Introduction

The behaviour of dogs has been of central interest to humanity for millennia–possibly since the early stages of dog domestication 15.000 years ago through human-directed selection for specific traits [1–3]. While dogs appear already in the work of early scientists such as Darwin [4] or Pavlov [5], it is only in the last 30 years that the research into dogs’ behaviour and cognition has been recognized for its intrinsic value, resulting in an explosion of studies focusing on dogs’ cognitive and human-directed capabilities [reviewed by 6,7], their temperament and personality traits [8], and the influence of domestication [reviewed in 9] and our own anthropogenic environment [e.g., as model for aging: 10] on behavioural development.

However, these efforts have primarily revolved around a relatively limited subset of the global dog population–those under immediate human care and control, such as pets, working, and shelter dogs [8]. While easy to access, this subset represents less than a quarter of the global dog population with the other 75–83% consisting of free-ranging dogs [11,12,]. These animals might or might not be owned but share the characteristic that their movement and mate-choice are largely unrestricted by human control [13], even though the majority still resides as street dogs that depend on human settlements for food and shelter [11].

Efforts to explore the behaviour of free-ranging dogs have so far predominantly relied on observational methods [e.g., 14,15, reviewed in 16], leading to important insights into these populations. While such studies are time-intensive and lack the experimental manipulation required to answer some of the questions addressed in the lab, they were so far chosen over experimental approaches due to several challenges that are associated with conducting more standardized test batteries for free-ranging dogs in their natural habitat: Beside the risk of aggression and disease status [e.g., rabies: 17], restraining the animals for a test would be unethical and obstruct the testing of natural behaviour. Hence, tests must rely on voluntary participation and transportable, flexible test setups. Established methodologies and apparatuses thus have to be adapted to the mostly unexplored street conditions, with some inevitable loss of direct comparability, given the different contexts in which animals are tested [18–20]. Finally, this lack of a stable test environment impedes standardisation and the variability and disturbance risk might distort results [21], particularly if the aim is to test the highly dynamic populations repeatedly [reviewed in 16,19,22].

Nonetheless, there are compelling reasons to experimentally assess the behaviour and personality of free-ranging dogs. For one, without studying a representative sample of the global dog population encompassing at least some of the vastly different life experiences and mating choice they may have, our knowledge about dogs’ cognitive abilities, human-animal interactions, the connection between genetics and personality, and beyond remains compromised [19, for more specific ideas see 21]. Secondly, it has been shown that, while the admixture of different breeds can be high in free-ranging populations [23], a large proportion is represented by genetically distinct populations that predate the recent strict artificial selection for breed formation [13,23]. Compared to our heavily selected modern breeds where domestication traits might have become uncoupled [24,25], those populations provide important insights for comparison with domestication research both from a genetic and ecological perspective [26]. Finally, free-ranging dogs’ pervasive presence in human settlements (e.g., 2,930 dogs/km² in Kathmandu, Nepal [27]) means that millions of people spend their daily life in dogs’ direct proximity. Increasing our understanding of their behaviour and behavioural profiles could therefore improve both human and animal welfare and aid in effective population control [21,26].

Some pioneering groups have already initiated work on free-ranging dogs, demonstrating that an experimental approach, particularly if involving short, one-off tests can indeed be conducted under the variable yet natural conditions in which they reside [18,20,28–30]. While presenting a valuable start, these tests have focused on single behavioural aspects in short tests [e.g., persistence in 18, pointing in 20], leaving the possibility of longer tests, test batteries, and repeated testing that is needed to explore behavioural profiles, largely unexplored.

Furthermore, critical information regarding the reliability of measures obtained from such tests on the streets remains scarce–information that is vital not only for singular behavioural testing but especially for comprehensive temperament assessments. Several aspects of reliability should therein be considered: the coding reliability, which describes how well the behaviours defined in the ethogram and the testing conditions allow for consistency in a rater’s observations over time (“intra-rater reliability”) and with others (“inter-rater reliability”), and the “test-retest reliability”, which assesses a test’s ability to capture the behavioural traits reliably. A reliable test would be expected to elicit the same behaviour in an individual again when being tested in the same test twice (“retest”). This is particularly important in terms of personality traits, which are defined as being consistent over time and context in an individual [8]. Test-retest reliability outcomes are often rather low in pet breed and shelter dogs, but some kind of consistency over time has often been reported [31,32]. Whether this also pertains to free-ranging dogs and/or under highly variable street conditions remains to be explored.

