Analysis of Facial Expressions Made While Watching a Video Eliciting Compassion

Abstract

The aim of the study was to describe the spontaneous facial expressions elicited by viewers of a compassionate video in terms of the respondents’ muscular activity of single facial action units (AUs). We recruited a convenience sample of 111 undergraduate psychology students, aged 18-25 years (M = 20.53; SD = 1.62) to watch (at home alone) a short video stimulus eliciting compassion, and we recorded the respondents’ faces using webcams. We used both a manual analysis, based on the Facial Action Coding System, and an automatic analysis of the holistic recognition of facial expressions as obtained through EmotionID software. Manual facial analysis revealed that, during the compassionate moment of the video stimulus, AUs 1 = inner-brow raiser, 4 = brow lowerer, 7 = lids tight, 17 = chin raiser, 24 = lip presser, and 55 = head tilt left occurred more often than other AUs. These same AUs also occurred more often during the compassionate moment than during the baseline recording. Consistent with these findings, automatic facial analysis during the compassionate moment showed that anger occurred more often than other emotions; during the baseline moment, contempt occurred less often than other emotions. Further research is necessary to fully describe the facial expression of compassion.

Keywords

compassion emotion facial expression analysis

Introduction

Only a few studies have attempted to describe the facial expression of compassion (Baránková, Halamová, Gablíková, Koróniová, & Strnádelová, 2019; Eisenberg et al., 1989; McEwan et al., 2014). Opinions greatly differ on what compassion is. Researchers have considered it to be an emotion (Haidt, 2003; Widen, Christy, Hewett, & Russell, 2011), or a self-transcendent emotion (Stellar et al., 2017), a sensitivity to act to alleviate suffering (Gilbert, 2017), or a complex state consisting of multiple components (Bloom, 2016; Strauss et al., 2016). We think compassion is a multidimensional construct. According to Strauss et al. (2016), compassion includes five main elements: recognizing suffering, understanding the universality of suffering, feeling empathy, tolerating personal discomfort stemming from the suffering of others, and the motivation to alleviate suffering. According to previous research (Baránková, Halamová, & Koróniová, 2019; Halamová, Baránková, Strnádelová, & Koróniová, 2018), the general population, whether from a helping profession or nonexpert background, defines compassion in terms of emotional, cognitive, behavioral, biological, and evaluative aspects. Even if compassion proves to be an emotion in its own right, emotion is an important aspect of the holistic construct of compassion. As Ekman and Friesen (2003) have stated, compassion involves the nonverbal reflection of emotions, mainly in the face. We sought to investigate how the emotional part of compassion is reflected in facial expressions by testing previous findings regarding the nature of compassionate facial expressions (Baránková et al., 2019).

Facial Expression of Compassion

Basic emotions appear to be universal, meaning that basic emotional expressions are recognizable by people all over the world (e.g., Ekman, 1989, 1994, 1998). Ekman and Friesen (2003) classified surprise, fear, disgust, happiness, anger, and sadness as the six basic emotions. Others (Keltner, 2005a, 2005b; Widen et al., 2011) have referred to the universality of other emotions, such as contempt, embarrassment, shame, or compassion. Of note, however, the paradigm of basic emotions has also been criticized. Russell and Feldman Barrett (1999) challenged the legitimacy of categorizing basic emotions, arguing that emotion is too blurry a construct to be clear in its meaning as discussed by various researchers. These authors assert that only the dimensions of pleasure and displeasure may be universal, as these concepts exist in all languages.

A few decades ago, Eisenberg et al. (1989) explored participants’ facial expressions in response to a compassionate stimulus, and, afterward, they defined the facial expression of compassion as follows: lowered and flattened eyebrows with wrinkles above the nose, higher lids slightly lifted, and lower part of the face relaxed alternatively with open mouth. Haidt and Keltner (1999) later attempted to describe the facial expression of compassion in the study by Eisenberg et al. (1989) using the following action units (AUs): 1 = inner-brow raiser, 4 = brow lowerer, and 57 = head forward. However, when tested on another participant sample, this expression description could not be recognized. McEwan et al. (2014) created a data set of photographs showing the facial expression of compassion. When compared with the expression described by Eisenberg et al. (1989), most of the photographs in the McEwan et al.’s (2014) data set had signs of a smile (see Ekman & Friesen, 2003; Ekman, Friesen, & Hager, 2002) but no other recognizable facial expression. Baránková et al. (2019) described the facial expression of compassion elicited by a short video stimulus using the following AUs from the Facial Action Coding System (FACS; Ekman, Friesen, & Hager, 1978/2002): 12 = lip-corner puller, 7 = lids tight, 43 = eyes closed, and 56 = head tilt right. According to the results of an automatic software analysis, during a compassionate moment, the emotions of sadness and anger were present more often than they were during the baseline moment.

There is a gap in the research on the facial expression of compassion despite many other compassion studies published. Compassion for others and self-compassion are constructs that are considered beneficial to mental and physical health (e.g., Beaumont & Hollins Martin, 2015; Desbordes et al., 2012; Gilbert & Procter, 2006; Jazaieri et al., 2013, 2014; McEwan et al., 2014; Mongrain, Chin, & Shapira, 2011). We know which situations may evoke compassion (Dossey, 2007; Valdesolo & DeSteno, 2011); however, we do not yet know whether there is a universal facial expression of compassion and, if so, what it looks like. Data sets on numerous emotional expressions are available for scientific or therapeutic purposes (e.g., Baveye, Dellandrea, Chamaret, & Chen, 2015; Kanade, Cohn, & Tian, 2009), but, so far, only one data set includes the facial expression of compassion (McEwan et al., 2014). The McEwan et al. (2014) compassion data set consists of expressions made by actors and actresses based on an imagined compassionate experience. To our best knowledge, there is no data set of spontaneous facial expressions of compassion. Neither are there any databases of stimuli for eliciting compassion. There are only databases containing mixed emotions, and these may include some stimuli for eliciting compassion (Samson, Kreibig, Soderstrom, Wade, & Gross, 2016). It is difficult to elicit compassion in a standardized way in the absence of a database of stimuli for evoking compassion.

There have been a few studies about the efficiency of automated facial expression, mainly in the area of affective computing (e. g. Cohn, Reed, Ambadar, Xiao, & Moriyama, 2004; Lien, Kanade, Cohn, & Li, 2000; Tian, Kanade, & Cohn, 2002). First attempts at automated facial expression were done in laboratory conditions, but there has now been automated facial expression recognition in the wild, with reliable naturalistic data now (e. g. Li, Deng, & Du, 2017).

Aim of the Study

The aim of this study was to describe spontaneous facial expressions, elicited by a compassionate video stimulus, in terms of the muscular activity of single facial-AUs through a manual analysis using the FACS and through holistic recognition of facial expressions of emotion using an automatic analysis via EmotionID software recognizing basic emotions plus contempt and neutral expression. As detailed later, this study differed from prior research in that we used different video to elicit compassion, different statistical analyses, and we asked participants to define the most compassionate moment and also incorporated contempt. Our research questions were as follows:

Which of the AUs in FACS occur significantly more frequently during the compassion moment than during the baseline moment?

Which combinations of the AUs in FACS occur significantly more frequently than other AU combinations during the compassion moment?

Which emotions detected by Emotion ID software occur significantly more frequently than other emotions during the compassion moment than during the baseline moment?

Which combinations of emotions detected by Emotion ID software are more likely to occur than other combinations of emotions during the compassion moment?

Method

Participants

Our research participants were a convenience sample of 111 respondents aged 18-25 years (M age = 20.53; SD = 1.62). Men accounted for 11.7%. This participant sample size was determined in part by our resources for engaging in a time-consuming manual analysis of each video in which each minute of facial expression in reacting to the compassion inducing video can take up to 100 minutes to analyze (Rosenberg, 2005). Participants were undergraduate psychology students who were offered credits for participating in the study. Participants signed an online informed consent form, making clear that they could cease participating in the research at any time and could remove their answers and videos if they wished.

