J. Anim. Breed. Genet. ISSN 0931-2668
ORIGINAL ARTICLE
Cluster and meta-analyses of genetic parameters for feed intake traits in growing beef cattle I.D.P.S. Diaz1,2, D.H. Crews Jr2 & R.M. Enns2 1 Department of Animal Sciences, State University of Sao Paulo, Sao Paulo, Brazil 2 Department of Animal Sciences, Colorado State University, Fort Collins, CO, USA
Summary
Keywords Genetic parameters; multivariate analysis; pooled estimates. Correspondence I.D.P.S. Diaz, College of Agricultural Science and Veterinary, Via de Acesso Prof. Paulo Donatto Castellane, s/n, S~ao Paulo, Brazil. Tel: +55 11 98273 06 48; Fax: +55 34 3822 0753; E-mail: [email protected] Received: 11 January 2013; accepted: 9 October 2013
A data set based on 50 studies including feed intake and utilization traits was used to perform a meta-analysis to obtain pooled estimates using the variance between studies of genetic parameters for average daily gain (ADG); residual feed intake (RFI); metabolic body weight (MBW); feed conversion ratio (FCR); and daily dry matter intake (DMI) in beef cattle. The total data set included 128 heritability and 122 genetic correlation estimates published in the literature from 1961 to 2012. The meta-analysis was performed using a random effects model where the restricted maximum likelihood estimator was used to evaluate variances among clusters. Also, a meta-analysis using the method of cluster analysis was used to group the heritability estimates. Two clusters were obtained for each trait by different variables. It was observed, for all traits, that the heterogeneity of variance was significant between clusters and studies for genetic correlation estimates. The pooled estimates, adding the variance between clusters, for direct heritability estimates for ADG, DMI, RFI, MBW and FCR were 0.32 0.04, 0.39 0.03, 0.31 0.02, 0.31 0.03 and 0.26 0.03, respectively. Pooled genetic correlation estimates ranged from 0.15 to 0.67 among ADG, DMI, RFI, MBW and FCR. These pooled estimates of genetic parameters could be used to solve genetic prediction equations in populations where data is insufficient for variance component estimation. Cluster analysis is recommended as a statistical procedure to combine results from different studies to account for heterogeneity.
Introduction Differences among animals in conversion of feed into body weight are important determinants of profit in a beef production system. Crews (2005) pointed out that more attention must be given to the cost inputs of beef production and how to reduce them. A trait, such as feed consumption, is a good example, because this typically is considered the greatest (>65%) variable cost in a beef production system. In this context, feed efficiency or utilization has been evaluated using numerous phenotypes, as described in detail by Archer & Barwick (1999). These ‘efficiency’ measures can be grouped into gross efficiency, partial efficiency © 2013 Blackwell Verlag GmbH
• J. Anim. Breed. Genet. 131 (2014) 217–226
of growth, maintenance efficiency, cow/calf efficiency and residual feed intake (RFI), among others, with the authors noting that more than 20 different measures have been described to some degree in the literature. Considering the importance of feed utilization, many studies have estimated genetic parameters for different measures of feed efficiency but most have done so with limited amounts of data. The relatively small amount of data found in the literature is due to the expense associated with measuring individual animal feed intake. Yet with the volume of studies, considerable variability between values of the parameters has been observed. Elzo et al. (2011) reported a 0.19 doi:10.1111/jbg.12063
I. D. P. S. Diaz et al.
Meta-analyses for feed intake traits
estimate for direct heritability of residual feed intake, whereas Hoque et al. (2009) reported a value of 0.49. Arthur & Herd (2008) attributed this apparent heterogeneity at least partially to the fact that all the feed intake measures used in beef cattle involve a series of complex of biological processes and interactions with the environment. In order to be able to compare performance across time and location, as required for large scale genetic prediction, it is important to attempt to control for environmental factors that affect feed intake and its utilization. These factors include age at start of test, sex of cattle, diet composition (e.g. energy density) and testing procedures. Results from recent studies on the duration of test and frequency of measurement (Archer et al. 1997; Wang et al. 2006) have formed the basis for developing standard test protocols (BIF 2010). Given the complexity of feed intake and utilization as well as sparse data resources, a meta-analysis of the published records could provide a more complete understanding of the genetic and environmental influences on these traits. To obtain a mean value that represents and summarizes the results from different studies and to overcome the heterogeneity between studies, a number of factors should be considered in the analyses such as the variability (variance) among the studies (Li & Begg 1994 and Li 1995). Meta-analysis has been used to combine results from different studies to increase the sample size (Koots et al. 1994; Giannotti et al. 2002, 2005). The aim of this study was to utilize a meta-analysis approach to combine published heritability and genetic correlation estimates among dry matter intake (DMI), average daily gain (ADG), metabolic body weight (MBW), feed conversion ratio (FCR) and residual feed intake (RFI) to produce genetic parameters useful for genetic evaluation in the absence of sufficient data to estimate these parameters.
Materials and methods Data
The data used in this study included 128 estimates of ^2 ) for dry matter intake (DMI), average heritability (h ij daily gain (ADG), feed conversion ratio (FCR), metabolic body weight (MBW) and residual feed intake (RFI) (i.e. traits commonly measured in bull performance tests) and 122 genetic correlations among those traits in a beef cattle system. The data were obtained from 50 studies published from 1961 to 2012 (see Appendix S1). The research represented original papers, theses, dissertations and conference proceed218
ings. The analyses in this study were divided into two parts: the first part was to perform a cluster analysis using all the heritability estimates on each trait and the second involved the meta-analyses to obtain a pooled estimate using the variance between studies. We chose not to perform the cluster analysis on genetic correlation ^rg estimates because of the total amount of values found, because the correlation between each pair of trait adds a very small amount of estimates and therefore impracticable to form clusters to each correlation. Cluster analysis
^2 ) from different studies for Heritability estimates (i) (h ij each trait were divided into clusters (j) within each trait by a cluster analysis approach. This analysis was chosen to gather within each cluster the most similar ^2 relying on information about the effects of fixed h ij variables of sex of the animals, breed, country of origin, method of analysis, age at start of the test, dura^2 tion of the test, sample size, type of feeding and the h ij from each study of each trait. To form the clusters, the hierarchical Ward method was chosen, where the main feature was to form groups that were homogeneous as possible (Sharma1996). In this approach, the distance between two clusters is defined as the sums of squares between the two clusters performed on all the variables. In each stage of the procedure of clustering, the internal (within cluster) sum of squares is minimized on all the partitions that could be obtained from the combination of two clusters from the previous stage. Euclidean distance was used as the similarity coefficient, which is considered the geometrical distance in a m-multidimensional space. Thisffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi distance was calculated using dAB ¼ ðx1A x1B Þ2 þ ðx2A x2B Þ2 þ . . . þ ðxmA xmB Þ2 , where A and B are the clusters and xij is the ith variable from the jth cluster. In our study, we choose the number of clusters by the evaluation of the tree diagram. The cluster analysis was conducted using STATISTICA 7.0 software (StatSoft, Inc 2004). The variables used in the method of clustering were transformed in unique labels (dummy variables) for categorical fixed effects (method of analysis, country, sex and breed). The resulting variables were as follows: X1 where a value of 1 was assigned if the study used purebred animals, and 0 if not; X2 where a value of 1 was assigned if location of the study were the United States and 0 if not; X3 where values of 1 for males, 2 for females and 3 for both were assigned based on animals in the study; and X4 where values of 1 were assigned studies implementing multiple-trait © 2013 Blackwell Verlag GmbH
• J. Anim. Breed. Genet. 131 (2014) 217–226
I. D. P. S. Diaz et al.
Meta-analyses for feed intake traits
analyses, 2 for bivariate analyses and 3 for univariate analyses. For the quantitative variables (number of observations – sample size, age at start of test, test ^2 estimate), a standardization was duration and the h ij performed. The variables were ‘standardized’ using the expression z = Xi min (x)/( max (x) min (x)) ^2 ; min where Xi represents the i-th value or the i-th h ij 2 ^ ; and max(x) (x) is the smallest value or the smallest h ij ^2 . is the largest value or the largest h ij After clustering the heritability estimates, the metaanalysis was performed.