In response to this research gap, our study endeavours to achieve a twofold objective. First, to pilot and evaluate the feasibility of a test battery specifically tailored to street dogs (i.e., free-ranging dogs living in and around human settlements), encompassing a range of behaviours commonly assessed in dog cognition or temperament tests and suggested to be relevant for different domestication hypotheses [tameness: 1, sociability: 2, deferential behaviour: 3, lack of aggression: 33]. These behaviours include human-directed approachability, conspecific sociability using a ’fake dog’ test, neophobia and exploration in a novel object test, dogs’ understanding of human communication in a pointing test and their use of human-directed communication (gazing) in a begging test, as well as their propensity to shy away from conflict in a tractability test [34–38]. Second, we assess the reliability of this test battery in the natural street setting, including inter-rater, intra-rater and test-retest reliability. We draw conclusions from the results on free-ranging dog testing overall, as well as how this test battery could be adapted to be more widely used in the field.

Methods

Procedure

The test battery consisted of six short behavioural subtests and three physiological samplings conducted in direct succession (Fig 1). The subtests were designed to resemble behaviour tests that have commonly been used to assess pet, shelter, or working dog temperament traits, dogs’ cognitive abilities, or traits that have been suggested as central during the domestication process (see description of the respective subtests). The three physiological samples (saliva and hair) were part of the test battery to assess the feasibility of sampling in such a setting and allow for later use of those samples in future exploration.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. The sequence of the employed test battery.

Subtests in light and physiological sample collections in dark grey.

https://doi.org/10.1371/journal.pone.0296509.g001

For the test-retest reliability measure, the test battery was conducted a second time (“retest”) with the same study subjects within an average time span of 46.28 days (min. 33, max. 76 days) which was similar to other test validation procedures [e.g., 39–42]. To this end, the location of the initial test was noted, and the area was frequently visited again after five to six weeks until the dog was found, identified with previous pictures, and tested again. It was noted if the dog was retested in a different location than the first test (i.e., not in visual vicinity of the previous location). The procedure in the first and retest was exactly the same as described below. For the object-based subtests (Fake dog and Novel object subtest, see below), two slightly different-looking versions were used respectively for the test and the retest to create novelty. The order of presentation of the two stimuli was randomized across subjects.

Initiation and general circumstances.

Two experimenters were involved in each testing occasion: E1 was always the same person (SC) and conducted the actual tests while E2 (three different people) video-recorded, helped set up the subtests, or distracted other dogs if necessary. Dogs were tested individually and had to be alone before and during the test. Tests were conducted between sunrise and 10 a.m. before the areas became too crowded. The experimenters accessed the study area by car and all dogs encountered alone were tested. If at any point during testing, the dog ran off more than 10 m, the subtest during which the dog left was terminated and E1 tried to lure the dog back. If the dog ran too far to be retrieved, or did not respond to the lure, the whole test was terminated. If anyone other than the experimenter approached and visibly distracted the dog (e.g., head turn), the test was interrupted until the criterium was restored and the test resumed. If at any point dogs showed active signs of aggression, the test was terminated (however, this never occurred). For each dog, location, time, weather, health status, and activity upon arrival of E1 (e.g., sleeping, walking, …) were noted.

The exact setup of each subtest is displayed in Fig 2, the procedure of each subtest is demonstrated in the supporting information (S1 Video).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. The set-up of each of the six subtests in the presented test battery.

The main experimenter (E1) is presented in black (or grey in later stages of the subtest), the assistant experimenter (E2) in white.

https://doi.org/10.1371/journal.pone.0296509.g002

Human approach test (subtest 1).

Target: Approachability, human-directed sociability/ docility, tameness, play. This test was modelled after similar tests used in widescale personality assessment tests (e.g., the Dog Mentality Assessment (DMA) in the Swedish Kennel Club [43] and has in part been used with Ethiopian village dogs [30]. Beside the assumption that human-directed sociability is an essential trait under selection during the domestication process [2], assessing the actual human-directed behaviour in a free-ranging population is provides valuable insights for disease- and population control.

