Article

# Comparison of Multitrait and Single-Trait Multiple Parity Evaluations by Monte Carlo Simulation

Authors:
To read the full-text of this research, you can request a copy directly from the author.

## Abstract

Three methods of analysis: 1) multiparity analysis with the first three parities analyzed as correlated traits, 2) multiparity analysis with all lactations analyzed as a single trait, and 3) first parity only analysis were compared on simulated data. A selection index was used to compute multiparity evaluations from the separate parity evaluations of the multitrait analyses. Twenty simulated populations, each of 8,500 cows, were generated by an algorithm that approximated the multitrait model. Populations were simulated with both random and yield-based culling of cows after first and second parity. Populations simulated with yield-based culling were analyzed both with the first parity records of all cows included and with an arbitrary one-third of first parity records deleted. First parity records of all cows were included in the analyses of the randomly culled populations. Accuracy of evaluation, estimated by correlations between true effects and evaluations and prediction error variances, was highest by the multitrait analysis and lowest by the first parity only analysis. Evaluations obtained by the two multilactation methods were nearly identical with random culling. Regression of effect on evaluation was close to unity for the multitrait evaluation; was only .94 and .90 for the all lactation single trait evaluation with random and yield culling, respectively; and was .80 for the index of sire effects on the first parity only analysis. Single-trait multilactation method may be preferred, as it is nearly as accurate as the multitrait method and easier computationally.

## No full-text available

... The R 2 of the simulated values was more than five-fold for the ML estimates, as compared to the LS estimates, but both were < 0.1. For an estimate to be unbiased, the regression of the true value on the estimate should be unity [21]. Both regressions of the simulated values on the estimates were < 0.5, but the regression on the ML estimate was nearly double the regression of the LS estimate. ...
Article
Full-text available
Abstract Estimates of quantitative trait loci (QTL) effects derived from complete genome scans are biased, if no assumptions are made about the distribution of QTL effects. Bias should be reduced if estimates are derived by maximum likelihood, with the QTL effects sampled from a known distribution. The parameters of the distributions of QTL effects for nine economic traits in dairy cattle were estimated from a daughter design analysis of the Israeli Holstein population including 490 marker-by-sire contrasts. A separate gamma distribution was derived for each trait. Estimates for both the α and β parameters and their SE decreased as a function of heritability. The maximum likelihood estimates derived for the individual QTL effects using the gamma distributions for each trait were regressed relative to the least squares estimates, but the regression factor decreased as a function of the least squares estimate. On simulated data, the mean of least squares estimates for effects with nominal 1% significance was more than twice the simulated values, while the mean of the maximum likelihood estimates was slightly lower than the mean of the simulated values. The coefficient of determination for the maximum likelihood estimates was five-fold the corresponding value for the least squares estimates.
... These values suggest that the RP model assumption, that each parity is genetically the same trait, is not correct. This is consistent with the literature which generally reports the three parities as separate traits (Weller, 1986;Schaeffer et al., 2000;Powell and Norman, 2006). Furthermore, the ANIM model would be more suitable than the SIRE model as the genetic parameter estimates are closer to the correct values and much more precise as shown by the lower standard deviations. ...
