Genetic analyses of herding traits in the Border Collie using sheepdog trial data

Introduction

Herding dogs are instrumental for the efficiency of large-scale sheep grazing systems. These dogs’ importance is exemplified by the study of Arnott et al. (2014) who estimated that over its lifetime, a herding dog in Australia performs work valued at approximately US$ 31 500, giving the owner a fivefold return on investment.

In Norway, two million sheep graze rangelands in summer and sheepdogs are a great help to farmers, especially when retrieving the sheep in autumn. In the 1950s, farmers learned of the Border Collie (BC) and its qualities, and imports began (Drabløs 2003). BCs are still imported from the UK and Ireland (Gjerjordet 2013), but most are now bred domestically.

To focus the BC breeding on working dog traits, the Norwegian Kennel Club (NKK) has assigned responsibility of breeding goals and programmes to the Norwegian Association of Sheep and Goat Breeders (NSG), which thus functions as the BC breed club, while the NKK maintains pedigrees and runs and records results from ordinary dog shows. Some 900 BCs are registered annually, and approximately 20% of them enter sheepdog trials (Gjerjordet 2013) based on International Sheep Dog Society (ISDS) rules. The main breeding goal is to preserve and improve herding skills and health (Norsk Sau og Geit, 2014). To be recommended for breeding by NSG, dogs must perform well in a trial, be healthy and have a good temperament (Norsk Sau og Geit 2014).

Hoffmann et al. (2002) and Arvelius et al. (2013) have studied herding behaviour in the BC, using limited-sized data sets from Germany and Sweden, estimating genetic parameters. In Norway, although sheepdog trial results have been available since 1993, no genetic parameters or breeding values have been estimated and selection is purely phenotypic.

This study aimed to evaluate the quality of Norwegian sheepdog trial data, estimate initial genetic parameters for herding traits and make recommendations for estimation of breeding values, as well as outlining an improved recording scheme.

Materials and methods

Norwegian sheepdog trials

Sheepdog trials, where a handler directs a dog that herds sheep around a field, form a competitive dog sport popular in many parts of the world. Rules and organization vary somewhat between countries; as in most other European countries, the Norwegian trials are based on the rules of the International Sheep Dog Society (ISDS) (International Sheepdog Society 2014), with some minor modifications regarding number of sheep and outrun distance. The following describe the current Norwegian rules for trials by NSG (Loftsgarden et al. 2014).

Sheepdog trials in Norway are divided into three main classes: 1, 2 and 3, where class 3 has six different variations. Thus, in total eight different class variants exist. For class 3, little data exist and the class will not be analysed or discussed in the remainder of this article. Class 1, the beginners’ class, is the easiest, while class 2 is more difficult. The classes include different elements with different maximum scores (Table 1).

Table 1. Trial classes 1 and 2, with trial elements and maximum score for each element, based on the Norwegian rules for sheepdog trials (Groseth et al. 2010)

Element	Class 1	Class 2
Outrun	20	20
Lift	10	10
Fetch	25	25
Driving	–	30
Driving in front of handler	20	–
Driving behind handler	10	–
Penning	15	15

The first element for the handler-dog team is the ‘outrun’ (Figure 1); the handler sends the dog along the right- or left-hand side of the course towards the sheep flock (the number of sheep varies; a minimum in Norway is four). The dog should run in an arc, ending up behind the sheep without disturbing them. The next element, ‘lift’, is when the sheep begin to move under the control of the dog. The dog will then bring the sheep in a straight line through gate 1 (Figure 1) and further towards the handler; this is called ‘fetch’ and is considered completed when the flock has circled behind the handler. Next is ‘driving’, where the dog must drive the sheep through gates 2 and 3, and then into the pen. In class 1, ‘driving’ is divided into two elements: ‘driving in front of handler’ and ‘driving behind the handler’. Here, the handler remains at the post until the sheep are at gate 2 and then walks in front of the sheep through the last gate and on to the pen. The element ‘penning’ is the dog moving all the sheep into the pen. Elements where the dog separates sheep from the flock are ‘shedding’, ‘single’ and ‘sorting’; these are not included in the trial classes studied here and will not be discussed further.