One of the main questions in a meta-analysis approach is to verify the homogeneity between studies. In this study, because the cluster analysis was performed previously, the homogeneity was tested between clusters and within clusters for each evaluated trait. The homogeneity was tested ^2 ¼ h ^2 ¼ . . . considering the null hypothesis of H0 : h 1þ 2þ 2 ^2 is the ^ where g is the number of clusters and h ¼h gþ iþ pooled estimate of the i-th cluster versus the alterna^2 estimates is tive hypothesis that at least one of the h iþ different from another. The test statistic was assumed to be distributed as chi-square with k – 1 df where g P ^2 h ^ 2 Þ2 , (Cochran, 1954): Qbetween ¼ wiþ ðh iþ þþ i¼1
Meta-analysis
When studies to be combined have different designs and have similar result structures and they provide estimates as the average differences between treatment groups and control groups, correlation coefficients and heritability estimates, these studies could be synthesized using a meta-analysis approach if there is information about the variance of the estimate and the sample size (N) used. (Hedges & Olkin 1985). For those studies that do not provide the variance, we cal^2 =N, culated with the following equations: ^s2ij ¼ 32h i 2 ^ calculated with animal or sire model when the hipwas ffiffiffiffi ^2 was calculated by regression and ^s2ij ¼ 2= N when h i of progeny on sire or regression progeny on dam (Koots et al. 1994). The meta-analysis of heritability estimates was performed in a stepwise manner after the cluster analysis. The first step involved the exploration of the data structure; the second was to evaluate the homogeneity of variance between clusters and within clusters which if the heterogeneity of variance was significant then the third step was performed where the variance between clusters was calculated; and the following last step was to obtain the pooled estimate. This variance was calculated as a way to include the variance between studies (Li & Begg 1994; Li 1995) and to overcome the problem of lack of power of the Q test (Higgins et al. 2003). Different box-plot graphics were constructed for each cluster within each trait to further explore the data. The mean, the standard error and the likely outliers were utilized to observe whether the different clusters are comparable with each other (in this case between clusters within each trait). Because a meta-analysis would only be valid if the normality and independence of the estimates to be combined are achieved (Hedges & Olkin 1985), we performed the Shapiro–Wilk test (StatSoft, Inc 2004) to quantify the normality of the estimates. © 2013 Blackwell Verlag GmbH
• J. Anim. Breed. Genet. 131 (2014) 217–226
wiþ ¼
mi P
^2 within each wij ; j is the number of h
j¼1
cluster and wij ¼ 1=^s2ij is the individual estimate mi mi ^2 ¼ P wij h ^2 ^2 = P wij ; and h weighting factor; h iþ ij iþþ j¼1
¼
g P i¼1
^2 = wiþ h iþ
g P i¼1
j¼1
wiþ , where ^s2i is the variance of each
^2 is the pooled cluster, wij is the inverse variance, h iþ 2 ^ heritability for each i cluster and hþþ is the pooled estimate between all groups for each trait. The null hypothesis was rejected considering the level of significance p, if Qbetween exceeded the critical value of 100(1– p)% of chi-square distribution with g–1 degrees of freedom. To test the homogeneity within clusters, the null hypothesis was Ho : h2i1 ¼ h2i2 ¼ . . . ¼ h2imiþ ¼ h2iþ where i = 1,2,….g where g is the total number of estimates of each cluster. The statistics used was: k P mi P ^2 h ^2 Þ2 , where k=m1+m2+…+ wiþ ðh Qwithin ¼ i¼1 j¼1
ij
iþ
mg, with a chi-square distribution with k–g degrees of freedom (Wang & Bushman 1999). These tests were made using macros Within & Wavgz (Wang & Bushman 1999) in SAS. When the presence of heterogeneity between clusters was observed, a random model was adopted in the meta-analysis procedure to account for variability between and within clusters. A fixed model was fitted when there was no significant heterogeneity between clusters, but there was variability only within clusters. In the random model, ^2 are different and could we considered that the h iþ ^2 þ ei þ ei ^ where be described as h2þþ ¼ h iþ ^2 represents the pooled esti^2 Nð ^2 ; s2 þ s2 Þ, h ~ h h þþ iþ ij þþ mate, where ei eNð0; ^s2ij Þ is the error related to each cluster and ei eNð0; s2 Þ is the error between clusters and s2 is the variance between clusters. For correlation estimates, the same interpretation and model ^2 . could be used placing ^rg in place of h iþ 219
I. D. P. S. Diaz et al.
Meta-analyses for feed intake traits
The s2 between clusters was calculated by REML method using macros and programmes published by Wang & Bushman (1999) developed in SAS. In this method, the following equation was used: k k k 2 P ^2 Þ2 s2 = P w 2 . ^ h ^s2 ¼ wi2 k1 ðh i þ i i i¼1
i¼1
After s estimation, the pooled heritability estimates for each trait were obtained by the following k mi k mi ^2 ¼ P P w h ^2 P P w where equation: h þ ij ij = ij 2
i¼1 j¼1
Koots et al. (1994) as well, where the ^rg values were transformed to the z quantity, where the zi value are considered to have a normal distribution with variance equal to varðzi Þ ¼ 1=ðni 3Þ, and ni is the number of observations of the i-th study. So, P zi =varðzi Þ2 ð1þrg Þ zi ¼ 0:5 ln ð1rg Þ and z pooled ¼ P 1=varðz Þ2 ; therefore, i
the
pooled
rg pooled ¼
estimate
ðe2x z pooled 1Þ . ðe2x z pooled þ1Þ
was
calculated
by:
In this case to combine the ^rg
values, the variances of z were used.
i¼1 j¼1
wij ¼ 1=ðs2ij þ ^s2 Þ and the pooled standard deviation sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi k P mi P ð1=wij Þ (Koots et al. was estimated as sþ ¼ 1=
Cluster discrimination
1994). For the genetic correlations among all the studies, the pooled values were obtained using the method of
When evaluating the tree diagram (Figure 1), it was possible to separate the estimates in two main clusters. Thus, for each of the five traits examined, two
i¼1 j¼1
Results
(b)
(a)
(d)
(c)
(e)
Figure 1 Tree diagram for 8 variables using Ward method by Euclidean distances considering sex of the animals; country of origin; method of evaluation; breed of the cattle; age at start of the test; number of observations (N) for (a) ADG, (b) DMI, (c) FCR, (d) RFI and (e) – MBW traits.
220
© 2013 Blackwell Verlag GmbH
• J. Anim. Breed. Genet. 131 (2014) 217–226
I. D. P. S. Diaz et al.
2.0 1.5 1.0 0.5 0.0 –0.5 –1.0 –1.5 –2.0
Meta-analyses for feed intake traits
(a)
Cluster 1 Cluster 2
2.0
(b)
Cluster 1 Cluster 2
1.5 1.0 0.5 0.0 –0.5 –1.0
h2ADG
Country Method Days_on_test N Age_start Sex Breed
Cluster 1 Cluster 2
2.0 (c) 1.5 1.0 0.5 0.0 –0.5 –1.0 –1.5 h2FCR Sex
Country Method Days_on_test Breed Age_start N 1.5 1.0
–1.5
2.0 1.5 1.0 0.5 0.0 –0.5 –1.0 –1.5 –2.0
h2DMI
Country Sex
Cluster 1 Cluster 2
(d)
h2RR
Method Days_on_test Breed Age_start N
Sex
Country Method Days_on_test Age_start Breed N
Cluster 1 Cluster 2
(e)
0.5 0.0 –0.5 –1.0 –1.5 –2.0 h2MBW Sex
Country Method Days_on_test N Age_start Breed
Figure 2 Groups of heritability estimates obtained considering sex of the animals; country of origin; method of evaluation; breed of the cattle; age at start of the test; number of observations (N) for (a) ADG, (b) DMI, (c) FCR, (d) RFI and (e) MBW traits.
distinct clusters were obtained (Figure 2); however, for each trait, the key variable responsible for the separation of clusters was different. The separation was based on the deviation from the mean where more discrepant values were those that deviate more from the centre of the graph. In this study, the variable that resulted in the greatest distance between the clusters was labelled the ‘key’ variable. For example, for ADG, the key variable was sex of the animals. Sex was also the key variable for RFI, which is not surprising given that ADG is included in the calculation of RFI. The sex of the animals was also the key variable followed by the variable country of origin for MBW. For DMI, the key variable was the group heritability estimates from each study. Considering FCR, the breed of the cattle was the key variable of the separation of the groups, © 2013 Blackwell Verlag GmbH
• J. Anim. Breed. Genet. 131 (2014) 217–226
followed by the method of evaluation and age at the start of the test. Exploratory analysis
The box-plots graphics (Figure 3) showed outliers and extreme values of heritability for all the traits but mostly for DMI for which the cluster analysis was able to separate estimates in two very distinct clusters corroborating the results in Figure 2. The evaluation of homogeneity between clusters and within clusters (Table 1), and between studies considering the genetic correlation estimates, was significant by the Q test (p < 0.01). It was observed that the Q values were high and significant. The lowest values for the Q statistics were found for the lowest values of degrees of freedom but could also have been 221
I. D. P. S. Diaz et al.
Meta-analyses for feed intake traits
(a)
(b)
(c)
(d)
(e)
Figure 3 Pooled heritability estimates and standard deviation for the traits: (a) average daily gain (ADG), (b) dry matter intake (DMI), (c) feed conversion ratio (FCR), (d) residual feed intake (RFI) and (e) metabolic body weight (MBW).
due to a lack of variability between the estimates of those clusters. The heterogeneity observed between studies suggests the implementation of the random model to obtain the pooled estimates. Pooled estimates
The pooled estimates of heritability were obtained for each cluster (Table 2) within each trait and for the clusters combined as well. The same approach was used for the genetic correlation estimates (Table 3) between traits. Heritability estimates were transformed to obtain the pooled estimates and their respective standard deviations, but the results are presented on the original scale. As expected, the values of the estimates differ between clusters for all traits indicating that the clus222
tering procedure is capturing differences among the study values, although all heritability estimates could be classified as moderate. Despite the small number of genetic correlation estimates obtained from different studies, the pooled estimates values are in the range expected by the authors (0.15–0.67). Most estimates were positive and with high magnitude with the exception of the genetic correlation between RFI and ADG (0.04) and ADG and FCR (0.15).
