Evaluation of Classifiers that Score Linear Type Traits and Body Condition Score Using Common Sires

Transcription

1 J. Dairy Sci. 85: American Dairy Science Association, Evaluation of Classifiers that Score Linear Type Traits and Body Condition Score Using Common Sires R. F. Veerkamp,* C. L. M. Gerritsen,* E. P. C. Koenen, A. Hamoen, and G. De Jong *Institute for Animal Science and Health, ID-Lelystad, PO Box 65, 8200 AB Lelystad, The Netherlands NRS, PO Box 454, 6800 AL Arnhem, The Netherlands ABSTRACT Subjective visual assessment of animals by classifiers is undertaken for several different traits in farm livestock, e.g., linear type traits, body condition score, or carcass conformation. One of the difficulties in assessment is the effect of an individual classifier. To ensure that classifiers rank animals consistently, i.e., the repeatability between classifiers and within classifier, genetic links across routinely scored observations may be used to validate scoring of individual classifiers. Eighteen classifiers of NRS scored 18 traits, and body condition for 91,589 first-lactation heifers, daughters of 601 sires. Genetic parameters were estimated in a series of bivariate analyses. In turn, observations of each individual classifier were trait 1 and all observations of all other classifiers were grouped as trait 2. Likelihoods were used to test whether additive genetic or residual variances for each classifier (trait 1) differed significantly from the grouped records (trait 2), and to test whether the genetic correlation between trait 1 and trait 2 was significantly smaller than unity. Arbitrary criteria were set to mark traits for individual classifiers when a significant deviation was found: genetic correlations of 0.40, and more than 15% deviation for the standard deviation. One classifier had relatively low heritabilities, but high genetic correlations with the others. This might indicate that the repeatability within classifier should be improved. Another classifier had high genetic correlations with the others, but his sire variances were significantly higher than average for most traits. For the genetic correlations, each classifier averaged 3.3 traits marked, ranging from 0 to 9. Overall feet and legs, rump width, central ligament, and foot angle received most marks (12 to 6 classifiers), but no disagreement existed on the definition (i.e., no mark) for body condition score, stature, rump angle, Received September 28, Accepted November 21, Corresponding author: R. F. Veerkamp; r.f.veerkamp@ id.wag-ur.nl. teat length, overall udder, and teat placement. These simple and cheap marks can be used in training sessions to improve the quality of the scoring system. (Key words: classification scheme, genetic selection, conformation, dairy cow) INTRODUCTION Subjective visual assessment of animals by classifiers is undertaken for several different purposes in livestock. For example, linear type classification in dairy cattle is routinely performed in many countries, and records are used, for example, for the prediction of longevity (Brotherstone and Hill, 1991; Vollema and Groen, 1997), BW (Veerkamp and Brotherstone, 1997; Koenen and Groen, 1998), udder health (Thomas et al., 1984; De Jong and Lansbergen, 1996), feet and leg problems (Boelling and Pollott, 1998), and calving ease (Dadati et al., 1985; Cue et al., 1990). For similar reasons, linear type classification is also used in horses (e.g., Koenen et al., 1995; Samoré et al., 1997; Molina et al., 1999), pigs (e.g., Van Steenbergen, 1990; Serenius et al., 2001) and sheep (e.g., De la Fuente et al., 1996; Cloete et al., 1998). Other health and management traits are also scored visually by classifiers for example, BCS (Jones et al., 1999; Koenen et al., 2001), locomotion (Boelling and Pollott, 1997), or carcass conformation score, and carcass fat score that are used when selecting beef animals (Bass et al., 1977; Amer et al., 1997; Van der Werf et al., 1998). One of the problems with type classification and these subjective scores is the effect of classifier. Classifiers differ in their mean score, and unofficial age adjustment, but also in the range of the scale that they use (Bowden, 1982; Fleuren, 1988). For breeding value estimation, most of these factors can be adjusted for in the model, or by preadjustment of the records. For example, for the phenotypic standard deviation within classifier (Brotherstone et al., 1990; NRS, 2000). Hence, these classifer effects are of little problem for animal breeders. Still phenotypic records are important for certificates, culling decisions, and selling of animals. There- 976

