Future application of an attention bias test to assess affective states in sheep

Keywords: anxiety, behaviour, cognitive bias, depression, euphoria, fear, livestock, Merino, ruminant, sheep, threat, welfare.

Consumers, producers, industry bodies and regulators are demanding greater standards of welfare in livestock industries (Australian Government 2008; Kauppinen et al. 2010; Red Meat Advisory Council Ltd 2015). A key component of welfare that needs to be considered is the emotional or affective states of animals. The term ‘affect’ describes an animal’s physiological and behavioural responses, that can vary in terms of intensity (arousal) and pleasantness or unpleasantness (valence) (Mendl et al. 2010). This can include short-term emotions that are triggered by specific events as well as longer-term moods. It is important that tools are available for researchers and producers to assess and benchmark the affective states of animals as part of a comprehensive welfare assessment.

One approach taken to measure affective states in animals is the assessment of affect-driven attention biases. An attention bias describes the tendency to process certain types of information before others, which can be altered by the affective state (Bar-Haim et al. 2007). In humans, predictable changes in attention towards certain types of stimuli have been used to determine affective states and the presence of clinical affective disorders. For example, increased attention to threats is attributed to anxious states and generalised anxiety disorder (Bradley et al. 1995, 1997; Bar-Haim et al. 2007; Cisler and Koster 2010).

Several studies have explored the potential for attention bias to provide a measure of affect in livestock (Crump et al. 2018). The methodologies used have ranged in their complexity and the degree of animal training required. Lee et al. (2016) presented a rapid ‘looking time’ task to assess attention bias in sheep, which may have more practical applications for welfare assessment on farm. A considerable number of studies have since been conducted to refine and further validate this methodology in sheep and to adapt the method for use in other livestock species. There is now an opportunity to collate these studies and critically examine the potential of this approach to assess affect in livestock.

The aim of this perspective article is to provide guidance on the following question: ‘which version of the attention-bias test methodology should I use and what will it tell me?’ To address this question, we summarise the literature that used variations of the test methodology described by Lee et al. (2016) in sheep, to examine its potential use as a practical measure of affect. The findings of each study are tabulated for a clear comparison of treated- and control-animal responses. We then make recommendations for future application of the methodology in welfare research and highlight key gaps that need to be addressed moving forward.

Crump et al. (2018) provided a comprehensive overview of the methods used in animals to assess affect-driven attention biases. One approach is the use of eye-tracking and looking-time tasks that measure fixation of the gaze on competing images of emotional stimuli (Hermans et al. 1999; Eizenman et al. 2003; Kellough et al. 2008). Looking-time tasks have been applied to primates for the assessment of attention biases, using methodologies similar to those used in humans (e.g. Bethell et al. 2012; Howarth et al. 2021). Both Vögeli et al. (2015) and Raoult and Gygax (2018) developed looking-time tasks for sheep to assess attention, where attention towards valenced video stimuli was determined on the basis of head orientation, ear postures and frontal brain activity assessed using functional near-infrared spectroscopy. These approaches showed promise for assessing attention bias, but requirements to confine or habituate sheep limits the practical application of the test for the purpose of welfare assessment.

Another approach used to assess attention bias in non-human animals includes foraging or threat perception tasks. Brilot and Bateson (2012) assessed attention bias in starlings, by measuring the extent to which birds were distracted from feeding by the sound of a conspecific alarm call. Key behaviours included vigilance (head up) and latency to feed after the alarm call. This approach has also been applied to other bird species including parrots (Cussen and Mench 2014) and chickens (Campbell et al. 2019a, 2019b, 2022; Anderson et al. 2021).

The attention-bias task developed for sheep by Lee et al. (2016) sat somewhere between the looking time and foraging tasks described above. Sheep were tested in a novel arena, where they were exposed to a threat (a live dog sitting quietly behind a window) for a period of 10 s. The window was then covered, and the dog was removed, then the sheep stayed in the test arena for a further 3 min. Attention was assessed by measuring duration looking towards the previous location of the dog, vigilance behaviour defined by having the head at or above shoulder height, and latency to eat from a familiar feed bowl containing pellets that was located centrally within the arena. Lee et al. (2016) pharmacologically validated the method using anxiolytic and anxiogenic drugs, finding that ‘Anxious’ sheep spent more time looking towards the dog, were more vigilant and had a longer latency to eat than did ‘Calm’ animals. Thus, the authors showed that the test could be used to measure biases in attention towards a threat, which were related to anxiety-like states in sheep.