Here, E1 stepped out of the car and whistled shortly to grab the dog’s attention. She then stepwise proceeded to 1) stand quietly looking at the subject for 30 s, 2) called the dog whilst making friendly gestures (snapping, tapping on leg) for another 30 s, 3) approached the dog for 30 s up to 2 m distance, 4) stood for another 30 s, 5) approached the dog to one body length distance and paused another 30 s, and 6) crouched for another 30 s all while continuing to call the dog over. If the dog approached to less than 10 cm of E1 during any of those phases, E1 kneeled and tried to pet the dog for 30 s. If tolerated, E1 used a small brush to collect hair samples. The dog was then invited to play tug of war with a rag toy (1 m long) for 30 s.

Fake dog test (subtest 2).

Target: Conspecific-directed sociability and aggression. Several studies support the idea that stuffed dogs can be used to reliably assess dogs’ initial behaviour towards conspecifics [34,44]. A reduction in conspecific aggression has been hypothesised as an important shift from wild to domestic types [45], and understanding intraspecific aggression in free-ranging dogs can provide important applications for the local population.

Hence, subjects were exposed to a “Giant Jack Russel Terrier” stuffed fake dog (brand: ‘Melissa & Douk’; wither height 28 cm; S2 File). The two used versions differed in ear posture (pointy or droopy), tail length (medium and long), and colour (one original, one wrapped in dark bags) in the test and in the re-test respectively, in a randomized order across subjects. The fake dog was positioned 5 m from the car behind a V-shape visual barrier with its opening towards the car. This visual barrier allowed only the approaching subject but no other surrounding dogs to see the fake dog. E1 guided the subject towards the fake dog (if needed with food) and quickly walked >5 m away as soon as the subject saw the fake dog. The subject’s behaviour was observed for 2 min.

Novel object test (subtest 3).

Target: Neophobia, exploration/novelty-seeking. Presenting a dog with an unknow object is commonly used to assess novelty-seeking and boldness [38,46], and explorative-behaviour is hypothesised to be lower in domesticated animals as part of the domestication syndrome [45].

To present an object that was truly novel to the dogs, we fixed a 90–100 cm high foil balloon with a wooden stick on top of a remotely controlled car (31x18x18 cm, Model: DEERC DE42 RC, S2 File). Two balloon forms were used (dinosaur, a number (zero, five) sprayed in neon green and blue) for test and retest and randomized between study subjects. The set up was the same as the Fake dog test. As soon as the dog saw the novel object, E2 started moving it back and forth for 1 min through remote control. The toy was then left inactive for another 1 min. The dog’s behaviour was observed for the full 2 min.

Pointing test (subtest 4).

Target: Dog-human communication skill, willingness to follow commands, cooperative communication [reviewed in 47,48]. Studies have shown that a wide array of species, including canids, can follow pointing cues [for review see 47]. Exploring this phenomenon in free-ranging dogs with differing levels of human experience will further our understanding of the role of life-experience vs. species predisposition (but see [20] for pioneering work in this regard).

The test started with four warm-up trials: E1 placed a baited bowl (12 cm diameter, 10 cm height) on the ground 2 m from the dog, standing 1 m away herself and verbally called the dogs attention. After the dog ate the food, the bowl was picked up, baited, and placed down again to familiarize the dog with the fact that the bowl contained retrievable food. The subtest was terminated if the dog did not approach the bowl within 30 s two times in a row. The warm-up trials were followed by twelve test trials: one empty and one baited bowl (both rubbed in food for smell, baited outside of the dog’s view) were placed 1.5 m apart 2 m from the dog. E1 stood equidistant behind the bowls, attracted the dog’s attention, and upon eye contact, pointed to the baited bowl for 2 s (momentary distal pointing). The dog was allowed to eat the food if it approached the correct bowl on the first choice, otherwise the bowls were immediately collected by E1. Each trial could be repeated twice if the dog did not choose within 30 s, followed by one repetition of the warm-up trial (i.e., only one bowl with food). If after the third repetition, the dog did not choose, the subtest was terminated. We used six different randomized sequences related to the positioning of the baited/pointed bowl to the right or left of the experimenter (e.g., RLRLLRLRRLLR) that were called out to E1 by E2.

