Subjects
The study included 36 Welsh mares (Equus caballus) aged 7.9 ± 2.4 years (mean ± SD) reared at the Animal Physiology Experimental Unit PAO (UEPAO, 37380 Nouzilly, France, DOI: 10.15454/1.55738963217 28955E12), INRAE. These mares lived in groups in indoor stalls (3-sided shelter, open on one side) bedded with straw with environmental enrichments and access to an outdoor area 12h a day. Fodder and water were available ad libitum. In this facility, a general health status check is performed twice a day by keepers and a more extensive clinical exam is performed at least once a year by a vet. No hearing or vision issues have been reported in the horses included in this study. These horses are used only for research purposes and are handled daily by humans. Thus, they have the opportunity to experience human emotions including joy and sadness expressed by caregivers and researchers.
Experimental setup
The experimental setup was similar to that used to investigate cross-modal recognition of human anger and joy by horses as well as cross-modal recognition of adults and children (Trösch et al. 2019a; Jardat et al. 2023). A horse was placed in the middle of the stall and attached with two loose ropes. Films were projected on two projection screens (1 × 2 m) placed in front of the horse on the right and left. For safety reasons, an assistant stayed with the horse to ensure that she did not panic or become entangled in the ropes. The assistant stood along the wall at the level of the horse; on the right of the horse for half the horses and on the left for the other half of the horses. They never interacted with the horse during the tests but remained still with their head down. They were instructed to stop the test if a horse panicked; no horses panicked during the experiment. The experiment was filmed by two cameras, one in front of each screen. The horses were equipped with a heart monitor system composed of a sensor placed on the horse and a wristwatch showing real-time heart rate values and recording beat-to-beat (RR) values (Polar Equine 149 RS800CX Science, Polar Oy, Finland). An overview camera (GoPro Hero Black) allowed the experimenter to follow the experiment from outside the stall and control the projected videos and the sounds accordingly.
Stimuli
The videos and voices were extracted from the GEMEP Core Set database (faces: 02sad, 02joy, 07sad, 07joy, 09sad and 09joy; voices: 02joy, 02sad, 06joy, 07joy, 07sad and 10joy - (Bänziger et al. 2012), a validated psychological database for experimental research on emotion perception. The horses were therefore not familiar with either the faces or the voices. Pairs of videos of the same woman expressing sadness and joy were projected simultaneously, one on each projection screen, while the voice of another woman expressing sadness or joy was played over a loudspeaker placed between the projection screens (Fig. 1). The faces were projected at a height of approximately 160 cm and were approximately twice the size of a real human face (as in Trösch et al. 2019; Jardat et al. 2022b). The films shown during the tests lasted 115 s and were divided into six sections of 15 s separated by 5 s intervals of black screen (Fig. 2). Each section was composed of sadness and joy videos of the same woman (presented simultaneously) accompanied by a voice, repeated seven times. Four different films were generated. In each film, three women appeared twice each (once on the right side and once on the left side), in different orders. The side of the congruent video (i.e., the video that matched the emotion of the voice) was semi-randomly distributed. The side that an emotion was shown on and the emotion expressed by the voice were also semi-randomly distributed. The horses were randomly assigned to view one of the four films, such that the order in which the women appeared and the side on which each woman expressed joy and sadness were counterbalanced among the horses.
Habituation phase
During the habituation phase, the horse was led to the middle of the stall facing the screen. She was attached loosely to two ropes, the assistant took their position, and the habituation phase began. Identical scenes of nature were projected on both screens, while the assistant monitored the horse’s heart rate on the Polar watch. If the horse moved such that her body formed an angle of less than 45° with one of the ropes, the assistant gently repositioned her to face the wall in the middle of the two projection screens. The criteria to consider that the horse was habituated to the setup were that the horse had to be calm (not neighing, pulling on the ropes, or trying to turn around and leave) and her heart rate had to have stayed below 80 bpm for two consecutive minutes (corresponding to a moderate physical effort or psychological stress - Visser et al. 2002; von Lewinski et al. 2013; Trösch et al. 2019b; Gouyet et al. 2023). These criteria allowed to make sure that the situation was not too aversive for the horse and that she was ready to pay attention to the stimuli. Once these criteria were met, the test phase began immediately. If these criteria were not met after five minutes, the session ended, and a new session was scheduled for the next day. Of the 36 horses initially included in the experiment, 21 met the criteria on day 1 and therefore continued with the test phase on the same day. Seven horses met the criteria on day 2 and continued with the test phase on that day. The remaining 8 horses did not meet the criteria on day 1 or day 2, indicating this setup was too aversive for these individuals. Therefore, these horses were excluded from the experiment at this stage. The 28 horses that successfully habituated moved forward to the test phase.