2 EVALUATION OF TYPE SCORING 977 fore, classifiers need to be trained regularly. For this purpose, in the Netherlands classifiers receive overviews of the mean and phenotypic standard deviation of their scores for each trait regularly. In addition to the mean and variation for each classifier, it is also important that all classifiers rank animals consistently, i.e., a consistent trait definition at all times. High repeatability within and between classifiers is also important for breeding value estimation since the statistical models that are used assume unity genetic correlations between scores (within and between classifiers). Bowden (1982) reported three studies that gave repeatability within classifier of between 0.56 and 0.82 and a repeatability across classifiers of In pigs, repeatabilities were 0.6 and 0.3 between and within classifier, respectively (Van Steenbergen, 1990). Validating the repeatability within and across classifiers is more difficult than validating the mean and variance, as it requires classifiers to score the same animals repeatedly. Therefore, international and national workshops are organized to standardize trait definitions between classifiers (e.g., Hewitt, 2000). Compared with organizing workshops, genetic parameters are expected to be an easier alternative. Genetic links across routinely scored records might be used to validate scoring of individual classifiers. The heritability is the observed correlation between relatives as a proportion of the correlation that would be found if all the variance was additive genetic (page 163 Falconer and Mackay, 1996). Therefore, the heritability estimated within each classifier can be used as criteria for the repeatability of scores within classifiers, albeit the optimum value is not unity but depends on the true heritability of each trait. Similarly, the genetic correlation between scores by different classifiers can be used as a measure of the repeatability between classifiers. Assuming classifiers score a random group of daughters of a bull and classifiers score the same trait, a genetic correlation of one between classifiers is expected. Among other reasons discussed later, the absence of a unity correlation between classifiers might be due to poor repeatability between classifiers. Given the relative simplicity of estimating genetic parameters on existing data compared with the organization of workshops with all classifiers, the objective of this study is to evaluate the use of genetic parameters for standardization of the linear type classification and body condition scoring in The Netherlands. Data MATERIAL AND METHODS Eighteen classifiers employed by the NRS scored 14 linear type traits, four descriptive traits, two farmer scored traits (character and milkability), and body condition for lactating heifers. Condition scoring was based on the system described by Lowman et al. (1973) and is described in more detail by Koenen et al. (2001). For the linear type traits and BCS a mean of 5 units (scale 1 to 9) and a standard deviation of 1.6 was the aim, whereas for the descriptive traits a standard deviation of 3.0 points and a mean of 80 points was the aim. Stature is measured in centimeters. Records on a total of 126,546 heifers with at least 50% Holstein genes scored according to the Black and White standard were collected between October 1998 and September Herds with fewer than 10 records or with daughters of fewer than three sires were excluded. Furthermore, it was required that sires had at least 15 daughters in the dataset. The final dataset consisted of 91,294 records on daughters of 601 sires in 5598 herds. Analysis Genetic parameters were estimated using ASREML (Gilmour et al., 2000) and a sire model, including the additive genetic relationships between sires (and their sires and grandsires) in the relationship matrix (n = 707). As no classifier was considered the gold standard, and given the limited number of records for most classifiers, correlations between two individual classifiers are highly sensitive to sampling errors. Therefore it was considered appropriate to compare each individual classifier as trait 1 with all other classifiers grouped as trait 2. Variance components were estimated in a series of bivariate analyses in which scores from each classifier were analyzed as the first trait and the combined scores of all other classifiers as the second trait. Thus, for all 21 traits and 18 classifiers, the following variances were estimated for the random sire and residual effects: Sire = σs 2 1 σ S12 σ S12 σs 2 ; Residual = σe σ 2 e 2 Where σ 2 s1 and σ 2 s2 = sire variance for trait 1 and 2, respectively; σ s12 = covariance between trait 1 and 2; σ 2 e1 and σ 2 e1 = residual variance for trait 1 and 2, respectively. No residual covariance existed between the records, as each animal was scored only once. The heritability for each classifier was calculated as 4 * σ 2 s1/(σ 2 s1 + σ 2 e1), and the genetic correlation between each classifier and the combined scores of all other classifiers was calculated as r g = σ s12 /σ s1 σ s2. Fixed effects were fitted across the two traits, and included breed with five Holstein levels (4/8 to 8/8 Holstein), year-month of calving (20 levels), herd (5598 levels), a quadratic regression

