The accuracy of breeding values strongly depends on the
population and herd structure, i.e., the number of animals considered in
genetic evaluations and the size of contemporary groups (CGs). Local breeds
are usually kept in small-sized family farms under alternative husbandry
conditions. For such herd structure, consideration of classical herd or
herd-test-day effects in CG modeling approaches implies only a few records
per effect level. In consequence, the present study aimed on methodological
evaluations of different herd clustering strategies, considering
social–ecological and herd characteristics. In this regard, we considered 19 herds keeping cows from the small local population of German Black Pied cattle (
Local cattle breeds contribute to genetic diversity and may carry favorable
alleles with relevance for future production systems and market
requirements, justifying efforts for the implementation of preservation
programs (Ajmone-Marsan et al., 2010; Toro et al., 2011). Numerous studies
(Toro et al., 2011; Fernández et al., 2011; Biscarini et al., 2015;
Mastrangelo et al., 2016; Cervantes et al., 2016) focused on strategies to
maintain genetic variability in endangered breeds and especially focused on
the minimization of inbreeding and genetic drift. A major feature of local
endangered cattle breeds is their ability for adaptation to harsh
environments (Fernández et al., 2011; Cervantes et al., 2016). In
consequence, local cattle breeds are mostly kept in alternative production
systems, reflecting a broad pattern of challenges such as limited food
resources or climatic stress (Halli et al., 2020). In Germany, the local
dual-purpose cattle breed German Black Pied cattle (
Accurate genetic evaluations in DSN are imperative, because several German breeding organizations sell semen from DSN sires worldwide. DSN and Holstein Friesian (HF) cattle are considered simultaneously in genetic evaluations, but their genetic connectedness is quite low, implying biased estimated breeding values (EBVs) when ignoring further genetic model improvements. In such context, Jaeger et al. (2019) suggested improved genetic evaluations for DSN through a widened population size, i.e., considering DSN cows from the Netherlands and from Poland.
In addition to the small DSN population size, accuracy of selection and genetic evaluations is hampered due to the small-sized herd structures. Classically, genetic evaluation models consider a herd effect, because the herd usually represents same management, feeding or husbandry conditions for all cows from the same herd. However, genetic (co)variance components and EBVs might be biased when creating specific management groups within herds or when applying preferential treatment for specific cow groups (König et al., 2005). Accordingly, Kennedy and Trus (1993) stretched the topic of contemporary groups (CGs) in genetic evaluations, which should represent the microenvironment as detailed as possible. The influence of the herd effect when modeling CGs is not constant throughout the year, implying consideration of a further time-dependent explanatory variable, i.e., via herd-year-season or herd-test-day modeling (Emmerling, 2000). However, such modeling approaches might be problematic in the case of small herd sizes, implying only a few cow records per CG, with detrimental impact on selection accuracy (Strabel et al., 2005; Pereira et al., 2018). Furthermore, human–animal relationships reflecting social characteristics determine herd effects. Social characteristics plus classical environmental conditions (e.g., climate, feeding resources) were considered when defining social–ecological systems for livestock classifications (Martin-Collado et al., 2014). In small-sized DSN family farms, Ebinghaus (2018) identified impact of the individual farm management and human–animal interactions on variations of disease incidences and production levels across herds (Ebinghaus, 2018).
Alternative modeling strategies to create CG were introduced by Strabel et al. (2005) and Vasconcelos et al. (2008). They grouped herds according to herd size or average milk yield per herd. Grouping herds according to herd, environmental or social characteristics points to the application of herd clustering strategies. Table 1 gives an overview of the clustering methods as applied in cattle populations. Key factors to allocate herds to herd clusters (HCs) were the overall production system (conventional or organic), herd size, breed, country or region (Blanco-Penedo et al., 2019; Ivemeyer et al., 2017). Tremblay et al. (2016) recommended that clustering approaches should consider variables reflecting herd particularities.
Overview for applied herd clustering approaches.