Contestants start each element with the maximum score, and the judges then deduct points as they assess the run. Excessive commands from the handler, sheep missing the gates and lack of control of the sheep are considered faults. The somewhat subjective criterion of ‘flow’ of the run is important for the total score. There are no rules regulating the number of judges at a trial, or on how to register the scores when more than one judge. There can be up to three judges, but most often, there is only one. When more than one judge is present, the scores from the judges are summarized or the judges will agree on a score. Each trial has a time limit set by the judge. If the run is not completed within the time cap, the elements not completed will be scored zero points.

Other circumstances can also result in a score of zero points in an element, such as disqualification (e.g. due to the dog biting the sheep) or withdrawal. The different causes behind a zero point score are not registered in the database. The least likely alternative is that a dog ends up with zero points on an element due to poor execution: if an element is carried out, it is usually rewarded with at least one point.

The distribution in the trial result data set for scores for the element ‘outrun’ in class 1 is shown in Figure 2; similar distributions were obtained for ‘outrun’ in class 2, and for ‘lift’ and ‘penning’ in both classes. ‘Driving’ in class 2 plus ‘fetch’ and ‘driving in front of handler’ in class 1 was more uniformly distributed, while ‘fetch’ in class 2 had a more normal distribution. All element distributions had a peak at zero.

Factors determining the level of difficulty of each run are class (element composition), type of trial (local, regional or national), fetch distance and number of sheep. The Norwegian rules for trials provide guidelines and minimum requirements but no specifications regarding fetch distance and number of sheep.

Statistical modelling

Initially, a univariate analysis was carried out, using data from class 1 only and treating the scores of the trial elements as continuous traits. The following general linear mixed animal model (full or reduced) was fitted for each element, including fixed effects assumed, a priori, to affect dog performance:

(1)

where Y_ijklmnopqrst is score of performance of a trial element of an individual dog in an event, μ the overall mean, sex_j the fixed effect of sex (male or female), year_k the fixed effect of the k^th year (18 levels: 1994 to 2011) and age_l the fixed effect of the l^th age group (three age levels: <2 years old, 2–4 years old and >4 years old). Further, type_m is the fixed effect of the m^th test type (five levels: m = 1–5, from local to national level), season_n is the fixed effect of the n^th testing season (four levels: winter, spring, summer or autumn) and distance_o is the fixed effect of the o^th fetching distance group (three levels: ≤100 m, 101–150 m and ≥151 m). Random effects are judge_p (119 different judges), event_q (2299 levels), a_r (random additive genetic effect of the r^th animal) and pe_s (random permanent environmental effect associated with the r^th animal). Finally, ijklmnopqr is the random residual variation not explained by the model.

The number of observations varies between elements; listed values are maximum in the total data set (see Table 2 for details on each element).

The univariate model [1] has the variance–covariance structure:

where A is the additive relationship matrix and I are identity matrices.

Judge and event were defined as random effects as subclasses had few records; to decide whether to include them in analyses, we compared full and reduced univariate models by -2ln likelihood testing (e.g. Lynch & Walsh 1998). First, a model without judge and event, then with judge and lastly with both judge and event was run; explanatory improvements were significant at the 0.01 level and led us to include both effects for analyses of all elements, both for class 1 and for class 2.

Finally, four bivariate analyses were carried out, one for each of the elements that was included in both class 1 and 2 (‘outrun’, ‘lift’, ‘fetch’ and ‘penning’). The same class 1 data as in the univariate analysis were used. Data from class 2 were selected based on the same principles as for class 1. The same independent model terms were used in the bivariate as in the univariate analysis (Equation 1). The expectations of random effects in this model were all zero with the following distributions:

For the bivariate analysis of ‘penning’, the model did not converge; this was resolved by a parameter reduction through assuming one common permanent environmental effect across trial classes.

All models were fitted using ASREML 3.0 (Gilmour et al. 2009). We used SAS/STAT version 9.3 software (SAS 2011) for data management and descriptive statistics.

Results

All fixed effects except fetch distance significantly (p < 0.05) affected at least one of the trial elements (Table 3).

Heritabilities ranged from 0.010 to 0.056, where penning had the lowest and ‘driving in front of handler’ had the highest heritability (both in class 1). Estimates of repeatability were low, ranging from 0.041 for ‘penning’ in class 1 up to 0.286 for ‘driving’ in class 2. All variances, heritabilities and repeatabilities are given in Table 4.

