Search

Estimating heritabilities with large genomic models by established methods such as restricted maximum likelihood (REML) or Bayesian via Gibbs sampling is computationally expensive. Alternatively, heritability can be estimated indirectly by method R and by maximum predictivity, referred to as MaxPred here, at a much lower computing cost. By method R, the heritability used for predictions with whole and partial data is considered the best estimate when the predictions based on partial data are unbiased relative to those with the complete data. By MaxPred, the heritability estimate is the one that maximizes predictivity. This study compared heritability estimation with genomic information using average information REML (AI-REML), method R and MaxPred. A simulated population was generated with ten generations of 5000 animals each and an effective population size of 80. Each animal had one record for a trait with a heritability of 0.3, a phenotypic variance of 10.0 and was genotyped at 50 k SNP. In method R, the heritability estimate is found when the expectation of a regression coefficient is equal to one. The regression is the EBV of selection candidates calculated with the whole dataset regressed on the EBV of candidates calculated from a partial dataset. In this study, we used the GBLUP framework and therefore, GEBV was calculated. The partial dataset was created by removing the last generation of phenotypes. Predictivity was defined as the correlation between the adjusted phenotypes of the selection candidates and their GEBV calculated from the partial data. We estimated the heritability for populations that included between three and 10 generations. In every scenario, predictivity increased as more data was used and was the highest at the simulated heritability. However, the predictivity for all data subsets and all heritabilities compared did not differ more than 0.01, suggesting MaxPred is not the best indication for heritability estimation. For the whole dataset, the heritability was estimated as 0.30 ± 0.01, 0.26 ± 0.01 and 0.30 ± 0.04 for AI-REML without genomics, AI-REML with genomics and method R with genomics, respectively. Heritability estimation with genomics by method R reduced timing by 83%, implying a reduction in computing time from 9.5 to 1.6 h, on average, compared to AI-REML with genomics. Method R has the potential to estimate heritabilities with large genomic information at a low cost when many generations of animals are present; however, the standard error can be high when only a few iterations are used., (© 2024 The Authors. Journal of Animal Breeding and Genetics published by John Wiley & Sons Ltd.)

The objective of this study was to determine whether the linear regression (LR) method could be used to validate genomic threshold models. Statistics for the LR method were computed from estimated breeding values (EBVs) using the whole and truncated data sets with variances from the reference and validation populations. The method was tested using simulated and real chicken data sets. The simulated data set included 10 generations of 4,500 birds each; genotypes were available for the last three generations. Each animal was assigned a continuous trait, which was converted to a binary score assuming an incidence of failure of 7%. The real data set included the survival status of 186,596 broilers (mortality rate equal to 7.2%) and genotypes of 18,047 birds. Both data sets were analysed using best linear unbiased predictor (BLUP) or single-step GBLUP (ssGBLUP). The whole data set included all phenotypes available, whereas in the partial data set, phenotypes of the most recent generation were removed. In the simulated data set, the accuracies based on the LR formulas were 0.45 for BLUP and 0.76 for ssGBLUP, whereas the correlations between true breeding values and EBVs (i.e. true accuracies) were 0.37 and 0.65, respectively. The gain in accuracy by adding genomic information was overestimated by 0.09 when using the LR method compared to the true increase in accuracy. However, when the estimated ratio between the additive variance computed based on pedigree only and on pedigree and genomic information was considered, the difference between true and estimated gain was <0.02. Accuracies of BLUP and ssGBLUP with the real data set were 0.41 and 0.47, respectively. This small improvement in accuracy when using ssGBLUP with the real data set was due to population structure and lower heritability. The LR method is a useful tool for estimating improvements in accuracy of EBVs due to the inclusion of genomic information when traditional validation methods as k-fold validation and predictive ability are not applicable., (© 2020 The Authors. Journal of Animal Breeding and Genetics published by Wiley-VCH GmbH.)