Eight studies have been conducted using the attention-bias test method described by Lee et al. (2016), or variations thereof, that applied pharmacological or environmental treatments to sheep prior to testing. Seven of these studies were conducted on Merino sheep at the same research station in Armidale, New South Wales (NSW), Australia, with the key results summarised in Table 1. The studies used sheep of varying ages ranging from 5 months to 7 years old and have used both male and female sheep. Overall, these studies have modelled chronic stress through environmental and pharmacological manipulation and have used pharmacological manipulations that attempted to model anxious, calm, depressed and euphoric-like states. Modifications made to the method over time included reducing the period of exposure to the threat and using a photograph of a conspecific in place of feed as a positive stimulus, to remove the potentially confounding effect of appetite on sheep responses. Hereafter, we broadly refer to methods using food as a positive stimulus as ‘the food method’ and methods using a conspecific photograph as the ‘photograph method’. The repeatability of the food method has been assessed in sheep (Monk et al. 2023). The food method has also been adapted to present a human as the threatening stimulus instead of a dog (Atkinson et al. 2022). Variations of the methodology have also been applied to cattle (Lee et al. 2018; Kremer et al. 2021), pigs (Luo et al. 2019; Verbeek et al. 2021), goats (Neave and Zobel 2020) and chickens (Campbell et al. 2019a, 2019b, 2022; Anderson et al. 2021), although this review will primarily focus on sheep. Notably, other studies have used similar methodologies, such as a fear test developed for dairy cattle by Welp et al. (2004) and an emotionality test for sheep developed by Torres-Hernandez and Hohenboken (1979).

Table 1.  Validation studies using variations of the attention-bias test for sheep described by Lee et al. (2016).

It is clear from the studies presented in Table 1 that induction of some affective states has an impact on sheep behaviour during the attention-bias test. Differences among induced affective-state groups were often strong in the attention-bias studies, but the effects relative to control groups were sometimes inconsistent. There are also some important differences in the methodologies that need to be considered. Here, we examine the results of the studies to discuss the merits and limitations of this approach for measuring attention bias in sheep and its potential ability to measure different types of affective states.

The most consistent findings across the studies shown in Table 1 support the method as an indicator of anxiety-like states in sheep. In three studies, treatment with the drug diazepam to decrease anxiety (anxiolytic) resulted in decreased attention to the dog window relative to control sheep. Use of the drug meta-chlorophenylpiperazine (mCPP) to increase anxiety (anxiogenic) resulted in an increased latency to eat or sniff the photograph in four studies and increased vigilance in three studies, relative to control animals. The effect of mCPP on attention bias was also replicated in cattle (Lee et al. 2018) and chickens (Campbell et al. 2019b). However, the differences between induced affective-state groups and control animals were sometimes inconsistent (Table 1). For example, Lee et al. (2016) reported a significant effect of mCPP on the duration looking towards the dog window, which was not replicated by Monk et al. (2018a). In contrast, Monk et al. (2018a) showed that sheep treated with mCPP spent significantly more time displaying vigilance behaviour than did control animals, but this effect was not observed by Lee et al. (2016). Overall, the test shows promise as a measure of anxious states in livestock, while highlighting behavioural variation that is not explained by the drug treatments alone. Examples of other sources of variation that might affect behaviour include the variable effect of drugs on individuals, variation in the animals’ moods prior to treatment and testing, individuals’ previous experiences or other aspects of animal temperament or personality.