Article
Full-text available
This study aimed to identify genetic evaluation models (GEM) to accurately select cattle for milk production when only limited data are available. It is based on a data set from the Pakistani Sahiwal progeny testing programme which includes records from five government herds, each consisting of 100 to 350 animals, with lactation records dating back to 1968. Different types of GEM were compared, namely: (1) multivariate v. repeatability model when using the first three lactations, (2) an animal v. a sire model, (3) different fixed effects models to account for effects such as herd, year and season; and (4) fitting a model with genetic parameters fixed v. estimating the genetic parameters as part of the model fitting process. Two methods were used for the comparison of models. The first method used simulated data based on the Pakistani progeny testing system and compared estimated breeding values with true breeding values. The second method used cross-validation to determine the best model in subsets of actual Australian herd-recorded data. Subsets were chosen to reflect the Pakistani data in terms of herd size and number of herds. Based on the simulation and the cross-validation method, the multivariate animal model using fixed genetic parameters was generally the superior GEM, but problems arise in determining suitable values for fixing the parameters. Using mean square error of prediction, the best fixed effects structure could not be conclusively determined. The simulation method indicated the simplest fixed effects structure to be superior whereas in contrast, the cross-validation method on actual data concluded that the most complex one was the best. In conclusion it is difficult to propose a universally best GEM that can be used in any data set of this size. However, some general recommendations are that it is more appropriate to estimate the genetic parameters when evaluating for selection purposes, the animal model was superior to the sire model and that in the Pakistani situation the repeatability model is more suitable than a multivariate.
... In a simulation study, Israel and Weller (2000) showed that sire misidentification reduces genetic gain by about 4% in a population with only 10% of incorrect sire assignation. Pedigree misidentification has a negative effect on estimation of the breeding value, which is the basis of modern animal genetic selection (Weller, 1986;Banos et al., 2001). Effectively, sire misidentification is known to reduce heritability estimates, while covariance between maternal and direct effects, and the variance of interaction between sire and year, are badly affected (Van Vleck, 1970;Lee and Pollack, 1997;Senneke et al., 2004). ...
Article
Dam-recorded information was checked out in a selection scheme of dairy goats in South-eastern Spain. A total of 388 dam-offspring verifications were molecularly achieved using nine microsatellite markers. Five of these markers (SR-CRSP-1, SR-CRSP-5, SR-CRSP-8, SR-CRSP-9 and INRA011c) were chosen from the literature according to their polymorphic information content (PIC). Three markers (ChirUCO2, ChirUCO4 and ChirUCO5) were cloned in the same nucleus, and the remaining one was the ETH10 bovine microsatellite. The number of alleles observed were 10, 7, 7, 14, 21, 8, 7, 7, and 7 for the markers SR-CRSP-1, SR-CRSP-5, SR-CRSP-8, SR-CRSP-9, INRA011c, ChirUCO2, ChirUCO4, ChirUCO5 and ETH10, respectively. The global exclusion probability accomplished was 0.9991. On the other hand, among the 388 verifications carried out, 71.9% (279) resulted compatible, while 16.2% (63) were clearly incompatible. Most of the incompatibilities (84.1%) were due to less than four markers. These results suggest that a high percentage of dam-registered information in the selection nucleus could be erroneous, due to the archaic system used for animal identification and for data transfer. Thus an automated method of identification and data reporting should be considered to reduce the high level of errors encountered.
... The R 2 of the simulated values was more than five-fold for the ML estimates, as compared to the LS estimates, but both were < 0.1. For an estimate to be unbiased, the regression of the true value on the estimate should be unity [21]. Both regressions of the simulated values on the estimates were < 0.5, but the regression on the ML estimate was nearly double the regression of the LS estimate. ...
Article
Full-text available
Estimates of quantitative trait loci (QTL) effects derived from complete genome scans are biased, if no assumptions are made about the distribution of QTL effects. Bias should be reduced if estimates are derived by maximum likelihood, with the QTL effects sampled from a known distribution. The parameters of the distributions of QTL effects for nine economic traits in dairy cattle were estimated from a daughter design analysis of the Israeli Holstein population including 490 marker-by-sire contrasts. A separate gamma distribution was derived for each trait. Estimates for both the alpha and beta parameters and their SE decreased as a function of heritability. The maximum likelihood estimates derived for the individual QTL effects using the gamma distributions for each trait were regressed relative to the least squares estimates, but the regression factor decreased as a function of the least squares estimate. On simulated data, the mean of least squares estimates for effects with nominal 1% significance was more than twice the simulated values, while the mean of the maximum likelihood estimates was slightly lower than the mean of the simulated values. The coefficient of determination for the maximum likelihood estimates was five-fold the corresponding value for the least squares estimates.
