. Data set 1: staged play with a target dog
2.1.1. Subjects
We paired a single medium-sized female dog with 33 play partners of differing sizes and breeds. This approach enabled the selection of test subjects that were bigger, smaller and the same size as the standardized play partner. Although this raises unavoidable issues of pseudoreplication, we did so for three reasons. First, we wished to ensure that all play partners were unacquainted with each other so that recorded rollovers could not be ascribed to any prior establishment of dominance. Second, and following this, by using a known dog with a placid temperament, we wanted to allow for the possibility that being paired with an unknown dog would elicit behaviors associated with dominance and subordination, while minimizing the risk of agonistic escalation. Third, by having a standard target dog for each of the subjects, it increased the chance that any differences among the dogs, due to differences in relative size to the play partner, was not due to possible individual differences in the partners used.
We used opportunistic sampling of owners at a pet store in Lethbridge, Alberta to recruit subjects. We asked the owner(s) if their dog was well-socialized (i.e., did it have experience of daycare centers or playgroups) and whether they could recall any instances in which the dog had been involved in aggression. Dogs with a history of overt fighting were excluded from the study. We obtained body weight and height data from the owners. All procedures were approved by the University of Lethbridge Animal Welfare Committee.
2.1.2. Procedure
Play sessions took place in an indoor laboratory containing a cordoned-off play arena measuring 5.5 m × 3.5 m. This area included a door that opened into a small room in which the target dog was placed before the session began. Each session involved only the target dog and a single partner. On arrival, owners brought their dogs into the laboratory, placed them in the arena and remained with them until they had settled in. Once the dogs were relaxed, the owners left the laboratory and the target dog was released from the holding room. Two researchers remained to record the session and to manage the dogs. A pilot study indicated that dyads that were left alone in the arena tended to focus their attention on the researchers and so did not interact with each other. Accordingly, to overcome this, we spoke to the dogs or petted them until they either started playing or had ignored each other for 5 min. Data collection took place between 6pm and 8pm, from June to August 2011.
We used a Samsung SD camcorder to document sessions and began filming once the dogs engaged with each other, either by coming face-to-face and making prolonged eye contact or when one of the dogs sniffed the other. Filming was ceased if the dogs had not interacted with each other for more than five consecutive minutes.
2.2. Data set 2: YouTube videos of playing dogs
To obtain video footage of unique pairs of playing dogs, we searched YouTube, using the keywords dogs and playing, and selected 20 videos in which the beginning and end of play bouts were clearly discernible. We selected 10 videos in which partners were of approximately the same size and 10 in which the two dogs were judged to be different in size relative to one another.
2.3. Data extraction
We used the Free Video to JPEG Converter (v. 5.0.6. build 221) software to reconfigure the video as individual frames (25 fps). When viewing the recorded sessions, we used the ethograms provided by Bekoff, 1972 and Bekoff, 1974, Horowitz (2009) and Ward et al. (2008). Playful interactions varied from rough-and-tumble play to the chasing of one dog by the other. In the latter case, we used the descriptions of canine body language by Abrantes (2005) and Handelman (2008) to distinguish between play and avoidance behavior.
The converted footage allowed quantitative estimates to be made of the number of rollovers performed by each subject, the frequency with which rollovers occurred, the duration of the play bout, as well as the duration of the supine phase of the rollover, in which the animal was on its back with all four feet off the substrate. Moreover, the videotaped material also allowed Eshkol-Wachman Movement Notation (Golani, 1976) to be used for detailed qualitative analysis of the contexts in which rollovers occurred. The analyses of the videotaped material were in three phases.
2.4. Data analysis
2.4.1. Qualitative analyses
Eschol-Wachman movement notation (EWMN) is a globographic system, designed to express relations and changes of relation between parts of the body, with the body (i.e., body and limb segments) treated as a system of articulated axes (Golani, 1976). An important feature of EWMN is that the same movements can be notated in several polar coordinate systems. The coordinates of each system are determined with reference to the environment, to the midline axis of the subject's body, and to the next proximal or distal limb or body segment. Its primary value in the study of social interactions is that the movement by one animal can be described as relative to the body of the other animal (Moran et al., 1981; and see Appendix A in Pellis et al., 2013 for a detailed outline of how the system is used to record the behavior of two interacting animals). By transforming the description of the same behavior from one coordinate system to the next, invariance in the behavior may emerge in some coordinates but not others (Golani, 1976).
We used EWMN to describe 20 rollovers occurring in playful interactions containing rollovers from data set 1 to identify when and how rollovers occurred. Then, EWMN was used to describe 20 rollovers in play fights from data set 2 were also notated as an independent evaluation of whether the contexts identified for the use of rollovers were common to a wider range of dogs. The categories of rollovers identified in this way were then subjected to quantitative analyses using the fuller data sets of available rollovers.
As the most subjective phase of the analysis was the EWMN descriptions, a procedure was implemented to ensure their objectivity. The principal observer (KN) produced all the EWMN scores. Twelve of the notated scores were randomly selected and read by another (SMP), who had not previously watched the video sequences. The reader then provided a verbal and written description of the sequences and these were compared to what the dogs actually did on the videotape. In each case, the naïve reader recreated the actual movements performed, confirming that the EWMN scores were an accurate descriptive representation of the behavior.
2.4.2. Quantitative analyses
For the analysis of data set 1 (play with a target dog), we analyzed only play partner data. As we found a strong positive correlation between the height and weight of the target dog and those partner dogs (N = 27) that did play (r = 0.85, N = 27, P < 0.001), we confine our presented analyses to height. Using weight did not affect the outcomes. We used logistic regression to determine the effect of body size and play bout duration on the probability that rollovers would occur during a bout. In the analysis of the duration of the supine phase of the rollover (hereafter ‘rollover duration’), where subjects contributed more than one datum to analysis, we ran generalized linear mixed models (GLMM) under restricted maximum likelihood estimation, with rollover duration as the dependent variable and subject identity as the random effect. We conducted all statistical tests using Stata 13 ( StataCorp, 2013) or JMP 10 ( SAS Institute, 2012). Tests were two-tailed, with alpha = 0.05. Again, data from the two data sets were analyzed separately so as to determine whether both converged on the same patterns.
3. Results
3.1. Qualitative analyses of rollovers
From the 40 sequences notated using EWMN, four distinct categories of rolling over were discernible. Given that few people are able to read EWMN scores, the types of rollovers are described in written form with photographs of sequences illustrating the main types. While several contexts were identified in which rolling over occurred, in none of the 40 interactions subjected to EWMN analysis, did any dog rollover in a manner that is consistent with submission. That is, no dog rolled over in response to an approach or aggressive action by the partner and did not remain passive in its back. A quantitative evaluation of all instances of rolling over was conducted to ascertain whether such submissive rollovers were ever present (see Section 3.3).
3.1.1. Offensive rollovers
During rough and tumble play, dogs compete to make contact with, and bite the nape and neck region of their partner (Fig. 1). To gain such access, the attacking dog may lunge forward and roll over onto its back, raising the head with open mouth, biting at the throat as the distance is narrowed. The roll occurs around the longitudinal axis and is coupled with a lateral component if the approach is from the side.