Economically important traits are usually complex traits influenced by genes, environment and genotype-by-environment (G × E) interactions. Ignoring G × E interaction could lead to bias in the estimation of breeding values and selection decisions. A total of 1,778 pigs were genotyped using the PorcineSNP80 BeadChip. The existence of G × E interactions was investigated using a single-step reaction norm model for growth traits of days to 100 kg (AGE) and backfat thickness adjusted to 100 kg (BFT), based on a pedigree-based relationship matrix (A) or a genomic-pedigree joint relationship matrix (H). In the reaction norm model, the herd-year-season effect was measured as the environmental variable (EV). Our results showed no G × E interactions for AGE, but for BFT. For both AGE and BFT, the genomic reaction norm model (H) produced more accurate predictions than the conventional reaction norm model (A). For BFT, the accuracies were greater based on the reaction norm model than those based on the reduced model without exploiting G × E interaction, with EV ranging from 0.5 to 1, and accuracy increasing by 3.9% and 4.6% in the reaction norm model based on A and H matrices, respectively, while reaction norm model yielded approximately 8.4% and 7.9% lower accuracy for EVs ranging from 0 to 0.4, based on A and H matrices, respectively. In addition, for BFT, the highest accuracy was obtained in the BJLM6 farm for realizing directional selection. This study will help to apply G × E interactions to practical genomic selection., (© 2020 Wiley-VCH GmbH.)

We investigated the effects of different strategies for genotyping populations on variance components and heritabilities estimated with an animal model under restricted maximum likelihood (REML), genomic REML (GREML), and single-step GREML (ssGREML). A population with 10 generations was simulated. Animals from the last one, two or three generations were genotyped with 45,116 SNP evenly distributed on 27 chromosomes. Animals to be genotyped were chosen randomly or based on EBV. Each scenario was replicated five times. A single trait was simulated with three heritability levels (low, moderate, high). Phenotypes were simulated for only females to mimic dairy sheep and also for both sexes to mimic meat sheep. Variance component estimates from genomic data and phenotypes for one or two generations were more biased than from three generations. Estimates in the scenario without selection were the most accurate across heritability levels and methods. When selection was present in the simulations, the best option was to use genotypes of randomly selected animals. For selective genotyping, heritabilities from GREML were more biased compared to those estimated by ssGREML, because ssGREML was less affected by selective or limited genotyping., (© 2018 Blackwell Verlag GmbH.)

We investigated the importance of SNP weighting in populations with 2,000 to 25,000 genotyped animals. Populations were simulated with two effective sizes (20 or 100) and three numbers of QTL (10, 50 or 500). Pedigree information was available for six generations; phenotypes were recorded for the four middle generations. Animals from the last three generations were genotyped for 45,000 SNP. Single-step genomic BLUP (ssGBLUP) and weighted ssGBLUP (WssGBLUP) were used to estimate genomic EBV using a genomic relationship matrix (G). The WssGBLUP performed better in small genotyped populations; however, any advantage for WssGBLUP was reduced or eliminated when more animals were genotyped. WssGBLUP had greater resolution for genome-wide association (GWA) as did increasing the number of genotyped animals. For few QTL, accuracy was greater for WssGBLUP than ssGBLUP; however, for many QTL, accuracy was the same for both methods. The largest genotyped set was used to assess the dimensionality of genomic information (number of effective SNP). The number of effective SNP was considerably less in weighted G than in unweighted G. Once the number of independent SNP is well represented in the genotyped population, the impact of SNP weighting becomes less important., (© 2017 Blackwell Verlag GmbH.)

The Algorithm for Proven and Young (APY) enables the implementation of single-step genomic BLUP (ssGBLUP) in large, genotyped populations by separating genotyped animals into core and non-core subsets and creating a computationally efficient inverse for the genomic relationship matrix (G). As APY became the choice for large-scale genomic evaluations in BLUP-based methods, a common question is how to choose the animals in the core subset. We compared several core definitions to answer this question. Simulations comprised a moderately heritable trait for 95,010 animals and 50,000 genotypes for animals across five generations. Genotypes consisted of 25,500 SNP distributed across 15 chromosomes. Genotyping errors and missing pedigree were also mimicked. Core animals were defined based on individual generations, equal representation across generations, and at random. For a sufficiently large core size, core definitions had the same accuracies and biases, even if the core animals had imperfect genotypes. When genotyped animals had unknown parents, accuracy and bias were significantly better (p ≤ .05) for random and across generation core definitions., (© 2017 The Authors. Journal of Animal Breeding and Genetics Published by Blackwell Verlag GmbH.)