3 978 VEERKAMP ET AL. on age at calving, and a cubic regression on DIM at the time that classification took place. A fixed effect for classifier was fitted for the second trait only. To test significant deviations from unity for the genetic correlation, the analysis was repeated to obtain the likelihood when the genetic correlation was fixed at unity. Twice the difference in log-likelihood was used as test statistic approximating a chi-square distribution with 1 degree of freedom. Constraining the heritability in the analysis was not feasible, and, therefore, the analysis was repeated twice, constraining σ 2 s1 = σ 2 s2 or σ 2 e1 = σ 2 e2 respectively. The likelihoods obtained were tested against the likelihood of the full model when sire and residual variance were different for trait 1 and trait 2. Some arbitrary criteria were used to mark trait by classifier combinations. These criteria were needed because classifiers that scored a large number of animals (ca. 7500) contributed a relatively large part of the genetic parameters. Estimates from classifiers that scored a few animals (ca. 1000) sometimes differed due to sampling. Criteria used for marking traits within classifier combinations were a genetic correlation 0.90, but the correlation had to be significantly different from unity. For the sire and residual standard deviation a significant deviation more than 15% from average score of all classifiers was marked. RESULTS For five traits, estimates for the genetic parameters per classifier are given in Table 1. The phenotypic standard deviation was 3.3 cm for stature, and for feet and legs, BCS, and the linear traits the phenotypic standard deviation was close to the breed standard (3.0 and 1.4 to 1.5, respectively). Heritability estimates for stature were high and all in a narrow range (0.50 to 0.59) for classifiers with more than 3000 heifers scored. For classifiers with fewer than 1000 observations, more variation in the heritabilities was observed. Most of the residual variances per classifier were significantly different from the combined score of the other classifiers, but none of the significantly different residual variances differed by more than 15% from the mean variance of all classifiers. Thus, due to the precision of the estimated residual components, significant differences were picked up that might be of relatively little importance. For stature, each classifier had a genetic correlation with the combined scores of all others close to unity, i.e., there was close agreement among classifiers. For overall feet and legs and foot angle, the mean heritability was lower compared with stature, and the heritability among classifiers varied more. Little agreement existed among classifiers when scoring these traits: most classifiers differed significantly from the combined score of all others. Similarly, a number of classifiers differed for their definition of rump width, although the heritability for rump width was relatively high compared with the heritability of overall feet and legs and foot angle. Body condition score is a trait recently introduced in The Netherlands, but no classifier had a correlation with the others below 0.90, and the heritability averaged The summary of the scoring of individual classifiers is given in Table 2. Using the arbitrary criteria that a heritability deviated more than two standard errors from the average heritability, three classifiers received five or six marks. For classifier Q the heritabilities for five out of six traits were lower than average, but only one genetic correlation differed significantly from unity. This might indicate that the repeatability within classifier should be improved, although there are probably too few heifers scored by this classifier to draw a firm conclusion. Results from classifier C were more surprising. Five heritabilities deviated positively and were more than 2 standard errors above average. For 12 traits the sire variance was significantly higher than average (more than 15% higher), whereas no residual variance was marked for any of the traits. However, trait definition agreed closely with the other classifiers. For classifier D, only two heritabilities deviated from average, but for a large number of the traits the residual and sire variance were significantly lower than average. Also, classifier D ranked heifers differently from the combined score of others for nine traits. Classifier H had similar variances and heritability as the others, but trait definition should be improved for six traits. A summary of the genetic parameters by trait (Table 3) indicates much agreement among classifiers about stature, BCS, rump angle, overall udder, teat placement, and teat length. There were no marks for the genetic correlations for these traits. The farmer scored character and milkability, so there should not be an effect of classifier. For the traits central ligament, foot angle, rump width and overall feet and legs up to 12 classifiers received a mark for the genetic correlation. For body depth, a relatively large number of classifiers received a mark for genetic variance. Three scored a higher and three scored a lower sire variance for body depth. DISCUSSION AND CONCLUSIONS In this study we used genetic parameters to evaluate precision and accuracy for scoring of linear type traits by individual classifiers. Parameters were estimated using genetic links across existing data. The advantages of this method are that biases (e.g., due to classifiers remembering cows they score previously) and costs