Discussion Cluster discrimination
The different values found between the heritability estimates for RFI, ADG and MBW due to the sex of the animals (Figure 2) could be related to the higher © 2013 Blackwell Verlag GmbH
• J. Anim. Breed. Genet. 131 (2014) 217–226
I. D. P. S. Diaz et al.
Meta-analyses for feed intake traits
Table 1 Q statistic to test the homogeneity between studies (clusters) and within cluster to genetic correlation and heritability estimates for average daily gain (ADG); dry matter intake (DMI); residual feed intake (RFI); metabolic body weight (MBW) and feed conversion ratio (FCR) Correlation estimates ADG,DMI
ADG,RFI
ADG,MBW
ADG,FCR
DMI,RFI
Causes of heterogeneity
D.F.
Q
D.F.
Q
D.F.
Q
D.F.
Q
D.F.
Q
Between studies
11
655.29*
16
327.01*
13
528.06*
15
727.88*
18
825.16*
DMI,MBW
DMI,FCR
RFI,MBW
RFI,FCR
MBW,FCR
Causes of heterogeneity
D.F.
Q
D.F.
Q
D.F.
Q
D.F
Q
D.F.
Q
Between studies
9
86.79*
25
976.25*
13
499.49*
20
906.05*
9
83.55*
Heritability estimates ADG
DMI
RFI
MBW
FCR
Causes of heterogeneity
D.F.
Q
D.F.
Q
D.F.
Q
D.F.
Q
D.F.
Q
Between clusters Within clusters Within cluster 1 Within cluster 2
1 30 11 19
18.63* 300.54* 71.76* 228.78*
1 27 19 8
32.78* 690.97* 414.88* 276.09*
1 44 23 21
10.40* 688.6* 393.54* 295.42*
1 29 20 9
28.89* 457.41* 392.70* 64.71*
1 47 36 11
16.06* 804.13* 732.80* 71.33*
*Significant by Q test, p > 0.01. Table 2 Pooled heritability estimates and the standard deviation obtained using REML for each cluster and for the clusters together Traits
Group
RFI h2*
MBW h2*
DMI h2*
ADG h2*
FCR h2*
1 2 Total
0.33 (0.04) 0.29 (0.02) 0.31 (0.02)
0.25 (0.02) 0.38 (0.05) 0.31 (0.03)
0.46 (0.03) 0.33 (0.03) 0.39 (0.03)
0.38 (0.05) 0.27 (0.03) 0.32 (0.04)
0.28 (0.02) 0.24 (0.04) 0.26 (0.03)
Table 3 Pooled heritability (diagonal) and genetic correlation estimates (off-diagonal) obtained by restricted maximum likelihood method Traits
RFI MBW DMI ADG FCR
RFI
MBW
DMI
ADG
FCR
0.31 (0.02)
0.25 (0.09) 0.31 (0.03)
0.67 (0.12) 0.36 (0.04) 0.39 (0.03)
0.04 (0.08) 0.33 (0.06) 0.38 (0.11) 0.32 (0.04)
0.65 (0.10) 0.24 (0.04) 0.34 (0.13) 0.15 (0.12) 0.26 (0.03)
average daily gain of males compared with females. Correspondingly, males’ genetic potential is higher and consequently so is the value of heritability from studies using only male data. According to Diaz et al. (2011a), the effect of sex on traits has been reported before, and it is a subject of discussion. Males and females may have different selection/ production objectives and therefore likely receive dif© 2013 Blackwell Verlag GmbH
• J. Anim. Breed. Genet. 131 (2014) 217–226
ferent treatment on farms with resulting environmental differences. Seemingly unimportant, because the selection takes place within sex, this difference in genetic parameters between males and females may gain importance as the genetic evaluation of animals is often performed considering additive genetic effects similar in both sexes. The observed difference found in this study between weights of males and females is consistent with other studies involving various beef cattle breeds (Diaz et al. 2011b; P egolo et al. 2011). The extreme points observed when the variable of ‘country of origin’ is evaluated for MBW, differences can be due to the fact that this study compared United States to all other countries. This was made because in most United States studies, MBW and ADG are computed using regression procedures. In this context, it is important to note that although these traits (FCR, MBW, RFI, DMI and ADG) are measured in controlled conditions and environments, the non-standardization of the tests across the central stations can increase the difference of the estimates in the same traits. According to Crews & Carstens (2012), this can also be due to the fact that serial weighing during test is preferred but there are still many test centres that record live weights only at the beginning and end of tests. As for intake, not all test stations use the same hardware/software for collection and (or) auditing of that data, although standard recommendations in 223
I. D. P. S. Diaz et al.
Meta-analyses for feed intake traits
North America do accommodate some of these hardware and software differences. Results from recent studies on the duration of test and frequency of measurement (Archer & Barwick 1999; Wang et al. 2006) have formed the basis for developing standard test protocols. The difference observed between clusters for DMI compared with other traits is the increased variation in heritability estimates. This trait had the largest number of estimates found in the literature (36 estimates). Dry matter intake variance is highly affected by the environment, mainly by different types of feed (National Research Council 2000) so most of these differences among heritability estimates could be related to different environmental variances. Another point that has to be considered is that these differences of heritability values, mainly for DMI, could be to differences on genetic variation among populations. Crossbred populations are more likely to have dominance effects on genetic variation; some studies may use animals with higher relationships than typical; however, these differences were higher and significant for DMI (Figure 2). Currently, some standard testing protocols further adjust DMI records to constant energy densities, but the majority of differences reported here among clusters would be as least partially accounted for by inclusion of appropriate contemporary group classification of DMI records (BIF 2010). It is well known that consumption of less-digestible, low-energy, high-fibre diets is affected by rumen fill and the feed passage rate through the animal. Meanwhile, consumption of highly digestible, high-energy, low-fibre feeds is controlled by the animal’s energy needs and by metabolic factors. These concepts may seem quite simple, but the factors that regulate DMI on concentrate rations are complex and are not fully understood. Because we are evaluating animals that had also pasture as a type of feed, we suggest that the type of feed offered, especially more palatable diets are more acceptable and therefore are more efficiently used by the animals than other types of feed. For FCR, the observed differences between clusters are primarily explained by the composition of the breed. Animals that have a higher percentage of Bos indicus breed composition had different conversion ratios when compared to predominantly Bos taurus animals. Schenkel et al. (2004) reported differences in FCR even among taurine breeds, where Simmental and Angus cattle had a higher FCR compared with Hereford. These are likely results in the differences among heritability estimates. 224
The method of evaluation (multiple-trait; univariate or bivariate) was also shown to have great importance in the analysis of feed efficiency traits. Studies have shown failure of convergence when all traits are evaluated together (Crews et al. 2010) such as when performing a multiple-trait analysis of RFI, ADG and MBW. These authors point out that this is a result of the interrelationships among traits such as when calculating RFI where DMI records are required. Age at start of the test also was significant in determining clusters for FCR. In this context, younger animals tend to have better feed efficiency (more efficient) than older animals, in other words, means that depends more on growth stage and diet of the animal. It takes more energy to gain weight through fat deposition than muscle and older animals on finishing rations deposit more fat. According to Freitas et al. (2006) older animals (with a greater degree of physiological maturity) have a greater efficiency gain for the synthesis of fat than younger animals, which have higher efficiency of protein gain and protein synthesis, so the feed conversion estimates become different depending on the age at which these animals are tested. Exploratory analysis
The observed outliers in the box-plot graphics (Figure 3) were included in this research, because all the authors explained those values in their studies. This procedure was also adopted in other meta-analysis studies (Giannotti et al. 2002, 2005). When the homogeneity of variance was tested between clusters, after using the Ward clustering method, it was expected that there would be heterogeneity among clusters (Table 1); however, it was not expected that there would be heterogeneity within clusters, because the aim of this method was to obtain the most homogeneous values within each cluster. Once the Ward clustering method tend to combine clusters with a small amount of information and also tended to produce groups with approximately the same number of observations, we believe that the obtained results are more dependent on the amount of information within each variable than the number of variables. Although increasing the number of variables, the method of Ward will group the clusters that have fewer observations. Pooled estimates
The estimates for all the traits for cluster 2 are higher than cluster 1 with exception of MBW. For MBW, © 2013 Blackwell Verlag GmbH