Begging test (subtest 5).

Target: human-directed communication, gazing behaviour, suggested to be help-seeking behaviour [35, reviewed by 36, but see for alternative explanations: 49,50]. Also a behaviour suggested as a trait under selection during domestication but the wide variability between individuals depending on human socialisation calls for further investigation with differentially socialised populations [51,52].

The dog was allowed to eat from the baited bowl three times. Afterwards, the bowl was visibly baited and placed out of reach but still visible on the car, a wall, or a tree. E1 stepped 1 m to the side and faced away from the bowl with a neutral facial expression. The dog’s behaviour was observed for 1 min.

Tractability test (subtest 6).

Target: Social inhibition, tractability/deferential behaviour (i.e., the ease to yield to higher command, conflict avoidance [3]). According to the deferential hypothesis [3], tractability has been a central selection factor in dog domestication. Furthermore, assessing dogs’ inclination to avoid conflicts with humans in a resource scenario may also have it uses in a more applied setting.

E1 gave the dog a piece of food, then took a big piece of meat, showed it to the dog, and visibly placed it on top of a small stool (40 cm height). E1 loudly stepped between dog and stool, stared the dog in the eyes and firmly said “No, it’s mine!”. If the dog attempted to get to the stool, E1 loudly stepped in their way and repeated “No” in a firm tone. After 1 min of this, E1 stepped 1 m to the side and stood still for 30 s while looking at the food. Finally, E1 turned around and looked away for 30 s. The test was finished as soon as the dog ate the food or after the last 30 s.

Video coding

The tests were recorded by an action camera (4K 24fps WiFi Sports Action Camera or Action Cam Jeemak WiFi) fixed on E1’s chest and a camera (Sony HDR-CX405 Full HD Camcorder, hand-held (subtest 1) or on a tripod (subtest 2–6)) handled by E2. BORIS (version 7.13.8, (Friard & Gamba, 2016)) was used for the video coding.

The ethogram aimed at capturing the target behaviours described for each subtest. We used a bottom-up approach by initially coding 55 variables with different modifiers (i.e., the person or object the behaviour was coded towards) that measured almost every potentially relevant aspect in the different subtests. The coded variables are displayed in Table 1 with a detailed ethogram in S1 Table. Social behaviours were based on the ethograms used at the Wolf Science Center of the University for Veterinary Medicine, Vienna [9], and to observe free-ranging dogs in Italy [15,53].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Coded behavioural variables and occurrence across subtests.

https://doi.org/10.1371/journal.pone.0296509.t001

This list of variables was narrowed down by 1) assessing their codability through the inter- and intra-rater analysis (for acceptable agreement see Statistical Analysis), 2) assessing the variables’ centrality in the subtest by analysing their frequency (i.e., excluding variables that were displayed in less than 10% of tests), and 3) assessing their consistency through the test-retest reliability. Only behavioural variables that passed step 1) and 2) went on to be analysed in step 3) to ensure that the test-retest reliability was based on reliably coded and not overly skewed data [54].

For the intra-rater reliability measures, a random subsample of 20% of the test videos (n = 13; balanced between test and retest, full test batteries and tests that were terminated halfway through) were re-coded by the initial rater after three weeks. For the inter-rater reliability, a random subsample of 20% of the full-test videos (n = 9) were coded independently by a second rater [for similar methods see 38,55]. Beforehand, the second rater was trained on three sample videos and the ethogram was further clarified where needed.

