no full separation of the mean environmental variance and the genetic variance in environmental variance, and equation M348 has to be known to interpret equation M349.

Fortunately, the models are sufficiently similar that their genetic parameters can be interconverted. The breeding values for environmental variance can be converted by equating the expectations of the second central moments of the environmental effects (see appendix b):

equation M350
(16)

[a first-order Taylor series approximation of Equation 16 is equation M351, illustrating the factor equation M352 between breeding values and the correction for the difference between equation M353 and equation M354]. The genetic variances can be converted by equating the fourth central moments of the environmental effects:

equation M355
(17)

(a first-order Taylor series approximation of Equation 17 is equation M356, showing a factor equation M357 between genetic variances and a correction for the difference between equation M358 and equation M359). Thus results obtained using the exponential model in data analysis could be converted using Equations 16 and 17 to the additive model and the deterministic equations derived in this study used to predict selection responses. Gavrilets and Hastings (1994) and Wagner et al. (1997) adopted slightly different multiplicative models to deal with genetic heterogeneity of environmental variance, but these are in essence very similar to the exponential model.

Multiple-regression framework:

In this study, a multiple-regression framework was used to predict breeding values and selection responses. Prediction equations were derived for incorporating phenotypic information of only one type, individual or family statistics, but the method can easily be extended to situations where phenotypic information is available from different kinds of relatives. For the common situation of optimal weighting of own performance and family information, most of the necessary elements in the prediction equation either have been derived here or can be derived straightforwardly using the same methods. Furthermore, the regression structure enables prediction of responses in mean and variance with different selection strategies using the classical selection index theory and extension to give optimal changes in mean and variance via a selection index (Hazel 1943). The framework presented can be used only for prediction of selection response after one generation of selection; due to buildup of gametic phase disequilibrium, genetic variance would decrease with directional selection, lowering selection responses (Bulmer 1971). Furthermore, gametic phase disequilibrium induces an unfavorable covariance between the additive genetic effects for mean and environmental variance, counteracting desired changes in mean and variance (Hill and Zhang 2004). Inclusion of the Bulmer effect was beyond the scope of this article, but could be implemented (Hill and Zhang 2004, 2005).

In the multiple-regression framework fixed effects are assumed to be known without error, but in practice they are estimated from the data, thereby reducing accuracy and selection response. Therefore, results in this study should be interpreted using an effective number of observations, which is lower than the actual number of observations. To predict breeding values in the presence of fixed effects on mean (e.g., herd effect) and variance (e.g., heterogeneity of variance between herds, environments with different stress levels), a model with genetically structured environmental variance can be used (SanCristobal-Gaudy et al. 1998; Sorensen and Waagepetersen 2003). Modeling of environmental heterogeneity of variance (e.g., between herds) has been reviewed by Foulley and Quaas (1995) and Hill (2004). To predict selection responses with genetic heterogeneity of environmental variance in environments differing in mean environmental variance (e.g., herds, stress levels), the present framework can be used by adjusting equation M360.

A disadvantage of the multiple-regression framework is that it relies on the assumption that the explanatory variables (x) are linearly related to the dependent variables (y), which is ensured when x and y are bivariate normally distributed. As a consequence, results may not be robust against deviations from normality, particularly when higher-order terms such as P3 are included in predictions. The multiple-regression framework was, however, robust against small deviations from normality induced by genetic heterogeneity of environmental variance. SanCristobal-Gaudy et al. (1998) and Sorensen and Waagepetersen (2003) also assumed multivariate normality in predictions of selection responses, but their approaches were more flexible in allowing for other distributions. Their approaches were not very different from those in this study, but some expressions were much more complex due to the use of the exponential model.

Multiple-regression methods would be useful to study the evolution of phenotypic variance in natural populations, to further develop analyses of Zhang and Hill (2005). Selection for reduced environmental variance could result in environmental canalization, a phenomenon of long-standing interest (e.g., Waddington 1942, 1960; reviews in Scharloo 1991, Debat and David 2001, and Flatt 2005). On the basis of our predictions, stabilizing selection cannot cause environmental canalization of traits within a few generations, but may do so eventually. With long-term canalizing selection, the question arises whether the limit of environmental variance is zero, whereas most quantitative traits in nature under stabilizing selection still exhibit environmental variance (e.g., Wagner et al. 1997). We know little about how levels of environmental variation are determined and maintained in nature in the face of stabilizing selection. Different mechanisms have been proposed (e.g., Wagner et al. 1997), such as introducing a cost for homogeneity or canalization (Zhang and Hill 2005). To further investigate long-term effects of natural selection on environmental variance, the current framework can be extended to include the effects of gametic phase disequilibrium, inbreeding, and mutation, analogous to effects of selection on the mean of traits.