• J. Anim. Breed. Genet. 131 (2014) 217–226
I. D. P. S. Diaz et al.
Meta-analyses for feed intake traits
cluster 1 had most of the estimates from the United States and therefore tended to be less variable and thus obtaining a lower pooled estimate. For DMI, cluster 1 was the group that had more animals fed with concentrate and for FCR, cluster 1 had a greater number of European purebreds while group 2 included crossbred animals, which may partially explain the similarity between estimates. The results reported here could enable researchers to use more appropriate and robust heritability estimates for genetic evaluation. These results provide more specific estimates with particular features compared with single generalized pooled estimates. If a pooled estimate was required for RFI from only male animals with a specific range in age at start of the test, these results from cluster 1 (Figure 1 and Table 2), could be used to obtain more optimal genetic (co)variance matrices for genetic evaluation. The pooled estimates for genetic correlations (Table 3) are consistent with studies that evaluated the same traits (Okanishi et al. 2008; and Crowley et al. 2011) with the exception of the genetic correlation between RFI and MBW (0.25). This is higher than some reported in the literature, which were close to zero. Typically, (i.e. in these studies) phenotypic RFI is calculated for animals with valid phenotypes for at least DMI, ADG and MBW, which through residualization does ensure phenotypic independence of RFI with the component regressors of ADG and MBW. However, this phenotypic independence of RFI with traits used to estimate its expected value given differing levels of body size and growth rate does not ensure genetic independence (e.g. Kennedy et al. 1993). Some authors (Kennedy et al. 1993; Crews, 2005; Pendley 2010) comment on methods to compute genetic RFI, which result in breeding values that are independent from those for ADG and MBW breeding values. Feed efficiency traits have gained emphasis in the last few years mainly because the objectives of production systems have been changed with increasing feed costs. The number of publications involving these traits has grown over the last few years but it is still considered relatively small with each study based on a relatively small number of observations. The metaanalysis approach in this study was able to gather all the information present in all studies found to obtain more accurate pooled estimates. Conclusions The cluster analysis approach allowed the formation of groups of results of similar studies, thus providing © 2013 Blackwell Verlag GmbH
• J. Anim. Breed. Genet. 131 (2014) 217–226
pooled estimates of heritability for specific groups, which can be used as a source for comparison of pooled estimates. These pooled genetic parameter estimates could be used to solve genetic prediction equations in populations where data are insufficient for variance components estimation. Future studies that are not suitable for within-population parameter estimation for use in genetic prediction and(or) the development of genetic improvement programmes could find the pooled estimates obtained from this metaanalysis useful to parameterize both genetic and residual dispersion matrices in mixed model equations. References Archer J.A., Barwick S.A. (1999) Economic analysis of net feed intake in industry breeding schemes. In: Association for the Advancement of Animal Breeding and Genetics, 13, Bunbury. Proceedings. Association for the Advance, Bunbury. Archer J.A., Arthur P.F., Herd R.M., Parnell P.F., Pitchford W.S. (1997) Optimum post weaning test for measurement of growth rate, feed intake and feed efficiency in British breed cattle. J. Anim. Sci., 75, 2024–2032. Arthur J.P.F., Herd R.M. (2008) Residual Feed intake in beef cattle. Rev. Bras. de Zootec., 37, 269–279. BIF (BEEF IMPROVEMENT FEDERATION). (2010) Guidelines for uniform beef improvement programs. In: W.D. Hohenboken (ed.). North Carolina Cooperative Extension Service, North Carolina State University, Raleigh, NC (available at: http://www.beefimprovement.org/ PDFs/guidelines/Guidelines-9th-Edition.pdf; accessed 12 March 2013). Crews D.H. Jr (2005a) Genetics of efficient feed utilization and national cattle evaluation: a review. Genet. Mol. Res., 4, 152–165. Crews D.H., Carstens G.E. (2012) Measuring Individual Feed Intake and Utilization in Growing Cattle. In: R.A. Hill (ed), Feed Efficiency in the Beef Industry. Wiley-Blackwell, Oxford, UK. Crews D.H. Jr, Pendley C.T., Carstens G.C., Mendes E.D.M. (2010) Genetic characterization of feed intake and utilization in performance tested beef bulls. Breeding and Genetics: feed intake utilization. J. Anim. Sci., 88 (E-Suppl), 126–130. Crowley J.J., Evans R.D., Hugh N.M., Kenny D.A., McGee M., Crews D.H. Jr, Berry D.P. (2011) Genetic relationship between feed efficiency in growing males and beef cow performance. J. Anim. Sci., 89, 3372–3381. Diaz I.D.P.S., Araujo Neto F.R., Marques L.A.M., Oliveira H.N. (2011a) Interacß~ao gen otipo x ambiente e caracterısticas pre-desmama em animais da racßa Simental em ~es de nascimento. Pesq. Agrop. Bras., 43, duas estacßo 323–333.
225
I. D. P. S. Diaz et al.
Meta-analyses for feed intake traits
Diaz I.D.P.S., Oliveira H.N., Bezerra L.F., L^ obo R.B. (2011b) Genotype by environment interaction in Nelore cattle from five Brazilian states. Genet. and Mol. Biol., 34, 435–442. Elzo M.A., Johnson D.D., Lamb G.C., Maddock T.D., Myer R.O., Riley D.G., Hansen G.R., Wasdin J.G., Driver J.D. (2011) Additive Genetic Parameters for Postweaning Growth, Feed Intake, and Ultrasound Traits in AngusBrahman Multibreed Cattle (available at: http://www. animal.ufl.edu/extension/beef/beef_cattle_report/2011/ documents/aelzoaddgenepara.pdf; accessed 4 April 2012). Freitas J.A., Queiroz A.C., Dutra A.R., Vieira R.A., Lana R.P., Leonel F.P., Henrique D.S., Lima A.V., Souza J.C. (2006) Efici^encia de utilizacß~ao da energia metabolizavel em bovinos Nelores puros e cruzados submetidos a quatro nıveis de concentrado na racß~ao. Rev. Bras. de Zootec., 35, 894–901. Giannotti J.D., Packer I.H., Mercadante M.E.Z. (2002) Meta-analise para estimativas de correlacß~ao genetica entre pesos ao nascer e desmama de bovinos. Scient. Agrıc., 59, 435–440. Giannotti J.D., Packer I.H., Mercadante M.E.Z., Lima C.G. (2005) Analise de Agrupamento para Implementacß~ao da Meta-Analise em Estimativas de Herdabilidade para Caracterısticas de Crescimento em Bovinos de Corte. Rev. Bras. de Zootec., 34, 1165–1172. Hedges L.V., Olkin I. (1985) Statistical Methods for MetaAnalysis. Academic Press, London, p. 369. Higgins J.P.T., Thompson S.G., Deeks J.J., Altman D.G. (2003) Measuring inconsistency in meta-analyses. BMJ, 327, 557–560. Hoque M.A., Hosono M., Oikawa T., Suzuki K. (2009) Genetic parameters for measures of energetic efficiency of bulls and their relationships with carcass traits of field progeny in Japanese Black cattle. J. Anim. Sci., 87, 99– 106. Kennedy B.W., Van der Werf J.H., Meuwissen T.H. (1993) Genetic and statistical properties of residual feed intake. J. Anim. Sci., 71, 3239–3250. Koots K.R., Gibson J.P., Smith C., Wilton J.W. (1994) Analyses of published genetic parameter estimates for beef production traits. 2. Phenotypic and genetic correlations. Anim. Breed. Abstr., 62, 826–853. Li Z. (1995) A multiplicative random effects model for meta-analysis with application to estimation of admixture component. Biometrics., 51, 864–873. Li Z., Begg C.B. (1994) Random effects for combining results from controlled and uncontrolled studies in
226
meta-analysis. J. American Stat. Assoc., 89, 1523– 1527. National Research Council (2000) Nutrients of Requirements for Beef Cattle, 7th edicß~ao: Update 2000. National Academic Press, Washington, DC, p. 249. Okanishi T., Shojo M., Katsuta T., Oyama K., Mukai F. (2008) Genetic analysis of residual feed intake and other performance test traits of Japanese Black cattle from revised protocol. J. Anim. Sci., 79, 291–296. Pegolo N.T., Albuquerque L.G., L^ obo R.B., Oliveira H.N. (2011) Effects of sex and age on genotype by environment interaction for beef cattle body weight studied using reaction norm models. J. Anim. Sci., 89, 3410– 3425. Pendley C.T. (2010) Genetic Parameters Estimation and dry Matter Intake Calculation as it Applies to Feed Utilization in Beef Cattle. MS thesis, Department of Animal Sciences, Colorado State University. Schenkel F.S., Miller S.P., Wilton J.W. (2004) Genetic parameters and breed differences for feed efficiency, growth, and body composition traits of young beef bulls. Can. J. Anim. Sci., 84, 177–185. Sharma S. (1996) Applied Multivariate Techniquies. John Wiley, NewYork, p. 493. StatSoft, Inc. (2004) STATISTICA (data analysis software system), version 7.www.statsoft.com. Wang M.C., Bushman B.J. (1999) Integration Results: Through Meta-Analytic Review Using SAS Software. SAS Institute, Cary, p. 400. Wang Z., Nkrumah J.D., Li C., Basarab J.A., Goonewardene L.A., Okine E.K., Crews D.H. Jr, Moore S.S. (2006) Test duration for growth, feed intake, and feed efficiency in beef cattle using the GrowSafe System. J. Anim. Sci., 84, 2289–2298.
Supporting Information Additional Supporting Information may be found in the online version of this article: Appendix S1. Studies used on meta-analysis; heritability estimates for residual feed intake (RFI), dry matter intake (DMI), metabolic body weight (MBW), feed conversion ratio (FCR) and average daily gain (ADG); number of genetic correlation (rg) estimates; number of animals (N); sex, country and breed of the animals evaluated, age at start and days on test and method of evaluation (1 – multiple-trait; 2 – univariate and 3 – bivariate).