Results

Test-retest reliability

The final test-retest reliability analysis was conducted for each variable by subtest and modifier, only taking the variables into account that 1) were not already dropped from the ethogram/tests throughout the testing (i.e., auto-grooming, sniffing the ground, play), 2) had an acceptable coding reliability, and 3) occurred in more than 10% of all tests. The resulting test-retest reliabilities are reported in Table 2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Summary of the test-retest reliability in all retained variables.

https://doi.org/10.1371/journal.pone.0296509.t002

The average test-retest reliability was poor with an average ICC of 0.23. To see whether the poor outcome was influenced by disturbances during the tests that altered the dogs’ behaviour, we conducted the same analysis but excluded all subtests in which a disturbance had been coded (i.e., when the dog clearly turned its head towards an interfering human/dog/other animal for more than three seconds during a running subtest and while the test subject was visible on the video). While this changed the reliability of some behaviours, the average test-retest reliability did not improve (Table 2; the exact statistical outputs of both analyses can be found in S3 Table).

To summarize by subtest, the Human approach test performed well with ten out of 15 behaviours showing significant test-retest reliability, resulting in an average test-retest reliability of 0.32. In the Fake dog test, only ‘tail between the legs’ had a moderate retest reliability, with ‘close proximity to the fake dog’ and ‘nose-mouth licking’ also being moderately reliable after tests with disturbances were removed. The average test-retest reliability was low (ICC = 0.06). Similarly, only ‘close proximity to E1 and E2’ were significant in the Novel object test, being joined by ‘nose-mouth licking’ after disturbed tests were excluded, resulting in an average test-retest reliability of 0.12. The average reliability of the Pointing test was significant with an average ICC of 0.51 with all subtest-specific variables, as well as ‘tail-wagging’, showing moderate test-retest reliability. The Begging test had a low test-retest reliability (average ICC = 0.16) with only the ‘close proximity to E1’ being significant across time when only looking at tests without disturbances. Lastly, ‘close proximity to E1’, ‘gazing at E1’, the ‘phase in which the dog ate the food’, ‘three-way gaze alternation between person and stool’ and ‘fleeing’ had significant reliability in the Tractability test, with the addition of ‘nose-mouth licking’ after disturbances were removed. The average test-retest reliability was 0.22.

Discussion

In this study, we presented what, to our knowledge, is the first attempt at testing free-ranging dogs in a behavioural test battery and report feasibility and reliability of the test measures. Testing the dogs in their natural environment was indeed feasible and the coding reliability was high, indicating that this (relative to previous studies) long and complex behaviour test battery can be conducted with street dogs. The range from good to poor test-retest reliability, on the other hand, reflects the challenges inherent in conducting behaviour tests with free-ranging dogs on the streets.

Although the subtests in this test battery were adapted from mostly ‘lab-based’ behaviour assessments [e.g., 34,38,46] [but see 20], standardizing the set up and testing our dog population in the streets turned out to be surprisingly manageable. Importantly, there were no terminations due to aggression towards people or equipment and most dogs completed all six subtests and sample collections despite an average test duration of 30 minutes, suggesting that street dogs (at least those of this population) can remain motivated throughout longer and more complex procedures than those used so far [e.g., 18,29,30,69].

Similarly, the concern that testing and filming in the variable street conditions would negatively impact the coding reliability was unfounded. Our results surpassed the target of a mean ICC of 0.7 with good average inter- and intra-rater reliability of 0.84 and 0.93 respectively. Contrary to our expectations, the coding reliability in this street test was thereby on par or even better than many pet and working dog studies [ICC 0.68 in meta-analysis by 8, ICC >0.9 in 38, ICC 0.77 in 62], suggesting that our chosen variables were not only well defined but could also be coded reliably despite the more challenging conditions. Similarly encouraging inter-rater results have been reported for a few other street dog tests as well [18,20], further highlighting that such results are achievable for street testing in general.

However, when looking at the reliability of the individual variables, a few did not meet the threshold. Notably, ‘medium proximity’, i.e., two body lengths distance from the experimenter, was the only variable which was low both in the inter- and intra-rater reliability analyses. With the constant adaptation of the camera angles to adjust for the variability of the street conditions, it might thus be best to stick to distances that are easier to identify such as the more reliable ‘close proximity’ (i.e., one body length distance). Additionally, ‘head dip’, ‘stand tall’, and ‘friendly’ all had poor agreement between the raters, which might be due to the complex definitions of these behaviours (S1 Table). While these definitions were adapted from previously used ethograms [16,70,71], the filming in street conditions likely made the subtle cues harder to detect. Given the poor coding reliability of those traits, they should thus be dropped from the presented ethogram unless a more reliable definition or training method can be found and validated, and care should be taken in the establishment of future tests to balance between the complexity of target behaviours and filming feasibility in the streets.