The purpose of this study was to examine accuracy of genomic selection via single-step genomic BLUP (ssGBLUP) when the direct inverse of the genomic relationship matrix (G) is replaced by an approximation of G(-1) based on recursions for young genotyped animals conditioned on a subset of proven animals, termed algorithm for proven and young animals (APY). With the efficient implementation, this algorithm has a cubic cost with proven animals and linear with young animals. Ten duplicate data sets mimicking a dairy cattle population were simulated. In a first scenario, genomic information for 20k genotyped bulls, divided in 7k proven and 13k young bulls, was generated for each replicate. In a second scenario, 5k genotyped cows with phenotypes were included in the analysis as young animals. Accuracies (average for the 10 replicates) in regular EBV were 0.72 and 0.34 for proven and young animals, respectively. When genomic information was included, they increased to 0.75 and 0.50. No differences between genomic EBV (GEBV) obtained with the regular G(-1) and the approximated G(-1) via the recursive method were observed. In the second scenario, accuracies in GEBV (0.76, 0.51 and 0.59 for proven bulls, young males and young females, respectively) were also higher than those in EBV (0.72, 0.35 and 0.49). Again, no differences between GEBV with regular G(-1) and with recursions were observed. With the recursive algorithm, the number of iterations to achieve convergence was reduced from 227 to 206 in the first scenario and from 232 to 209 in the second scenario. Cows can be treated as young animals in APY without reducing the accuracy. The proposed algorithm can be implemented to reduce computing costs and to overcome current limitations on the number of genotyped animals in the ssGBLUP method., (© 2015 Blackwell Verlag GmbH.)

This work studied differences between expected (calculated from pedigree) and realized (genomic, from markers) relationships in a real population, the influence of quality control on these differences, and their fit to current theory. Data included 4940 pure line chickens across five generations genotyped for 57,636 SNP. Pedigrees (5762 animals) were available for the five generations, pedigree starting on the first one. Three levels of quality control were used. With no quality control, mean difference between realized and expected relationships for different type of relationships was ≤ 0.04 with standard deviation ≤ 0.10. With strong quality control (call rate ≥ 0.9, parent-progeny conflicts, minor allele frequency and use of only autosomal chromosomes), these numbers reduced to ≤ 0.02 and ≤ 0.04, respectively. While the maximum difference was 1.02 with the complete data, it was only 0.18 with the latest three generations of genotypes (but including all pedigrees). Variation of expected minus realized relationships agreed with theoretical developments and suggests an effective number of loci of 70 for this population. When the pedigree is complete and as deep as the genotypes, the standard deviation of difference between the expected and realized relationships is around 0.04, all categories confounded. Standard deviation of differences larger than 0.10 suggests bad quality control, mistakes in pedigree recording or genotype labelling, or insufficient depth of pedigree., (© 2014 Blackwell Verlag GmbH.)

In single-step genomic evaluation using best linear unbiased prediction (ssGBLUP), genomic predictions are calculated with a relationship matrix that combines pedigree and genomic information. For missing pedigrees, unknown selection processes, or inclusion of several populations, a BLUP model can include unknown-parent groups (UPG) in the animal effect. For ssGBLUP, UPG equations also involve contributions from genomic relationships. When those contributions are ignored, UPG solutions and genetic predictions can be biased. Options to eliminate or reduce such bias are presented. First, mixed model equations can be modified to include contributions to UPG elements from genomic relationships (greater software complexity). Second, UPG can be implemented as separate effects (higher cost of computing and data processing). Third, contributions can be ignored when they are relatively small, but they may be small only after refinements to UPG definitions. Fourth, contributions may approximately cancel out when genomic and pedigree relationships are constructed for compatibility; however, different construction steps are required for unknown parents from the same or different populations. Finally, an additional polygenic effect that also includes UPG can be added to the model., (© 2013 Blackwell Verlag GmbH.)