© 2013 Blackwell Verlag GmbH
• J. Anim. Breed. Genet. 131 (2014) 217–226
ORIGINAL ARTICLE
Cluster and meta-analyses of genetic parameters for feed intake traits in growing beef cattle I.D.P.S. Diaz1,2, D.H. Crews Jr2 & R.M. Enns2 1 Department of Animal Sciences, State University of Sao Paulo, Sao Paulo, Brazil 2 Department of Animal Sciences, Colorado State University, Fort Collins, CO, USA
Summary
Keywords Genetic parameters; multivariate analysis; pooled estimates. Correspondence I.D.P.S. Diaz, College of Agricultural Science and Veterinary, Via de Acesso Prof. Paulo Donatto Castellane, s/n, S~ao Paulo, Brazil. Tel: +55 11 98273 06 48; Fax: +55 34 3822 0753; E-mail: [email protected] Received: 11 January 2013; accepted: 9 October 2013
A data set based on 50 studies including feed intake and utilization traits was used to perform a meta-analysis to obtain pooled estimates using the variance between studies of genetic parameters for average daily gain (ADG); residual feed intake (RFI); metabolic body weight (MBW); feed conversion ratio (FCR); and daily dry matter intake (DMI) in beef cattle. The total data set included 128 heritability and 122 genetic correlation estimates published in the literature from 1961 to 2012. The meta-analysis was performed using a random effects model where the restricted maximum likelihood estimator was used to evaluate variances among clusters. Also, a meta-analysis using the method of cluster analysis was used to group the heritability estimates. Two clusters were obtained for each trait by different variables. It was observed, for all traits, that the heterogeneity of variance was significant between clusters and studies for genetic correlation estimates. The pooled estimates, adding the variance between clusters, for direct heritability estimates for ADG, DMI, RFI, MBW and FCR were 0.32 0.04, 0.39 0.03, 0.31 0.02, 0.31 0.03 and 0.26 0.03, respectively. Pooled genetic correlation estimates ranged from 0.15 to 0.67 among ADG, DMI, RFI, MBW and FCR. These pooled estimates of genetic parameters could be used to solve genetic prediction equations in populations where data is insufficient for variance component estimation. Cluster analysis is recommended as a statistical procedure to combine results from different studies to account for heterogeneity.
Introduction Differences among animals in conversion of feed into body weight are important determinants of profit in a beef production system. Crews (2005) pointed out that more attention must be given to the cost inputs of beef production and how to reduce them. A trait, such as feed consumption, is a good example, because this typically is considered the greatest (>65%) variable cost in a beef production system. In this context, feed efficiency or utilization has been evaluated using numerous phenotypes, as described in detail by Archer & Barwick (1999). These ‘efficiency’ measures can be grouped into gross efficiency, partial efficiency © 2013 Blackwell Verlag GmbH
• J. Anim. Breed. Genet. 131 (2014) 217–226
of growth, maintenance efficiency, cow/calf efficiency and residual feed intake (RFI), among others, with the authors noting that more than 20 different measures have been described to some degree in the literature. Considering the importance of feed utilization, many studies have estimated genetic parameters for different measures of feed efficiency but most have done so with limited amounts of data. The relatively small amount of data found in the literature is due to the expense associated with measuring individual animal feed intake. Yet with the volume of studies, considerable variability between values of the parameters has been observed. Elzo et al. (2011) reported a 0.19 doi:10.1111/jbg.12063
I. D. P. S. Diaz et al.
Meta-analyses for feed intake traits
estimate for direct heritability of residual feed intake, whereas Hoque et al. (2009) reported a value of 0.49. Arthur & Herd (2008) attributed this apparent heterogeneity at least partially to the fact that all the feed intake measures used in beef cattle involve a series of complex of biological processes and interactions with the environment. In order to be able to compare performance across time and location, as required for large scale genetic prediction, it is important to attempt to control for environmental factors that affect feed intake and its utilization. These factors include age at start of test, sex of cattle, diet composition (e.g. energy density) and testing procedures. Results from recent studies on the duration of test and frequency of measurement (Archer et al. 1997; Wang et al. 2006) have formed the basis for developing standard test protocols (BIF 2010). Given the complexity of feed intake and utilization as well as sparse data resources, a meta-analysis of the published records could provide a more complete understanding of the genetic and environmental influences on these traits. To obtain a mean value that represents and summarizes the results from different studies and to overcome the heterogeneity between studies, a number of factors should be considered in the analyses such as the variability (variance) among the studies (Li & Begg 1994 and Li 1995). Meta-analysis has been used to combine results from different studies to increase the sample size (Koots et al. 1994; Giannotti et al. 2002, 2005). The aim of this study was to utilize a meta-analysis approach to combine published heritability and genetic correlation estimates among dry matter intake (DMI), average daily gain (ADG), metabolic body weight (MBW), feed conversion ratio (FCR) and residual feed intake (RFI) to produce genetic parameters useful for genetic evaluation in the absence of sufficient data to estimate these parameters.
Materials and methods Data
The data used in this study included 128 estimates of ^2 ) for dry matter intake (DMI), average heritability (h ij daily gain (ADG), feed conversion ratio (FCR), metabolic body weight (MBW) and residual feed intake (RFI) (i.e. traits commonly measured in bull performance tests) and 122 genetic correlations among those traits in a beef cattle system. The data were obtained from 50 studies published from 1961 to 2012 (see Appendix S1). The research represented original papers, theses, dissertations and conference proceed218
ings. The analyses in this study were divided into two parts: the first part was to perform a cluster analysis using all the heritability estimates on each trait and the second involved the meta-analyses to obtain a pooled estimate using the variance between studies. We chose not to perform the cluster analysis on genetic correlation ^rg estimates because of the total amount of values found, because the correlation between each pair of trait adds a very small amount of estimates and therefore impracticable to form clusters to each correlation. Cluster analysis
^2 ) from different studies for Heritability estimates (i) (h ij each trait were divided into clusters (j) within each trait by a cluster analysis approach. This analysis was chosen to gather within each cluster the most similar ^2 relying on information about the effects of fixed h ij variables of sex of the animals, breed, country of origin, method of analysis, age at start of the test, dura^2 tion of the test, sample size, type of feeding and the h ij from each study of each trait. To form the clusters, the hierarchical Ward method was chosen, where the main feature was to form groups that were homogeneous as possible (Sharma1996). In this approach, the distance between two clusters is defined as the sums of squares between the two clusters performed on all the variables. In each stage of the procedure of clustering, the internal (within cluster) sum of squares is minimized on all the partitions that could be obtained from the combination of two clusters from the previous stage. Euclidean distance was used as the similarity coefficient, which is considered the geometrical distance in a m-multidimensional space. Thisffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi distance was calculated using dAB ¼ ðx1A x1B Þ2 þ ðx2A x2B Þ2 þ . . . þ ðxmA xmB Þ2 , where A and B are the clusters and xij is the ith variable from the jth cluster. In our study, we choose the number of clusters by the evaluation of the tree diagram. The cluster analysis was conducted using STATISTICA 7.0 software (StatSoft, Inc 2004). The variables used in the method of clustering were transformed in unique labels (dummy variables) for categorical fixed effects (method of analysis, country, sex and breed). The resulting variables were as follows: X1 where a value of 1 was assigned if the study used purebred animals, and 0 if not; X2 where a value of 1 was assigned if location of the study were the United States and 0 if not; X3 where values of 1 for males, 2 for females and 3 for both were assigned based on animals in the study; and X4 where values of 1 were assigned studies implementing multiple-trait © 2013 Blackwell Verlag GmbH
• J. Anim. Breed. Genet. 131 (2014) 217–226
I. D. P. S. Diaz et al.
Meta-analyses for feed intake traits
analyses, 2 for bivariate analyses and 3 for univariate analyses. For the quantitative variables (number of observations – sample size, age at start of test, test ^2 estimate), a standardization was duration and the h ij performed. The variables were ‘standardized’ using the expression z = Xi min (x)/( max (x) min (x)) ^2 ; min where Xi represents the i-th value or the i-th h ij 2 ^ ; and max(x) (x) is the smallest value or the smallest h ij ^2 . is the largest value or the largest h ij After clustering the heritability estimates, the metaanalysis was performed.