Measurement Instruments

Manual analysis of facial expressions: FACS

We used the FACS for the manual analysis of facial expressions (Ekman et al., 1978/2002) because it is the most used frequently coding system for analyzing facial expressions. FACS is based on the anatomy of facial muscles. The detailed manual created by Ekman et al. (1978/2002) contains all the changes in facial muscles that are visible to the naked eye. Every facial expression is the contraction of one or more AUs or action descriptors (hereafter in this study, we will refer to both as AUs). These AUs describe facial muscle and head and eye movements. A single facial expression can consist of one or a combination of AUs. The FACS is dependent on video recordings that can be played slowly and repeatedly and on the facial expression coding work of certified human coders (Ekman & Keltner, 1997), consisting in this study of three certified FACS coders whose independent detections of expression detection were subjected to intercoder reliability analysis. In our study, only occurrence of AUs was coded without intensity coding.

Automatic analysis of facial expressions: EmotionID software

We used EmotionID software for the automatic emotion expression analysis. Emotion ID is based on two main principles: webcam computer data and machine learning (Gablíková & Halamová, 2016). Viewers of facial expressions of the six basic emotions (anger, disgust, fear, happiness, sadness, and surprise) as well as contempt and neutral expression can recognize these emotions with 95% accuracy (Martinkovič in Baránková et al., 2019). Emotion ID software is designed around two approaches to emotion detection: geometric and appearance (e.g., Cohn & De la Torre, 2015). Software algorithms calculate the probability that a facial expression represents a given emotion based on facial change ratings from 0 to 1 occurring at every second of the video in real time (e.g., Lewinski, den Utyl, & Butler, 2014). According to Gablíková and Halamová (2016) Emotion ID software is a product of machine learning on thousands of pictures showing specific facial expressions of emotion that were rated first by human coders. Software EmotionID is a tool for time-efficient automated facial expression recognition, and it is available for research purposes for free.

Compassionate Video Stimulus

All authors of the study unanimously selected the short video clip (83 seconds) for evoking compassion. We also contacted multiple external experts on compassion regarding the appropriateness of the video stimulus. In line with previous findings (Baránková & Halamová, 2018; Baránková et al., 2019), we chose a shorter video with no words to prevent the participants from becoming distracted by reading subtitles or listening to the characters talking. The video meets the theoretical definition of compassion and shows a short story of an orphaned little girl in a junkyard who needs money for food.

Procedure

To establish proper conditions for the spontaneous elicitation of emotion, the participants were asked to watch the video alone, ideally at home (Hess, Banse, & Kappas, 1995). Following the informed consent process, participants answered sociodemographic questions. They were then asked to watch a short video while their faces were being recorded, and they were told that, afterward, they would answer questions about their experience of the video. Prior to recording participants’ facial expressions, we tested and calibrated the webcam to the light conditions and conducted face detection. The participants were then shown the short (83 seconds) video inducing compassion. The video (without words) was about an orphaned little girl in a junkyard. At the beginning of the video, a 3-second black screen enabled researchers to capture the participants’ neutral facial expression at baseline. These facial expression recordings were sent to a secure server where they were analyzed automatically via EmotionID software and manually by the three certified FACS coders.

Most Compassionate Moment in the Video Stimulus

After the participants watched the video, we asked them which moment in the video they found to be the most emotional. The statistically significant moment, χ²(1) = 73.89, p < .001, was one in which the little girl drank the last of the bottled water she had found in the junkyard. Then, we asked at which moment the participants felt the most compassion. In this case, two moments were statistically significant: the same moment when the little girl drank the last of the water, χ²(1) = 45.77, p < .001, and another moment showing a detailed close-up of the girl’s face, χ²(1) = 13.27, p˂.001. For the purposes of our subsequent analysis, we chose, as the most compassionate moment in the video (for a comparison to the initial three baseline seconds of a black screen), the moment when the participants felt the most emotional.

Data Analysis

We analyzed the FACS and EmotionID data using R version 3.5.1. (R Core Team, 2018) and the brms statistical package (Bürkner, 2017, 2018).

FACS data: manual coding

Bayesian statistical models are growing in popularity and are increasingly used in psychology research (Feinberg & Gonzalez, 2012; Wagenmakers et al., 2018) because they are able to avoid many of the problems inherent in classical (frequentist) statistical procedures, such as requirements for very large samples in order to apply the central limit theorem, issues with variance equality and balanced designs for group comparisons, the absence of outliers, and restrictions concerning the distribution of errors. Nonparametric alternatives to classical inferential statistics can cope with some of these problems, but they usually do so at the cost of reducing statistical power. Bayesian statistical models are available in many statistical software programs (e.g., Mplus and R), and modern computers substantially reduce computational effort and time associated with these models.

First, we needed to specify which model(s) was most appropriate for handling our data. For each participant, there were 59 AUs rated independently by three raters on two occasions: the baseline episode and the compassion episode. Therefore, there were 354 observations per participant and 39,294 total observations (354 observations × 111 participants). This meant that we had to fit the multilevel model to compensate for the possibility of a participant giving the same responses, for the same AUs, as judged by the same raters. We could not assume independence of errors for the individual responses, particularly for the AUs and raters. Consequently, our model had to include three group effects (or random effects in frequentist terminology): individual participant effects, AU effects, and rater effects. Second, these responses were coded as 0 (absence) or 1 (presence), they followed a binomial (i.e., Bernoulli) distribution. Third, the predictor (population effect or fixed effect in frequentist terminology) is one factor with two levels because all the respondents were rated during the baseline and compassion episodes.

We fitted several Bayesian models of increasing complexity to these data: (a) Model 0, with intercept-only and no-group effects; (b) Model 1, with a factorial predictor (population effects) for baseline versus compassion moments; (c) Model 2, with ID (respondents and group effect) to compensate for 354 identifications per respondent, and rater (group effect) to compensate for three raters; (d) Model 3, with ID, rater, and AUs (group effect) to compensate for 59 AUs; (e) Model 4a with differences at the different moments (baseline vs. compassion) for particular AUs (random slopes in frequentist terminology) to test whether some AUs were present more frequently during the compassion moment than during the baseline moment; and (f) to compensate for the excessive number of zeros, we fitted several models to take into account such distributions: Model 4b is the same as Model 4a but with a negative binomial distribution, Model 4c is the same as Model 4a but with a zero-inflated binomial distribution, and Model 4d is the same as Model 4a but with a zero-inflated negative binomial distribution.

Among several methods for selecting the best fitting Bayesian model, the most frequently used are deviance information criterion (Spiegelhalter, Best, Carlin, & van der Linde, 2002), widely applicable information criterion (WAIC; Watanabe, 2010), and leave-one-out (LOO) cross-validation (Vehtari, Mononen, Tolvanen, Sivula, & Winther, 2016). However, from the Bayesian viewpoint, deviance information criterion is not recommended because it is not an unbiased estimate of the true generalization utility (Piironen & Vehtari, 2017). As WAIC is asymptotically equal to LOO, and much easier to compute, we used WAIC to select the best fitting model—the model with the largest expected log pointwise predictive density (ELPD) or the smallest WAIC (WAIC is ELPD multiplied by −2 to obtain a result on the deviance scale). The model with the smallest WAIC has the highest predictive accuracy of the compared models.