Article
Data obtained from the INTERGIS for Jersey breeders participating in the National Dairy Cattle Performance and Progeny Testing Scheme were analysed. The registered and grade records commenced in test year 1977 through 1992 (incomplete) and consisting of lactations between 240 and 300 days, were considered as normal completed records. Five production traits were evaluated, ie. milk, fat and protein yields and fat and protein percentages for first and second lactation records. Variance components and resulting heritability estimates were obtained by DFREML procedures using 45 240 first and 27 414 second lactation records. The heritability estimates for first lactation milk, fat and protein yields and fat and protein percentages were 0.35, 0.35, 0.34, 0.57 and 0.58 respectively. Corresponding estimates for second lactation traits were 0.29, 0.28, 0.28, 0.53 and 0.56. Repeatability estimates, calculated as interclass correlations when only two records per animal are available, were 0.73, 0.72, 0.74, 0.71 and 0.74, respectively, for the production traits considered. The heritability estimates were in agreement with literature results, indicating that selection for all five traits would be effective. It was suggested that an increase in environmental effects may be partly responsible for lower estimates of heritability of second lactation traits. The results indicate that selection of dairy bulls can be based on first parity records.
Article
The effects of evaluating milk yield in the first three lactations by a single-trait animal model or by a repeatability animal model instead of by the true multitrait animal model were investigated using stochastic simulation. The models were compared both in terms of how accurately they predicted genetic trend when applied to the same dataset and regarding difference in true genetic progress when selecting on breeding values predicted by these models. A breeding structure resembling the Icelandic dairy cattle population was used for the simulation.The single-trait model underpredicted the true genetic trend in second and third lactations quite heavily but only by 2.3% in the first lactation, while the repeatability model overpredicted the true genetic trend in the first three lactations by 9.4% on average. The multitrait animal model, however, estimated genetic trends with a high degree of accuracy.When selection was made on all three lactations with equal economic weights, the multitrait model was only about 2.5% superior to the repeatability model in terms of true genetic trend.The multitrait animal model was applied to estimate genetic trend regarding milk production using real data from the Icelandic dairy cattle population. A total of 61 621 animals were included, of which 38 014 were cows with production records. The genetic trend for milk yield was estimated at 10.4, 10.0 and 9.3 (kg per year) in each of the three first lactations respectively.
Article
Segregating quantitative trait loci can be detected via linkage to genetic markers. By selectively genotyping individuals with extreme phenotypes for the quantitative trait, the power per individual genotyped is increased at the expense of the power per individual phenotyped, but linear-model estimates of the quantitative-locus effect will be biased. The properties of single- and multiple-trait maximum-likelihood estimates of quantitative-loci parameters derived from selectively genotyped samples were investigated using Monte-Carlo simulations of backcross populations. All individuals with trait records were included in the analyses. All quantitative-locus parameters and the residual correlation were unbiasedly estimated by multiple-trait maximum-likelihood methodology. With single-trait maximum-likelihood, unbiased estimates for quantitative-locus effect and location, and the residual variance, were obtained for the trait under selection, but biased estimates were derived for a correlated trait that was analyzed separately. When an effect of the QTL was simulated only on the trait under selection, a “ghost” effect was also found for the correlated trait. Furthermore, if an effect was simulated only for the correlated trait, then the statistical power was less than that obtained with a random sample of equal size. With multiple-trait analyses, the power of quantitative-trait locus detection was always greater with selective genotyping.
Article
The effect of pedigree errors on estimated breeding value and genetic gain for a sex-limited trait with heritability of 0.25 was evaluated. Ten populations of 100,000 milking cows were simulated with correct paternity identification for all animals, and 10 populations were simulated with 10% incorrect paternal identification. The initial populations consisted of 100,000 unrelated individuals, and simulations were continued for 20 yr. The BLUP genetic evaluations were computed every year by an animal model analysis for each complete population. Estimated breeding values for the populations with 10% incorrect paternity were biased, especially in the later generations. Genetic gains were 4.3% higher with correct paternity identification. Reduction of pedigree errors by paternity confirmation of daughters of test sires by DNA microsatellites may result in considerable economic benefits, depending on the cost of testing in each country.