One of the main questions in a meta-analysis approach is to verify the homogeneity between studies. In this study, because the cluster analysis was performed previously, the homogeneity was tested between clusters and within clusters for each evaluated trait. The homogeneity was tested ^2 ¼ h ^2 ¼ . . . considering the null hypothesis of H0 : h 1þ 2þ 2 ^2 is the ^ where g is the number of clusters and h ¼h gþ iþ pooled estimate of the i-th cluster versus the alterna^2 estimates is tive hypothesis that at least one of the h iþ different from another. The test statistic was assumed to be distributed as chi-square with k – 1 df where g P ^2 h ^ 2 Þ2 , (Cochran, 1954): Qbetween ¼ wiþ ðh iþ þþ i¼1
Meta-analysis
When studies to be combined have different designs and have similar result structures and they provide estimates as the average differences between treatment groups and control groups, correlation coefficients and heritability estimates, these studies could be synthesized using a meta-analysis approach if there is information about the variance of the estimate and the sample size (N) used. (Hedges & Olkin 1985). For those studies that do not provide the variance, we cal^2 =N, culated with the following equations: ^s2ij ¼ 32h i 2 ^ calculated with animal or sire model when the hipwas ffiffiffiffi ^2 was calculated by regression and ^s2ij ¼ 2= N when h i of progeny on sire or regression progeny on dam (Koots et al. 1994). The meta-analysis of heritability estimates was performed in a stepwise manner after the cluster analysis. The first step involved the exploration of the data structure; the second was to evaluate the homogeneity of variance between clusters and within clusters which if the heterogeneity of variance was significant then the third step was performed where the variance between clusters was calculated; and the following last step was to obtain the pooled estimate. This variance was calculated as a way to include the variance between studies (Li & Begg 1994; Li 1995) and to overcome the problem of lack of power of the Q test (Higgins et al. 2003). Different box-plot graphics were constructed for each cluster within each trait to further explore the data. The mean, the standard error and the likely outliers were utilized to observe whether the different clusters are comparable with each other (in this case between clusters within each trait). Because a meta-analysis would only be valid if the normality and independence of the estimates to be combined are achieved (Hedges & Olkin 1985), we performed the Shapiro–Wilk test (StatSoft, Inc 2004) to quantify the normality of the estimates. © 2013 Blackwell Verlag GmbH
• J. Anim. Breed. Genet. 131 (2014) 217–226
wiþ ¼
mi P
^2 within each wij ; j is the number of h
j¼1
cluster and wij ¼ 1=^s2ij is the individual estimate mi mi ^2 ¼ P wij h ^2 ^2 = P wij ; and h weighting factor; h iþ ij iþþ j¼1
¼
g P i¼1
^2 = wiþ h iþ
g P i¼1
j¼1
wiþ , where ^s2i is the variance of each
^2 is the pooled cluster, wij is the inverse variance, h iþ 2 ^ heritability for each i cluster and hþþ is the pooled estimate between all groups for each trait. The null hypothesis was rejected considering the level of significance p, if Qbetween exceeded the critical value of 100(1– p)% of chi-square distribution with g–1 degrees of freedom. To test the homogeneity within clusters, the null hypothesis was Ho : h2i1 ¼ h2i2 ¼ . . . ¼ h2imiþ ¼ h2iþ where i = 1,2,….g where g is the total number of estimates of each cluster. The statistics used was: k P mi P ^2 h ^2 Þ2 , where k=m1+m2+…+ wiþ ðh Qwithin ¼ i¼1 j¼1
ij
iþ
mg, with a chi-square distribution with k–g degrees of freedom (Wang & Bushman 1999). These tests were made using macros Within & Wavgz (Wang & Bushman 1999) in SAS. When the presence of heterogeneity between clusters was observed, a random model was adopted in the meta-analysis procedure to account for variability between and within clusters. A fixed model was fitted when there was no significant heterogeneity between clusters, but there was variability only within clusters. In the random model, ^2 are different and could we considered that the h iþ ^2 þ ei þ ei ^ where be described as h2þþ ¼ h iþ ^2 represents the pooled esti^2 Nð ^2 ; s2 þ s2 Þ, h ~ h h þþ iþ ij þþ mate, where ei eNð0; ^s2ij Þ is the error related to each cluster and ei eNð0; s2 Þ is the error between clusters and s2 is the variance between clusters. For correlation estimates, the same interpretation and model ^2 . could be used placing ^rg in place of h iþ 219
I. D. P. S. Diaz et al.
Meta-analyses for feed intake traits
The s2 between clusters was calculated by REML method using macros and programmes published by Wang & Bushman (1999) developed in SAS. In this method, the following equation was used: k k k 2 P ^2 Þ2 s2 = P w 2 . ^ h ^s2 ¼ wi2 k1 ðh i þ i i i¼1
i¼1
After s estimation, the pooled heritability estimates for each trait were obtained by the following k mi k mi ^2 ¼ P P w h ^2 P P w where equation: h þ ij ij = ij 2
i¼1 j¼1
Koots et al. (1994) as well, where the ^rg values were transformed to the z quantity, where the zi value are considered to have a normal distribution with variance equal to varðzi Þ ¼ 1=ðni 3Þ, and ni is the number of observations of the i-th study. So, P zi =varðzi Þ2 ð1þrg Þ zi ¼ 0:5 ln ð1rg Þ and z pooled ¼ P 1=varðz Þ2 ; therefore, i
the
pooled
rg pooled ¼
estimate
ðe2x z pooled 1Þ . ðe2x z pooled þ1Þ
was
calculated
by:
In this case to combine the ^rg
values, the variances of z were used.
i¼1 j¼1
wij ¼ 1=ðs2ij þ ^s2 Þ and the pooled standard deviation sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi k P mi P ð1=wij Þ (Koots et al. was estimated as sþ ¼ 1=
Cluster discrimination
1994). For the genetic correlations among all the studies, the pooled values were obtained using the method of
When evaluating the tree diagram (Figure 1), it was possible to separate the estimates in two main clusters. Thus, for each of the five traits examined, two
i¼1 j¼1
Results
(b)
(a)
(d)
(c)
(e)
Figure 1 Tree diagram for 8 variables using Ward method by Euclidean distances considering sex of the animals; country of origin; method of evaluation; breed of the cattle; age at start of the test; number of observations (N) for (a) ADG, (b) DMI, (c) FCR, (d) RFI and (e) – MBW traits.
220
© 2013 Blackwell Verlag GmbH
• J. Anim. Breed. Genet. 131 (2014) 217–226
I. D. P. S. Diaz et al.
2.0 1.5 1.0 0.5 0.0 –0.5 –1.0 –1.5 –2.0
Meta-analyses for feed intake traits
(a)
Cluster 1 Cluster 2
2.0
(b)
Cluster 1 Cluster 2
1.5 1.0 0.5 0.0 –0.5 –1.0
h2ADG
Country Method Days_on_test N Age_start Sex Breed
Cluster 1 Cluster 2
2.0 (c) 1.5 1.0 0.5 0.0 –0.5 –1.0 –1.5 h2FCR Sex
Country Method Days_on_test Breed Age_start N 1.5 1.0
–1.5
2.0 1.5 1.0 0.5 0.0 –0.5 –1.0 –1.5 –2.0
h2DMI
Country Sex
Cluster 1 Cluster 2
(d)
h2RR
Method Days_on_test Breed Age_start N
Sex
Country Method Days_on_test Age_start Breed N
Cluster 1 Cluster 2
(e)
0.5 0.0 –0.5 –1.0 –1.5 –2.0 h2MBW Sex
Country Method Days_on_test N Age_start Breed
Figure 2 Groups of heritability estimates obtained considering sex of the animals; country of origin; method of evaluation; breed of the cattle; age at start of the test; number of observations (N) for (a) ADG, (b) DMI, (c) FCR, (d) RFI and (e) MBW traits.
distinct clusters were obtained (Figure 2); however, for each trait, the key variable responsible for the separation of clusters was different. The separation was based on the deviation from the mean where more discrepant values were those that deviate more from the centre of the graph. In this study, the variable that resulted in the greatest distance between the clusters was labelled the ‘key’ variable. For example, for ADG, the key variable was sex of the animals. Sex was also the key variable for RFI, which is not surprising given that ADG is included in the calculation of RFI. The sex of the animals was also the key variable followed by the variable country of origin for MBW. For DMI, the key variable was the group heritability estimates from each study. Considering FCR, the breed of the cattle was the key variable of the separation of the groups, © 2013 Blackwell Verlag GmbH
• J. Anim. Breed. Genet. 131 (2014) 217–226
followed by the method of evaluation and age at the start of the test. Exploratory analysis
The box-plots graphics (Figure 3) showed outliers and extreme values of heritability for all the traits but mostly for DMI for which the cluster analysis was able to separate estimates in two very distinct clusters corroborating the results in Figure 2. The evaluation of homogeneity between clusters and within clusters (Table 1), and between studies considering the genetic correlation estimates, was significant by the Q test (p < 0.01). It was observed that the Q values were high and significant. The lowest values for the Q statistics were found for the lowest values of degrees of freedom but could also have been 221
I. D. P. S. Diaz et al.
Meta-analyses for feed intake traits
(a)
(b)
(c)
(d)
(e)
Figure 3 Pooled heritability estimates and standard deviation for the traits: (a) average daily gain (ADG), (b) dry matter intake (DMI), (c) feed conversion ratio (FCR), (d) residual feed intake (RFI) and (e) metabolic body weight (MBW).
due to a lack of variability between the estimates of those clusters. The heterogeneity observed between studies suggests the implementation of the random model to obtain the pooled estimates. Pooled estimates
The pooled estimates of heritability were obtained for each cluster (Table 2) within each trait and for the clusters combined as well. The same approach was used for the genetic correlation estimates (Table 3) between traits. Heritability estimates were transformed to obtain the pooled estimates and their respective standard deviations, but the results are presented on the original scale. As expected, the values of the estimates differ between clusters for all traits indicating that the clus222
tering procedure is capturing differences among the study values, although all heritability estimates could be classified as moderate. Despite the small number of genetic correlation estimates obtained from different studies, the pooled estimates values are in the range expected by the authors (0.15–0.67). Most estimates were positive and with high magnitude with the exception of the genetic correlation between RFI and ADG (0.04) and ADG and FCR (0.15).