Effects on prediction of analysing a multi-line chicken population as one line were evaluated. Body weight records were provided by Cobb-Vantress for two lines of broiler chickens. Phenotypic records for 183 695 and 164 149 broilers and genotypic records for 3195 and 3001 broilers were available for each line. Lines were combined to create a multi-line population and analysed using a single-step procedure combining the additive relationship matrix and the genomic relationship matrix (G). G was scaled using allele frequencies from each line, the multi-line population, or 0.5. When allele frequencies were calculated from each line, distributions of diagonal elements were bimodal. When allele frequencies were calculated from the multi-line population, the distribution of diagonal elements had one peak. When allele frequency 0.5 was used, the distribution was bimodal. Genomic estimated breeding values (GEBVs) were predicted using each allele frequency. GEBVs differed with allele frequency but had ≥ 0.99 correlations with GEBVs predicted with correct allele frequencies. Means of each line and differences in mean between the lines differed based on allele frequencies. Assumed allele frequencies have little impact on ranking within line but larger impact on ranking across lines. G may be used to evaluate multiple populations simultaneously but must be adjusted to obtain properly scaled estimates when population structure is unknown., (© 2011 Blackwell Verlag GmbH.)

Genomic evaluations can be calculated using a unified procedure that combines phenotypic, pedigree and genomic information. Implementation of such a procedure requires the inverse of the relationship matrix based on pedigree and genomic relationships. The objective of this study was to investigate efficient computing options to create relationship matrices based on genomic markers and pedigree information as well as their inverses. SNP maker information was simulated for a panel of 40 K SNPs, with the number of genotyped animals up to 30 000. Matrix multiplication in the computation of the genomic relationship was by a simple 'do' loop, by two optimized versions of the loop, and by a specific matrix multiplication subroutine. Inversion was by a generalized inverse algorithm and by a LAPACK subroutine. With the most efficient choices and parallel processing, creation of matrices for 30 000 animals would take a few hours. Matrices required to implement a unified approach can be computed efficiently. Optimizations can be either by modifications of existing code or by the use of efficient automatic optimizations provided by open source or third-party libraries., (© 2011 Blackwell Verlag GmbH.)

This study explored distributions of diagonal elements of genomic relationship matrix (G), evaluated the utility of G as a diagnostic tool to detect mislabelled animals in a genomic dataset and evaluated the effect of mislabelled animals on the accuracy of genomic evaluation. Populations of 10 000 animals were simulated with 60 000 SNP varying in allele frequency at each locus between 0.02 and 0.98. Diagonal elements of G were distributed with a single peak (mean = 1.00 ± 0.03) and ranged from 0.84 through 1.36. Mixed populations were also simulated: 7 000 animals with frequencies of second alleles ranging from 0.02 through 0.98 were combined with 1750 or 7000 animals with frequencies of second alleles ranging from 0.0 through 1.0. The resulting distributions of diagonal elements of G were bimodal. Body weight at 6 weeks was provided by Cobb-Vantress for broiler chickens, of which 3285 were genotyped for 57 636 SNP. Analysis used a combined genomic and pedigree relationship matrix; G was scaled using current allele frequencies. The distribution of diagonal elements was multimodal and ranged from 0.54 to 3.23. Animals with diagonal elements >1.5 were identified as coming from another chicken line or as having low call rates. Removal of mislabelled animals increased accuracy by 0.01. For the studied type of population, diagonal elements of G may be a useful tool to help identify mislabelled animals or secondary populations., (© 2011 Blackwell Verlag GmbH.)