Article
Standard animal model programs can be modified to include the effect of a quantitative gene, even if only a fraction of the population is genotyped. Five methods to estimate the effect of a diallelic quantitative gene affecting a quantitative trait were compared to a standard animal model (model I) on simulated populations, based on mean squared errors and bias. In models II, III, and IV complete linkage between a single genetic marker and the quantitative trait gene was assumed. In models II and III the elements of the incidence matrix for the gene effect were 0 or 1 for genotyped individuals, and the probabilities of the possible candidate gene genotypes for individuals that were not genotyped. In model III segregation analysis was used to compute these probabilities. If only some of the cows were genotyped, the model III estimates were nearly unbiased, while model II underestimated the simulated effects. When only sires were genotyped, model II overestimated the simulated effect. In models V and VI two markers bracketing the quantitative gene with recombination frequencies of 0.1 and 0.2 with the quantitative gene were simulated, and the algorithm of Whittaker et al. (1996) was used to derive estimates of gene effect and location. In model V marker allele effects were included in the animal model analysis. In model VI, the model I genetic evaluations were analyzed. Model V estimates for both effect and location of the quantitative gene were unbiased, while model VI estimates were only 0.25 of the simulated effect.
Article
Full-text available
The first and second lactation records of New York artifically sired Holstein cows were analyzed to determine the effect of culling after the first lactation on sire evaluation based on both first and second lactation records. Results indicated that weighting first and second records according to number of records per cow, repeatability, and heritability evaluated sires almost identically with the method which uses the average of a daughter's first and second records. Even with a pronounced differential culling rate after the first lactation, there was no evidence of a differential bias in valuating sires of different genetic merit based on first and second lactation records.
Article
Full-text available
Variances of errors of prediction for sire evaluations which included only first records and for those with records of all lactations were compared for bulls of Ayrshire, Guernsey, Jersey, and Brown Swiss breeds used by artificial insemination with daughters having Dairy Herd Improvement records processed at the New York Dairy Records Processing Laboratory. The model for best linear unbiased prediction included fixed effects of sire group and herd-year-season of freshening and random effects of sires within group, sire-by-herd interaction (to account for environmental correlation among paternal sisters), cow within sire and herd, and residual. Variances of solutions for group effects were generally small relative to variances of prediction errors for sire effects. Using all lactation records, however, reduced variances of group solutions by 7 to 14% for groups of sires used artificially and by 20 to 24% for groups used in natural service. Use of all lactation records decreased the variance of prediction error of the sire solutions so that 15 daughters per sire with all lactations gave accuracy equivalent to 25 daughters using only first records; use of all lactation records with 25 daughters gave accuracy equivalent to 40 daughters with only first records. Genetic progress per year from selection of bulls to sire daughters would be expected to be 10 to 15% greater with use of all lactation records than with use of only first lactation records. The comparable increase from selection of bulls to sire replacement bulls would be 3 to 10%. These theoretical increases must be weighed against possible biases from use of records other than first lactation.
Article
SUMMARY Analytical methods have been derived for comparing alternative sire evaluation methods. The criteria used for comparison are unbiased- hess and prediction error variance. The conse- quences of ignoring certain fixed effects, of ignoring certain random effects, of computing as though random effects are fixed, of discard- ing part of the data, and of using data resulting from selection are evaluated analytically.