Discussion Cluster discrimination
The different values found between the heritability estimates for RFI, ADG and MBW due to the sex of the animals (Figure 2) could be related to the higher © 2013 Blackwell Verlag GmbH
• J. Anim. Breed. Genet. 131 (2014) 217–226
I. D. P. S. Diaz et al.
Meta-analyses for feed intake traits
Table 1 Q statistic to test the homogeneity between studies (clusters) and within cluster to genetic correlation and heritability estimates for average daily gain (ADG); dry matter intake (DMI); residual feed intake (RFI); metabolic body weight (MBW) and feed conversion ratio (FCR) Correlation estimates ADG,DMI
ADG,RFI
ADG,MBW
ADG,FCR
DMI,RFI
Causes of heterogeneity
D.F.
Q
D.F.
Q
D.F.
Q
D.F.
Q
D.F.
Q
Between studies
11
655.29*
16
327.01*
13
528.06*
15
727.88*
18
825.16*
DMI,MBW
DMI,FCR
RFI,MBW
RFI,FCR
MBW,FCR
Causes of heterogeneity
D.F.
Q
D.F.
Q
D.F.
Q
D.F
Q
D.F.
Q
Between studies
9
86.79*
25
976.25*
13
499.49*
20
906.05*
9
83.55*
Heritability estimates ADG
DMI
RFI
MBW
FCR
Causes of heterogeneity
D.F.
Q
D.F.
Q
D.F.
Q
D.F.
Q
D.F.
Q
Between clusters Within clusters Within cluster 1 Within cluster 2
1 30 11 19
18.63* 300.54* 71.76* 228.78*
1 27 19 8
32.78* 690.97* 414.88* 276.09*
1 44 23 21
10.40* 688.6* 393.54* 295.42*
1 29 20 9
28.89* 457.41* 392.70* 64.71*
1 47 36 11
16.06* 804.13* 732.80* 71.33*
*Significant by Q test, p > 0.01. Table 2 Pooled heritability estimates and the standard deviation obtained using REML for each cluster and for the clusters together Traits
Group
RFI h2*
MBW h2*
DMI h2*
ADG h2*
FCR h2*
1 2 Total
0.33 (0.04) 0.29 (0.02) 0.31 (0.02)
0.25 (0.02) 0.38 (0.05) 0.31 (0.03)
0.46 (0.03) 0.33 (0.03) 0.39 (0.03)
0.38 (0.05) 0.27 (0.03) 0.32 (0.04)
0.28 (0.02) 0.24 (0.04) 0.26 (0.03)
Table 3 Pooled heritability (diagonal) and genetic correlation estimates (off-diagonal) obtained by restricted maximum likelihood method Traits
RFI MBW DMI ADG FCR
RFI
MBW
DMI
ADG
FCR
0.31 (0.02)
0.25 (0.09) 0.31 (0.03)
0.67 (0.12) 0.36 (0.04) 0.39 (0.03)
0.04 (0.08) 0.33 (0.06) 0.38 (0.11) 0.32 (0.04)
0.65 (0.10) 0.24 (0.04) 0.34 (0.13) 0.15 (0.12) 0.26 (0.03)
average daily gain of males compared with females. Correspondingly, males’ genetic potential is higher and consequently so is the value of heritability from studies using only male data. According to Diaz et al. (2011a), the effect of sex on traits has been reported before, and it is a subject of discussion. Males and females may have different selection/ production objectives and therefore likely receive dif© 2013 Blackwell Verlag GmbH
• J. Anim. Breed. Genet. 131 (2014) 217–226
ferent treatment on farms with resulting environmental differences. Seemingly unimportant, because the selection takes place within sex, this difference in genetic parameters between males and females may gain importance as the genetic evaluation of animals is often performed considering additive genetic effects similar in both sexes. The observed difference found in this study between weights of males and females is consistent with other studies involving various beef cattle breeds (Diaz et al. 2011b; P egolo et al. 2011). The extreme points observed when the variable of ‘country of origin’ is evaluated for MBW, differences can be due to the fact that this study compared United States to all other countries. This was made because in most United States studies, MBW and ADG are computed using regression procedures. In this context, it is important to note that although these traits (FCR, MBW, RFI, DMI and ADG) are measured in controlled conditions and environments, the non-standardization of the tests across the central stations can increase the difference of the estimates in the same traits. According to Crews & Carstens (2012), this can also be due to the fact that serial weighing during test is preferred but there are still many test centres that record live weights only at the beginning and end of tests. As for intake, not all test stations use the same hardware/software for collection and (or) auditing of that data, although standard recommendations in 223
I. D. P. S. Diaz et al.
Meta-analyses for feed intake traits
North America do accommodate some of these hardware and software differences. Results from recent studies on the duration of test and frequency of measurement (Archer & Barwick 1999; Wang et al. 2006) have formed the basis for developing standard test protocols. The difference observed between clusters for DMI compared with other traits is the increased variation in heritability estimates. This trait had the largest number of estimates found in the literature (36 estimates). Dry matter intake variance is highly affected by the environment, mainly by different types of feed (National Research Council 2000) so most of these differences among heritability estimates could be related to different environmental variances. Another point that has to be considered is that these differences of heritability values, mainly for DMI, could be to differences on genetic variation among populations. Crossbred populations are more likely to have dominance effects on genetic variation; some studies may use animals with higher relationships than typical; however, these differences were higher and significant for DMI (Figure 2). Currently, some standard testing protocols further adjust DMI records to constant energy densities, but the majority of differences reported here among clusters would be as least partially accounted for by inclusion of appropriate contemporary group classification of DMI records (BIF 2010). It is well known that consumption of less-digestible, low-energy, high-fibre diets is affected by rumen fill and the feed passage rate through the animal. Meanwhile, consumption of highly digestible, high-energy, low-fibre feeds is controlled by the animal’s energy needs and by metabolic factors. These concepts may seem quite simple, but the factors that regulate DMI on concentrate rations are complex and are not fully understood. Because we are evaluating animals that had also pasture as a type of feed, we suggest that the type of feed offered, especially more palatable diets are more acceptable and therefore are more efficiently used by the animals than other types of feed. For FCR, the observed differences between clusters are primarily explained by the composition of the breed. Animals that have a higher percentage of Bos indicus breed composition had different conversion ratios when compared to predominantly Bos taurus animals. Schenkel et al. (2004) reported differences in FCR even among taurine breeds, where Simmental and Angus cattle had a higher FCR compared with Hereford. These are likely results in the differences among heritability estimates. 224
The method of evaluation (multiple-trait; univariate or bivariate) was also shown to have great importance in the analysis of feed efficiency traits. Studies have shown failure of convergence when all traits are evaluated together (Crews et al. 2010) such as when performing a multiple-trait analysis of RFI, ADG and MBW. These authors point out that this is a result of the interrelationships among traits such as when calculating RFI where DMI records are required. Age at start of the test also was significant in determining clusters for FCR. In this context, younger animals tend to have better feed efficiency (more efficient) than older animals, in other words, means that depends more on growth stage and diet of the animal. It takes more energy to gain weight through fat deposition than muscle and older animals on finishing rations deposit more fat. According to Freitas et al. (2006) older animals (with a greater degree of physiological maturity) have a greater efficiency gain for the synthesis of fat than younger animals, which have higher efficiency of protein gain and protein synthesis, so the feed conversion estimates become different depending on the age at which these animals are tested. Exploratory analysis
The observed outliers in the box-plot graphics (Figure 3) were included in this research, because all the authors explained those values in their studies. This procedure was also adopted in other meta-analysis studies (Giannotti et al. 2002, 2005). When the homogeneity of variance was tested between clusters, after using the Ward clustering method, it was expected that there would be heterogeneity among clusters (Table 1); however, it was not expected that there would be heterogeneity within clusters, because the aim of this method was to obtain the most homogeneous values within each cluster. Once the Ward clustering method tend to combine clusters with a small amount of information and also tended to produce groups with approximately the same number of observations, we believe that the obtained results are more dependent on the amount of information within each variable than the number of variables. Although increasing the number of variables, the method of Ward will group the clusters that have fewer observations. Pooled estimates
The estimates for all the traits for cluster 2 are higher than cluster 1 with exception of MBW. For MBW, © 2013 Blackwell Verlag GmbH
