We tested 114 (68 female and 46 male) healthy adult domestic pet dogs (Canis familiaris) of various breeds (see Table S1 for detailed information). All subjects had been living with their human family for at least six months and were used to interacting with new people and being in unfamiliar environments. From this sample, 91 (56 females and 35 males) dogs were included in the final dataset according to the following criteria. In the observation phase any dog that showed signs of stress, did not look at both actors, or did not watch the interaction for at least five seconds was excluded. This last criterion was selected as this time has been shown to be sufficient for dogs to acquire information for emotion processing (Albuquerque et al. 2016). Six dogs were excluded in this stage. The next stage considered any interference during testing, e.g. sudden noises from the outside, owner interacting with the dog in any way. Seventeen further dogs were excluded at this point.
Ethical approval was granted by the University of São Paulo ethics committee and all procedures complied with the ethical guidance for the use of animals by the International Society for Applied Ethology. We obtained signed informed consent from owners for all subjects. Dogs were monitored throughout the entire experimental session to ensure that they did not exhibit signs of stress. Dogs had their owners close to them at all times and were only handled by them.
We used a controlled yet naturalistic setting where subjects had the opportunity to witness a social interaction between two completely unfamiliar humans, which involved differential indirect display of emotional expressions.
Dogs observed two female Caucasian actors of the same age who always wore the same clothing (within and between subjects), had their hair wrapped in a bun, wore no make-up nor jewellery and had no distinct marks visible on their faces or bodies. The actors underwent an extensive training period prior to the beginning of the study to ensure similarity in the expression of emotion and synchronisation of the movements (within and between dogs).
Each subject was tested in a single experimental session, which took place at the Institute of Psychology of the University of São Paulo (Brazil) in a rectangular experimental room (185 cm × 308 cm). All testing was done in the same inside space, which was marked for each object and person position and controlled for not varying between subjects, during the entire study. At one end of the room, there was a small table against the wall which had two bowls and a set of three identical black opaque discs (they were produced for the purpose of this experiment and the dogs had never seen, touched or interacted with them) on it. Two stools were placed one on each side of the table and a digital video camera was positioned below the table to give a clear frontal view of the dog (see Fig. 1). At the opposite end of the room, there were marks on the floor showing the positioning of the dog and the owner, the experimenter and another video camera.
Our experiment consisted of two phases, an observation phase and a response phase (see Fig. 1 for an overview of the experimental set up and the online supplement for additional experimental details).
The actors were initially standing still, quiet with neutral facial and body expression beside a table, upon which two bowls with desirable food were placed. We used small pieces of dried-meat treats to bait the bowls. The amount of treats, as well as the size of the pieces, varied according to the size of the animal, and was aimed at motivating the dog to approach the bowls. None of the dogs were food deprived. The only instruction provided to the owners was not to feed their dogs just before bringing them to the test setting.
During the observation phase, dogs were positioned two metres away from the actors/demonstrators. The interaction then proceeded as follows. One actor (giver) turned and picked up one of three identical opaque black discs placed on a table (see information about the experimental room below), handed it to the other actor (receiver) and returned to her initial position. The aim was to create a giver-receiver dynamic. This was used because it is a common and natural type of interaction between two people that dogs are used to watching and because it would allow us to tease apart whether the dogs were making their choices based on the motivation/role of the humans, or, based on the humans’ emotional state. The three black discs were chosen to guarantee intrinsically neutral objects.
The giver remained neutral throughout the entire session. The receiver turned to receive the disc from the giver, held it, looked at it and then reacted to it. The reaction could be positive (happy), negative (angry) or neutral. The receiver turned back to her initial position while holding and looking at the disc and exhibiting the designated emotional expression (Figure S1). Then, while continuing to react emotionally, she looked ahead, turned towards the table and placed the object back with the set of discs. The emotional reaction was completely silent. This was repeated three times, each with a different disc.
The entire interaction lasted 30 s (Videos S1–S3). We used silent displays to control for variation in the demonstration of the interaction between subjects, which can easily occur in a naturalistic setting when using vocalisations or emotionally charged sounds that are not pre-recorded. Pre-recorded sounds suffer from a loss in natural information and sonority, possibly disrupting acoustic perception and was therefore not used (Fukuzawa et al. 2005). The entire demonstration was inspired by real life social interactions between two human beings, which can be extremely nuanced across contexts and can occur with and without the display of emotions (see videos S1–S3).
After the third exchange, each actor sat on a stool and remained looking down (reading a paper) with a neutral body posture and facial expression. Upon sitting, the situation could proceed in two distinct ways: (i) each actor held a baited bowl next to their body and made the food directly accessible to the dogs (direct access) or (ii) the actors left the baited bowls on the table. These were stacked one on top of the other thus creating what was apparently a “single” bowl. The food remained unreachable (indirect access). Position of each actor (left or right), role of each actor (giving or receiving), emotional display (positive, negative or neutral) and food accessibility (direct or indirect access to the food) was randomised between subjects.