Article
Records from the North Carolina State University dairy where each of 130 fe- males produced second records regardless of first lactation yield provided data to compare sire evaluations from three mixed model procedures: single trait eval- uations of first and second lactations (Model 1); evaluations of both lactations together including a random component for cow effects (Model 2); and a multi- trait procedure where first and second lactation evaluations were calculated simultaneously (Model 3). Relationships among the 45 sires were included in all models. Culling was simulated at intensi- ties of 10, 20, and 30% on deviation of first lactation from population means. Variation in Model 2 evaluations was least affected by increases of culling intensity. Evaluations by second lactation from Model 1 were most affected by culling. The effect of culling on variability of sire evaluations by Model 3 was not large but was dependent on the genetic correlation assumed between first and second lacta- tions. Expected values for correlations between sire evaluations on first and second lactations were derived and tested.
Article
All lactation records are included in sire evaluations for milk components by mixed model methodology by the United States Department of Agriculture. To in- crease accuracy of these evaluations, pro- cedures were investigated for processing information from different parities as correlated traits rather than as a repeated single trait. The current model for mature- equivalent records, which includes fixed group, random sire, fixed herd-year- season, and random cow effects, was modified to include a separate sire effect for each of first three parities. Residual variances for different parities were as- sumed uncorrelated and of equal variance and, therefore, could be factored out of mixed model equations. Cows and herd- year-seasons were absorbed prior to itera- tion. Sire equations then were augmented by addition of direct product of the in- verse of the relationship matrix and the inverse of the matrix of genetic variance among parities. Sire solutions were ob- tained by block iteration, with blocks consisting of group equations and three equations for each sire. Evaluations were computed for 591 Ayrshire sires with 13,551 daughter records for milk and fat yields and 447 Guernsey sires with
Article
A procedure is presented for rapidly computing the diagnonal elements of a large numerator relationship matrix, say $\mathbf{A}$. It also generates a lower triangular matrix, say $\mathbf{L}$, defined such that $\mathbf{LL'}$ = $\mathbf{A}$. Given the diagonal elements of $\mathbf{L}, \mathbf{A}^{-1}$ may be easily computed from a list of sires and dams.
Article
Sire evaluation methods that invoke best linear unbiased prediction have not utilized information available from known relationships among sires. Principles for doing this have been known, but computations were too costly, involving the inverse of a large numerator relationship matrix. In a method for writing this inverse rapidly without computing the relationship matrix, all relationships can be used with little additional labor. Also records of female ancestors can be incorporated easily. Further, the number of equations for groups to account for genetic trend and for differences among subpopulations is reduced.
Article
Problems associated with selected rec- ords in methods of sire evaluation have been examined, and multiplicative factors for adjustment have been developed to incorporate later lactation records in sire evaluation procedures where a first record is missing. All Holstein records in the Dairy Herd Improvement Association files at Beltsville, Maryland, with freshen- ing dates between 1960 and 1973 were used to quantify the amount of selection bias in later records that do not have first-lactation information available. Dif- ferences for lactation groupings show the amount of bias from selection in these records. The model included an overall mean, effects due to the herd-year-season of freshening, and opportunity groups that were formed depending on the num- ber of later lactations for individual cows. These muttiplicative factors can be used to adjust approximately 47% of cows in the files that do not have first-lactation information. These records can be incor- porated, if needed to increase accuracy or increase comparisons, in the Best Linear Unbiased Prediction method or the Modi- fied Contemporary Comparison method of sire evaluation.
Article
Heritabilities for milk and fat produc- tion of Holstein cows were highest at .35 and .33 for first lactation and decreased to .21 and .20 for fifth lactation. Genetic correlations between consecutive lacta- tions were above .9, except for .75 between first and second, and decreased as lactations were farther apart. Relation- ships between production by cows and their daughters tended to be higher when both animals were in the same herd than when they were in different herds. The cow lactation most valuable in predicting a daughter lactation was generally the one with the same number; this finding suggests that production in different lactations is controlled by some different genes. The third lactation of a cow was most highly related to lifetime pro- duction of the daughter. First lactation was not as useful in estimating genetic merit as is accepted generally. The results provide little support for weighting lactations differentially for sequence.