Data included 90,242,799 test day records from first, second and third parities of 5,402,484 Holstein cows and 9,326,754 animals in the pedigree. Additionally, daily temperature humidity indexes (THI) from 202 weather stations were available. The fixed effects included herd test day, age at calving, milking frequency and days in milk classes (DIM). Random effects were additive genetic, permanent environment and herd-year and were fit as random regressions. Covariates included linear splines with four knots at 5, 50, 200 and 305 DIM and a function of THI. Mixed model equations were solved using an iteration on data program with a preconditioned conjugate gradient algorithm. Preconditioners used were diagonal (D), block diagonal due to traits (BT) and block diagonal due to traits and correlated effects (BTCORR). One run included BT with a 'diagonalized' model in which the random effects were reparameterized for diagonal (co)variance matrices among traits (BTDIAG). Memory requirements were 8.7 Gb for D, 10.4 Gb for BT and BTDIAG, and 24.3 Gb for BTCORR. Computing times (rounds) were 14 days (952) for D, 10.7 days (706) for BT, 7.7 days (494) for BTDIAG and 4.6 days (289) for BTCORR. The convergence pattern was strongly influenced by the choice of fixed effects. When sufficient memory is available, the option BTCORR is the fastest and simplest to implement; the next efficient method, BTDIAG, requires additional steps for diagonalization and back-diagonalization.

Genetic parameters for daily feed intake (DFI, g/day) and daily gain (DG, g/day) were estimated using records of 1916 Duroc boars from electronic feeder stations. Management was limited and resulted in varied ranges of age and weight on test. Boars were housed in 102 pens, each equipped with one feeder, and allowed ad libitum feeding. Weekly averages of DFI and DG were used due to large variation in daily records. Six traits were defined as DFI and DG during 85-106 (period 1), 107-128 (period 2) and 129-150 days of age (period 3). A six-trait model included age as a linear and a quadratic covariate for DFI and a linear covariate for DG with a fixed effect of year-week-pen and random effects of litter, additive genetic animal and permanent environmental animal. Variance components were estimated by a Bayesian approach using Gibbs sampling algorithm. Estimates of heritability for respective periods were 18%, 12% and 10% for DFI and 21%, 11% and 10% for DG. Genetic correlations between DFI and DG in the same period were 0.70, 0.73 and 0.32 for the respective periods. DFI and DG obtained from automatic feeders can be analysed to reveal variation across testing periods by using weekly averages when many monthly averages are incomplete.

This study examined the utility of serial weights from FIRE (Feed Intake Recording Equipment, Osborne Industries, Inc., Osborne, KS, USA) stations for an analysis of daily gain. Data included 884 132 body weight records from 3888 purebred Duroc pigs. Pigs entered the feeder station at age 77-149 days and left at age 95-184 days. A substantial number of records were abnormal, showing body weight close to 0 or up to twice the average weight. Plots of body weights for some animals indicated two parallel growth curves. Initial editing used a robust regression, which was a two-step procedure. In the first step, a quadratic growth curve was estimated assuming small or 0 weights for points far away from the curve; the process is iterative. In the second step, weights more than 1.5 SD from the estimated growth curve were treated as outliers. The retained body weight records (607,597) were averaged to create average daily weight (170,443) and then used to calculate daily gains (152,636). Additional editing steps included retaining only animals with >or=50 body weight records and SD of the daily gain

Social genetic relationships among average daily gain (ADG, g) and feeding pattern as daily feed intake (DFI, g), daily feeder occupation time (DOT, min), and daily feeding rate (DFR, g/min) were examined using records of 547 Duroc boars. Single-trait animal models were fitted differently for traits, including or excluding social genetic effects, random or fixed pen effects, with covariates of pen sizes and initial age or weight. Genetic parameters for feeding pattern were estimated by restricted maximum likelihood. Six sets of parameters for ADG based on literature estimates were used due to difficulty in untangling confounded effects. Positive and negative signs of direct-social genetic covariances were interpreted as heritable cooperation and competition, respectively. Dominant and subordinate pigs were classified as pigs with higher direct and social genetic values, respectively. Correlations of estimated breeding values between ADG and DFI, DOT, and DFR were 0.46, 0.04 and 0.29 for dominant pigs. Given heritable cooperation, subordinate pigs tended to increase feed intake (r = 0.36) and eating rate (r = 0.25). Given heritable competition, subordinate pigs fail to compensate for the competition with decreased feed intake (r = -0.53). The slow eating rate (r = -0.31) was considered as a consequence of eating during less busy hour of feeding.