Once we had selected the most accurate model, we checked its convergence. There are various diagnostic tools for assessing the successful convergence of Bayesian models: (a) Gelman–Rubin diagnostics (Gelman & Rubin, 1992)—Rhat value for all parameters should be close to 1 at convergence, and values substantially above 1 (larger than 1.10) indicate a lack of convergence; and (b) Heidelberger–Welch diagnostics (Heidelberger & Welch, 1983), which uses Cramer–Mises statistics to test whether the sampled values come from a stationary distribution (requiring us to calculate the half-width test); (c) Geweke diagnostics (Geweke, 1992), testing the equality of the means of the first part (10%) and last part (50%) of the Markov chain (concluding equal means if the distribution is stationary, with Z- scores from Geweke statistics above 1.96 or below −1.96 indicating a lack of stationarity); (d) Raftery–Lewis diagnostics (Raftery & Lewis, 1995), which test the accuracy of the estimation of quantile q (usually q = 0.025) and require as a minimum length the sample size for a chain with no correlation between consecutive samples (Estimate I or the dependence factor summarizes the extent to which autocorrelation inflates the required sample size, and values larger than 5 indicate too strong an autocorrelation); and (e) visual inspection of traceplots.

Each model was fitted with four Markov chains, each with 2,500 iterations, and 2,500 burn-in (discarded) iterations, so the overall number of sampling iterations was 10,000. Higher posterior density (credible) intervals (95%) are reported for each parameter. As the model link is logit, we reported the odds ratio (OR) for the population parameter—the increase/decrease in the probability of AUs occurring more frequently during the compassion moment compared with the reference group (the baseline moment). Our hypothesis was not that the occurrence of AUs would increase substantially in the compassion episode, but that some AUs would occur more frequently than others, and that the frequency of some AUs would be higher in the compassion episode versus the baseline episode. In other words, our hypothesis was about group effects and not population effect.

EmotionID data: automatic coding

For emotional ID data, we took measurements for each participant n several times and obtained 48 identifications (three identifications per eight emotions, both during the baseline moment and the compassion moment). This meant we had to fit a multilevel model that would compensate for participants giving the same responses. The dependent variable is a continuous and bounded variable defined at unit interval [0,1), which provides a measure of the proportions of emotions identified by the EmotionID software. Given the large number of observations with zero values and none at the upper bound, we carried out a zero-inflated beta multilevel regression model to study the factors affecting the differences among the emotions. This model consists of three submodels: (a) one estimating the probability of an identification being equal to zero (parameter α), (b) the second estimating the proportion of emotions where the identification is greater than zero and less than one (parameter μ), and (c) the final submodel making it possible to model precision while taking skewness and heteroscedasticity into account (parameter φ). We used the logit link function to define the zero-inflated and proportion submodels. We selected this link to specify the relationship between the linear predictor and the response variable because the results are easy to interpret in terms of the ORs, as is the case with a logistic regression model. Finally, we used the log link for the precision submodel. To compensate for the fact that each participant was measured several times (and provided several observations), we included the participant’s ID as the group effect (or random effect, in frequentist terminology).

For predicting the outcome, we used the same set of explanatory variables for the three submodels and included emotions, the difference between baseline moment and compassion moment and the interaction between the two. The first variable was emotions (eight emotions, an eight-level factor with neutral as the reference level), the second variable was the contrast between the baseline and compassion moments (a dummy variable indicating the difference between baseline and compassion, a two-level factor), and the third variable was the interaction between them (eight interactions, an eight-level factor with neutral at the baseline as the reference level).

We fitted several Bayesian models of increasing complexity: (a) Model 1 with predictors and group effect, where the zero-inflation and precision parameters were estimated only for the predictors, not for group effect; (b) Model 2 with predictors and group effect, where the zero-inflation and precision parameters were estimated for the predictors, but only the zero-inflation parameter was estimated for the group effect; (c) Model 3 with predictors and group effect, where the zero-inflation and precision parameters were estimated for the predictors, but only the precision parameter was estimated for the group effect; and (d) Model 4 with predictors and group effect, where the zero-inflation and precision parameters were estimated for the predictors and group effect. Although the WAIC is asymptotically equal to LOO, Vehtari, Gelman, and Gabri (2017) have demonstrated that LOO is more robust in finite cases with weak priors or outliers. Therefore, we used LOO to select the best fitting model—the model with the largest ELPD or the smallest LOO (LOO information criterion is ELPD multiplied by −2 to obtain a result on the deviance scale). The model with the smallest LOO had the highest predictive accuracy of the models we compared. In addition, we used the model stacking method for predictive distributions (Yao, Vehtari, Simpson, & Gelman, 2018) to compare the weights of the models.

Each model was fitted with four Markov chains, each with 5,000 iterations, and 5,000 burn-in (discarded) iterations, so the overall number of sample iterations was 20,000. Higher posterior density (credible) intervals (95%) are reported for each parameter. As the model link is logit, we report the OR for the population parameters.

Results

Manual Analysis of FACS Data

Model section

In Table 1, we report the parameters of all the models, and Table 2 shows that Model 4c (binomial with zero inflation) had the lowest WAIC information criterion, indicating that its predictive accuracy was the highest among the models. All the simpler models (Model 0, Model 1, Model 2, and Model 3) were less accurate, and so were the models with other distributions (negative binomial). As Model 4c, with the zero-inflation parameter (accounting for the excessive number of zeros), was most accurate, we retained it and proceeded to diagnose convergence associated with it. The fact that it was the most accurate of the models by no means entailed successful model convergence, as it might have only been the best of a group of very bad models. All the relevant convergence diagnostic tools had to be checked before we could begin interpreting its parameters.

Table 1.

Parameters of Models.

	Population effects	Group effects	Distribution
Model 0	Intercept only	–	Binomial
Model 1	Moments	–	Binomial
Model 2	Moments	ID + raters	Binomial
Model 3	Moments	ID + raters + AUs	Binomial
Model 4a	Moments	ID + raters + moments\|AUs	Binomial
Model 4b	Moments	ID + raters + moments\|AUs	Negbinomial
Model 4c	Moments	ID + raters + moments\|AUs	ZI binomial
Model 4d	Moments	ID + raters + moments\|AUs	ZI negbinomial

Note. AU = action unit; ZI = zero inflation.

Table 2.

WAIC Information Criteria for Models.

	WAIC (SE)	ELPD_WAIC (SE)	ELPD_WAIC_diff
Model 0	11,421.1 (239.4)	−5,710.6 (119.7)	−2,182.6
Model 1	11,418.6 (239.3)	−5,709.3 (119.7)	−2,181.3
Model 2	11,255.5 (236.3)	−5,627.8 (118.1)	−2,099.8
Model 3	7,380.9 (175.5)	−3,690.4 (87.7)	−162.4
Model 4a	7,108.4 (171.7)	−3,554.2 (85.9)	−26.2
Model 4b	7,671.8 (179.9)	−3,835.9 (90.0)	−307.9
Model 4c	7,056.0 (169.4)	−3,528.0 (84.7)	0
Model 4d	7,672.7 (179.9)	−3,836.4 (89.9)	−308.4

Note. WAIC = widely applicable information criterion; SE = standard error; ELPD = expected log pointwise predictive density.

Diagnostics of the selected model

The Gelman–Rubin diagnostics showed that all parameters had Rhat values of 1.00 except the variance of raters, which had a value of 1.01, indicating excellent convergence. All the parameters passed the Heidelberger–Welch diagnostics in the stationarity tests and only eight of the 126 parameters failed the half-width mean tests. All except four of the 126 parameters passed the Geweke diagnostics as well. The Raftery–Lewis diagnostics showed that no dependence factor reached five. A visual inspection of the traceplots also showed excellent model convergence. Thus, we can conclude that this model converged successfully, and its estimation of parameters is accurate.

Parameters of the selected model

We next inspected the parameter estimations of the selected model (see Table 3). The SD of ID, accounting for individual differences among respondents, was 0.92 meaning that the variance was 0.85. The best way to interpret this value is to compute the intraclass correlation (ICC) or the amount of variance explained by individual differences among variables in the class. The ICC for the ID was 0.04, meaning 4% of the explained variance was accounted for by individual differences among respondents. The SD of the raters was 0.52, and variance of this parameter was 0.27. The ICC for the raters was 0.01, meaning only 1% of the explained variance was due to differences among the raters; interrater reliability was relatively high. There were substantial differences among the AUs, as the ICC for this parameter was 0.79, meaning that 79% of the explained variance was due to differences among the AUs. Because there was very little general difference between the baseline moment and the compassion moment (the remaining parameter), AU differences account for practically all the explained variance.