Evidence for genetic heterogeneity of environmental variance:

Although the tools for evaluating breeding strategies to change the mean and the size of environmental variance are now available, the whole exercise would just be a theoretical game if equation M361. As reviewed in the Introduction, there is empirical evidence that genotypes differ in environmental variance, but it is not abundant. To compare results of different studies analyzing field data we use Equation 17, because some are based on the exponential genetic model (SanCristobal-Gaudy et al. 1998, 2001; Sorensen and Waagepetersen 2003; Ros et al. 2004) and some on the additive genetic model (Rowe et al. 2006). The measure of heritability (equation M362) developed in this study (Equation 13) and the genetic coefficient of variation for environmental variance equation M363, denoted “evolvability” (Houle 1992), are used to compare results from different studies (Table 9). Heritabilities of environmental variance were low, in the range 0.02–0.05 as used in this study, and GCVE's were large, in the range of 0.30–0.58 (excluding 0). Note that equation M364 is close to equation M365. The low values of equation M366 show that a large amount of information is necessary to estimate equation M367 accurately, but the high values of equation M368 show that there is substantial opportunity for genetic change.

TABLE 9
Comparison of literature estimates of genetic variance in environmental variance

Other evidence of the existence of genetic heterogeneity of environmental variance can come from selection experiments. With genetic heterogeneity of environmental variance, environmental variance would decrease with stabilizing selection and increase with disruptive selection. In most studies (e.g., Rendel et al. 1966; Cardin and Minvielle 1986), however, only changes in phenotypic variance are reported, which are not separated into changes in genetic and environmental variance. Interpretation of any selection experiment is complicated by possible changes in genetic variance due to gene frequency change, which cannot be predicted from simple base population parameters. Under infinitesimal model assumptions, effects of gene frequency change can be ignored and those due to gametic phase disequilibrium can be predicted. Due to negative gametic phase disequilibrium, genetic variance is expected to decrease with stabilizing selection and increase with disruptive selection (Bulmer 1971). In agreement with this expectation, Kaufman et al. (1977) found substantial decreases in genetic and environmental variance with stabilizing selection in T. castaneum and Scharloo et al. (1972) observed substantial increases in genetic and environmental variance with disruptive selection in D. melanogaster, indicating a substantial genetic variance in environmental variance. Sorensen and Hill (1983), however, found large increases only in genetic variance with disruptive selection in D. melanogaster.

Even though some studies analyzing field data or selection experiments show existence of genetic heterogeneity of environmental variance, it could still be due to statistical artifacts, e.g., due to confounding genetic and environmental effects on variance or violation of the infinitesimal model assumption. If genetic heterogeneity of environmental variance is a truly biological phenomenon, it could be due to scaling, genetic variance in environmental sensitivity, or a combination of both. Traits seem to have a rather constant CV, even when the mean changes dramatically due to selection (Hill and Bünger 2004). A constant CV would require a correlation of unity between mean and standard deviation, which has not been found in analysis of field data (Sorensen and Waagepetersen 2003; Ros et al. 2004; Rowe et al. 2006). Genetic heterogeneity of environmental variance can arise from genetic differences in environmental sensitivity (Falconer and Mackay 1996; Lynch and Walsh 1998). When genotypes perform under variable environmental conditions, which are unknown to the researcher, genetic differences in response to environmental conditions may be observed as genetic heterogeneity of environmental variance.

Genetic heterogeneity of environmental variance is a complicated phenomenon and there is not yet abundant evidence of its existence. The results in this study may help in understanding the consequences of genetic heterogeneity of environmental variance on phenotypes and the methods can help in designing selection experiments and in evaluating breeding strategies or the effects of natural selection that change both the mean and the variance. Genetic heterogeneity of environmental variance may indeed be exploited to breed more “robust” or “stable” genotypes.