While these outcomes paint a positive picture for street dog testing so far, the feasibility in terms of capturing the desired traits cannot be claimed without looking at the test’s ability to elicit the same behaviours consistently over time. The presented test battery resulted in an average test-retest reliability of 0.23, which, according to [60], is regarded as poor reliability. While the definition of personality traits is firmly bound to their consistency over time and across contexts [8], capturing such behavioural consistency has been an issue for many test batteries. A meta-analysis of 31 studies revealed an average trait consistency of 0.43 in pet, shelter and working dogs, which was considered significant although the studies within displayed highly variable results ranging from -0.73 to 1.00 [31]. We found a similarly high variability in the test-retest reliability of the behaviours we assessed, with subtest reliability varying between a moderate ICC of 0.51 for the pointing test and a poorly reliable 0.06 for the fake dog test. Capturing consistent personality or behavioural scores over time is hence an issue not unique to the testing of street dogs in their natural habitat.

Intriguingly, most of the variables that were reliable over time were human-directed behaviours, including proximity to and tail wagging at the experimenter, body contact, and time of first approach. Likewise, the subtests in which the experimenter constantly interacted with the dog (i.e., human approach, pointing, tractability test) had the highest average test-retest reliability over time. Our results are consistent with other studies that report high reliability and strong heritability for sociability and biddability in pet dogs [e.g., 38,46,72] as well as dogs’ performance during pointing tasks [73,74], which might tentatively point to a genetic basis for these traits [2]. Simultaneously, while it has been shown that dogs already follow such pointing cues as pups [75] and that their performance is stable regardless of keeping condition, time spent with the owner, or training [73], we cannot rule out that the individuals’ life experience with human food provisioning played a part in these stable outcomes (as human socialisation seems to play a big part in pointing performance in other species [e.g., wolves: 75, goats: 76]).

More important in terms of street dog testing, however, is that the discrepancy in reliability between these social tests and the other subtests (fake dog, novel object, and begging test in particular) might be driven by the former’s high interactivity, leaving the animal little time to be distracted. Contrary to pet dogs, free-ranging dogs have been shown to gaze at an inattentive person less than at an attentive person [29]. Moreover, testing in the free-ranging dogs’ natural habitat means testing in highly variable and frequently disturbed environments. We tried to account for this by excluding all subtests with coded disturbances in a second analysis, however this did not improve reliability, possibly because we underrepresented the disturbances to our study subjects. Indeed, only six of the 35 dogs were alone in both tests (i.e., initial test and retest). For all others, additional dogs appeared during the test(s) and had to be ‘distracted away’ by the second experimenter. The appearance of other dogs is an unavoidable occurrence, and whilst testing can continue in such cases, it might explain why the subtests in which the dog had more time to focus on their environment without the direct engagement by the experimenter (fake dog, novel object, begging) showed a lower retest reliability. Whether behaviours like gazing and yawning and the entire begging test (which does not only lack interaction but also novelty) can reliably reflect a dog’s behaviour under such circumstances is hence questionable and requires further rigorous testing in alternative test setups on the streets.

To address the disturbance problem on the streets, we propose four not mutually exclusive approaches: 1) Have at least one additional person on the team that is solely responsible for distracting approaching dogs out of sight. 2) Allow for longer between-subtest-phases if the disturbance is temporary and let the dog return to an attentive state rather than pushing through the test. However, one should keep in mind that this changes the time dogs spend in the overall test procedure and carefully consider an influence on the respective goal. 3) If the distraction is substantial, terminating the test and either repeating (though the measures from the initial testing should be used for the repeated subtests) or resuming the test on a different day might be an option, depending on the respective test set up, goal, and familiarity with the population. 4) Meticulously coding any distraction might allow for correction of the disturbances, or at least provide deeper understanding of their influence on the behaviour, based on which decisions for future testing can be made.