Table 3.

Parameter Estimations of the Selected Model.

Population parameter	Estimation	EE	HPD CI 5%–95%	OR
Intercept	−5.09	0.54	−6.18 to −4.03	0.01
Moments	−0.41	0.32	−1.05 to 0.21	0.66
Zero-inflation parameter	0.34	0.03	0.28 to 0.40	N/A
Group parameter
ID	0.92	0.09	0.75 to 1.12	N/A
Raters	0.52	0.55	0.08 to 2.15	N/A
Baseline AUs	2.75	0.40	2.07 to 3.64	N/A
Compassion AUs	2.85	0.41	2.18 to 3.77	N/A
Correlation	0.87	0.05	0.77 to 0.94	N/A

Note. ID = respondents (N = 111); AUs = action units (N = 59); EE = estimation error; HPD CI = highest posterior density credible interval; OR = odds ratio; N/A = not applicable.

We also compared the most frequently occurring AUs (relative to other AUs) during the compassion moment and the testing moment (compassion vs. baseline) during which most compassion-related AUs occurred. It seemed entirely possible that an AU might occur more often than other AUs during the compassion moment but still not display a significant increase in the compassion moment compared with the baseline moment. On the other hand, the occurrence of an AU could significantly increase during the compassion moment compared with the baseline moment but still not be identified as occurring more often than the other AUs during the compassion moment. Of the 59 AUs, 18 AUs were identified as occurring significantly more often than the others during the compassion moment: 1, 2, 4, 5, 7, 12, 14, 17, 24, 25, 26, 43, 45, 55, 56, 61, 62, and 64 (see Figure 1). Some of these occurred more frequently in the compassion moment than during the baseline moment: 1, 4, 7, 17, 24, 43, and 55.

Figure 1.

Occurrence of AUs in compassion episode. AU = action unit.

Automatic Analysis of the Emotion ID

Model selection

In Table 4, we report the parameters in all the models, and Table 5 shows that Model 4 (zero-inflation and precision parameters are estimated for population effects and group effect) had the lowest LOO information criterion, indicating that its predictive accuracy was the highest among the models. As all the simpler models (Model 1, Model 2, and Model 3) were less accurate, we retained Model 4 and proceeded to diagnose convergence for the same reasons as was necessary for the FACS model selection.

Table 4.

Parameters of Models.

	Population effects	Group effect	Zero inflation	Precision
Model 1	Emotions × Time	ID	Emotions × Time	Emotions × Time
Model 2	Emotions × Time	ID	Emotions × Time + ID	Emotions × Time
Model 3	Emotions × Time	ID	Emotions × Time	Emotions × Time + ID
Model 4	Emotions × Time	ID	Emotions × Time + ID	Emotions × Time + ID

Table 5.

LOO Information Criteria for Models.

	LOO IC (SE)	ELPD_LOO (SE)	ELPD_LOO_diff	Weight (%)
Model 1	−959.4 (164)	479.7 (82)	N/A	0
Model 2	−2,248.8 (178.5)	1,124.4 (89.3)	1,289.37	1.6
Model 3	−1,143.6 (166.1)	571.8 (83.1)	−1,105.15	0
Model 4	−2,438 (181.3)	1,219 (96)	189.21	98.4

Note. LOO = leave-one-out; LOO IC = LOO information criterion; SE = standard error; ELPD = expected log pointwise predictive density; N/A = not applicable.

Diagnostics of the selected model

The Gelman–Rubin diagnostics showed that all the parameters had Rhat values of 1.00, indicating excellent convergence. All the parameters except one passed the Heidelberger–Welch diagnostics in the stationarity tests and only five of the 51 parameters failed the half-width mean tests. All except two of the 51 parameters passed the Geweke diagnostics as well. The Raftery–Lewis diagnostics showed that no dependence factor reached five, so the model had converged successfully using this procedure as well. A visual inspection of the traceplots also showed that the convergence of this model was excellent. Thus, the model had converged successfully, and its estimation of parameters was accurate. The effect size—Bayesian R square (the variance of the predicted values divided by the variance of the predicted values plus the variance of the errors, see Gelman, Goodrich, Gabry, & Vehtari, 2018)—was quite high at 0.55.

Parameters of the selected model

We next inspected the parameter estimations of the selected model. Table 6 summarizes the parameter estimates together with the corresponding standard errors, 95% credible interval, and ORs where applicable.

Table 6.

Parameter Estimates, EEs, 95% CI, and OR.

	Estimate	EE	95 % CI		OR
α (zero inflated)
Anger	5.18	0.32	4.56	5.81	177.68
Disgust	4.74	0.32	4.13	5.38	114.43
Fear	5.84	0.34	5.19	6.51	343.78
Happiness	6.24	0.35	5.56	6.94	512.86
Sadness	4.32	0.31	3.72	4.94	75.19
Surprise	4.17	0.31	3.58	4.77	64.72
Contempt	4.57	0.31	3.97	5.20	96.54
Time2	0.38	0.33	−0.27	1.03	1.46
Anger:Time2	−0.18	0.42	−0.99	0.65	0.84
Disgust:Time2	−0.67	0.41	−1.46	0.13	0.51
Fear:Time2	−0.17	0.44	−1.03	0.69	0.84
Happiness:Time2	−0.07	0.46	−0.97	0.83	0.93
Sadness:Time2	−0.33	0.40	−1.12	0.46	0.72
Surprise:Time2	−0.36	0.40	−1.13	0.43	0.70
Contempt:Time2	−0.50	0.41	−1.28	0.31	0.61
Constant	−3.44	0.32	−4.09	−2.82	0.03
σ (group effect ID)	1.77	0.15	1.50	2.09	–
μ (proportion)
Anger	−3.98	0.30	−4.55	−3.30	0.02
Disgust	−2.82	0.22	−3.26	−2.39	0.06
Fear	−4.54	0.25	−5.01	−4.03	0.01
Happiness	−4.14	0.29	−4.68	−3.56	0.02
Sadness	−2.19	0.21	−2.60	−1.77	0.11
Surprise	−2.03	0.21	−2.44	−1.63	0.13
Contempt	−3.33	0.22	−3.75	−2.90	0.04
Time2	0.11	0.14	−0.16	0.39	1.12
Anger:Time2	0.87	0.36	0.18	1.58	2.39
Disgust:Time2	0.04	0.27	−0.48	0.56	1.04
Fear:Time2	−0.03	0.37	−0.75	0.70	0.97
Happiness:Time2	−0.79	0.40	−1.50	−0.01	0.45
Sadness:Time2	0.24	0.27	−0.29	0.77	1.27
Surprise:Time2	−0.35	0.26	−0.87	0.17	0.71
Contempt:Time2	−1.22	0.29	−1.79	−0.65	0.30
Constant	1.24	0.13	1.00	1.50	3.46
σ (group effect ID)	0.62	0.10	0.44	0.82	–
φ (precision)
Anger	2.09	0.36	1.36	2.77	–
Disgust	0.99	0.23	0.55	1.44	–
Fear	2.96	0.31	2.32	3.53	–
Happiness	2.34	0.36	1.60	3.02	–
Sadness	0.14	0.19	−0.24	0.51	–
Surprise	0.11	0.18	−0.23	0.46	–
Contempt	1.29	0.24	0.81	1.75	–
Time2	−0.01	0.13	−0.27	0.24	–
Anger:Time2	−1.22	0.40	−2.02	−0.45	–
Disgust:Time2	−0.31	0.27	−0.84	0.23	–
Fear:Time2	−0.22	0.46	−1.13	0.68	–
Happiness:Time2	0.98	0.54	−0.07	2.04	–
Sadness:Time2	−0.32	0.24	−0.79	0.16	–
Surprise:Time2	0.15	0.25	−0.34	0.63	–
Contempt:Time2	1.56	0.34	0.89	2.21	–
Constant	−0.08	0.11	−0.30	0.14	–
σ (group effect ID)	0.56	0.07	0.44	0.70	–

Note. n = 88. EE = estimation error; CI = credible interval; OR = odds ratio.

