Subject and preliminary training
Our subject was an 11-year-old female cat, called Ebisu, living with her owner in Ichinomia (Japan). The owner, Fumi Higaki, is a professional dog trainer, experienced with the use of the Do as I Do method to train dogs. She reported that the cat had always been exceptionally motivated for food, a condition that allowed her to train the cat relatively easily by applying the same operant conditioning-based training methods commonly used for dogs. Ebisu was trained by her owner with the Do as I Do protocol to match her behaviour to actions that were demonstrated by a human (Topál et al. 2006; Fugazza and Miklósi, 2014). The Do as I Do training took place in the pet shop of the owner, where the cat lived, between May and September 2019. The training protocol is based on Topál et al. (2006) and Fugazza and Miklósi (2014) and it involves two steps. First, the subject is trained by operant conditioning techniques to match her behaviour to three familiar actions (i.e. actions that the cat was previously trained to perform on verbal cue) demonstrated by her owner, on command “Do it!”. Second, this command is generalized to three other familiar actions, after which the “Do it!” command can be used as a rule for the subject to reproduce novel actions presented by the demonstrator.
The actions that the owner trained and taught to the cat to imitate during the Do as I Do training included: spin, stand up on the hind legs, touch a wobbling toy with a paw, open a little plastic drawer, bite a rubber string and lay down (Table 1). The training included overall 21 training sessions of 3–10 trials each (Table 2). After the 21st training session, the owner taught the cat two new actions using the “action matching rule” (i.e., demonstrating the action and giving the “Do it!” command): open a sliding lid and climb with forearms on a book. When the owner demonstrated opening a sliding lid—i.e. sliding to a side the lid of a stainless-steel container—Ebisu successfully slid the lid to a side on the first trial. When the owner demonstrated climbing with forearms on a book, at first, the cat touched it with one paw without moving from her sitting position. In the next 4 similar demonstrations, the cat did the same; therefore, the owner placed the book further from the cat (at approximately 30 cm from her). On the 1st occasion with this new set-up, Ebisu placed her forearms on the book, thereby matching the demonstration (Table 1).
In March 2019, the cat was diagnosed with renal disease (stage 3), but the owner reported that her motivation for food and her physical condition had been apparently good and stable until September 2019, when the cat started to show a decreased appetite and a lower level of activity.
The tests were conducted in December 2019 in the pet shop where the training took place, in the evenings, when the shop was closed to the public. In each test, a cardboard box (27 × 18 × 10 cm) was displayed on a table (180 × 90 × 72 cm, the same table that had also been used during training). The owner was standing in front of the table and provided cues to induce the cat to go on it and to sit in front of her. The owner and the cat were facing each other, and the box was placed laterally at 50 cm from the cat’s position (Fig. 1).
At the beginning of each trial, the owner attracted the cat’s attention using food, petting and vocalizations and induced the cat to sit and stay in front of her using signals known by the cat. As soon as the owner noticed that the cat was looking at her (which could be further encouraged by vocalizations), the owner performed the demonstration, then returned to her starting position facing the cat and gave the “Do it!” command. The cat’s behaviour was then observed for maximum 20 s (or until the cat performed an action). To avoid any possible inadvertent cueing on the part of the owner, the owner looked straight ahead while giving the “Do it!” command (Fugazza and Miklósi, 2014). We also note that the two-action method intrinsically controls for cues that may direct the animal to a given direction or object because, even a directional cue inviting the subject to move towards the object would not provide information as to what action to perform on it.
Following the two-action procedure (Akins and Zentall 1996), two actions of similar difficulty (A and B) were defined in advance but, while in previous work where the two-action method was applied, only one of these actions was ever demonstrated to a given subject (between-subject design), we demonstrated both actions—in different test sessions—to our subject (within-subject design).
We used the two following actions on the box: Action A: the owner raised her right hand and touched the box with it; Action B: the owner bent down to rub her face on the box (Fig. 2 and video S1). The two object-related actions (A and B) were chosen to represent tasks of similar difficulty (simple interactions with objects), and similar performance between the two actions confirmed this (see Results). While touching an object with paw was a trained action (but the object on which the demonstration was done was novel), rubbing the face against something had never been trained.