• J. Anim. Breed. Genet. 131 (2014) 217–226
I. D. P. S. Diaz et al.
Meta-analyses for feed intake traits
cluster 1 had most of the estimates from the United States and therefore tended to be less variable and thus obtaining a lower pooled estimate. For DMI, cluster 1 was the group that had more animals fed with concentrate and for FCR, cluster 1 had a greater number of European purebreds while group 2 included crossbred animals, which may partially explain the similarity between estimates. The results reported here could enable researchers to use more appropriate and robust heritability estimates for genetic evaluation. These results provide more specific estimates with particular features compared with single generalized pooled estimates. If a pooled estimate was required for RFI from only male animals with a specific range in age at start of the test, these results from cluster 1 (Figure 1 and Table 2), could be used to obtain more optimal genetic (co)variance matrices for genetic evaluation. The pooled estimates for genetic correlations (Table 3) are consistent with studies that evaluated the same traits (Okanishi et al. 2008; and Crowley et al. 2011) with the exception of the genetic correlation between RFI and MBW (0.25). This is higher than some reported in the literature, which were close to zero. Typically, (i.e. in these studies) phenotypic RFI is calculated for animals with valid phenotypes for at least DMI, ADG and MBW, which through residualization does ensure phenotypic independence of RFI with the component regressors of ADG and MBW. However, this phenotypic independence of RFI with traits used to estimate its expected value given differing levels of body size and growth rate does not ensure genetic independence (e.g. Kennedy et al. 1993). Some authors (Kennedy et al. 1993; Crews, 2005; Pendley 2010) comment on methods to compute genetic RFI, which result in breeding values that are independent from those for ADG and MBW breeding values. Feed efficiency traits have gained emphasis in the last few years mainly because the objectives of production systems have been changed with increasing feed costs. The number of publications involving these traits has grown over the last few years but it is still considered relatively small with each study based on a relatively small number of observations. The metaanalysis approach in this study was able to gather all the information present in all studies found to obtain more accurate pooled estimates. Conclusions The cluster analysis approach allowed the formation of groups of results of similar studies, thus providing © 2013 Blackwell Verlag GmbH
• J. Anim. Breed. Genet. 131 (2014) 217–226
pooled estimates of heritability for specific groups, which can be used as a source for comparison of pooled estimates. These pooled genetic parameter estimates could be used to solve genetic prediction equations in populations where data are insufficient for variance components estimation. Future studies that are not suitable for within-population parameter estimation for use in genetic prediction and(or) the development of genetic improvement programmes could find the pooled estimates obtained from this metaanalysis useful to parameterize both genetic and residual dispersion matrices in mixed model equations. References Archer J.A., Barwick S.A. (1999) Economic analysis of net feed intake in industry breeding schemes. In: Association for the Advancement of Animal Breeding and Genetics, 13, Bunbury. Proceedings. Association for the Advance, Bunbury. Archer J.A., Arthur P.F., Herd R.M., Parnell P.F., Pitchford W.S. (1997) Optimum post weaning test for measurement of growth rate, feed intake and feed efficiency in British breed cattle. J. Anim. Sci., 75, 2024–2032. Arthur J.P.F., Herd R.M. (2008) Residual Feed intake in beef cattle. Rev. Bras. de Zootec., 37, 269–279. BIF (BEEF IMPROVEMENT FEDERATION). (2010) Guidelines for uniform beef improvement programs. In: W.D. Hohenboken (ed.). North Carolina Cooperative Extension Service, North Carolina State University, Raleigh, NC (available at: http://www.beefimprovement.org/ PDFs/guidelines/Guidelines-9th-Edition.pdf; accessed 12 March 2013). Crews D.H. Jr (2005a) Genetics of efficient feed utilization and national cattle evaluation: a review. Genet. Mol. Res., 4, 152–165. Crews D.H., Carstens G.E. (2012) Measuring Individual Feed Intake and Utilization in Growing Cattle. In: R.A. Hill (ed), Feed Efficiency in the Beef Industry. Wiley-Blackwell, Oxford, UK. Crews D.H. Jr, Pendley C.T., Carstens G.C., Mendes E.D.M. (2010) Genetic characterization of feed intake and utilization in performance tested beef bulls. Breeding and Genetics: feed intake utilization. J. Anim. Sci., 88 (E-Suppl), 126–130. Crowley J.J., Evans R.D., Hugh N.M., Kenny D.A., McGee M., Crews D.H. Jr, Berry D.P. (2011) Genetic relationship between feed efficiency in growing males and beef cow performance. J. Anim. Sci., 89, 3372–3381. Diaz I.D.P.S., Araujo Neto F.R., Marques L.A.M., Oliveira H.N. (2011a) Interacß~ao gen otipo x ambiente e caracterısticas pre-desmama em animais da racßa Simental em ~es de nascimento. Pesq. Agrop. Bras., 43, duas estacßo 323–333.
225
I. D. P. S. Diaz et al.
Meta-analyses for feed intake traits
Diaz I.D.P.S., Oliveira H.N., Bezerra L.F., L^ obo R.B. (2011b) Genotype by environment interaction in Nelore cattle from five Brazilian states. Genet. and Mol. Biol., 34, 435–442. Elzo M.A., Johnson D.D., Lamb G.C., Maddock T.D., Myer R.O., Riley D.G., Hansen G.R., Wasdin J.G., Driver J.D. (2011) Additive Genetic Parameters for Postweaning Growth, Feed Intake, and Ultrasound Traits in AngusBrahman Multibreed Cattle (available at: http://www. animal.ufl.edu/extension/beef/beef_cattle_report/2011/ documents/aelzoaddgenepara.pdf; accessed 4 April 2012). Freitas J.A., Queiroz A.C., Dutra A.R., Vieira R.A., Lana R.P., Leonel F.P., Henrique D.S., Lima A.V., Souza J.C. (2006) Efici^encia de utilizacß~ao da energia metabolizavel em bovinos Nelores puros e cruzados submetidos a quatro nıveis de concentrado na racß~ao. Rev. Bras. de Zootec., 35, 894–901. Giannotti J.D., Packer I.H., Mercadante M.E.Z. (2002) Meta-analise para estimativas de correlacß~ao genetica entre pesos ao nascer e desmama de bovinos. Scient. Agrıc., 59, 435–440. Giannotti J.D., Packer I.H., Mercadante M.E.Z., Lima C.G. (2005) Analise de Agrupamento para Implementacß~ao da Meta-Analise em Estimativas de Herdabilidade para Caracterısticas de Crescimento em Bovinos de Corte. Rev. Bras. de Zootec., 34, 1165–1172. Hedges L.V., Olkin I. (1985) Statistical Methods for MetaAnalysis. Academic Press, London, p. 369. Higgins J.P.T., Thompson S.G., Deeks J.J., Altman D.G. (2003) Measuring inconsistency in meta-analyses. BMJ, 327, 557–560. Hoque M.A., Hosono M., Oikawa T., Suzuki K. (2009) Genetic parameters for measures of energetic efficiency of bulls and their relationships with carcass traits of field progeny in Japanese Black cattle. J. Anim. Sci., 87, 99– 106. Kennedy B.W., Van der Werf J.H., Meuwissen T.H. (1993) Genetic and statistical properties of residual feed intake. J. Anim. Sci., 71, 3239–3250. Koots K.R., Gibson J.P., Smith C., Wilton J.W. (1994) Analyses of published genetic parameter estimates for beef production traits. 2. Phenotypic and genetic correlations. Anim. Breed. Abstr., 62, 826–853. Li Z. (1995) A multiplicative random effects model for meta-analysis with application to estimation of admixture component. Biometrics., 51, 864–873. Li Z., Begg C.B. (1994) Random effects for combining results from controlled and uncontrolled studies in
226
meta-analysis. J. American Stat. Assoc., 89, 1523– 1527. National Research Council (2000) Nutrients of Requirements for Beef Cattle, 7th edicß~ao: Update 2000. National Academic Press, Washington, DC, p. 249. Okanishi T., Shojo M., Katsuta T., Oyama K., Mukai F. (2008) Genetic analysis of residual feed intake and other performance test traits of Japanese Black cattle from revised protocol. J. Anim. Sci., 79, 291–296. Pegolo N.T., Albuquerque L.G., L^ obo R.B., Oliveira H.N. (2011) Effects of sex and age on genotype by environment interaction for beef cattle body weight studied using reaction norm models. J. Anim. Sci., 89, 3410– 3425. Pendley C.T. (2010) Genetic Parameters Estimation and dry Matter Intake Calculation as it Applies to Feed Utilization in Beef Cattle. MS thesis, Department of Animal Sciences, Colorado State University. Schenkel F.S., Miller S.P., Wilton J.W. (2004) Genetic parameters and breed differences for feed efficiency, growth, and body composition traits of young beef bulls. Can. J. Anim. Sci., 84, 177–185. Sharma S. (1996) Applied Multivariate Techniquies. John Wiley, NewYork, p. 493. StatSoft, Inc. (2004) STATISTICA (data analysis software system), version 7.www.statsoft.com. Wang M.C., Bushman B.J. (1999) Integration Results: Through Meta-Analytic Review Using SAS Software. SAS Institute, Cary, p. 400. Wang Z., Nkrumah J.D., Li C., Basarab J.A., Goonewardene L.A., Okine E.K., Crews D.H. Jr, Moore S.S. (2006) Test duration for growth, feed intake, and feed efficiency in beef cattle using the GrowSafe System. J. Anim. Sci., 84, 2289–2298.
Supporting Information Additional Supporting Information may be found in the online version of this article: Appendix S1. Studies used on meta-analysis; heritability estimates for residual feed intake (RFI), dry matter intake (DMI), metabolic body weight (MBW), feed conversion ratio (FCR) and average daily gain (ADG); number of genetic correlation (rg) estimates; number of animals (N); sex, country and breed of the animals evaluated, age at start and days on test and method of evaluation (1 – multiple-trait; 2 – univariate and 3 – bivariate).
© 2013 Blackwell Verlag GmbH
• J. Anim. Breed. Genet. 131 (2014) 217–226