Estimating the probability of zero identification

Table 3 shows that the probability of zero identification in all the emotions was very high compared with the neutral baseline period; the neutral expression was identified most frequently during the baseline period. The probability of zero identification was neither higher nor lower in the compassion phase versus the baseline phase.

Estimating the proportion of emotion identification and precision

The proportions of emotion identifications followed the zero-inflation submodel with some modifications: there is of course a substantially lower probability that any emotion would be identified more often than the neutral expression; but anger was identified more often in the compassion than in the baseline phase, and happiness and contempt were identified less often in the compassion than in the baseline phase. Regarding the precision parameter submodel, eight of the 17 were large precision parameters, having to do with heteroscedasticity and skewness, justifying the inclusion of this parameter in the model. Comparing differences among the emotions during the compassionate moment (see Table 7), anger, disgust, sadness, and surprise were identified more often than fear, happiness, and contempt (see Figure 2).

Table 7.

Parameter Estimates, SEs, 95% CI, and OR.

	Estimate	SE	95% CI		OR
α (zero inflated)
Anger	5.37	0.34	4.72	6.06	214.86
Disgust	4.33	0.31	3.73	4.96	75.94
Fear	6.13	0.36	5.44	6.87	459.44
Happiness	6.70	0.39	5.96	7.48	812.41
Sadness	4.24	0.31	3.64	4.86	69.41
Surprise	4.03	0.31	3.44	4.65	56.26
Contempt	4.33	0.31	3.73	4.96	75.94
Constant	−3.29	0.33	−3.94	−2.66	0.04
σ (group effect ID)	2.08	0.20	1.72	2.49	–
μ (proportion)
Anger	−3.03	0.34	−3.69	−2.38	0.05
Disgust	−2.85	0.26	−3.38	−2.35	0.06
Fear	−4.98	0.33	−5.61	−4.31	0.01
Happiness	−5.18	0.33	−5.80	−4.50	0.01
Sadness	−2.22	0.25	−2.71	−1.73	0.11
Surprise	−2.59	0.25	−3.08	−2.11	0.08
Contempt	−4.77	0.22	−5.20	−4.34	0.01
Constant	1.47	0.15	1.17	1.78	4.35
σ (group effect ID)	0.66	0.13	0.41	0.93	–
φ (precision)
Anger	0.66	0.33	0.26	1.63	–
Disgust	0.74	0.25	0.15	1.03	–
Fear	3.17	0.42	2.36	3.52	–
Happiness	3.50	0.45	2.01	3.30	–
Sadness	0.06	0.20	−0.24	0.42	–
Surprise	0.42	0.21	−0.18	0.46	–
Contempt	3.07	0.26	0.88	1.72	–
Constant	−0.22	0.13	−0.35	−0.01	–
σ (group effect ID)	0.64	0.09	0.48	0.82	–

Note. n = 88. SE = standard error; CI = credible interval; OR = odds ratio.

Figure 2.

Differences among emotions in the compassionate phase.

Discussion

The goal of our study was to describe the most common facial expression to be elicited by a compassionate video stimulus. We studied changes in mimic muscular activity (AUs) manually by three certified human coders using FACS for manual coding (on a consensus basis) and using EmotionID software for automatic coding by artificial intelligence. The manual analysis gave us an overview of the AUs involved in the facial expression during what the participants considered the most compassionate moment in the video. We obtained the probability of the seven basic emotions (anger, contempt, disgust, fear, happiness, sadness, and surprise) and the neutral expression occurring during the most compassionate moment by automatic analysis. Our aim was also to test previous findings (Baránková et al., 2019) with a different participant sample and a different video for eliciting compassion and characterizing facial expressions. The results are discussed later in the context of past findings on the facial expressions of emotion.

Manual Analysis of Facial Expressions With FACS

Of all the 59 AUs investigated in our analysis, the following AUs occurred more often, relative to other AU’s during the compassionate phase of the video: 1 = inner-brow raiser, 2 = outer-brow raiser, 4 = brow lowerer, 5 = upper-lid raiser, 7 = lids tight, 12 = lip-corner puller, 14 = dimpler, 17 = chin raiser, 24 = lip presser, 25 = lips part, 26 = jaw drop, 43 = eyes closure, 45 = blink, 55 = head tilt left, 56 = head tilt right, 61 = eyes left, 62 = eyes right, and 64 = eyes down. Of these, the following AUs 1 = inner-brow raiser, 4 = brow lowerer, 7 = lids tight, 17 = chin raiser, 24= lip presser, 43 = eye closure, and 55 = head tilt left occurred more often during the compassionate moment than during the baseline moment.

In this study, we identified more AUs as important during the compassionate versus baseline moments than did prior researchers (Baránková et al., 2019). In both our study and Baránková et al. (2019), two AUs, 7 = lids tight and 43 = eye closure, occurred more often during the compassionate moment. The eye AUs were found consistently in both studies. In the upper part of the face, we also found two AUs, 1 = inner-brow raiser and 4 = brow lowerer. The combination of these two AUs produces a kind of countermovement in which the inner part of the brows rise while the brows are also lowered, a movement combination that also appears in the expression of sadness. Altogether, these movements produce a horseshoe shape above the eyebrows (Ekman & Friesen, 2003; Ekman et al., 2002). We also found that AU 7 = lids tight occurred more often than other AUs during the compassionate moment more often during the compassionate moment than the baseline moment. According to the FACS manual (Ekman et al., 2002), this AU is connected with the prototypical expression of anger. Hypothetically, this AU might convey concentration or intensity (Rozin & Cohen, 2003). That intensity is an emotional communication that might be shared in both angry and compassionate moments, as if to say to others, “I’m serious about this.”

Compared with previous research (Baránková et al., 2019), we found differences in the inclination of the head. We found AU 55 = head tilt left prevalent in compassion, whereas the previous study had identified AU 56 = head tilt right. In our study, both these AUs (AU 55 and AU 56) were displayed more often than other AUs during the compassionate moment. It is possible that this movement is connected to brain hemisphere preference or to how the person in the video stimulus was positioned, but it is clear from both analyses that head inclination plays a role in the facial expression of compassion.

The main difference between our findings and those from the previous study is that we did not find AU 12 = lip-corner puller to be as prevalent as in other research. In our study, AU 12 occurred more frequently than other AUs during the compassionate moment but not more frequently during the compassionate than during the baseline moment. The compassion-related lower face AUs we found were completely different from those described in Baránková et al. (2019). The full complement of lower face AUs exhibited more often than other AUs during the compassionate moment among our participants were AU 12, AU 14, AU 17, AU 24, AU 25, and AU 26. Each of these AUs is connected with a different facial expression of basic emotion. AU 12 = lip-corner puller is associated with a smile and AU 14 = dimpler with contempt. AU 17 is associated with the expression of disgust and AU 24 with anger. AUs 25 = lips part and 26 = jaw drop are associated with the facial expression of surprise or fear. We concluded that the AUs in the upper face stayed consistent, but the AUs in the lower face varied, meaning that these AUs are found in variants of the expressions.

EmotionID Analysis

Using the EmotionID software, we analyzed the probability of the basic emotions (anger, disgust, fear, happiness, sadness, surprise, contempt, and neutral) occurring in facial expressions. Comparing the compassionate moment with the baseline moment, we found that anger occurred more often than other emotions during the compassionate moment and that contempt occurred more often than other emotions during the baseline moment. Contempt can be perceived as the opposite of compassion. In some cases, when it is used to convey pity to someone, it is interpreted as contemptuous (Cartwright, 1988). The difference between compassion and pity is mainly one of pitying from a position of power, when we may see them as weaker than ourselves (Cartwright, 1988). The fact that our video stimulus elicited contempt less often than other emotions during the compassionate moment supports this assumption and serves to double validate the participants’ selection of the most compassionate moment.