The objective of this study was to evaluate different HC strategies including a sample of DSN herds and some HF herds, with the aim of defining appropriate CGs in genetic evaluations. Created clusters will be described based on descriptive statistics for HC variables, as well as on detailed analyses for cow traits in respective HCs using mixed model applications. In a last step, we evaluated the impact of HC modeling on genetic parameter estimates and on accuracies of genetic evaluations.
Herd data were collected via face-to-face interviews and farm
characterizations between September 2017 and March 2018, considering 19 DSN
herds, 10 HF herds and one mixed herd keeping both breeds. The 19 DSN herds
reflected “pure” DSN herds with DSN gene percentages larger than 87.5 %,
considering the algorithm for gene percentage calculations as developed by
Jaeger et al. (2018). Also the DSN cows in the mixed herd were pure DSN. The
HF herds were chosen to consider a genetically related breed, but with
opposite breeding goals, production levels and farming systems. The
participating herds were located in three major geographic regions from
Germany: (1) intensive grazing systems on coastal marshlands, (2) large-scale
farms (indoor system) in one region of former East Germany and (3) small-scale family farms in the middle of Germany (semi-intensive grazing
systems with maximal 5 h grazing per day). The altitudes of farms in the
three regions were 14.68, 84.00 and 200.47 m, respectively, and the
latitudes were 53
The farm visits for structured interviews and herd characterizations
comprised 45 to 90 min per farm. The survey and visual herd observations
included quantitative and qualitative information with regard to general
herd characteristics, the feeding regime, the housing system, the husbandry
practices, the herd and pasture management, herd fertility and health status
as well as the management of calves, heifers and dry cows. Social components
addressed, e.g., the herd manager's education, the expenditure of time used
for dairy cattle farming, the family status, the number and age of the
children and the number of farm employees. In total, the herd
characterization comprised 117 variables (26 quantitative and 91 qualitative
variables) as indicated in the Supplement (Table S1). Answers were possible via
multiple choice but also included open questions and required specific
numeric values in some cases (Supplement Table S1). Quantitative variables
were scaled by a
The previously applied HC approaches (as summarized in Table 1) focused on
one specific method. In the present study, we compared four different HC
methods, especially from the perspective of HC consideration in genetic
evaluations. The following four different HC methods were applied: (i) agglomerative hierarchical clustering (AHC), (ii) partition around medoids
(PAM), (iii) fuzzy clustering (FZC) and (iv) a clustering of variables
combined with agglomerative hierarchical clustering (CoVAHC). All clustering
analyses were conducted in R version 4.0.2 (R Core Team, 2020) and applying
the packages “cluster” (Maechler et al., 2018) and “ClustOfVar” (Chavent
et al., 2017). According to Pimenta et al. (2017), herd variables indicating
limited variation or strong correlations with other variables, were deleted.
After herd variable editing, 106 variables (23 quantitative and 83
qualitative variables) remained for the ongoing cluster analyses. Based on
the mixed data types (nominal, ordinal, (a)symmetric binary, metric), the
Gower distance (Gower, 1971) modified by Struyf et al. (1996), was used to
calculate the dissimilarity matrix. The overall average silhouette width
(ASW) as defined by Rousseeuw (1987) was used to evaluate the clustering
approaches and to identify the optimal number of HCs. The silhouette width
ranges between
We varied
Start partition: each variable is one start cluster. Two clusters will be merged to a new partition when the dissimilarity is the
smallest, so that the loss of homogeneity of the new cluster is minimal.
This merging step is repeated until each variable is grouped with another
variable or partition. End partition: all start clusters form one complete cluster.
The CoV focuses on two aspects. The first is merging closely related variables by
grouping them into partitions that maximize the homogeneity criterion, which
is defined by the sum of squared Pearson correlations for quantitative
variables and correlation ratios for qualitative variables. If all
quantitative variables and all qualitative variables in a cluster are
correlated (or anti-correlated) or the correlation ratios are equal to 1,
the homogeneity criterion is maximized. The second aspect focuses on the
definition of a synthetic variable of each cluster by a principal component
approach for mixed data. Afterwards, the values of the synthetic variables
are used via AHC to cluster the herds (Chavent et al., 2012; Brida et al.,
2014).