The purpose of this study is to present guidelines in selection of statistical and computing algorithms for variance components estimation when computing involves software packages. For this purpose two major methods are to be considered: residual maximal likelihood (REML) and Bayesian via Gibbs sampling. Expectation-Maximization (EM) REML is regarded as a very stable algorithm that is able to converge when covariance matrices are close to singular, however it is slow. However, convergence problems can occur with random regression models, especially if the starting values are much lower than those at convergence. Average Information (AI) REML is much faster for common problems but it relies on heuristics for convergence, and it may be very slow or even diverge for complex models. REML algorithms for general models become unstable with larger number of traits. REML by canonical transformation is stable in such cases but can support only a limited class of models. In general, REML algorithms are difficult to program. Bayesian methods via Gibbs sampling are much easier to program than REML, especially for complex models, and they can support much larger datasets; however, the termination criterion can be hard to determine, and the quality of estimates depends on a number of details. Computing speed varies with computing optimizations, with which some large data sets and complex models can be supported in a reasonable time; however, optimizations increase complexity of programming and restrict the types of models applicable. Several examples from past research are discussed to illustrate the fact that different problems required different methods.

Breeding value prediction for dairy goats in Germany is still based on herd mate comparison within breeding society. The objective of this study was to estimate genetic parameters for milk yield based on a test day model. For the analysis 35,308, 30,551 and 23,640 test day records from lactations 1, 2 and 3 from 5079, 4118 and 3132 animals, respectively, were used. The data between 1987 and 2003 were obtained from six German breeding societies. The multiple trait (lactations 1, 2 and 3) repeatability model (RPT) included the fixed effects of breeding society-breed-herd-year, litter size, lambing season, and days in milk of third-order Legendre polynomials nested within herd-year, and the random effects of animal additive and permanent environment. The three-trait random regression model (RR) also included the random regressions based on second-order Legendre polynomials for animal additive and permanent environmental effects. Heritability estimates in RPT were 0.27 +/- 0.02, 0.20 +/- 0.02 and 0.37 +/- 0.02 for the first, second and third lactation, respectively. The genetic correlation between the first and second lactation was 0.69, between the second and third lactation 0.79, and between the first and third lactation 0.45. Heritability estimates from the RR in the first and second lactations decreased from the beginning to the end of the lactation, with average values of 0.28 and 0.27, respectively. Estimates in the third lactation showed a maximum in the middle of lactation, averaging 0.37. Genetic correlations between the first and second lactation averaged 0.64, between the second and third lactation 0.72, and between the first and third lactation 0.46. Despite the small data set and restricted relationship structure the estimates were reasonable with the exception of estimates from the third lactation, which seemed inflated. RR could be used for genetic evaluation of dairy goats in Germany.

A multi-breed model was presented for the genetic evaluation of growth traits in beef cattle. In addition to the fixed effects, random direct and maternal genetic effects, and random maternal permanent environmental effects are considered; the model also fits direct and maternal heterosis and direct and maternal breed-of-founder (BOF) x generation group effects using a Bayesian approach that weights prior literature estimates relative to information supplied by the dataset to which the model will be applied. The multi-breed evaluation procedures also allow the inclusion of external evaluations for animals of other breeds. The multi-breed model was applied to a dataset provided by the American Gelbvieh Association. Different analyses were conducted by varying the weights given to the prior literature relative to the information provided by the dataset. Large differences were observed for the heterosis estimates, the BOF x generation group effect estimates, and the predicted breeding values across breeds due to the weights posed on prior literature estimates versus estimates derived directly from data. However, the rankings within breed were observed to be relatively robust to the different weights on prior information.