Article
Accuracy of sire evaluations can be increased by a mixed model procedure that analyzes information from the first several parities as information from correlated traits because genetic cor- relations between yields from different parities are less than one. To increase usefulness of sire evaluations on individual parities, an economic model was developed to obtain factors for weighting informa- tion from the first three parities in a sire selection index. Factors in the model were genetic correlations between yields for different parities, conception rate, probabilities of cow survival, mature- equivalent yield factors, probability of female calves, age at first calving, calving interval, minimum attractive rate of return, planning horizon, milk price, costs variable with milk yield, and sire evalua- tions on the first three parities. Genetic correlations, survival rates, minimum attractive rate of return, and planning horizon were varied. Changes of minimum attractive rate of return and planning horizon had greater effect on net present value of a sire's semen than on relative weights of the three parities in the economic index. A rate of return of .1 and a planning horizon of 10 yr resulted in weights of .38 for first-, .21 for second-, and .41 for third-lactation evaluations.
Article
Mixed linear models are assumed in most animal breeding applications. Convenient methods for computing BLUE of the estimable linear functions of the fixed elements of the model and for computing best linear unbiased predictions of the random elements of the model have been available. Most data available to animal breeders, however, do not meet the usual requirements of random sampling, the problem being that the data arise either from selection experiments or from breeders' herds which are undergoing selection. Consequently, the usual methods are likely to yield biased estimates and predictions. Methods for dealing with such data are presented in this paper.
Article
Performance in first lactation has been the standard of evaluation for most genetic studies with dairy cattle. First records are available sooner on more cows and are less susceptible to error from selection, injury, previous days dry, and mastitis than are later records. However, first records have been considerably less than perfect in predicting traits of lifetime performance, which should be the primary selection objective in dairy cattle. Later records provide additional information for more accurate sire evaluations and should be a better index of health and resistance to mastitis than first lactations. The economic importance of later records relative to first is that there are more and actual yields are higher. Herds with more older cows can reduce commitment of resources to heifer rearing, and heifers can be selected more intensely prior to first calving than herds with more cows of younger ages. Some sire evaluation systems ignore later records because of computational expense, potential selection bias, and difficulty of age adjustment. Later records may contain useful information relative to lifetime profitability of sire progeny groups. Development of proper methodology for utilizing information available in later records appears fruitful for research.
Article
Genetic evaluations were computed for milk and fat for 3,181 Alpine, 1,039 LaMancha, 4,455 Nubian, 1,449 Saanen, and 1,546 Toggenburg bucks. These evaluations were based on 58,562 lactations of 43,913 does that kidded from 1976 through 1982. Best linear unbiased predictions were computed with relationships among multiherd bucks and information from all lactations included. An interaction between herd and sire was included in the model. Evaluations were computed across breed, which allowed does of different breeds to be herdmates. Bucks were grouped by breed and herd usage (single herd versus multiherd). Correlations between evaluations computed with and without relationships were only .84 to .88, which indicates that relationships had a significant effect. Evaluations of 2,491 bucks with Repeatability 15% or more were released to the industry.
Article
The data consisted of the first two records for milk and fat yield of 677,800 daughters of 200 widely used Holstein sires. From these data, 10, 20, and 30% of second records were eliminated for least yield of milk in first lactation. Best Linear Unbiased Prediction evaluations of sires were obtained separately for both records and seconds only for culled and unculled groups and for all first records. Evaluations from second records were affected by culling with standard deviations for milk evaluations declining 58 kg as elimination of second records increased from 0 to 30%. Correlations between first and second milk evaluations declined from .84 to .70 as culling increased. Evaluations by both records showed little effect of culling with standard deviations declining only slightly with culling, and correlations among the evaluations close to unity. Adjustment of second evaluations for selection appeared to remove much effect of selection. Best Linear Unbiased Prediction evaluations and those based on the same daughters from Modified Contemporary Comparison procedures showed similar effects of culling and adjustment for culling and ranked bulls nearly identically.