In this regard, due to the largest ASW, we considered the four HCs to be created
by the CoVAHC application (details are presented in Sect. 3.1). Cow traits
were from the recording years 2017 and 2018. The number of cows per HC was
as follows: HC1 of 1091 cows, HC2 of 64 cows, HC3 of 1059 cows and HC4 of 3324 cows. Production data considered 55 181 repeated test-day records from
5538 cows (19 964 records from 1947 DSN cows and 35 217 records from
3591 HF cows) for milk yield (Mkg), protein yield (Pkg), fat yield (Fkg)
protein percentage (P%), fat percentage (F%), somatic cell sore (SCS)
and fat-to-protein (FPR) ratio from the first to the third lactations. Female
fertility traits included the interval from calving to first insemination
(CFI) and the success of a first insemination (SFI). In this regard, we
considered 6100 observations for CFI from 4562 cows (2548 records form
1871 DSN cows, and 3552 records from 2691 HF cows) and 7333 first
inseminations for SFI from 5119 cows (3119 records form 2096 DSN cows,
and 4214 records from 3023 HF cows). The udder health indicator somatic
cell count (SCC) was log transformed into SCS
Linear mixed models were applied to assess the effect of defined HC on the
cow test-day traits: Mkg, Pkg, Fkg, P%, F%, SCS and FPR. All
calculations were performed with R version 4.0.2 (R Core Team, 2020) and
applied the package “emmeans” (Lenth, 2020). This package was also used
to calculate least squares means (LSMs) for traits within HCs and to test
for corresponding significant differences. The respective model (model 1) was
defined as follows:
A linear mixed model was applied to CFI, and a generalized linear mixed
model with a logit-link function was applied to SFI. Effects in the respective model (model 2)
were the same for both female fertility traits and defined as follows:
Genetic evaluations for test-day milk yield considered phenotypic data from
calving years 2012 to 2018 from the 5538 cows with 55 181 test-day
records (35 217 records from DSN and 19 964 records from HF) from the first
three lactations. The estimation of genetic parameters and breeding values
was carried out for a test-day model (model 3) with a herd-test-day
(HTD) or herd-cluster-test-day (HCTD) effect and for
an alternative test-month model (model 4) with a herd-test-month
(HTM) or herd-cluster-test-month (HCTM) effect. Again, we
considered the four HCs from the CoVAHC approach. For the genetic parameter
estimations with the DMU software package (Madsen and Jensen, 2013), the
following linear animal models were defined:
The first step was to determine the optimal number of HCs. Figure 1 shows the ASW for the four clustering methods, indicating a wide range from 0.015 (FZC with 10 HC) to 0.510 (CoVAHC with four HCs). Such huge variation for ASW displays the differences in separation efficiency of the different approaches. The ASW was highest (0.510) when creating four HCs and applying CoVAHC clustering. Such desired value for ASW for selected clustering procedures is in agreement with Gorgulu (2010), Ivemeyer et al. (2017) and Guiomar et al. (2018). For the remaining clustering approaches AHC, PAM and FZC, the ASW was quite stable in dependency of HC variations, but generally, ASW obviously declined for more than four HCs (Fig. 1).
Average silhouette width for the different numbers of herd clusters considering the following clustering approaches: agglomerative hierarchical clustering (AHC), partition around medoids (PAM), fuzzy clustering (FZC), clustering of variables combined with agglomerative hierarchical clustering (CoVAHC); dotted lines are smoothed conditional means; the vertical dotted line indicates the chosen number of herd clusters for ongoing studies.