The higher occurrence of participants’ anger expression during the compassionate moment can be explained in several ways. First, respondents may have been distressed by the video. If they had reacted emotionally to the girls suffering this may have induced feelings of aversion and a desire to move away from the uncomfortable situation (Singer & Klimecki, 2014). Alternatively, the video could have prompted anger toward those responsible for the suffering girl’s situation. As stated by Catarino, Gilbert, McEwan, and Baião (2014), compassion can have two facets. Submissive compassion is connected with shame-based caring, anxiety, or depression, while genuine compassion is not. Both our FACS and automatic analyses identified AUs 4 and 24, which are associated with the facial expression of anger. Keltner and Cordaro (2015) considered the contraction of the brow muscles (AU 4) to be a sign of concentration, and Eisenberg et al. (1989) stated that, in their opinion, compassion (in their work on sympathy) was a form of concentration directed outward. Thus, recognizing others’ suffering may lead to concentration directed at this experience by others.

During the compassionate moment, anger, disgust, sadness, and surprise occurred more often than fear, happiness, and contempt. We have already discussed the occurrence of anger during the compassionate moment, and in the research by Baránková et al. (2019), anger was more frequently present during the compassionate moment. We assumed that there was no connection between the occurrence of disgust and surprise during the compassionate moment and the compassionate reaction itself. These emotions could have been reactions to the video stimulus scene—perhaps disgust at the junkyard and the waste and surprise at the inadequacy of the girls’ situation.

Eisenberg et al. (1989) thought that compassion (sympathy) was connected to sadness. They assumed that a sad facial expression could reflect sympathetic responding, in their words “reflection of other-oriented empathic sadness rather than … egoistic personal sadness … ” (Eisenberg et al., 1989, p. 59). Also Ekman and Cordaro (2011) have pointed out that there is no specific distinction between the facial expressions in the sadness and compassion facial expression families, perhaps explaining why people struggle to recognize the facial expression of compassion (Haidt & Keltner, 1999; Keltner & Shiota, 2003).

Summary of Manual and Automatic Analysis

If we try to create a distinct facial expression out of the AUs that occurred more often during the compassionate moment than during the baseline moment, the final expression obtained is quite similar to the one obtained by Eisenberg et al. (1989) and by Haidt and Keltner (1999). The expression Haidt and Keltner (1999) used in their study contains AUs: 1 = inner-brow raised, 4 = brow lowerer, and 57 = head forward. In our study, the head was more frequently tilted to the left as mentioned, and AUs 1 and 4 occurred together. Eisenberg et al. (1989) noted AU 7 = lids tight, which also features in our results. A combination of AUs 1 and 4 is typical of the expression of sadness. Haidt and Keltner (1999) concluded that this could be why people often confuse the expression of sadness with the expression of compassion. The upper face AUs were consistent with the previous research (Baránková et al., 2019; Eisenberg et al., 1989; Haidt & Keltner, 1999). Anger (or concentration) occurred more often during the compassionate moment, and this is consistent with the previous research (Baránková et al., 2019) and with our manual analysis showing AUs 4 and 7 as prevalent with compassion.

The stimulus video used by Eisenberg et al. (1989) showed a woman sitting in a hospital room describing a car accident in which her two children had sustained injuries. It showed a situation of single suffering and worry. Similarly, our stimulus was a situation of distress. The moment the participants identified as being most compassionate and most emotional was the moment the girl drank water from a waste plastic bottle. By contrast, in the study by Baránková et al. (2019), participants considered the most poignant part of the video to be the moment the little boy stole drugs for his sick mother. A stranger saw the situation and helped the child in his distress. Despite the distressing situation of suffering, the bystander’s behavior facilitated the experience of compassion in all its components (Strauss et al., 2016) in the video viewer. The participants could feel for this bystander expressing compassion toward the little boy. In this study, it may have been more difficult for the participants to feel compassion, as there was no scene in which a character modeled compassion directly, just a situation of distress that could have motivated participants to feel more distress than compassion. As stated by Singer and Klimecki (2014), the two empathic responses to suffering are compassion and empathic distress. It is more likely that our participants experienced the distress aspect of empathy at seeing the poor girl in the junkyard. We think that this is connected to the greater occurrence of AUs associated with the expressions of anger, disgust, and sadness in our study.

The results of the manual and automatic analysis were relatively consistent. The AUs which occurred more often than others during the compassionate moment and that occurred more often during the compassionate moment than during the baseline movement were 1 = inner-brow raiser, 4 = brow lowerer, 7 = lids tight, 17 = chin raiser, 24 = lip presser, and 55 = head tilt left. We used these AUs to create three variants of compassionate facial expressions which differed in the lower face AUs while the upper face AUs remained the same: AUs 1 + 4 + 7 + 17 + 24 + 55 (Figure 3), AUs 1 + 4 + 7 + 17 + 55 (Figure 4), and 1 + 4 + 7 + 24 + 55 (Figure 5).

Figure 3.

AUs 1 + 4 + 7 + 17 + 24 + 55.

Figure 4.

AUs 1 + 4 + 7 + 17 + 55.

Figure 5.

AUs 1 + 4 + 7 + 24 + 55.

Further Research and Limitations

The contradictory results from the most recent research (Baránková et al., 2019; Eisenberg et al., 1989; Haidt & Keltner, 1999; McEwan et al., 2014) and our study indicates a further need to test the results and modify the research goals based on the current findings. In future research, different samples could be used, not ones comprising psychology undergraduates or helping professionals. The way the head inclines in the compassionate expression is not clear from the FACS analysis. There is therefore a need to explore how head inclination varies in response to suffering. This could be verified using different stimuli to elicit compassion. Using multiple stimuli, we could also study variation in a single expression or if there are universal signs. As mentioned by Ekman and Cordaro (2011), compassion is specific in that it is often displayed only toward those close to us. We do not know what compassion looks like when expressed in close relationships and in more distant ones. It is also important to be able to distinguish a compassionate expression from a sad expression. Just as research has been conducted on the meaning of compassion (Halamová et al., 2018), we also need to distinguish the facial expressions of compassion. As the facial expression of compassion can be blocked out of a fear of compassion (Gilbert et al., 2012), the link between fear of compassion and potential to respond compassionately should be explored too.

We are aware of several limitations in our research. The research sample was fairly homogeneous, consisting of undergraduate psychology students. In the future, a larger and more representative sample, ideally from different cultures, should be used. The manual coding of the facial-AUs was conducted by certified coders who were familiar with this research area, so it is possible their expectations may have distorted their views during the data analysis. The research was conducted via an online questionnaire, and the participants were advised to complete the form alone at home, but as we could not control the conditions we could not be sure that is what they did. In the future, it would be beneficial to develop a set of stimuli that reliably elicit compassion, and a set of photos showing the facial expression of compassion.

Conclusion

In this study, we used the FACS manual coding system (Ekman et al., 2002) and automatic EmotionID software to analyze participants’ facial expressions in response to a compassionate scene in a video stimulus. In the FACS analysis, we found that AUs 1 = inner-brow raiser, 4 = brow lowerer, 7 = lids tight, 17 = chin raiser, 24 = lip presser, 43 = eye closure, and 55 = head tilt left occurred more often than other AUs during the compassionate moment; and we found that these AUs occurred more often during the compassionate than during the baseline moment. Automatic analysis using EmotionID software revealed that anger, disgust, sadness, and surprise occurred more often than fear, happiness, and contempt during the compassionate moment. Of these emotions, anger occurred more often during the compassionate moment than during the baseline moment, while contempt occurred more often during the baseline moment than during the compassionate moment.