Even after considering the disturbance problem, the consistency of the dogs’ behaviour in the fake dog and novel object tests was notably low. This might seem surprising since both tests are regularly used in dog temperament assessments [e.g., 34,63]. However, we could not find any reference to the retest reliability of fake dog tests. Moreover, even highly standardised studies exploring novel object tests and novelty-seeking and fearlessness traits often showed poor test-retest reliability (i.e., ICC<0.50) [e.g., Novelty-seeking: 0.48 in 38, Environmental Sureness: 0.16 in 62]. Accordingly, it has been proposed that test-retest outcomes are sensitive to the novelty of the object and situation due to learning effects [46]. Indeed, the fact that the free-ranging dogs explored the fake dog significantly more in the initial test while showing less fear in the retest supports the assumption that the animals partially remembered the initial encounter and adapted their behaviour accordingly. Although we made efforts to modify the appearance of the fake dog and novel object for the retest, the situation, procedure, and set-up remained the same and seem to have lacked the novelty needed to “fool” the animals twice within just six weeks. Nevertheless, it needs to be stressed that the street dogs frequently displayed appropriate conspecific, startling, and investigation behaviour in the initial presentation. Additionally considering that those subtests have been validated with fake vs. real dogs in case of the fake dog test [34,44] and deemed reliable despite similar retest problems in case of the novel object [e.g., 38,46], we suggest that these two subtests are still suitable to test the street dogs when only a single exposure to the test is needed. Additionally for the fake dog test, the social behaviours towards the stimulus were mostly exhibited in the first half of the subtest, suggesting that it is the initial reaction to the fake dog that is the most representative of conspecific behaviours. A shorter test duration and/or movement in the fake dog [e.g., 44] could thus be considered to decrease not only the influence of distractions, but also the risk that neophobia and boldness traits replace the conspecific behaviours as soon as the dogs realize that the fake dog is not in fact real.

At the same time, these two subtests act as a reminder of the importance of appropriate between-test-intervals, particularly for test batteries investigating behavioural profiles. With free-ranging dogs, the balance between waiting out potential learning effects and retesting early enough to find a considerable number of test subjects again can be difficult, especially if the study population is not well-known. Our substantial retesting rate of 72% within an average of six weeks was hence surprising, though the considerable variation between the individuals (min. 33, max. 76 days) has to be acknowledged. Meticulous notes on appearance, location, and particular behaviours were crucial in this effort. While working with a well-known population would streamline such efforts, our results show that it is not a prerequisite. In addition, using fewer memorable cues (e.g., the V-shaped wooden board in the Fake dog and Novel object subtest) might reduce learning effects and allow for a shorter between-test-interval with better outcomes than presented here.

Beside these two probable explanations for the low test-retest reliability in some subtests, an additional–though highly speculative–explanation might be that free-ranging dogs could be less genetically fixed in some behavioural traits than modern breed dogs and thus behave less consistently in general. This idea might not be too farfetched, considering that free-ranging dogs are still subject to free mate choice and the instability of their environment [13]. One might therefore speculate that behavioural plasticity for certain traits (e.g., adaptiveness to the highly dynamic group compositions, weather conditions, or anthropogenic changes) might be a valuable and thus widely retained trait in free-ranging dogs to deal with those constant environmental changes. Meanwhile, the directional selection for certain reliable breed characteristics in modern breed dogs [77,78] and the resulting considerable loss of genetic variation [79] might have decreased behavioural plasticity in our modern breeds. Indeed, the study of reaction norms in wild fish and birds provides cautious evidence that behavioural plasticity (i.e., intra-individual variability) can have a heritable component [80,81]. And though a meta-analysis targeting the question of whether behaviour-selected working-breed dogs were more fixed in their behaviour than non-selected non-working breed dogs could not find any support for this hypothesis [31], it was reported that an unstable rearing environment does not seem to have an effect on the consistency of behaviour, further highlighting the stability of certain traits in breed dogs [73]. Exploring how this compares to free-ranging dogs is an exciting avenue for further research.