In the Supplement (Fig. S1), ASWs for individual herds are shown. The different
herd numbers as presented on the
The evaluations of HCs in the present study focused on aspects with relevance for data recording and for genetic evaluations. The collected herd characteristics in this study represented different types of data, i.e., qualitative or quantitative. In order to overcome such obstacles, previous studies (Toro-Mujica et al., 2012; Riveiro et al., 2013; Ivemeyer et al., 2017; Blanco-Penedo et al., 2019) applied principal component analysis (PCA) to translate the categorical data structure indirectly into quantitative variables. These studies suggested a PCA due to the pronounced variation as identified among the most important principal components. To prevent possible biases through indirect transformations, Struyf et al. (1996) suggested a modified Gower distance. As a further method for handling mixed data types, Chavent et al. (2012) applied CoVAHC. They favored this approach over PCA, because more information can be taken into account when clustering the elements (herds). Furthermore, in contrast to PCA, orthogonality of the principal components is not required (Kuentz-Simonet et al., 2017).
The four HCs formed by CoVAHC, which are shown in Fig. 2, differ in multiple farm characteristics (Table 2), which in turn were used to describe the HC. The two breeds (DSN and HF) were clearly separated, meaning that HF herds only appeared in HC4. Overall, 91 % of herds from HC4 represented HF, and only 9 % were DSN herds. The percentage of DSN herds in HC1, HC2 and HC3 was 100 %. Such herd allocation based on herd characteristics indicates that the evolutionarily closely related DSN and HF breeds (Biedermann et al., 2005) are kept in different production systems representing a different herd management. The DSN are mostly kept in low input or grassland systems (Jaeger et al., 2018), but HF mostly in free-stall farms applying all available modern management instruments especially with regard to feeding strategies (e.g., feeding of total mixed rations) (König et al., 2005). Accordingly, Ivemeyer et al. (2017) clearly separated HF from local breeds with small population size such as original Angler cattle or DSN. Tremblay et al. (2016) only considered herds in automatic milking systems. Despite the same milking technology, they identified obvious differences in production pattern, feeding and management characteristics between small (Jersey, Guernsey, Ayrshire) and large populations (HF, brown Swiss).
Dendrogram of the herds merged to herd clusters by clustering of variables combined with agglomerative hierarchical clustering (CoVAHC). Each block with specific pattern represents one HC.
Percentages and values of main herd characteristics
displaying significant
It is interesting to note that both the HC1 and HC2 clusters included a mixture of conventional and organic herds, despite the substantial differences in legal regulations for both farming types. The organic farms from the present study base their feeding, breeding and management strategies on the guidelines for organic farming as defined by the European Union which are less strict than national German organic programs. Some particularities are defined in the basic principles for organic farming (IFOAM, 2020), especially addressing the breeding focus on longevity. Nevertheless, the allocation of organic as well as conventional herds to HC1 and HC2 suggests that there is an overlap of environmental conditions such as climatic impact between these two main classes (organic and conventional) affecting livestock production. Sorge et al. (2016) investigated herd management practices in organic and conventional dairy herds in Minnesota. They made similar conclusions, i.e., indicating that management decisions are diverse and herd specific, and do not strongly depend on the overall farming type organic or conventional. All of the HF herds as allocated to HC4 were conventional herds.
The majority of cows are kept in cubicle stables (HC1: 77 % of the herds, HC3: 100 % of the herds, HC4: 100 % of the herds) (Table 2). HC2 includes all herds with tie stables (16.7 % of all herds or 1.8 % of the cows), which usually have less than 20 cows. As identified for HC1, all cows in HC2 have access to pasture. In contrast, only one-third of the high-yielding herds (HC3 and HC4) reflect grazing systems. Accordingly, Müller-Lindenlauf et al. (2010) reported that herd productivity decreases with increasing length of the grazing period.
The cluster process (CoVAHC) separated the more traditional herds (herds in HC1 and HC2) from the more modern dairy herds (herds in HC3 and HC4). The level of digitization in animal housing, especially in dairy cattle farms, was defined as a major factor explaining herd and cow trait differences (Büscher, 2019). Herds allocated to HC2 did not use modern digital infrastructure (Table 2). Also, in HC1, the proportion of herds using a herd management software was comparatively low with 31 %. With regard to feeding strategies, herds from HC1 and HC2 use a very simple feed ration with only a few components, and they do not consider systematic feed analyses. In contrast, all herds in HC3 and HC4 base their management decisions on digital supporting systems. Also, the feeding rations are optimized considering scientific aspects and the needs of the cows. In total, 64 % of the herds from HC4 feed on a ration with a broad variety of ingredients.