Sheep treated with morphine to induce a euphoric-like state did not display an attention bias toward or away from the threat or a conspecific photograph presumed to be perceived as positive during testing (Monk et al. 2019b, 2020). This could suggest that morphine did not model positive affect or that the test could not discriminate positive affect induced by morphine. The effect of morphine on sheep behaviour in a food-based attention-bias test has not been examined. Sheep treated with the anxiolytic drug diazepam were often labelled as being in a calm-like state, which might be considered a positively valenced state. However, due to the nature of the attention-bias test involving social isolation and novelty, we propose it is likely that all animals tested were in a relatively high arousal, negatively valenced state, irrespective of their assigned pharmacological treatment. The drugs expected to induce positively valenced states may have been partially or completely over-ridden by the emotional response sheep have to isolation in the novel test environment and the threat of predation. As such, the ‘Calm’ sheep may have been less anxious than those in the ‘Anxious’ groups, but were not necessarily in a calm state (Fig. 1). The approach may still be useful as a measure of positive affect in species, breeds or individuals for which isolation is a less aversive stressor, or where the stress associated with testing and isolation is reduced through modification of the method or arena, as discussed further in the section ‘Refinement of the test arena’. However, in its current form, we do not believe that this method can provide a measure of true positive affect in sheep.

Fig. 1.  Diagram depicting the (a) intended and (b) potential positions of pharmacologically treated sheep in the affective space, delineated by axes of valence and arousal. Positions depicted in (b) are not intended to be accurate, but rather to exemplify the potential mismatch between intended and actual affective outcomes due to environmental stressors or other external factors.

Monk et al. (2018b) observed a significant effect of the depressant drug para-chlorophenylalanine (pCPA) on attention bias to threat, suggesting that the test may be sensitive to depressive-like states. Both pCPA and mCPP, inducing depressive-like and anxiety-like states respectively, resulted in increased vigilance compared with control animals. It was therefore proposed that, in the context of the test methodology, increased vigilance provides a measure of negative valence, which is consistent with the other attention-bias studies in sheep and cattle (Lee et al. 2018; Monk et al. 2018a, 2018b). Negative valence was also expected to result in increased attention towards the threat, which was the case for the depressed treatment group; however, the anxious treatment group unexpectedly spent less time looking towards the threat than did control animals. This finding contrasts with studies using the food method, where anxious sheep showed increased attention towards the threat. However, the tendency for anxious sheep to pay less attention towards the threat than for control animals using a conspecific photograph was also supported by Monk et al. (2019b). It was suggested that increased attention to the conspecific photo as a social stimulus aligned with the strong flocking instinct of sheep when faced with the threat of predation (Lynch et al. 1992). Together, this highlights the importance of considering a range of behavioural responses during testing to gain a more complete picture of affect and the need to carefully validate methods using different stimuli as attention biases are highly context specific. While the results of Monk et al. (2018b) are promising, further validation is needed using other models of depression in the attention-bias test.

Chronic stress induced through an environmental model using lying deprivation was found to reduce vigilance and result in a quicker approach to a feed reward, in contrast with the authors’ expectations (Verbeek et al. 2019). A similar unexpected ‘positive’ response has also been observed in judgement-bias tasks following stress in sheep (Doyle et al. 2010; Sanger et al. 2011; Guldimann et al. 2015). Potential explanations for this included release from stressful conditions generating a positive mood or a general increase in motivation for rewarding stimuli under chronic stress conditions. Monk et al. (2019a) modelled chronic stress by using a pharmacological model, by administering synthetic adrenocorticotropic hormone to induce an exogenous stress response, which had no impact on attention bias compared with control animals (Table 1). The pharmacological model used by Monk et al. (2019a) suggested that the attention bias observed in sheep exposed to lying deprivation may not be explained by cortisol response alone. Together, the findings of Verbeek et al. (2019) and Monk et al. (2019a) suggested that the test may not be sensitive to changes in affective state resulting from induced chronic stress and that further work is needed to understand the effect of chronic stress on affective state in livestock.

Finally, Atkinson et al. (2022) adapted the method to assess attention to a human threat, by swapping the dog for a human, but otherwise following the protocol outlined by Monk et al. (2018a). Prior to testing, they applied two different treatments involving different types of human–animal interactions over a period of 7 weeks, with the aim to reduce human-directed fear in weaned lambs through either habituation (low intensity, predictable human behaviour) or stress inoculation (moderate intensity, active, unpredictable human behaviour). Neither intervention was shown to affect attention bias towards the human. It is difficult to determine what type of affective state may have been induced by the human exposure treatments. No pharmacological validation studies have been applied to a method using a human as the threatening stimulus.