4 EVALUATION OF TYPE SCORING 979 Table 1. Estimates for the heritability (h 2 ), residual (E), and genetic (A) standard deviation per classifier (and overall) and the genetic correlation between scores by each individual classifier and the combined scores of all other classifiers. Significant levels for the variances indicate if a model assuming the individual classifiers variance to be equivalent to the variance of the combined scores of all other classifiers gives a lower likelihood. Significant levels for the genetic correlations indicate if genetic correlations differ from unity. Stature Overall feet and legs Foot angle Rump width BCS Classifier # h 2 E A r g h 2 E A r g h 2 E A r g h 2 E A r g h 2 E A r g A *** ** *** ** *** ** *** 1.0* 0.90*** *** 0.6** 0.93** B *** 2.7* *** 1.5** 0.89*** *** ** *** *** 0.9* 0.92*** C *** *** 1.9*** 0.88*** *** 0.7*** 0.95** *** 1.0*** 0.96*** *** 0.9*** 1.00 D *** 2.3* *** *** *** *** *** 0.5*** 0.80*** *** 0.7* 0.98* E *** * *** *** 0.6* *** 1.0* 0.79*** *** 1.0** 0.99 F * *** 1.4* 0.89*** *** * *** *** *** *** G ** ** *** 0.4** *** 1.2*** *** *** H * * *** *** *** * I *** 2.3* *** ** *** *** *** *** 1.0* 0.94*** J *** 2.8* *** 1.4* 0.91*** *** 0.7* 0.87*** *** 1.3*** 0.96*** *** * K * *** *** *** *** 0.5*** 0.71*** *** L *** *** ** *** 0.7* 0.80*** *** *** *** M ** * * * 0.6* * *** * N * *** ** *** *** *** * *** 1.0* 0.99 O * ** *** * * *** P * * * *** Q * *** *** * R * 0.83* *** Overall ***P < **P < *P < 0.05.

5 980 VEERKAMP ET AL. Table 2. Number of traits per classifier that had sire or residual variances significantly different from the combined score of the other classifiers or where the genetic correlation differed significantly from unity. Additionally, the number of traits per classifier where there was large deviation of the genetic parameters from the average score of the other classifiers, i.e., the h 2 deviated more than two standard errors, the residual (E) or sire (A) variances deviated more than 15% (higher =+; lower = ), or the genetic correlation (r g ) was 0.90 or less. Traits were marked when these large differences were significantly different using likelihood test. E A r g h 2 Classifier 2* SE <15%> P < 0.05 Marked + Marked <15%> P < 0.05 Marked + Marked 0.90 P < 0.05 Marked A B C D E F G H I J K L M N O P Q R associated with large-scale training sessions (where all classifiers score a large number of cows repeatedly) might be reduced. In this study, the combined scores of all other classifiers were used as gold standard, as there is no objective measure of the gold standard available. Comparing each classifier with the scores of all others is based on the principle that if a trait is scored consistent (within and between classifiers) then there is agreement on the trait definition. Thus the scoring should be accurate. The drawback of grouping all others together is that good classifiers are compared against a group including the worst classifiers, assuming that within and between genetic correlation for classifications within this group are unity. As a result, the heritability of the combined score is likely to be lower than the average of the individual ones. In this study therefore, the heritability of each classifier was compared with the average of the heritabilities for each classifier alone (i.e., trait 1). Using the scores of the head classifier might have been an alternative gold standard, but the disadvantage is that sampling variance of the genetic parameters increases due to a limited number of records in the gold standard. A better solution might be to select a panel of classifiers that scores each trait closest to the breed standard. The results of a first analysis as presented in this study might help to establish such a panel of classifiers, by grouping classifiers with a high heritability and genetic correlation. Another alternative might be to estimate genetic correlations between all classifi- ers and use a clustering technique to identify groups of classifiers. Genetic parameters should be interpreted carefully. The number of observations per classifier ranged from 814 to Therefore, accuracy of the estimated genetic parameters differs among classifiers. Neither a significance test nor the magnitude of the genetic parameters is sufficient to judge a trait classifier combination. In this study both criteria were used, and arbitrarily a deviation of 15% and a genetic correlation below 0.90 were