Efforts to identify the facial expression of compassion are proceeding slowly. From the findings of previous research on this topic (Baránková et al., 2019; Eisenberg et al., 1989; Haidt & Keltner, 1999; McEwan et al., 2014), we can conclude that the compassion expression resembles the facial expression of sadness to some extent. Eisenberg et al. (1989) explicitly pointed out that compassion as attention turned outward is connected with some kind of sadness but not with personal, egoistic sadness. Similarly, Ekman and Cordaro (2011) have stated that there are no clear differences between the emotional families of compassion and sadness. Our contribution to this topic is in description of specific facial and head movements connected with compassion. Results show the connection between expressions of compassion and sadness, or anger associated with concentration. Therefore, future research should capture and distinguish the differences in the facial expressions of compassion, sadness, and anger or concentration.

Footnotes

Acknowledgments

The authors thanks Nicola Petrocchi for his consultation about research design and video stimulus.

Notes

ORCID iD

Júlia Halamová

References

Baránková

Halamová

(2018). Validating video stimulus for eliciting compassion: A pilot study. In G. Đigić (Ed.), Psychology in and around us. Proceedings of the 13th Days of Applied Psychology 2017 (pp. 283–292). Niš, Serbia: University of Nis.

Baránková

Halamová

Gablíková

Koróniová

Strnádelová

(2019). Analysis of spontaneous facial expression of compassion elicited by the video stimulus. Československá psychologie, 63(1), 75–89.

Baránková

Halamová

Koróniová

(2019). Non-expert views of compassion: Consensual qualitative research using focus groups. Human Affairs, 29(1), 58–69.

Baveye

Dellandrea

Chamaret

Chen

(2015). Liris-accede: A video database for affective content analysis. IEEE Transactions on Affective Computing, 6(1), 43–55.

Beaumont

Hollins Martin

C. J.

(2015). A narrative review exploring the effectiveness of compassion-focused therapy. Counselling Psychology Review, 30(1), 21–32.

Bloom

(2016). Against empathy: The case for rational compassion. New York, NY: Harper Collins.

Bürkner

P. C.

(2017). brms: An R package for Bayesian multilevel models using Stan. Journal of Statistical Software, 80(1), 1–28. doi:10.18637/jss.v080.i01.

Bürkner

P. C.

(2018). Advanced Bayesian multilevel modeling with the R Package brms. The R Journal, 10(1), 395–411.

Cartwright

D. E.

(1988). Schopenhauer’s compassion and Nietzsche’s pity. Schopenhauer Jahrbuch, 69, 557–567.

10.

Catarino

Gilbert

McEwan

Baião

(2014). Compassion motivations: Distinguishing submissive compassion from genuine compassion and its association with shame, submissive behavior, depression, anxiety and stress. Journal of Social and Clinical Psychology, 33(5), 399–412.

11.

Cohn

J. F.

De la Torre

(2015). Automated face analysis for affective computing. In Calvo

D´Mello

Gratch

Kappas

(Eds.), The Oxford handbook of affective computing (pp. 131–150). New York, NY: Oxford University Press.

12.

Cohn

J. F.

Reed

L. I.

Ambadar

Xiao

Moriyama

(2004, October). Automatic analysis and recognition of brow actions and head motion in spontaneous facial behavior. In 2004 IEEE International Conference on Systems, Man and Cybernetics (IEEE Cat. No. 04CH37583) (Vol. 1, pp. 610–616). Piscataway, NJ: IEEE.

13.

Desbordes

Negi

L. T.

Pace

T. W. W.

Wallace

B. A.

Raison

C. R.

Schwartz

E. L.

(2012). Effects of mindful-attention and compassion meditation training on amygdala response to emotional stimuli in an ordinary, non meditative state. Frontiers in Human Neuroscience, 6, 292. doi:10.3389/fnhum.2012.00292

14.

Dossey

(2007). Compassion. Explore: The Journal of Science & Healing, 3(1), 1–5.

15.

Eisenberg

Fabes

R. A.

Miller

P. A.

Fultz

Shell

Mathy

R. M.

Reno

R. R.

(1989). Relation of sympathy and personal distress to prosocial behavior: A multimethod study. Journal of Personality and Social Psychology, 57(1), 55–66.

16.

Ekman, P., Friesen, W. V., & Hager, J. C. (2002). Facial Action Coding System. Manual and Investigator’s Guide. Salt Lake City, UT Research Nexus.

17.

Ekman

(1989). The argument and evidence about universals in facial expressions of emotion. In Wagner

Manstead

(Eds.), Handbook of social psychophysiology (pp. 144–164). New York, NY: John Wiley & Sons.

18.

Ekman

(1994). All emotions are basic. In Ekman

Davidson

(Eds.), The nature of emotion (pp. 15–19). New York, NY: Oxford University Press.

19.

Ekman

(1998). Universality of emotional expression? A personal history of the dispute. In C.

Darwin

(Ed.), The expression of the emotions in man and animals (3rd ed., pp. 363–393). New York, NY: Oxford University Press.

20.

Ekman

Cordaro

(2011). What is meant by calling emotions basic. Emotion Review, 3(4), 364–370.

21.

Ekman

Friesen

W. V.

(2003). Unmasking the face: A guide to recognizing emotions from facial clues. San Jose, CA: ISHK.

22.

Ekman

Friesen

V. W.

Hager

J. C.

(2002). Facial Action Coding System: The manual on CD-ROM (Investigator´s Guide). Salt Lake City, UT: Research Nexus.

23.

Ekman

Keltner

(1997). Universal facial expressions of emotion: An old controversy and new findings. In Segerstrale

Molnár

(Eds.), Nonverbal communication: Where nature meets culture (pp. 27–45). Hillsdale, NJ: Lawrence Erlbaum Associates.

24.

Feinberg

F. M.

Gonzalez

(2012). Bayesian modeling for psychologists: An applied approach. In Cooper

Camic

P. M.

Long

D. L.

Panter

A. T.

Rindskopf

Sher

K. J.

(Eds.), APA handbook of research methods in psychology, vol. 2. Research designs: Quantitative, qualitative, neuropsychological, and biological (pp. 445–464). Washington, DC: American Psychological Association. doi:10.1037/13620-024

25.

Gablíková

Halamová

(2016). Tváre mužov a žien keď sa nikto nepozerá – analýza rozdielov v prejave hnevu. Psychologica, 44, 75–84.

26.

Gelman

Goodrich

Gabry

Vehtari

(2018, December). R-squared for Bayesian regression models, The American Statistician. doi:10.1080/00031305.2018.1549100

27.

Gelman

Rubin

D. B.

(1992). Inference from iterative simulation using multiple sequences, Statistical Science, 7, 457–511.

28.

Geweke

(1992). Evaluating the accuracy of sampling-based approaches to calculating posterior moments. In Bernado

J. M.

Berger

J. O.

Dawid

A. P.

Smith

A. F. M.

(Eds.), Bayesian statistics 4. (pp. 169–193) Oxford, England: Clarendon Press.

29.

Gilbert

(2017). Compassion: Definitions and controversies. In Gilbert

(Ed.), Compassion: Concepts, research and applications (pp. 3–15). New York, NY: Routledge.

30.

Gilbert

McEwan

Gibbons

Chotai

Duarte

Matos

(2012). Fears of compassion and happiness in relation to alexithymia, mindfulness, and self-criticism. Psychology and Psychotherapy: Theory, Research and Practice, 85, 374–390. doi:10.1111/j.2044-8341.2011.02046.x

31.

Gilbert

Procter

(2006). Compassionate mind training for people with high shame and self-criticism: Overview and pilot study of a group therapy approach. Clinical Psychology & Psychotherapy: An International Journal of Theory & Practice, 13(6), 353–379.

32.

Haidt, J. (2003). “The Moral Emotions” (PDF). In Davidson, Richard; Scherer, Klaus; Goldsmith, H. (eds.). Handbook of Affective Sciences (pp. 852–870). Oxford University Press.

33.