The percentage of herds using natural service sires substantially differed among the HC (Table 2). The proportion was highest in HC1 with 77 %, followed by HC2 with 60 % but was quite low in HC4 (18 %). Herds from HC3 only considered artificial insemination. From a breeding perspective, Yin et al. (2014) identified utilization of natural service sires as a major characteristic when comparing organic with conventional farm types or DSN with HF herds.
With regard to social characteristics, mainly young farmers are responsible for the herd management in large-scale herds. Such a finding is in agreement with a comprehensive study across European dairy cattle herds (Blanco-Penedo et al., 2019). In HC4, the average age of herd managers was 37 years (Table 2). Nevertheless, the quite young farmers had substantial 20 years' experience in managing large-scale cow herds. The older farmers (average: 51.8 years with 37.2 years of agricultural experience) mainly managed the smaller herds (median: 15 cows) with tie stables (i.e., the herds from HC2).
The comparison of LSM for test-day traits revealed significant differences
(
Least square means and corresponding standard errors of test day and fertility traits in the first three lactations for four herd clusters created with CoVAHC.
Least square means in the same row with different superscripts a, b, c or d are statistically significant different at
Cows from HC3 had the highest P% (3.7 %), but the remaining HCs did not
differ significantly (
HC2 comprised the herds with the highest SCS (average SCS: 3.35) (Table 3). In herds from HC2, all cows are housed in isolated tie stables with quite high air temperature and humidity. Inadequate hygiene and climate management contributed to impaired immune responses due to toxic gases (Barkema et al., 1999), resulting in increased SCS. Most of these herds (60 %) used an alternative dry-off management without antibiotic treatments. In contrast, the cows from the remaining HC are predominantly kept in loose housing and cold stalls, and the dry-off management is mostly based on antibiotic applications. The optimal climatic husbandry conditions plus preventive veterinary treatments might be an explanation for lower SCS in HC1 (2.99), HC3 (2.94) and HC4 (2.77) compared to HC2. Doherr et al. (2007) associated herd size with an increasing risk for clinical mastitis. In contrast, we identified lower SCS in HCs representing the large-scale herds. Doherr et al. (2007) and Ivemeyer et al. (2011) reported significant breed effects on SCS, indicating impaired udder health for HF cows. In our study, HC4 comprised all HF herds, and SCS in HC4 was lowest. Thus, the herd management and climatic conditions might have a stronger impact on the udder health status than herd size or breed (Barkema et al., 2015).
A FPR larger than 1.5 is an indicator for subclinical ketosis (Heuer et al., 1999). Surprisingly, the high productive DSN herd in HC3 considering large percentages of concentrates in the feeding ration is characterized by a significantly lower FPR (1.11) compared to the DSN herds from HC1 (1.21) and HC2 (1.16). The influence of the breed on FPR as described by Ivemeyer et al. (2019) was not confirmed in the present study, because the FPR in the “HF cluster” (HC4) was 1.16.
The LSM for SFI in HC1 (67 %) and HC2 (66 %) was significantly higher than in HC3 (53 %) and HC4 (45 %). The high proportion of natural matings in HC1 and HC2 (77 % and 60 %, respectively) explains such differences. Andersen-Ranberg et al. (2005) and Löf et al. (2012) reported a longer voluntary waiting period for a first insemination after calving in high-yielding herds. In our study, cows from HC1 and HC3 displayed the shortest CFI with 78.8 and 76.2 d, respectively, but both HC differed significantly with regard to milk yield (HC1: 15.8 kg vs. HC3: 29.0 kg).