There is currently no way to directly measure affect in another living being, so there is no gold standard to which we can compare when validating new methods and models. Instead, measures and models can be incrementally validated against each other by drawing on human literature and by comparing a range of environmental and pharmacological models against a range of behavioural, physiological and neurological indicators. Environmental manipulations can be used to alter affective state in a way that more closely matches natural conditions. However, it can be unclear which affective states are being induced and the induced affective states are not always maintained during behavioural testing (Doyle et al. 2010; Sanger et al. 2011). Pharmacological models have an advantage as they can remain active during testing, be applied in a standardised manner and be easily paired with appropriate controls such as saline injections (Doyle et al. 2015). However, they are generally targeted towards a limited number of neurophysiological pathways and may not reflect naturally occurring affective states. They can also have unwanted side effects, such as the abnormal behaviours observed in sheep treated with mCPP, including head, tail and body shaking or ataxia (Doyle et al. 2015; Monk et al. 2018a). A relatively small number of studies have been conducted to find appropriate pharmacological models of affective states in livestock species and there is also often limited information available on the pharmacokinetic pathways of drug models. Further studies are required to understand the appropriate drugs, dose rates and dosing schedules for pharmacological models to have the desired outcome and to reduce unwanted side effects. While there are limitations for both environmental and pharmacological affective-state models, each can provide valuable information and a variety of models should be used to validate new welfare assessment methods.

To determine how best to interpret and apply an attention-bias test, it is important to understand to what extent it is affected by emotions, moods or trait affect. Emotions are short-term states triggered by specific events, while moods occur over a longer time frame and are less context specific (Mendl et al. 2010; Kremer et al. 2020). Trait affect describes the propensity of an animal to experience a particular affective state, as an aspect of animal personality (Boissy and Erhard 2014), where personality traits are patterns of behavioural responses that are consistent across time and/or contexts (Réale et al. 2007). Measures of emotions may be best applied in research to determine the immediate impact of certain events or environments on welfare. Measures of moods may be applied in both research and on-farm welfare assessment to measure the cumulative effect of recent events on an animal’s affective state. Measures of trait affect have the potential to be applied as a selection tool to identify animals with a less anxious or fearful personality. If the test is readily confounded by factors such as noise or weather, its practical application for welfare assessment may be limited.

Although the attention-bias studies have confirmed that the attention-bias test is to some extent influenced by affective state (Table 1), the extent to which it is affected by emotions, moods, personality and other factors remains unclear. Across three repetitions of an attention-bias test, Monk et al. (2023) showed consistency of vigilance behaviour in adult ewes by using the food method presented by Monk et al. (2018a). In cattle, Kremer et al. (2021) observed relationships between vigilance in an attention-bias test and a fearfulness personality trait that was characterised on the basis of behaviours across an open-field, novel object and runway test. Lee et al. (2018) demonstrated a tendency for more nervous cattle, as measured through flight speed and crush score, to show increased vigilance in an attention-bias test. Together these findings suggest that vigilance behaviour in the attention-bias test may be strongly driven by an underlying trait or aspect of personality, and thus may be considered to indicate the propensity of an animal to experience negative affect. However, further studies are still needed to examine the implications of increased vigilance during attention-bias testing for welfare outcomes more broadly.

Similar relationships were found between personality and attention to the threat during an attention-bias test in cattle (Lee et al. 2018) and consistency was observed in putatively ‘fearful’ and ‘attentive’ personality traits derived across two repetitions of an attention-bias test in goats (Neave and Zobel 2020). In cattle, Kremer et al. (2021) observed some relationships between attention to the threat and personality traits, as well as consistency in feeding behaviour between two repetitions of an attention-bias test. However, they did not observe consistency in threat-directed behaviours across the test repetitions and Monk et al. (2023) observed poor repeatability of looking and feeding behaviours over three repeated attention-bias tests in sheep. The inconsistency observed across repeated tests suggests that looking and feeding behaviours may be more strongly driven by transient affective states or other undetermined factors. Verbeek et al. (2019) demonstrated a positive response during attention-bias testing following a lying deprivation treatment, with reduced vigilance and latency to eat, which may suggest that the test is more sensitive to short-term emotions after release from a stressful condition, rather than a negative mood that the condition was expected to induce. However, further studies are needed to confirm this suggestion and to rule out other potential effects such as an increased motivation for rewarding stimuli (Verbeek et al. 2019).