Once the interactions between actors were completed, the response phase started. During the response phase, the dogs were released and their spontaneous behaviour was recorded for 30 s (videos S1–S3). We coded dogs’ choices along with visual exploration, body orientation, position in the room and sniffing behaviour.
Actors never looked at, nor interacted with, the dogs in either phase of the experiment. Each dog was tested only once to control for habituation and learning effects and ensure independence of the data. Of the analysed sample, 27 dogs were tested in the positive condition (happy receiver, neutral giver), 33 in the negative conditions (angry receiver, neutral giver) and 31 in the neutral conditions (neutral receiver, neutral giver); 48 dogs were tested in the indirect context and 43 in the direct context.
The primary variable used to determine dogs’ preference for interaction upon release was the dog’s Choice, which could be one of the actors, straight to the table or no approach. We analysed the choice behaviour as both binary (choosing or not making a choice) and nominal (specific choice) data, to test whether valence and food accessibility affected the decision to make a choice and the specific choices made (approach actors, table or other).
In addition, we analysed the gazing behaviour of the dogs in the different conditions, as determined by their head orientation towards a particular person’s upper body, which is potentially a more informative region and thus predicted to be associated with information gathering. We further examined looking (head orientation to any part of the persons present or elsewhere), body orientation (as per looking), position in the room (location relative to particular individuals or specific areas of the room) and sniffing (any part of the scene—see Table 1 for details of all behaviours) using multivariate and univariate analyses of variance. Even though gazing and looking are visual exploration behaviours, they are conceptually and operationally different. Gazing is the looking behaviour towards the upper half of the person’s body, whilst looking refers to the looking behaviour in general, that is, towards any direction. Thus, while gazing could only be directed at the humans present in the experimental area (two actors, owner and experimenter), looking could occur towards humans, the objects and to any empty space.
The videos obtained from the two digital video cameras were synchronised to allow robust coding of the behaviours and behavioural responses, which was done using frame-by-frame and real time speeds in Solomon Coder Beta (www.solomoncoder.com). We coded the dogs’ first choices as well as the seven exploratory behaviour categories listed in Table 1: approaching, gazing, looking, body orientation, positioning in the room, sniffing and touching.
Overall, we used a non-parametric approach to test choices and approaches (frequency variables) and multivariate and univariate models of analysis of variance to test the other categories (duration variables). The significance level was always 0.05. All analyses were conducted in IBM SPSS Statistics. Touching was a rare event and, therefore, no inferential analysis was conducted for this variable. A second researcher coded all behaviours for 25% of the videos, which were drawn randomly from the total sample of analysed subjects. Both primary and secondary coders were blind to the emotional expressions and experimental conditions. Inter-observer reliability measures yielded very high and significant correlation as well as concordance values. See tables S2 and S3 for details.
For the other categories, we used the relative durations of each variable (e.g. duration of looking at the emotional actor divided by the total duration of looking) in three different approaches to investigate the effect of emotional group and context (direct or indirect) on the responses towards: (1) the emotional actor and the unemotional actor (N = 60); (2) others, e.g. the owner and the experimenter (N = 91); and (3) the actors combined, regardless of their emotional status in all trials (N = 91). All assumptions for univariate and multivariate analysis of variance models were met and only very well adjusted models were used. When necessary, to meet the assumptions of each model, outliers were excluded (see Supplementary Materials for details). Since there were three levels within the factor group (positive, negative and neutral) in models 2 and 3, we used Scheffe post hoc tests to assess where were the differences between the levels of each model.
We initially focussed on the impact of emotion on choice. Second, we included the emotional group (positive or negative) and food accessibility (direct or indirect access) as fixed factors and the responses directed at the neutral actor (giver) and the emotional actor (receiver) as dependent variables in the model. Because we wanted to compare the responses in relation to positive and negative emotional expressions (see Fig. 2), we did not include the trials when there was no emotion exhibited by either actor.
The third analysis considered the potential effects of emotional group and food accessibility on the responses directed at others (i.e. owner, experimenter, out) and so examined all trials, i.e. positive, negative and neutral trials (n = 91).
The fourth analysis examined whether the two factors, “type of exchange between the actors” and “accessibility of food”, affected how dogs behaved towards the two actors who had been involved in the social interaction. We included emotional group and food accessibility as factors and the response directed towards the actors (combined, regardless of being neutral or emotional) as a dependent variable. Accordingly, we analysed the trials from all conditions, i.e. positive, negative and neutral trials (n = 91).
As each subject was assigned to only one emotional condition and dogs were tested only once, the data were independent and so subjects were not considered as random factors in the models (see Supplementary Materials for more details).