A last consideration regarding the test itself pertains to the exclusion of the play component in the Human approach test and what that teaches us about the need to consider one’s study subject in all its facets. Play is frequently used in temperament tests to assess sociability [32,37,43], and the retainment of play behaviour into adulthood has also been suggested as a consequence of domestication in several species [reviewed in 45,82]. When applying this subtest with the street dogs, the low engagement rate and frequent fearful behaviour towards the toy might be explained by the negative or lack of experience such unowned free-ranging scavengers had with pulling on an object with a human [22]. Similarly, [18] suggested that free-ranging dogs’ lower persistence in comparison to pet dogs was likely due to their need to preserve energy and a lack of experience with human-manipulated items. On the other hand, our anecdotal observation with free-ranging populations indicates multiple incidents of conspecific and object play. Whether play behaviour does indeed persist into adulthood in free-ranging dogs in the conspecific context or solitary play, but not with a human partner, or whether our choice of method was just ill-considered, remains to explored and emphasizes how carefully the subject population, its ecology and life experience must be considered despite our extensive experience with the species itself.

Finally, two important limitations need to be addressed in terms of free-ranging dog testing as well as the representativeness of this study. First, the reliance on voluntary participation when testing adult free-ranging dogs creates a natural bias for more social animals, especially in longer and more interactive tests, and might be particularly challenging for feral populations (i.e., not reliant on human settlements). This bias is somewhat reduced in our Human approach test since the only selection criteria for starting a test was that the dog was seemingly alone. Hence, even the dogs that ran away fearfully upon approach were analysed as part of our dataset. The fact that we were able to retest more than half of these ‘escaping’ individuals and they manifested the same behaviour on second testing provides evidence that this measurement was indeed reliable. Nevertheless, these individuals were lost for all the following subtests and are therefore not part of the majority of the reliability analysis. At the same time, constraining a free-ranging animal would undoubtedly alter its natural behaviour in a test situation. Explorable options to test such dogs could be the use of shorter tests with little human presence wherever possible (e.g., reaction- or exploration-based tests) or testing populations that now live in a fenced shelter but genetically belong to the clade of free-ranging dogs and had spent most of their lives roaming freely. However, the influence of recent life experiences and social context in the shelter on the target behaviour have to be strongly considered [21]. Secondly, the fact that our study population has previously been recognized as exceptionally friendly [19] did not allow us to analyse the reliability of aggressive and many stress-related behaviours. For example, only 16.7% of our sample fled immediately and none showed aggression, while more than 50% of an Ethiopian village dog sample fled upon approach and 11% reacted aggressively (though one needs to consider that these dogs were not always tested alone or outside [30]). While expanding the presented test battery to other populations would initiate the characterisation of different free-ranging dog populations and what may affect differences in their behavioural profiles, we acknowledge that these limitations are likely difficult to be overcome. Extreme fear or the slightest sign of human-directed aggression will necessarily lead to aborting the test (particularly considering the risk of lethal disease, [e.g., 17]). And while using a fake dog rather than a real dog was indeed an attempt to test for the presence of aggression in a safe way, collecting additional observational data (regarding both human and conspecific directed behaviour) of the same individuals may be a complementary method allowing for multiple measures of specific behavioural traits.

In conclusion, the easy implementation, successful participation of most street dogs, and the high coding reliability presented here confirm that free-ranging dogs can be tested in longer test batteries with complex ethograms despite the challenging natural conditions. The wide variability in the test-retest reliability across subtests demonstrate that the management of disturbances, the choice of set up and test-retest intervals, and careful considerations of the subjects’ life experience and ecology are central challenges when testing free-ranging dogs in their natural habitat. At the same time, the acceptable reliability in some of the subtest demonstrates that under well-chosen conditions, testing those populations is possible and opens a vast array of intriguing pathways to explore [22]. Bearing in mind the suggestions made in the discussion, going forward the test battery presented here can be a valuable method to test questions about cognitive traits and behavioural profiles in such an understudied population as free-ranging dogs.

Supporting information

Acknowledgments

We thank Professor Ikhlass El Berbri for her support and collaboration in this project. Additionally, we thank the rest of our team in Morocco (in particular Andreas Berghänel, Haytem Bouchri, Manon Delaunay, and Magdalena Juskaite) without whom the project would not have been possible. We are also grateful to the population and authorities in the Sous-Massa region for supporting our field work. Finally, we want to extend our gratitude to the two reviewers for their valuable feedback and suggestions.