Halamová

Baránková

Strnádelová

Koróniová

(2018). Consensual qualitative research on free associations for compassion and self-compassion. Human Affairs, 28(3), 253–270.

34.

Heidelberger

Welch

P. D.

(1983). Simulation run length control in the presence of an initial transient. Operations Research, 31(6), 1109–1144. doi:10.1287/opre.31.6.1109

35.

Hess

Banse

Kappas

(1995). The intensity of facial expression is determined by underlying affective state and social situation. Journal of Personality and Social Psychology, 69(2), 280.

36.

Jazaieri

Jinpa

G. T.

McGonigal

Rosenberg

E. L.

Finkelstein

Simon-Thomas

… Goldin

P. R.

(2013). Enhancing compassion: A randomized controlled trial of a compassion cultivation training program. Journal of Happiness Studies, 14, 1113–1126. doi:10.1007/s10902-012-9373-z

37.

Jazaieri

McGonigal

Jinpa

Doty

J. R.

Gross

J. J.

Goldin

P. R.

(2014). A randomized controlled trial of compassion cultivation training: Effects of mindfulness, affect, and emotion regulation. Motivation and Emotion, 38, 23–35. doi:10.1007/s11031-013-9368-z

38.

Kanade

Cohn

J. F.

Tian

(2009). Comprehensive database for facial expression analysis. In Proceedings of the Fourth IEEE International Conference on Automatic Face and Gesture Recognition (FG′00), Grenoble, France (pp. 484–490). Piscataway, NJ: IEEE.

39.

Keltner

(2005a). Signs of appeasement: Evidence for the distinct displays of embarrassment, amusement, and shame. In Rosenberg

E. L.

Ekman

(Eds.), What the face reveals (2nd ed., pp. 133–157). New York, NY: Oxford University Press.

40.

Keltner

(2005b). The forms and functions of embarrassment. In Rosenberg

E. L.

Ekman

(Eds.), What the face reveals (2nd ed., pp. 158–160). New York, NY: Oxford University Press.

41.

Keltner

Cordaro

D. T.

(2015). Understanding multimodal emotional expressions: Recent advances in basic emotion theory. Emotion Researcher, 1–17. Retrieved from: http://emotionresearcher.com/wp-content/uploads/2015/08/Keltner-and-Cordaros-Original-Paper-With-Changed-Text-Highlighted.pdf

42.

Keltner

Shiota

M. N.

(2003). New displays and new emotions: A commentary on Rozin and Cohen (2003). Emotion, 3(1), 86–91. doi:10.1037/1528-3542.3.1.86

43.

Lewinski

den Utyl

T. M.

Butler

(2014). Automated facial coding: Validation of basic emotion and FACS AUs in FaseReader. Journal of Neuroscience, Psychology, and Economics, 7(4), 227–236. doi:10.1037/npe0000028

44.

Deng

(2017). Reliable crowdsourcing and deep locality-preserving learning for expression recognition in the wild. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 2852–2861). Piscataway, NJ: IEEE.

45.

Lien

J. J. J.

Kanade

Cohn

J. F.

C. C.

(2000). Detection, tracking, and classification of action units in facial expression. Robotics and Autonomous Systems, 31(3), 131–146.

46.

McEwan

Gilbert

Dandeneau

Lipka

Maratos

Paterson

K. B.

Baldwin

(2014). Facial expressions depicting compassionate and critical emotions: The development and validation of a new emotional face stimulus set. PLoS One, 9(2), 1–8. doi:10.3171/journal.pone.0088783

47.

Mongrain

Chin

J. M.

Shapira

L. B.

(2011). Practicing compassion increases happiness and self-esteem. Journal of Happiness Studies, 12, 963–981. doi:10.1007/s10902-010-9239-1

48.

Piironen

Vehtari

(2017). Comparison of Bayesian predictive methods for model selection. Statistics and Computing, 27(3), 711–735. doi:10.1007/s11222-016-9649-y

49.

Raftery

A. E.

Lewis

S. M.

(1995). The number of iterations, convergence diagnostics and generic Metropolis algorithms. In Gilks

W. R.

Spiegelhalter

D. J.

Richardson

(Eds.), Practical Markov Chain Monte Carlo. (pp. 115–130) London, England: Chapman and Hall.

50.

R Core Team (2018). R: A language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing. Retrieved from https://www.R-project.org/

51.

Rosenberg

E. L.

(2005). The study of spontaneous facial expressions in psychology. In Rosenberg

E. L.

Ekman

(Eds.), What the face reveals (2nd ed., pp. 3–18). New York, NY: Oxford University Press.

52.

Rozin

Cohen

A. B.

(2003). High frequency of facial expressions corresponding to confusion, concentration, and worry in an analysis of naturally occurring facial expressions of Americans. Emotion, 3(1), 68.

53.

Russell

J. A.

Feldman Barrett

(1999). Core affect, prototypical emotional episodes, and other things called emotion: Dissecting the elephant. Journal of Personality and Social Psychology, 76(5), 805–819.

54.

Samson

A. C.

Kreibig

S. D.

Soderstrom

Wade

A. A.

Gross

J. J.

(2016). Eliciting positive, negative and mixed emotions states: A film library for affective scientists. Cognition and Emotion, 30, 827–856. doi:10.1080/02699931

55.

Singer

Klimecki

(2014). Empathy and compassion. Current Biology, 24(8), R875–R878. doi:10.1016/j.cub.2014.06.054

56.

Spiegelhalter

D. J.

Best

N. G.

Carlin

B. P.

van der Linde

(2002). Bayesian measures of model complexity and fit. Journal of the Royal Statistical Society B, 64, 583–639. doi:10.1111/1467-9868.00353

57.

Stellar

J. E.

Gordon

A. M.

Piff

P. K.

Cordaro

Anderson

C. L.

Bai

… Keltner

(2017). Self-transcendent emotions and their social functions: Compassion, gratitude, and awe bind us to others through prosociality. Emotion Review, 9(3), 200–207.

58.

Strauss

Talyor

B. L.

Kuyken

Baer

Jones

Cavanagh

(2016). What is compassion and how can we measure it? A review of definitions and measures. Clinical Psychology Review, 47, 15–27. doi:10.1016/j.cpr.2016.05.004

59.

Tian

Y. L.

Kanade

Cohn

J. F.

(2002, May). Evaluation of Gabor-wavelet-based facial action unit recognition in image sequences of increasing complexity. In Proceedings of Fifth IEEE International Conference on Automatic Face Gesture Recognition (pp. 229–234). Piscataway, NJ: IEEE.

60.

Valdesolo

DeSteno

(2011). Synchrony and the social tuning of compassion. Emotion, 1(2), 262–266.

61.

Vehtari, A., Gelman, A., & Gabry, J. (2017). Practical Bayesian model evaluation usingleave-one-out cross-validation and WAIC. Statistics and Computing, 27(5), 1413–1432.

62.

Vehtari

Mononen

Tolvanen

Sivula

Winther

(2016). Bayesian leave-one-out cross-validation approximations for Gaussian latent variable models. Journal of Machine Learning Research, 17(103), 1–38.

63.

Wagenmakers

E-J.

Marsman

Jamil

Verhagen

Love

… Morey

R. D.

(2018). Bayesian inference for psychology. Psychonomic Bulletin & Review, 25(1), 35–57. doi:10.3758/s13423-017-1343-3

64.

Watanabe

(2010). Asymptotic equivalence of Bayes cross validation and widely applicable information criterion in singular learning theory. Journal of Machine Learning Research, 11, 3571–3594.

65.

Widen

S. C.

Christy

A. M.

Hewett

Russell

J. A.

(2011). Do proposed facial expressions of contempt, shame, embarrassment, and compassion communicate the predicted emotion? Cognition and Emotion, 25(5), 898–906.

66.

Yao

Vehtari

Simpson

Gelman

(2018). Using stacking to average Bayesian predictive distributions. Bayesian Analysis, 13(3), 917–1007. doi:10.1214/17-BA1091