Overall, these studies suggest that the attention-bias test is not only state-sensitive, but may also indicate trait affect, in a behaviour-dependent manner and in the absence of treatments that modulate affective state. Consideration of vigilance, looking and feeding behaviours independently may provide information on both trait and state affect within a single test. Importantly, however, it is likely that emotions, moods and personality interact and work together in some way to shape the responses of sheep during testing. Further studies should aim to manipulate emotion and mood independently, prior to attention-bias testing, so as to determine which of these aspects of affective state most strongly drive animal responses in the test.

Attention biases are highly context specific and so the choice of threatening stimulus should be carefully considered (Zvielli et al. 2014; Pergamin-Hight et al. 2015). In all the studies listed in Table 1, brief exposure to a predator threat (dog) was used as a threatening stimulus. Removal of the threat after a short time served two purposes. The first was to reduce the intensity of the threat that might otherwise prevent the animals from displaying attention towards other stimuli or the environment. The presence of a live dog in other behavioural tests is shown to be highly aversive for sheep, reducing and even eliminating the occurrence of exploratory behaviours (Torres-Hernandez and Hohenboken 1979; Beausoleil et al. 2005). The second was to remove the actual threat to the sheep, so that we could examine anxious states rather than fear states. The behavioural and physiological responses of fear and anxiety are largely the same but differ in the context of an actual versus an unknown threat respectively (Steimer 2002).

Brief exposure to a live dog has also been used as a threatening stimulus in an attention-bias test for both steers (Lee et al. 2018) and goats (Neave and Zobel 2020). While shown to be effective, the use of a live dog introduces more variation during testing, given the challenge of standardising the dog’s behaviour. For dairy calves, Kremer et al. (2021) used a dog statue in conjunction with the scent of dog urine and audio of a dog growling as a threatening stimulus. Although they did not validate before testing that the dog model was perceived as threatening by heifers, behaviours displayed during the test were consistent with it being perceived as threatening. In pigs, Luo et al. (2019) used a combined visual and auditory threat of a squeaky door moving up and down to show a flashing light for 10 s, while Verbeek et al. (2021) used a 15 s audio recording of an aggressive dog barking. Other attention-bias test paradigms for sheep have shown variable success using video images (Raoult and Gygax 2018) and acoustic stimuli (Raoult and Gygax 2019) to represent predator threats and conspecifics. Other potential threats might include an air puff such as used by Salvin et al. (2020) in a startle test for sheep, or startling movements such as the opening of an umbrella (Coulon et al. 2011; Neave and Zobel 2020).

Alternatively, a human could be used as a threatening stimulus in sheep that are not accustomed to human handling. Atkinson et al. (2022) applied an attention-bias test to sheep using a human as the threat. The test was unable to differentiate sheep that had undergone different levels of human exposure to induce habituation or stress-inoculation, although it remains unclear whether this was due to a lack of sensitivity of the attention-bias test or the treatments not having the desired outcome on human-directed fear. Humans and even human-like models have been used as a fear-eliciting stimulus in behavioural tests such as the arena test, which induces conflict between approaching humans and conspecifics (e.g. Vandenheede and Bouissou 1994; Bouissou and Vandenheede 1995; Forkman et al. 2007) and are shown to be less aversive to sheep than dogs (Beausoleil et al. 2005). Importantly, the attention-bias test paradigm is known to be context specific, and the response that an individual sheep has to a dog threat may not be comparable to their response to a human. The induction of negative affect due to human interaction is something many producers may want to measure and reduce, either through improved management and environment or selective breeding programs. Conversely, increased attention and vigilance towards predator threats may be desirable in many extensive production environments where predators represent an actual danger to sheep (Dwyer 2009). Thus, in a context-specific test paradigm such as the attention-bias test, it is important to consider the production context and reasons for measuring attention and vigilance when choosing a threat.

To be able to categorise how the individuals in the test arena divide their attention, an alternative positive attractant can be used alongside the threatening stimulus. Most attention-bias tests in sheep have used food as the positive stimulus (Table 1, Atkinson et al. 2022), where measures such as duration eating and latency to eat are used to indicate attention. Other studies have used conspecific photographs (Table 1), videos of conspecifics (Vögeli et al. 2015; Raoult and Gygax 2018) or audio of sheep bleating (Raoult and Gygax 2019), where attention is measured through behaviours such as looking, vigilance and ear postures. As acute stress responses typically involve allocation of resources away from non-essential functions such as feeding behaviour (Sherwood et al. 2005), using food as a positive stimulus provides a clear contrast against the predator threat. However, this contrast may become less clear depending on the testing context and the level of hunger experienced by an individual. Fraser and Duncan (1998) described how negative affect evolves from a ‘need situation’ where action is required for survival or reproductive success. In contrast, positive affect evolves from an ‘opportunity situation’ where performance of certain pleasurable behaviours such as play occur only when the cost of performing such behaviours is low. Feeding has the potential to fall under either category depending on the context.