The results from the bivariate analysis show that additive genetic correlations for the four elements were positive, but with considerable standard errors (Table 5). They were low for ‘outrun’ and ‘lift’, and higher for ‘fetch’ and ‘penning’, indicating that the elements common to the two classes cannot be considered the same traits. There were rather high correlations for the effect of permanent environment, which implies that dogs that get high scores in class 1 do so in class 2 as well. The judge correlations are high, meaning that each judge seems to judge the elements similarly in both classes.

Discussion

In general, we found low repeatability and low heritability estimates, ranging from 0.041 to 0.286 and from 0.010 to 0.056, respectively. These parameters reflect the variation in the traits studied in classes 1 and 2, given a random sample of dogs born were given the opportunity to be trained and tested (no preselection), and without pedigree-related preferential treatment. These assumptions have been studied for horses by Olsen et al. (2012), who found close to no difference in cross-validation between only utilizing phenotypes from trials and a bivariate approach with an additional binomial phenotype of ‘animals tested or not’ to account for preselection. In a study similar to ours, Hoffmann et al. (2003) found similar heritabilities, ranging from 0.001 to 0.130. These estimates are lower than those found in other studies of the constitutionally alike (Coppinger & Schneider 1995) hunting behaviour of dogs (Karjalainen et al. 1996; Liinamo et al. 1997; Brenøe et al. 2002; Lindberg et al. 2004; Arvelius & Klemetsdal 2013). In a genetic analysis of herding behaviour in BCs, Arvelius et al. (2013) found high heritabilities when analysing herding trait characterization scores, that is not competition results.

A considerable amount of the data material could not be used in the analysis because the quality of the data is poor. The original data material consisted of some 46 000 records, but after weeding out erroneous data only half was left. After deleting the records from class 3, we removed dogs with errors in the pedigree, as well as records with zero point scores, the latter which may have many different meanings, making it a non-consistent phenotype. In the end, only approximately a third of the data remained.

A most vital improvement would be to register the reason behind a zero point score. It is important to distinguish between withdrawal, exceeding the time limit, and disqualification, especially if disqualification was due to the dog biting the sheep, as this behaviour is unacceptable in the BC and is likely to be heritable (Coppinger & Schneider 1995). When more than one judge is present at a trial, the number of judges evaluating each participant is not recorded or whether the judges score individually or agree upon a common score. Trials with more than one judge were thus excluded from the analysis. It is further problematic that the rules for trials are not standardized in terms of numbers of sheep and time limits and that decided levels are not registered. Another problem is the manual registration of results from each trial, performed by local managers, without sufficient quality assurance. It is important to make national registration protocols that decrease the frequency of erroneous information in the data sets and ensure that all information is submitted correctly, in a usable format. ID of handler, for instance, is a factor that should be included in analyses, but it was impossible to use due to imprecise registrations.

The lower repeatability compared to previous studies reflects the aforementioned suboptimal evaluation system. As pointed out, the evaluated trial elements are not well defined. Also, one element can consist of more than one behavioural trait; this means that two dogs may get the same score on an element even if executing it in very different manners. Further, the distributions of scores were unfavourable, calling for an objectively defined and neutral use of the whole scale, for example by adopting a linear scoring system.

The heritabilities are low, resulting in a low potential genetic response from phenotypic selection. If basing selection on breeding values and with the present average of 6.3 observations per animal, the accuracy of breeding values will increase two to three times as will the genetic response. In practical breeding, however, selection needs to include information from class 3 trials, as this is where the highest performing and thus most sought-after (as working dogs and sires) dogs are found. An alternative would be to specify an accumulated phenotype defined as ‘the highest class for which a dog qualifies’, but at present there are no established criteria for promotion to higher classes. When analysing each element in class 1 and 2 using bivariate models, we obtained a maximum of 11 EBVs per animal. The analysis did not utilize information between traits within each class or between different traits in different classes. The fact that no convergence could be achieved for the full bivariate model analysis of ‘penning’ implies that despite the large amount of data, there is not more information available in the data material; hence, we did not pursue further analyses.