Due to her disease, Ebisu’s motivation for food had reduced significantly in the period when we conducted the test (December 2019). This forced us to expose her to a limited amount of trials per test occasion and to a limited number of test occasions as well. Every test session consisted of 3 trials in which the same action (either A or B) was demonstrated, every time followed by the “Do it!” command. We planned to run 6 test sessions of 3 trials each, one test session per day with an inter-test interval of minimum 1 day, maximum 2 days. In half of the test sessions, we demonstrated action A and in the other half, we demonstrated action B. The action demonstrated (A or B) was semi-randomized, so that the same action would not be demonstrated in more than 2 sessions in a row. We kept the number of sessions and trials limited to ensure that the cat would remain motivated. Due to unforeseen circumstances, on one day, it was not possible to test the cat (the alarm of the pet shop got activated, the cat was frightened and hid behind a closet for the whole test). Therefore, we did not test the cat on that day and, instead, we ran two test sessions in the next test occasion, two days later. Ebisu’s health condition did not allow further testing; therefore, we had to refrain from carrying out more test sessions and from testing her on other actions.
Data collection and analysis
Video recordings of the cat’s behaviour following the ‘Do it!’ command were used to investigate whether she matched the demonstrated action. Out of the 18 trials, 2 trials were excluded from analysis, one because the cat did not watch the demonstration, and the other because the cat responded by touching the object with both her face and her paw (this happened in the first trial and this trial was excluded from the action matching analysis, but not from the object matching analysis).
We carried out two analyses. Although our aim was testing presence of action matching, we observed that, unexpectedly, in some trials, the cat matched the demonstrated movements but did not perform them on the object (see results). Therefore, we carried out two separate analysis, one for action matching and one for object matching.
To ensure unbiased coding, videos of the cats’ behaviour after the “Do it!” were watched by the coder without knowing what action had been demonstrated. The cat’s behaviour was noted and then compared to the description of the demonstrations to determine action matching. For the action matching analysis, the response of the cat was considered as matching the demonstrated action if the cat performed a movement similar to the one demonstrated by the owner, using a matching body part: when the human demonstrator raised a hand to touch the object, the cat’s response was considered as matching if it raised a front paw either to touch an object or mimicking the movement of raising a hand/paw without touching an object (both coded as “paw action”—action A); when the human demonstrator rubbed her face on the object, the cat’s response was considered as matching if she rubbed her face on something, or if she mimicked the face-rubbing movement without touching any object (both coded as “face action”—action B). If the cat performed any other than the demonstrated action, including action A when action B was demonstrated or vice versa, an action that differed from action A and B, or no action at all, action matching was considered unsuccessful. For the object matching analysis, the response of the cat was considered as matching the object if the cat interacted with any body part with the object touched by the demonstrator.
Statistical analyses were carried out using the R statistical environment (v. 3.4.2, R Development Core Team 2017). First, we calculated whether the cat matched the demonstrations above chance using Binomial tests, setting chance level at 0.5 because there were two actions that were demonstrated. However, we note that this is a conservative analysis, because the cat could potentially perform any behaviour, not only the two demonstrated actions (so that the applied 0.5 chance level is an upper limit).
Then, we followed the analysis of the two-action procedure described by (Akins and Zentall 1996), but adapted it to our experimental design and the single binary response variable we had for each test. In this analysis, we focused on the tests in which the cat performed action A or B (15 trials) and introduced a new variable ('action A') which was coded as ‘1’ if the action performed by the cat was A and ‘0’ if the performed action was B. Action A (binary response variable) was then analysed in binomial Generalized Linear Models (GLM, R package ‘lme4’, Bates et al., 2014) with demonstrated action as an explanatory variable (factor with two levels: A or B). Initial models of action matching included test session and trial, but these variables were kept in the final model only if they had a significant effect (based on AIC values). We used likelihood ratio tests (LRT) to investigate the effects of explanatory variables; we report χ2 and P-values of likelihood ratio tests of models including and excluding the explanatory variable. If action matching was observed, as opposed to other alternative processes, such as stimulus enhancement or goal emulation (Zentall 2006), we expected the demonstrated action to explain whether action A or B was performed by the cat.
For the object matching analysis, we calculated whether the cat interacted with the object more often than what was expected by chance using a Binomial test, considering the two possibilities of whether or not the cat interacted with the object.