Acknowledgments

H.M. thanks Johan van Arendonk, Bart Ducro, and Roel Veerkamp for valuable comments on earlier versions of the manuscript. We thank Ian White for statistical advice. European Animal Disease Genomics Network of Excellence for Animal Health and Food Safety is acknowledged for financial support to a visit of H.M. to Edinburgh and the Biotechnology and Biological Sciences Research Council for research support to W.G.H.

APPENDIX A: DERIVATION OF ELEMENTS IN THE P- AND G-MATRICES

Selection on a single phenotype:

The elements in the P- and G-matrices were derived as follows, using the Roman E to denote expectation and the italic E to denote the environmental deviation equation M500 and noting that equation M501:

equation M502
equation M503

Similarly equation M504, equation M505, and equation M506.

Selection based on a group of relatives:

equation M507
equation M508
equation M509
equation M510
equation M511
equation M512
equation M513
equation M514
equation M515

where k, l, m, and n are different relatives within the family.

equation M516
equation M517
equation M518

and similarly equation M519, equation M520, and equation M521.

APPENDIX B: SIMILARITIES BETWEEN EXPONENTIAL AND ADDITIVE GENETIC MODELS

The breeding values for environmental variance were converted from the exponential model (Equation 15) to the additive model (Equation 1) by equating the second central moments of the environmental effects of both models resulting in Equation 16:

equation M522

The genetic variance in environmental variance was converted from the exponential model to the additive model by equating the fourth central moments of the environmental effects of both models resulting in Equation 17 (e.g., Stuart and Ord 1994):