An alternative to using the present sheepdog trials would be to adopt the non-competition herding trait characterization score (HTC) scheme that was evaluated in Sweden between 1989 and 2003 (Arvelius et al. 2013). Using HTC data, Arvelius et al. (2013) estimated herding trait heritabilities as high as 0.14–0.50. In HTC, the focus was on behavioural herding traits, in contrast to the sheepdog trials where complete elements or ‘tasks’, that is the result of the dog's work, are evaluated. In the sheepdog trials, the element ‘lift’ is an evaluation of how adept the dog is at getting the sheep to start moving (the dog is ‘good’ or ‘bad’). In contrast, the HTC trait lift describes, in a linear manner (from 5 points= ‘running straight through the flock’ to 0 points= ‘stopping/moving slowly’), how the dog behaves during execution. The intensity of the expression of the traits is thus evaluated on an objective scale; also, the evaluation is carried out over a longer period. In HTC, other traits such as affability towards humans were evaluated in addition to herding behaviours. Both the HTC and the Norwegian sheepdog trials score ‘outrun’ and ‘lift’, but the estimated heritabilities were substantially lower in the Norwegian data, likely due to the limitations of the trial data and the more precise trait definitions in HTC. With the established genetic ties between BC populations in Sweden and Norway (Norsk Sau og Geit 2014), both countries could benefit from using similar evaluation schemes (Arvelius & Klemetsdal 2013).

Herding behaviour is in essence predatory behaviour to which sheep, belonging to a prey species, are sensitive (Houpt & Willis 2001; Spady & Ostrander 2008). The basic dog predatory behaviour motor pattern is described in similar manners by Scott & Fuller (1965) and Coppinger & Coppinger (2001), and can be presented as a sequence of behaviours: orient → eye-stalk → chase → grab-bite → kill-bite → dissect → consume (Coppinger & Coppinger 2001). These are behaviours that are largely instinctual and that can be modified by artificial selection (Scott & Fuller 1965; Coppinger & Coppinger 2001). Centuries of artificial selection for changing this motor pattern have resulted in changes in intensity and display of the different sub-behaviours, but not in the basic structure and order (Scott & Fuller 1965; Coppinger & Coppinger 2001) of the motor pattern. Comparing the behaviour of the BC to the full predatory sequence, the eye-stalk and chase have been intensified, and the grab- and kill-bite strongly diminished (Coppinger & Schneider 1995). It is apparent that the selection for behaviour of different hunting dogs (e.g. Pointers and Retrievers) can also be described as modification of the same predatory behaviour sequence. This suggests that recording and breeding for traits that are based on the predatory motor pattern would make biological sense, and should give better and more objective information than generated by the current sheep dog trial scheme.

It might be difficult to get people to participate at a new test if this is introduced as a separate test, especially when the aspect of competition is not a motivational factor. It would be an advantage to implement new traits that can be judged in conjunction with the current trials. The following proposal of new trait definitions is based on traits from the predatory motor pattern and traits from the HTC. Traits from the HTC that would be potentially useful to adopt are those with high heritability and that can be evaluated during a sheepdog trial, primarily outrun, lift, working distance and ability to keep the flock together. Traits from the predatory motor pattern that we consider the most important for the BC are intensity of chase, shape of chase, intensity of eye-stalk and the welfare-relevant frequency of biting sheep (bite–grip). In addition, the transition thresholds between different motor pattern stages should also be considered, as this is a crucial factor determining the trainability of the BC according to Coppinger & Coppinger (2001). These 11 traits can all be judged during the current trials. All new traits, except ‘outrun’ and ‘lift’, should be judged across the elements of the current trials, in a standardized, precise and objective way (Diederich & Giffroy 2006; Arvelius et al. 2013).

With improved information in the data, implementing EBVs will have greater potential benefits than in the current situation. The majority of the elements of the sheepdog trials are comprised of eye-stalk and chase behaviours found in the predatory sequence, and we suggest that selection based on EBVs should be especially beneficial for ‘lift’, as this element is most similar to and limited to one of the predatory motor patterns, namely eye-stalk. ‘Outrun’ can also be a candidate for EBV selection, because it might be equivalent to orient of the predatory motor pattern, although less obvious.

However, to increase the accuracy and improve the value of sheepdog trials in breeding there is definitely a need to redefine traits as described above. We recommend defining new traits based on the HTC and the predatory motor pattern. These new traits should be used as breeding traits, and may be scored across the elements that make up the current trial system, which should be kept in place to stimulate participation in the genetic evaluation scheme.