Across the attention-bias studies using food, Lee et al. (2016) provided sheep with ad libitum access to pasture overnight prior to testing, while others withheld or limited access to food (Monk et al. 2018a, 2023; Atkinson et al. 2022). Given that feeding behaviour may arise from either a ‘need’ or ‘opportunity’ situation, relating to negative or positive affective states respectively (Fraser and Duncan 1998), it follows that the clearest interpretation of feeding behaviour as a contrast against the threat of predation would occur when the test sheep are not hungry. However, a complete absence of hunger may reduce the likelihood that sheep are willing to feed during the test, thus increasing the number of animals that fail to eat and limiting the ability of this measure to distinguish individuals. This may have been the case for Lee et al. (2016) where 85% of the control sheep failed to eat during the test, although other factors could have also contributed to a lack of feeding. Finding the right balance between hunger and satiation prior to testing may be difficult and presents a potential avenue for further validation of the test. Identifying feeds that are most rewarding and palatable for any given species may also be useful to increase positive interest in food during testing.

It is also important to control variation in appetite within a cohort during testing. Across all the attention-bias studies, sheep were housed in yards without feed while attention-bias testing was undertaken, which may have resulted in increased hunger over the course of the day. Atkinson et al. (2022) attempted to account for this by providing a half ration overnight to sheep that would be tested in the afternoon, while sheep tested in the morning were fasted. It is recommended that a similar approach is adopted for all further research using the food method to standardise hunger as much as possible across the cohort being tested.

Monk et al. (2018b) changed the positive stimulus from a food reward to a photograph of a conspecific to remove the potential influence of appetite on behaviour during testing. This removed appetite as a confounding factor, but sheep no longer had a strong incentive to become non-vigilant to feed. Additionally, a shift in attention towards a social stimulus represents an important strategy for sheep to cope with the threat of predation through flocking behaviour (Dwyer 2004; Wemelsfelder and Farish 2004). Thus, duration looking towards the positive stimulus alone may not be enough to discriminate positive affiliative motivations from flocking behaviour and therefore may not indicate the valence of the affective state of an animal being tested without also considering other behavioural responses. Overall, it is suggested that using food as a positive stimulus allows for a clearer interpretation of behaviour than does the photograph method and is the preferred approach in a context where appetite is not expected to confound results.

It is also important to consider the sensory capabilities of a species when selecting any stimulus, whether it is positive or negative. When selecting models, videos and photographs, researchers should consider the visual acuity of the target species and their abilities to perceive colour, luminance, depth and motion (Winters et al. 2015). Likewise, auditory and olfactory capabilities must be considered when using sounds and scents. Photographs or models of conspecifics and threats may not always be perceived by sheep in the expected way. For example, Franklin and Hutson (1982) found that the use of a taxidermy sheep as an attractant was unsuccessful as test sheep showed fear responses to the taxidermy model rather than affiliative responses. Together, these findings once again have highlighted the need to carefully validate the stimuli used for attention-bias tests.