equation M523

References

  • Bulmer, M. G., 1971. The effect of selection on genetic variability. Am. Nat. 105: 201–211.
  • Cameron, N. D., 1997. Selection Indices and Prediction of Genetic Merit in Animal Breeding. CAB International, Wallingford, UK.
  • Cardin, S., and F. Minvielle, 1986. Selection on phenotypic variation of pupa weight in Tribolium castaneum. Can. J. Genet. Cytol. 28: 856–861.
  • Clay, J. S., W. E. Vinson and J. M. White, 1979. Heterogeneity of daughter variances of sires for milk yield. J. Dairy Sci. 62: 985–989.
  • Debat, V., and P. David, 2001. Mapping phenotypes: canalization, plasticity and developmental stability. Trends Ecol. Evol. 16: 555–561.
  • Falconer, D. S., and T. F. C. Mackay, 1996. Introduction to Quantitative Genetics. Pearson Education Limited, Essex, UK.
  • Falconer, D. S., and A. Robertson, 1956. Selection for environmental variability of body size in mice. Z. Indukt. Abstammungs-Vererbungsl. 87: 385–391. [PubMed]
  • Flatt, T., 2005. The evolutionary genetics of canalization. Q. Rev. Biol. 80: 287–316. [PubMed]
  • Foulley, F. L., and R. L. Quaas, 1995. Heterogeneous variances in Gaussian linear mixed models. Genet. Sel. Evol. 27: 211–228.
  • Gavrilets, S., and A. Hastings, 1994. A quantitative-genetic model for selection on developmental noise. Evolution 48: 1478–1486.
  • Hazel, L. N., 1943. The genetic basis for constructing selection indexes. Genetics 28: 476–490. [PMC free article] [PubMed]
  • Henderson, C. R., 1984. Applications of Linear Models in Animal Breeding. University of Guelph, Guelph, Ontario, Canada.
  • Hill, W. G., 2004. Heterogeneity of genetic and environmental variance of quantitative traits. J. Ind. Soc. Agric. Stat. 57: 49–63.
  • Hill, W. G., and L. Bünger, 2004. Inferences on the genetics of quantitative traits from long-term selection in laboratory and domestic animals. Plant Breed. Rev. 24(2): 169–210.
  • Hill, W. G., and X. S. Zhang, 2004. Effects of phenotypic variability of directional selection arising through genetic differences in residual variability. Genet. Res. 83: 121–132. [PubMed]
  • Hill, W. G., and X. S. Zhang, 2005 Erratum Hill and Zhang (2004). Genet. Res. 86: 160.
  • Hohenboken, W. D., 1985. The manipulation of variation in quantitative traits: a review of possible genetic strategies. J. Anim. Sci. 60: 101–110.
  • Houle, D., 1992. Comparing evolvability and variability of quantitative traits. Genetics 130: 195–204. [PMC free article] [PubMed]
  • Kaufman, P. K., F. D. Enfield and R. E. Comstock, 1977. Stabilizing selection for pupa weight in Tribolium castaneum. Genetics 87: 327–341. [PMC free article] [PubMed]
  • Lynch, M., and B. Walsh, 1998. Genetics and Analysis of Quantitative Traits. Sinauer Associates, Sunderland, MA.
  • Mackay, T. F. C., and R. F. Lyman, 2005. Drosophila bristles and the nature of quantitative genetic variation. Philos. Trans. R. Soc. B. 360: 1513–1527. [PMC free article] [PubMed]
  • Rendel, J. M., B. L. Sheldon and D. E. Finlay, 1966. Selection for canalization of the scute phenotype. II. Am. Nat. 100: 13–31.
  • Robertson, F. W., and E. C. R. Reeve, 1952. Heterozygosity, environmental variation and heterosis. Nature 170: 286. [PubMed]
  • Ros, M., D. Sorensen, R. Waagepetersen, M. Dupont-Nivet, M. SanCristobal et al., 2004. Evidence for genetic control of adult weight plasticity in the snail helix aspersa. Genetics 168: 2089–2097. [PMC free article] [PubMed]
  • Rowe, S. J., I. M. S. White, S. Avendano and W. G. Hill, 2006. Genetic heterogeneity of residual variance in broiler chickens. Genet. Sel. Evol. 38: 617–635. [PMC free article] [PubMed]
  • SanCristobal-Gaudy, M., J. M. Elsen, L. Bodin and C. Chevalet, 1998. Prediction of the response to a selection for canalisation of a continuous trait in animal breeding. Genet. Sel. Evol. 30: 423–451.
  • SanCristobal-Gaudy, M., L. Bodin, J. M. Elsen and C. Chevalet, 2001. Genetic components of litter size variability in sheep. Genet. Sel. Evol. 33: 249–271. [PMC free article] [PubMed]
  • Scharloo, W., 1991. Canalization: genetic and developmental aspects. Annu. Rev. Ecol. Syst. 22: 65–93.
  • Scharloo, W., A. Zweep, K. A. Schuitema and J. G. Wijnstra, 1972. Stabilizing and disruptive selection on a mutant character in Drosophila. IV. Selection on sensitivity to temperature. Genetics 71: 551–566. [PMC free article] [PubMed]
  • Sorensen, D., and R. Waagepetersen, 2003. Normal linear models with genetically structured residual variance heterogeneity: a case study. Genet. Res. 82: 207–222. [PubMed]
  • Sorensen, D. A., and W. G. Hill, 1983. Effects of disruptive selection on genetic variance. Theor. Appl. Genet. 65: 173–180. [PubMed]
  • Stuart, A., and J. K. Ord, 1994. Kendall's Advanced Theory of Statistics: Distribution Theory. Arnold, London.
  • Van Vleck, L. D., 1968. Variation of milk records within paternal-sib groups. J. Dairy Sci. 51: 1465–1470.
  • Waddington, C. H., 1942. Canalization of development and the inheritance of acquired characters. Nature 150: 563–565.
  • Waddington, C. H., 1960. Experiments on canalizing selection. Genet. Res. 1: 140–150.
  • Wagner, G. P., G. Booth and H. Bagheri-Chaichian, 1997. A population genetic theory of canalization. Evolution 51: 329–347.
  • Zhang, X. S., and W. G. Hill, 2005. Evolution of the environmental component of the phenotypic variance: stabilizing selection in changing environments and the cost of homogeneity. Evolution 59: 1237–1244. [PubMed]

Articles from Genetics are provided here courtesy of Genetics Society of America
PubReader format: click here to try

Formats:

Related citations in PubMed

See reviews...See all...

Cited by other articles in PMC

See all...

Links

Recent Activity

Your browsing activity is empty.

Activity recording is turned off.

Turn recording back on

See more...