Properties of random regression models using linear splines (RRMS) were evaluated with respect to scale of parameters, numerical properties, changes in variances and strategies to select the number and positions of knots. Parameters in RRMS are similar to those in multiple trait models with traits corresponding to points at knots. RRMS have good numerical properties because of generally superior numerical properties of splines compared with polynomials and sparser system of equations. These models also contain artefacts in terms of depression of variances and predictions in the middle of intervals between the knots, and inflation of predictions close to knots; the artefacts become smaller as correlations corresponding to adjacent knots increase. The artefacts can be greatly reduced by a simple modification to covariables. With the modification, the accuracy of RRMS increases only marginally if the correlations between the adjacent knots are > or =0.6. In practical analyses the knots for each effect in RRMS can be selected so that: (i) they cover the entire trajectory; (ii) changes in variances in intervals between the knots are approximately linear; and (iii) the correlations between the adjacent knots are at least 0.6. RRMS allow for simple and numerically stable implementations of genetic evaluations with artefacts present but transparent and easily controlled.

Summary: Computing properties of better derivative and derivative-free algorithms were compared both theoretically and practically. Assuming that the log-likelihood function is approximately quadratic, in a t-trait analysis the number of steps to achieve convergence increases as t(2) in 'better' derivative-free algorithms and is independent of that number in 'better' derivative algorithms. The cost of one step increases as t(3) . Consequently, both classes of algorithms have a similar computational cost for single-trait models. In multiple traits, the computing costs increase as t(3) and t(5) , respectively. The derivative-free algorithms have worse numerical properties. Four programs were used to obtain one-, two-, and three-trait REML estimates from field data. Compared to single-trait analyses, the cost of one run for derivative-free algorithms increased by 27-40 times for two traits and 152-686 times for three traits. A similar increase in rounds of iteration for a derivative algorithm reached 5 and 21, and 1.8 and 2.2 in canonical transformation. Convergence and estimates of derivative algorithms were more predictable and, unlike derivative-free algorithms, were much less dependent on the choice of priors. Well-implemented derivative REML algorithms are less expensive and more reliable in multiple traits than derivative-free ones. ZUSAMMENFASSUNG: Vergleich von Rechen (Computing) merkmalen von abgeleiteten und ableitungsfreien Algorithmen zur Varianzkomponentenschätzung mittels REML Rechenmerkmale von verbesserten ableitungsfreien und Algorithmen, die Ableitung benutzen, werden theoretisch und praktisch verglichen. Unter der Annahme einer ungefähr quadratischen log-likelihood Funktion, nimmt in der Analyse von t Merkmalen die Zahl der Rechenschritte bis zu Konvergenz mit t(2) in 'besseren' ableitungsfreien Algorithmen zu und ist davon unabhängig von dieser Zahl in der 'besseren' Ableitung. Die Kosten je Schritt steigen mit t(3) . Daher haben beide Berechnungsarten für Einzelmerkmale ähnliche Rechenkosten. Bei mehreren Merkmalen steigen die Kosten mit t(3) bzw. t(5) und ableitungsfreie Algorithmen haben schlechtere numerische Eigenschagten. Vier Programme haben für ein-, zwei- und drei-Merkmale REML Schätzungen von Felddaten erzeugt. Im Vergleich zu Ein-Merkmal Analysen stiegen Kosten für einen Lauf bei ableitungsfreien Algorithmen um das 27-40 fache bei zwei- und um das 152-686 fache bei drei-Merkmalen. Die Steigerungen je Lauf bei auf Ableitung beruhenden Algorithmen waren 5-21 fach und 1.8 und 2.2 fach bei kanonischer Transformation. Konvergenz und Schätzwerte von Algorithmen mit Ableitung waren besser vorhersagbar und weniger von der Wahl der priors beeinflußt. Gut ausgestattete REML Methoden, die Ableitungen benutzen, sind ökonomischer und verläßlicher bei Mehrmerkmalsproblemen als ableitungsfreie., (1994 Blackwell Verlag GmbH.)