The attention-bias methods listed in Table 1 present two stimuli, which are presumed to have contrasting emotional valence qualities, with the dog being perceived negatively and either food or a conspecific photograph being perceived positively. Other studies of attention bias in livestock have described a necessity to balance dually presented stimuli with regards to their presentation times and intensities, such that the stimuli differ only in emotional valence (Raoult and Gygax 2019). The stimulus presentation times used in the attention-bias test developed by Lee et al. (2016) are not balanced between the positive and negative stimuli, nor have they been in any variation of the method used thereafter. Further, it is difficult to determine and balance the intensity of a threatening stimulus when compared with a feed reward. In the context of this test paradigm, an attention bias is interpreted as increased attention towards a given stimulus relative to other tested animals. This is opposed to increased attention towards a given stimulus relative to other the stimuli presented. We argue that by comparing behavioural responses among and not within individuals, balancing of the positive and negative stimulus durations and intensity is not essential, so long as the test remains consistent for all tested animals in a population. Nevertheless, the presentation of stimuli for different durations has the potential to introduce new confounding factors that may affect animal responses. For example, spatial memory or learning may confound animal responses if a test subject no longer associates the previous location of the dog with the threat of a dog and does not localise their attention accordingly. This potentially confounding factor is important to consider when using pharmacological models that may have an impact on spatial memory or learning, as may be the case for diazepam (Brioni and Arolfo 1992; Sasaki-Hamada et al. 2013). It may be worthwhile exploring options to balance stimulus presentation times in the attention-bias test in a way that does not increase the intensity of the threat, such as by using a photograph of a dog instead of or in conjunction with a live dog.

To measure a bias in attention, we first need to be able to accurately quantify attention, which can be difficult in livestock species. Measures such as vigilance defined by having the head at or above shoulder height and latency to feed provide a very crude measure of attention compared with the eye-tracking studies used in humans and primates. A key problem with removing the visual threat of a dog after a short period of time is that sheep can no longer localise their gaze towards the threatening stimulus itself, only to the last known location of the threat. While some attention-bias studies have shown that this approach can be effective (Lee et al. 2016; Monk et al. 2018b), the use of a threatening stimulus that remains visually present for the entire test duration would be likely to allow for a clearer characterisation of visual attention. Importantly, however, in the presence of an actual threat, the test may be considered as a measure of fearful rather than anxious states. Consideration is needed as to whether this distinction is functionally important, given that studies in humans tend to focus on anxiety states with regards to attention bias, rather than fear. The test duration must also be carefully considered as the animals’ responses may become extinct as the threat is not further reinforced throughout the test period (Erhard et al. 2006).

Irrespective of the stimuli used, measuring direction of looking with binocular vision in a species with a wide visual field may not effectively characterise direction of attention. Expanding the definition of attention to incorporate other sensory modalities, and adjusting the stimuli presented accordingly, may help determine to which stimuli sheep are allocating their attention with a greater accuracy. Auditory stimuli have been used in an attention-bias test for sheep developed by Raoult and Gygax (2019). In their study, the direction of attention towards contrasting audio stimuli was determined by the orientation of the head while a sheep was restrained; however, to be considered attentive, sheep also needed to have their heads up in an alert position and their ears in a non-passive posture (i.e. the ears were both forward, both backward or asymmetrical). Incorporation of ear posture to the ethogram used in the attention-bias test could help better define direction of attention and may also give an indication of the affective state itself (Reefmann et al. 2009; Boissy et al. 2011; Lee et al. 2018). The collection of ear postures may be more meaningful if auditory cues were used instead of or in addition to the visual stimuli presented during testing. However, observations of ear postures is a labour-intensive and time-consuming process which would limit the practical application of the method, unless using an automated tracking system such as the one developed for sheep by Vögeli et al. (2014). Overall, modifications to the ethogram and stimuli, alongside the use of automated ear- and/or gaze-tracking technologies, may help to more clearly characterise attention in sheep and make the test more practical to apply.

Modifications to the attention-bias test arena may also allow for a clearer assessment of the direction in which attention is being directed. The original method positioned food in a way that allowed sheep to continue looking in the direction of the threat while feeding. Further, during the observation of video footage, the authors anecdotally noticed that the sheep may be remaining alert and attentive to their surroundings while their heads are lowered to a non-vigilant position and even while they are feeding. To prevent this from occurring, the food could be positioned against the wall opposite the threatening stimulus, where the photograph was positioned during later studies, or following a design similar to that used by Kremer et al. (2020) where the food was positioned in the corner of the test arena. Alternatively, a small visual barrier could be created between the food and the threatening stimulus so that sheep cannot remain visually attentive towards the threat while feeding or becoming non-vigilant. This approach was taken by Welp et al. (2004) when measuring vigilance towards a human in dairy cattle. Importantly, however, as a more fearful species, removing the ability of sheep to remain somewhat vigilant while feeding could reduce the number of sheep that are willing to feed during testing and may therefore limit the ability of the test to detect affective states.