Reliabilities of EBVs considering the whole population (cows and sires) and only for sires are given in Table 4. Statistical models including a herd-cluster-test-day or a herd-cluster-test-month effect increased the reliability in the entire population by 3.1 % and 3.5 %, respectively. Regarding sires, the increase was 5.1 % (herd-cluster test day) and 5.7 % (herd-cluster test month). Hence, the detailed consideration of environmental conditions and herd characteristics contributed to an increase of EBV reliabilities in the range from 3 % to 6 %, reflecting the postulations by Zwald et al. (2003) and Osorio-Avalos et al. (2015). The herd cluster instead of the herd effect for the CG modeling contributed to increased heritability estimates from 0.23 (herd-test-day effect and herd-test-month effect, respectively) to 0.36 (herd-cluster-test-day effect) or to 0.38 (herd-cluster-test-month effect).
Heritabilities (
The creation of herd-test-day effects in genetic evaluations for local breeds with small population size generally implies a limited number of records or animals in CG. This, in turn, can lead to biased genetic calculations (Strabel and Szwaczkowski, 1999). As an alternative, HC or HC-test-month effects were suggested as CG effects in genetic–statistical modeling approaches (Vasconcelos et al., 2008). The general aim of CG creation is to depict environmental conditions influencing cow traits as detailed as possible (Kuehn et al., 2007; Osorio-Avalos et al., 2015). In this regard, a clear differentiation among HCs is imperative, as realized when applying the CoVAHC clustering approach. CoVAHC generated four different HCs with obvious herd similarities within HCs, implying 496 different CGs in herd-cluster-test-day models (model 3). A classical herd-test-day modeling approach would generate 603 CGs, with only a few records for some effect levels. A model based on CoVAHC HC avoided the problem of weakly occupied effect levels. The minimal number was three records per CG when considering the herd-cluster-test-day effect in model 3. Consequently, a genetic evaluation based on CoVAHC HC will contribute to an increase in the effective number of daughters (Tosh and Wilton, 1994). In this regard, Schmitz (1990) addressed the positive impact on breeding value accuracies, especially for breeds with a small population size.
A final severe issue in genetic evaluations is the computation time, which is generally time consuming in classical test-day models with a large number of herd-test-day effects. In this regard, models 3 and 4 on a HC basis were superior over classical herd-test-day or herd-test-month models, with on average 5 %–10 % reductions in computation time.
The superiority of the CoVAHC approach over the AHC, PAM and FZC methods for herd clustering in the local DSN population could be clearly demonstrated. In this regard, German DSN herds were clearly allocated to different HCs based on broad spectra of social–ecological and herd characteristics. Hence, we postulate also correct herd groupings in other German cattle populations, when considering similar descriptors for herd characterization and when applying CoVAHC clustering. Other clustering methods as developed for other fields of science including AHC, PAM and FZC are not appropriate (due to the obvious herd misclassifications) for animal breeding objectives. Utilization of herd clusters instead of single herds is suggested in genetic evaluations for breeds with a small population size kept in small-sized herds with a limited number of contemporaries. The suggestion is based on the observed increased EBV reliabilities and heritabilities. The clustering approach for herd allocation with corresponding ongoing genetic evaluations is an alternative also for large-sized populations, when creating different feeding or management groups in the same herd.
The housing and treatment of the animals were carried out in accordance with the German national regulations. The study was restricted to routine on-farm observations considering only data as used for national genetic evaluations plus data sets for herd characterization. All presented methods were non-invasive. Therefore, they did not cause the included animals pain, suffering or harm in compliance with the German Animal Welfare Act § 7.
The data that support the findings of this study are available from the authors upon reasonable request.
The supplement related to this article is available online at:
SK and KB designed the experiment and supervised the research. SK supported JH in writing and data validation. JH was responsible for farm characteristics and cow trait recording. JH and KB prepared the data and conducted the statistical analyses. All authors read and approved the manuscript.
The authors declare that they have no conflict of interest.
The authors thank the participating farms.
The project is supported by funds of the German Government's Special Purpose Fund held at Landwirtschaftliche Rentenbank (grant no. 313-06.01-28-RZ-3.061). This open-access publication was funded by Justus Liebig University Giessen.
This paper was edited by Antke-Elsabe Freifrau von Tiele-Winckler and reviewed by Dirk Hinrichs and Eva Strucken.