Test phase
Immediately after the horse met the habituation criteria, videos of women expressing sadness and joy were projected (see the Stimuli section). The conditions were the same as those during the habituation phase. If the assistant needed to reposition the horse, the experiment was paused during a black screen interval and resumed once the assistant had repositioned the horse and was back in place. This was a rare event that occurred only five times during the whole experiment (in one trial for two horses and three trials for another horse, over a total of 168 trials for all horses). The assistant was unaware of the side on which each video appeared. At the end of the test, the horse was led directly back to its stall.
Behavioral and physiological analyses
The videos of the test phase were analyzed with BORIS software (Friard and Gamba 2016) by a first coder. The videos did not contain sound and the screens were not visible in the videos, so the coder was blind to the side on which each emotion appeared and the location of the congruent video. A horse was considered to look at the right screen when her left eye was visible from the camera placed under the right screen, and vice versa. She was considered to be attentive to the right screen when her right eye could also be seen from the camera placed under the right screen (i.e., both eyes visible from this camera), with both of her ears oriented forward; and vice versa. Thirty percent of the videos (9 individuals) were analyzed again by a second coder to evaluate scoring reliability. Interclass correlation coefficients (ICC) and their 95% confidence intervals were calculated; they were equal to 0.822[0.75,0.88] for the looking duration and 0.91[0.86,0.93] for the duration of being attentive to the videos, which are considered good to excellent reliabilities, respectively (Koo and Li 2016). From the analysis of the first coder, the following variables were quantified: for each section, the total looking duration and duration of being attentive to each video, the number of looks and the duration of the first look toward each video, and the latency between the start of a section and the first look toward each video.
RR data were extracted from the Polar recordings. A visual correction was applied to eliminate artifactual beats (as recommended by von Borell et al. 2007, see Supplementary Information (SI) – Figure S1). RR values were converted to heart rate values and the difference in heart rate (in beats per minute—bpm) between the last and first 5 s of each section was calculated. Data were missing for three individuals due to technical issues with the heart rate monitor. Therefore, the heart-rate analysis included data from 25 individuals.
Statistical analyses
All statistical analyses were performed using R 4.1.2 (R Core Team 2021), and figures were created using the packages ggplot2 (Wickham 2016), ggpubr (Kassambara 2020) and survminer (Kassambara et al. 2021). Significant differences were considered at p ≤ 0.05 and tendencies were considered at p ≤ 0.1.
Behavioral analysis
The durations of looking at and of being attentive to each video, the number of looks, and the duration of the first look toward each video were analyzed according to the congruence of the video with generalized linear mixed models (GLMMs) from the package glmmTMB (Brooks et al. 2017), using linear, Tweedie and Poisson distributions as appropriate for the given variable (see Table 1). The latency to look at the video was investigated according to the congruence of the video with the sound, and to the emotion displayed in the video, using survival analyses from the package coxme (Therneau 2022). This allowed us to include in the analysis horses that did not look at one of the videos at all during a section.
For each of the five response variables, an initial model was constructed to assess the interaction effect of video-sound congruence with the emotion displayed in the video. The side of the watched video was added to control for potential side biases. Horse identity was added as a random effect to account for individual variation in paired data, as each horse was tested six times. Distributions, within-group variance and homoscedasticity of the residuals for the GLMM were checked using the package DHARMa (Hartig 2021). For the survival analysis, outliers, the linear assumption and the proportionality assumption were evaluated graphically. If a horse did not look at one of the videos during the test, the latency to look at this video was set as censored (this happened 56 times). The best version of each model was selected according to model comparison via two-tailed ANOVA with the null model and simpler models (without the interaction, then without each variable of interest). The selected models are presented in Table 1 (see SI - Table S1 for the detailed results of each ANOVA).
Additionnally, outliers were identified using the boxplot method and the GLMM analyses were repeated (Aguinis et al. 2013), showing they did not influence the results (see SI – Table S2). Moreover, the influence of the experimenter manipulations on a few horses (see Methods) was checked, revealing it did not influence the results (see SI – Table S3).
The duration of the first look toward each video was also analyzed using a binomial transformation. For each horse, the variable was set at 1 if the first look (in the first section) toward the incongruent video was longer in duration than the first look (in the first section) toward the congruent video, and was set at 0 if it was not. The number of horses that looked more at the incongruent video was compared to chance using a one-tailed binomial test (binom.test with p=0.5 and n=28).
Heart-rate analysis
Heart-rate variation was investigated with GLMMs from the package glmmTMB (Brooks et al. 2017), using a generalized Poisson distribution (family=genpois). A model was constructed to assess the effect of the voice heard. Horse identity was added as a random effect to account for individual variation in paired data, as each horse was tested six times. Distributions, within-group variance and homoscedasticity of the residuals were checked using the package DHARMa (Hartig 2021). The model was selected according to model comparison via two-tailed ANOVA with the null model (Table 1, see SI - Table S4 for the detailed results of the ANOVA).