A number of modifications could be made to the test arena and method to improve its practical application, standardisation and interpretation. The first is to have sheep enter through a narrow chute rather than a large gate used in the current method, to standardise the angle at which sheep enter the arena and, consequently, the angle and time that they see the stimuli. The second is to explore options for the automation of behavioural analysis using on-animal sensors or video-analysis software, as manual video observations are a time-consuming and labour-intensive process that may limit the test’s application to larger populations of animals. The third is to modify the test method or arena to allow attention bias to be assessed using existing handling or housing facilities on-farm. Currently, the time and equipment required to conduct attention-bias testing are likely to limit its application to research settings. Together, automation of behavioural annotation and adapting the method to existing handling facilities might allow the method to be applied as a welfare assessment tool on-farm.

Modifications to the method that reduce the fear-eliciting nature of the test environment and isolation may allow for a clearer assessment of attention and potentially allow for assessment of positive affect without the confounding effects of fear and stress. This could be undertaken by adapting the test to existing housing facilities if applicable to the production system, or by using habituation periods to reduce the stress caused by being in a novel environment; however, the latter will reduce the practical application of the test. Live conspecifics could be introduced to reduce the effect of social isolation on a test animal. However, the test arena would need to be carefully designed so that the conspecifics are not also exposed to the threatening stimulus, to reduce the potential effect of social contagion (Salvin et al. 2020). Alternatively, it may be useful to have sheep spend a short period of time in the test arena to get a baseline of behaviour prior to exposure to the valenced stimuli. Rather than remove the fear-eliciting elements of the test, this may allow researchers to account for individual variation in fearfulness as a covariate in the analysis. This could be standardised as a set time period (Verbeek et al. 2021) or could be based on feeding behaviour, whereby the sheep is exposed to a threat only after eating, following a design similar to that used in starlings (Brilot and Bateson 2012) or goats (Neave and Zobel 2020); however, the latter design would exclude any sheep that are not willing to feed in the novel environment. Due to the highly context-specific nature of attention biases, further modifications to the method or stimuli used during attention-bias testing should be carefully validated, taking into consideration the basic ethology of the species being tested.

The attention-bias test method is still new and further research is needed to properly answer this question. However, on the basis of the discussion above, we can make some recommendations moving forward. It is suggested that measures of interest in food more clearly represent a shift in attention away from the threat of predation than do methods that use a conspecific stimulus. Thus, in the absence of treatments that have a large influence on appetite, we recommend using the methodology presented by Monk et al. (2018a) or a variation thereof, that uses food as the positive stimulus instead of a photograph. Hunger should be standardised across the cohort being tested as best as possible. We believe that in this context, vigilance behaviour can provide an indication of trait anxiety, fearfulness or negative affect more broadly, while other measures of attention such as looking duration and feeding may be more sensitive to transient changes in anxiety-like states. The method does not appear to be appropriate for measuring positive affect in a prey species such as sheep, without further modifications to the method or arena. Changes in mean behavioural responses were evident across repeated attention-bias tests in sheep, as the sheep habituated to the novel test environment (Monk et al. 2023). Therefore, it is suggested that all animals being tested should have the same prior experience with the attention-bias test, to ensure a valid comparison of individual responses. Overall, the attention-bias test provides another valuable tool for researchers to better understand the impact that management practices and the environment have on livestock welfare.

However, it is important to note here that the method has been applied only to Merino sheep raised under similar conditions, and that all but one study have been conducted on the same research station, using the same or similar dogs as a threatening stimulus. Further studies are still needed to explore the relative influence of emotions, moods and personality on animal responses in differing populations and contexts, to enable a clearer interpretation of behaviour during the attention-bias test. The specific effects of age and sex have not been examined and there is currently not enough data available to draw meaningful conclusions on the potential impact that these factors have on attention bias. Modifications to the ethogram or test arena discussed throughout this review could be made to more clearly characterise the direction of attention towards the chosen stimuli during testing, and to automate the collection of behavioural data for a more practical application of the test. There is also a need to further validate the pharmacological models used across these studies to ascertain their effect on affect in sheep and other non-human animals.

Data sharing is not applicable as no new data were generated or analysed during this study.

Dana Campbell is an Associate Editor of Animal Production Science but was blinded from the peer-review process for this paper.

This research did not receive any specific funding.