chosen to indicate important differences. These criteria might differ depending on the classification scheme and traits scored, and only when other type classification schemes are analyzed, can more realistic benchmarks be set. In this study, we used four parameters for each classifier trait combination, i.e., heritability, sire and residual variance, and the genetic correlation. The importance of interpreting all parameters is also highlighted by some individual examples. Problems for classifier D (i.e., within repeatability is good, but variances are too low and correlations with others were poor) were not identified when only heritabilities were estimated. Classifier H should focus on the ranking of heifers compared with others. Following these examples, we suggest evaluating each trait for each classifier following the diagram in Figure 1. Evaluation obviously starts with the mean score for each classifier, i.e., the mean should be close to the trait standard (5 for linear traits and 80 for descriptive traits). Secondly, the genetic

6 EVALUATION OF TYPE SCORING 981 Table 3. Classifiers per trait that had sire or residual variances significantly different from the combined score of the other classifiers or where the genetic correlation differed significantly from unity. Additionally, the number of classifiers per trait where there was large deviation of the genetic parameters from the average score of the other classifiers, i.e., the h 2 deviated more than two standard errors, the residual (E) or sire (A) variances deviated more than 15% (higher =+; lower = ), or the genetic correlation (r g ) was 0.90 or less. Traits were marked when these large differences were significantly different using likelihood test. E A r g h 2 2* SE <15%> P < 0.05 Marked + Marked <15%> P < 0.05 Marked + Marked 0.90 P < 0.05 Marked Stature Character Milkability BCS Teat placement Rump angle Teat length Overall udder Udder depth Type Overall conformation Rear height Fore attachment Rear legs side view Body depth Angularity Chest width Central ligament Foot angle Rump width Overall feet and legs

7 982 VEERKAMP ET AL. Figure 1. Evaluation process for trait by classifier combination using genetic parameters. standard deviation should not be lower than the average. If the genetic standard deviation is lower, this could be due to the scale used (measured by the phenotypic standard deviation), due to poor within classifier repeatability (a low heritability) or both. Finally, if the genetic correlation is close to unity (accounting for sampling errors using the likelihood test), it might be expected that the classifier scored the trait accurately. This scheme would help to improve scoring of individual classifiers. However, surprisingly, some classifiers (e.g., C and J) score consistently higher sire components than average, and had less deviation for the residual variance or the genetic correlation. This finding suggests that these classifiers score more accurately than others; however, it might be speculated that these classifier unconsciously use additional information when scoring some traits. They may, for example, recognize daughters of some bulls and the type scores associated with these bulls. The results for milkability and character (Table 3) may suggest that these traits are scored very accurately, but this might be misleading. These traits are scored by the farmers (during the herd visit) so no classifier effect is expected. Apart from scoring by individual classifiers, there are other reasons why the genetic parameters might deviate from average. A major reason might be that an individual classifier scores closer to the agreed trait than the average of all others. Another reason might be that each classifier scores different subsets of the population and genotype by environment interactions might exist. For example, differences in the housing system and foot trimming practices between parts of the country may affect the results. Also, single classifiers might score more daughters from Red-and-White herds, whereas others score daughters of the same bulls in Black-and-White herds. Preferential mating might then cause some classifiers to differ from the combined score of the others. For this reason, it might have been better to use the full pedigree information in an animal model analysis. However, given the large number of analyses involved, preference was given to the faster sire model. Also, we expect this effect to be relatively minor as some of the traits, e.g., stature, had close to unity correlations for all classifiers. Previously, in The Netherlands, support to classifiers was given by comparing the mean and phenotypic variance. However, using the evaluation procedure presented in this study, classifiers can be informed about their consistency of scoring a trait and about their trait definition as well. This will help to improve the quality of scoring for a relatively low cost compared with the organization of workshops. Similar analysis across classification schemes might help to improve international conversions of type traits. Apart from type classification, this method can be used also for evaluation of other classification schemes, e.g., meat quality and fat scoring in slaughterhouses, or the subjective evaluation of health traits. REFERENCES Amer, P. R., G. C. Emmans, and G. Simm Economic values for carcass traits in UK commercial beef cattle. Livest. Prod. Sci. 51: Bass, J. J., C. Robinson, and F. Colomer Rocher Value of conformation in New Zealand beef grading. Proc. N.Z. Soc. Anim. Prod. 37: Boelling, D., and G. E. Pollott The genetics of feet, legs and locomotion in cattle. Anim. Br. Abstr. 65:1 11. Boelling, D., and G. E. Pollott Locomotion, lameness, hoof and leg traits in cattle: II. Genetic relationships and breeding value. Livest. Prod. Sci. 54: Bowden, V Type classification in dairy cattle: a review. Anim. Breeding Abstr. 50:

8 EVALUATION OF TYPE SCORING 983 Brotherstone, S., and W. G. Hill Dairy herd life in relation to linear type traits and production. 2. Genetic analyses for pedigree and non-pedigree cows. Anim. Prod. 53: Brotherstone, S., C. M. McManus, and W. G. Hill Estimation of genetic parameters for linear and miscellaneous type traits in Holstein-Friesian dairy cattle. Livest. Prod. Sci. 26: Cloete, S. W. P., A. J. Scholtz, J. M. Ten Hoope, P. J. A. Lombard, and M. C. Franken Ease of birth in relation to pelvic dimensions, litter weight and conformation of sheep. Small Rum. Res. 31: Cue, R. I., H. G. Monardes, and J. F. Hayes Relationships of calving ease with type traits. J. Dairy Sci. 73: Dadati, E., B. W. Kennedy, and E. B. Burnside Relationships between conformation and reproduction in Holstein cows: type and calving performance. J. Dairy Sci. 68: De Jong, G., and L. Lansbergen Udder health index: selection for mastitis resistance. Proceedings of the International workshop on genetic improvement of functional traits in cattle, Gembloux, Belgium, January INTERBULL Bull. 12: De la Fuente, L. F., G. Fernandez, and F. San Primitivo A linear evaluation system for udder traits of dairy ewes. Livest. Prod. Sci. 45: Falconer, D. S., and T. F. C. Mackay Introduction to Quantitative Genetics, 4th ed. Longman, Harlow, UK. Fleuren, W Subjective scoring of foot angle [In Dutch]. Den Bosch Agricultural College, pp 27. Gilmour, A. R., B. R. Cullis, S. J. Welham, and R. Thompson ASREML. Program user manual. NSW Agriculture, Orange Agricultural Institute, Forest Road, Orange, NSW, 2800, Australia. Hewitt, D Type harmonisation working group report of World Holstein Friesian Federation 5th Classifiers Workshop. Pages in 10th World Holstein-Friesian Conference, Sydney. Jones, H. E., I. M. S. White, and S. Brotherstone Genetic evaluation of Holstein Friesian sires for daughter condition-score changes using a random regression model. Anim. Sci. 68: Koenen, E. P. C., and A. F. Groen Genetic evaluation of body weight of lactating Holstein heifers using body measurements and conformation traits. J. Dairy Sci. 81: Koenen, E. P. C., A. E. Van Veldhuizen, and E. W. Brascamp Genetic parameters of linear scored conformation traits and their relation to dressage and show jumping performance in the Dutch Warmblood riding horse population. Livest. Prod. Sci. 43: Koenen, E. P. C., R. F. Veerkamp, P. Dobbelaar, and G. De Jong Genetic analysis of body condition score of lactating Dutch Holstein and Red-and-White heifers. J. Dairy Sci. 84: Lowman, B. G., N. A. Scott, and S. H. Somerville Condition scoring of cattle. Bulletin no. 6, East of Scotland College of Agriculture, Edinburgh, Scotland. Molina, A., M. Valera, R. Dos Santos, and A. Rodero Genetic parameters of morphofunctional traits in Andalusian horse. Livest. Prod. Sci. 60: NRS Handboek Koninklijk Nederlands Rundvee Syndicaat, Arnhem. Samoré, A. B., G. Pagnacco, and F. Miglior Genetic parameters and breeding values for linear type traits in the Haflinger horse. Livest. Prod. Sci. 52: Serenius, T., M. L. Sevon-Aimonen, and E. A. Mantysaari The genetics of leg weakness in Finnish Large White and Landrace populations. Livest. Prod. Sci. 69: Thomas, C. L., W. E. Vinson, R. E. Pearson, F. N. Dickinson, and L. P. Johnson Relationships between linear type scores, objective type measures, and indicators of mastitis. J. Dairy Sci. 67: Van der Werf, J. H. J., E. H. Van der Waaij, A. F. Groen, and G. De Jong An index for beef and veal characteristics in dairy cattle based on carcass traits. Livest. Prod. Sci. 54: Van Steenbergen, E. J Relevance of exterior appraisal in pig breeding. Ph.D. Thesis. Wageningen University, Wageningen, The Netherlands. Veerkamp, R. F., and S. Brotherstone Genetic correlations between linear type traits, food intake, live weight and condition score in Holstein Friesian dairy cattle. Anim. Sci. 64: Vollema, A. R., and A. F. Groen Genetic correlations between longevity and conformation traits in an upgrading dairy cattle population. J. Dairy Sci. 80: