Trait emotional empathy: IRI.

Davis developed the Interpersonal Reactivity Index (IRI), which comprises four 7-item subscales that estimate perspective-taking (PT), fantasy (FS), empathic concern (EC), and personal distress (PD). The EC subscale assesses an individual’s warmth, compassion, and concern for others. The PD subscale estimates an individual’s discomfort and anxiety when witnessing others’ negative experiences. The PT subscale reflects an individual’s tendency (or ability) to adopt another individual’s perspective (or point of view), which corresponds to cognitive empathy [16]. The IRI (particularly the EC subscale) remains the most prevalent measure of empathy [15]; overall, the IRI is a multidimensional measure of empathy [16, 17]. Davis presumed that the EC and PD subscales estimated the emotional/affective component of empathy. However, the PD subscale was significantly correlated with the EC subscale (r = 0.380, p = .0002) and the AQ (r = 0.392, p < .0001) in our samples. The statistical models including the IRIPD, IRIEC, and AQ were easily overfitted, preventing the modeling of random slopes of the emotional and presentation condition and negatively affecting the generalizability of our statistical results [82]. Therefore, we included only the EC subscale in our statistical analysis.

Autistic traits: AQ.

Assuming that the autism spectrum is a continuum that covers the entire population (with and without an ASD diagnosis), the AQ measures individual autistic traits. The 50 questions assess five areas: social skills, attention-switching, attention to detail, communication, and imagination [50].

Subjective experiential ratings: The Russel “affect grid”.

Subjective experiential ratings are used to estimate the extent of emotional contagion in the present study. We instructed the participants about the ratings based on the Russel “affect grid” [76]. The instructions read: “When viewing the images, please rate the emotion you feel. Please rate the emotion in two dimensions: (1) Pleasure-displeasure, which is about the quality of the emotion, and (2) arousal, which is about the energy of the emotion. Please input the number you find most suitable (1–9) in the sequence of ‘pleasure-displeasure’ and ‘arousal’ using the typewriter keys”. Further explanation of valence and arousal using the affect grid was then provided: “The center of the square represents the neutral emotion, which is neither positive nor negative; it is the ordinary emotion you experience. The upper right corner represents the emotion of excitement and happiness. The lower left corner represents the emotion of depression, sadness, or worry. The upper left corner represents stress and nervousness. The lower right corner represents tranquility and relaxation.”

Facilities.

For the 50 participants of the behavioral experiment, each model and participant faced a prompter. A concealed VIXIA HF R800 camera (Canon Inc., Tokyo, Japan) was placed behind the mirror of each prompter (Fig 1). The models viewed “live-relay images” of the participants. Presentation software (Neurobehavioral Systems, Inc., Berkeley, CA, USA) was used to run the experiments on a Precision T3500 computer (Dell Inc., Round Rock, TX, USA) with the Windows 7 Professional operating system (Microsoft Corp., Redmond, WA, USA). Serial signals were sent to an SL-41C switcher (Imagenics Co., Ltd., Tokyo, Japan) to switch the visual input of the participant’s prompter between the computer and the model’s camera. Presentation software transmitted sounds to the models using a Pasonomi TWS-X9 Bluetooth earphone (Shenzhen City Meilianfa Technology Co., Ltd., Shenzhen, China), instructing them to produce dynamic facial expressions. The switcher maintained the resolution (1,280 × 960) and the height-to-width ratio of the video output of the presentation software and live-relay images. A BrainAmp ExG MR amplifier (Brain Products, Munich, Germany) and BrainVision Recorder software (Brain Products) were used to record EMG from the ZM and CS of participants. The setting for the 44 participants of the neuroimaging experiment is identical, except that the participants lay in the scanner viewing the image relayed by the switcher.

Paradigm and procedures.

As described previously [13], a 2 × 2 study design was used, with factors of presentation (video vs. live) and emotion (positive vs. negative), resulting in four conditions (Fig 3). After obtaining written informed consent and the EMG electrodes were attached, the participant and model engaged in conversation for 3 minutes through the prompter. The conversation demonstrated to the participants that the prompter system transmitted live images of the model. In the behavioral experiment (50 participants), the conversation was followed by eight practice passive-viewing trials, 60 passive-viewing trials, four practice rating trials, and 16 rating trials. In the neuroimaging experiment (44 participants), there were 200 passive-viewing trials, four practice rating trials, and 16 rating trials. The practice trials gave participants an understanding of what to expect and allowed them to ask any questions before data acquisition began. The relative proportions of passive viewing (EMG) trials and rating trials aligned with our prior assumption about the effect size and signal-to-noise ratio of spontaneous facial mimicry and ratings. The EMG would have a smaller effect size and lower signal-to-noise ratio than ratings, thus requiring more repetitions per participant. However, participants might experience a loss of attention or interest if the passive viewing period persists for too long. The total testing time per session for the study’s behavioral and neuroimaging versions was less than 1 hour.

During the passive-viewing component, the participants fixated on a cross (behavioral experiment: mean inter-trial interval = 2,604 ms; range: 2,000–3,750 ms; neuroimaging experiment: mean inter-trial interval = 4,475 ms; range: 2,950–10,150 ms) until they were shown the announcement of either “Video” or “Real Person” for 1 s (1.25 s in the neuroimaging experiment). The instruction enabled a clear audience effect [59, 60, 63]. Participants did not need to de-ambiguate or infer a live vs. pre-recorded situation by relying on subtle cues such as microexpressions or environmental inconsistencies, which do not guarantee successful detection/judgment of live performances. Also, otherwise, it would have been difficult to ensure whether participants were really in mental states appropriate for live interactions. Without instructions, it would be more difficult to claim whether an effect (if any) was attributable to low-level processing rather than social cognition. During the video trials, one pre-recorded video clip was presented immediately after the instructions. During the live trials, the models performed dynamic facial expressions according to the instructions in their earphones. The screen displayed the fixation cross again after the facial expressions had been shown. In the behavioral experiment, after eight practice trials, 15 passive-viewing trials were completed per condition for a total of 60 trials, with a break after 32 trials (8 trials per condition). In the neuroimaging experiment, after eight practice trials, there were five runs of 40 trials (10 trials per condition) for a total of 200 trials. To avoid movement-associated EMG artifacts, and to detect spontaneous, automatic facial mimicry without top-down emotional processing, as in previous studies [83, 84], no rating was requested during the passive-viewing trials. For each participant, the sequence of conditions during the trial was pseudo-randomized, and the presentation sequence of the pre-recorded videos per condition was randomized. EMG data were collected during the passive-viewing trials.

After the passive-viewing trials, participants received instructions about and explanations of the subjective experiential ratings. During the subjective rating trials [76], each trial began with the described passive-viewing component; valence (pleasure-displeasure) and arousal ratings were completed sequentially after the participant had viewed dynamic facial expressions (Fig 3). Four practice rating trials were followed by 16 test-rating trials, during which no EMG data were acquired. For video conditions, the clips used in the passive viewing trials (15 per emotional condition) and rating trials (4 per emotional condition) were not repeated. After the ratings had been completed, the electrodes were removed; participants then completed the AQ and IRI questionnaires. Seventy-two participants (all participants tested in 2020 and 2021 and one participant in 2019) also completed the State-Trait Anxiety Inventory [85] and the Interaction and Audience Anxiousness Scale [86] (the resulting data were not analyzed in the present study). The participants were remunerated for their participation and then dismissed.

Pre-recorded videos.

As described previously [13, 14], four female models each recorded more than 20 video clips of positive and negative dynamic facial expressions (Fig 2). The models were instructed to display happy, smiling faces for positive expressions. They were instructed to display angry, frowning faces with a protruding mouth for negative expressions. The models’ appearances were consistent between the video clips and live performances. An apron covered their clothes, and hairpins were used to maintain their hairstyles. The clips lasted 3 s; they featured a neutral expression, gradual dynamic changes, and maximal emotional facial expression, all sustained for 1 s. Two smiling and two frowning clips were used for the positive-video and negative-video conditions in practice passive-viewing trials; 15 clips per video condition (50 clips per video condition in the neuroimaging study) were used in the actual passive-viewing trials; one clip per video condition for the practice ratings, and four clips per video condition were used for the actual ratings. During passive-viewing trials, no clips were repeated within a single condition. The positive-live and negative-live conditions did not use pre-recorded clips; the models’ live performances were relayed. Two models in the behavioral study gave written informed consent (as outlined in the PLOS consent form) to publish their images (Figs 2 and 3).

Stimuli validations.

We recorded and visually inspected the live performances of the models during the passive-viewing trials to ensure valid performances of dynamic facial expressions. In the behavioral study, one positive trial for two participants and one negative trial for two other participants were excluded from the analysis because they either were not performed correctly or were performed after a noticeable delay. Video recordings of the live performances in the 2019 sessions were validated by naïve participants, whose valence and arousal ratings were not different from the ratings for prerecorded stimuli. The naïve participants could not differentiate between the prerecorded stimuli and live-performance recordings [13]. Video recordings of the live performances and pre-recorded videos in the 2021 neuroimaging study were validated using OpenFace [87, 88], showing that the action unit 12 (AU12, lip corner puller, corresponding to the ZM) dynamic change was significantly weaker in the positive-live than in the positive-video trials. AU4 (brow lowerer, corresponding to the CS) dynamic change was significantly weaker in the negative-live than in the negative-video trials [14]. This increased our certainty that the observed interaction between emotion and presentation condition was not due to enhanced emotional expression during the live performances.

Because some people might suspect that the models’ negative expressions did not clearly distinguish between angry (as instructed) and sad, we performed another post hoc rating study on sadness and anger using a Likert scale (range: 1–5). Sixteen negative video clips (four per presentation condition [prerecorded or live]) were presented to 28 naïve female participants (mean ± SD age = 27.64 ± 3.07 years; range: 20–30 years). The negative stimuli were highly rated in terms of both anger (mean ± SD = 3.36 ± 0.61), and sadness (mean ± SD = 3.01 ± 0.85), but a paired t-test of the participant-wise mean rating values showed that participants generally perceived more anger than sadness in the negative stimuli (t = 2.55, degrees of freedom [df] = 27, p = 0.017). The distribution of the anger and sadness ratings did not violate the assumption of normality (anger ratings: Shapiro-Wilk W = 0.96, p = 0.39; sadness ratings: Shapiro-Wilk W = 0.96, p = 0.30).

EMG data pre-processing.

For the neuroimaging experiment, we first inspected video recordings of participants inside the scanner, and excluded runs in which the participant closed her eyes for more than half of the time for EMG analysis (see [14] for details). Raw EMG data from the neuroimaging study were first pre-processed using the BrainVision Analyzer 2 software. MR gradient artifacts were removed based on the pulse trigger, and data were down-sampled to 500 Hz. Raw EMG data were pre-processed using the EEGLAB MATLAB toolbox (version 2019.1; Swartz Center for Computational Neuroscience, San Diego, CA, USA). A notch filter was applied at around 60 Hz and multiples of 60 Hz; a high-pass filter was applied at 20 Hz, and a low-pass filter was applied at 500 Hz. The data were then screened for movement artifacts. For each trial, the signal was detrended and baseline-corrected for low-frequency drift. The baseline for this preprocessing step was defined as the mean value from 3 s before to 1 s after stimulus onset, which was the end of the model’s neutral expression immediately before the dynamic facial expression change. All data points in each trial were shifted by the per-trial baseline mean value. The data points were then rectified to retrieve the amplitudes of oscillation. The absolute or rectified values (oscillation amplitudes) plus one were natural-log-transformed to correct for the right-skewness of the raw data distribution. Trial-wise, the mean differences in the EMG natural log-transformed oscillation amplitudes between the neutral (0–1 s after stimulus onset) and maximal phases (2.5–3.5 s after stimulus onset) of the dynamic facial expression served as dependent variables (facial muscular responses) during statistical analysis.

Statistical analyses.

Linear mixed effect (LME) models, model comparisons, and influence diagnostics were performed using R software (version 4.3.1; R Core Team; Vienna, Austria) and the packages lme4 1.1–34, lmerTest 3.1–3, HLMdiag 0.5.0, and emmeans 1.8.7, along with the optimizer BOBYQA. The dependent variables included valence ratings, arousal ratings, and the ZM and CS responses. For each LME model, the emotional condition, presentation condition, mean-centered IRIEC, AQ, and their interactions were fixed effects, whereas “subject” and “Type” (behavioral vs. neuroimaging study) were random factors. The reference level for the emotional condition was “Negative,” and the reference level for the presentation condition was “Video.” Model complexity, in terms of the number of random effects (n(n+1)/2 model parameters to be estimated), was increased in a stepwise manner; each model was compared with the previous (i.e., less complicated) model using F-tests [lmerTest::anova()] until the more complex model was no longer significantly superior to the simpler one. The model iteration process was considered complete when the model estimation resulted in a singular fit, indicating overfitting or failure to converge. The most complex model showing convergence was chosen. Initially, random intercepts for participants were included; the inclusion of by-subject random slopes for the effects of the emotional condition, presentation condition, their interaction term, and random intercepts for the study type depended on the results of the model comparison process. The random intercept for the study type was included in the model whenever it did not cause a singular fit to account for the potential effect of the study type.

The homogeneity of residual variance is an a priori assumption when seeking valid inferences from LME models [89, 90]. We first performed the robust LME (robustlmm 3.2) to check the extent of bias caused by highly influential data points in the model without diagnostics. In the robust approach [91], each item’s robustness weight is estimated; highly influential data points are given a low robustness weight. This approach could limit the bias caused by highly influential data points, while retaining all data points in the model. However, information criteria (e.g., Akaike’s information criterion and tests based on log-likelihood statistics), Satterthwaite and Kenward-Roger approximation of the degrees of freedom, and estimation of effect sizes and their confidence intervals are, for the time being, unavailable in the robust approach. Therefore, we used the Satterthwaite approximation of the degrees of freedom of the non-robust estimation of the complete data set to calculate p-values for the robust estimation [92]. When the robust and non-robust estimations showed compatible significant effects, we report and discuss the results based on the non-robust estimation of the complete data set. Results of the robust estimation are presented in the supplementary tables S1, S3, S5, and S7 Tables. To show whether participants of the neuroimaging study responded differently, we also included results of the robust estimation of the 50 participants in the behavioral study in the supplementary tables S2, S4, S6, and S8 Tables.

When the robust and non-robust estimations showed incompatible significant effects (as in CS responses), we performed upward residual and influence analyses using the R package HLMdiag [93]. Trials were excluded if they had an absolute standardized residual larger than 3, or a Cook’s distance greater than 1.5 × the interquartile range above the third quartile (Q₃ + 1.5 × IQR as implemented in the HLMdiag package). After the reduced data had been refitted, the second-level Cook’s distance was checked, as suggested in the HLMdiag package [93], and highly influential participants were excluded. We present and discuss the results of the CS responses for which model diagnostics were performed.

The Satterthwaite approximation was used to evaluate the df values. Effect size estimation of the fixed effects’ semi-partial R-squared (variance-explained) was conducted using the R package r2glmm 0.1.2 and the Nakagawa and Schielzeth approach [94–96]. According to Cohen [81], the effect sizes for partial r² values of regression models are considered small at 0.0196 (f² = 0.02), medium at 0.13 (f² = 0.15), and large at 0.26 (f² = 0.35), but see also [97]. Interactions between emotional conditions (positive vs. negative) and presentation conditions (live vs. video) were evaluated using a simple-effects analysis via the emmeans function. Simple-slope analyses were performed to investigate significant interactions among individual differences, emotional conditions, and presentation conditions using the emtrends function in the R package emmeans. The “pick-a-point” approach [98] was used for two-way interactions involving one continuous and one categorical independent variable, as follows. Simple-effect contrasts were estimated using the emmeans function at mean-centered continuous variable levels from -3 SD to 3 SD in steps of 1 SD, with comparisons between two categorical variable levels; the Šidák method was used to correct for multiple comparisons. The emmeans package uses a reference grid approach to estimate/predict marginal means based on a fitted model [99], rather than based directly on the data. The purpose of this package was to provide an interpretation of the fitted model (please see the Basics vignette for the emmeans package). Because a significant main effect involved in a significant interaction could be biased, we refrained from reporting such main effects in the Results section. Readers can examine the simple-effect/simple-slope results to determine whether the significant interaction biased the main effect. Marginal effects of interaction terms involving individual differences (continuous variables) were visualized using the ggemmeans function in the ggeffects 1.1.2 R package. Two-tailed p-values are reported throughout the text.

Valence ratings.

The final LME model of valence ratings included the emotional conditions, the presentation conditions, the mean-centered IRIEC and AQ scores, and their interactions with the emotion and presentation conditions as fixed effects. It also included the random intercept and random slopes for the emotional and presentation conditions and their interactions by the random factor “subject” (Table 1), as well as the random intercept by the random factor “Type.” Adding the random intercept over the factor “Type” significantly improved the model fitting (model comparison p = 8.992e-05). We noted a significant interaction between the emotional conditions and presentation conditions. Simple-effect analysis according to the emotional condition showed that the interaction was attributable to a significant effect of positive-live (mean = 7.16, standard error [SE] = 0.267, 95% confidence interval [CI] = [5.21, 9.11]) being rated as more positive than positive-video (mean = 6.64, SE = 0.264, 95% CI = [4.58, 8.17]; difference = 0.519, SE = 0.102, df = 91, t = 5.076, p < 0.0001*) conditions; while negative-live (mean = 3.63, SE = 0.259, 95% CI = [1.29, 5.96]) was rated as more negative than negative-video (mean = 3.79, SE = 0.260, 95% CI = [1.50, 6.07]; difference = -0.160, SE = 0.065, df = 91, t = -2.469, p = 0.0154*) conditions. Simple-effect analysis according to the presentation condition showed that the differences between both the positive-live and negative-live conditions (difference = 3.54, SE = 0.162, df = 91, t = 21.787, p < 0.0001*) and the positive-video and negative-video conditions (estimate = 2.86, SE = 0.149, df = 91, t = 19.187, p < 0.0001*) were significant.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Statistical summary of valence ratings.

https://doi.org/10.1371/journal.pone.0290765.t001

We found a two-way interaction between the IRIEC and the emotional conditions. The 95% CI of the IRIEC slope was above zero under the positive condition (slope = 0.066, SE = 0.018, df = 91.6, 95% CI = [0.030, 0.101]) but below zero under the negative condition (slope = -0.058, SE = 0.016, df = 91.6, 95% CI = [-0.090, -0.026]; slope difference = 0.124, SE = 0.028, df = 91, t = -4.384, p < 0.0001*). We performed a further simple-effect contrast study comparing the positive and negative conditions at mean-centered IRIEC levels from -3 SD to 3 SD in steps of 1 SD, using the Šidák method to correct for multiple comparisons. Individuals with a middle-to-high trait emotional empathy (IRIEC levels of the mean minus two SD and above) tended to feel more positive when observing positive stimuli and feel more negative when watching negative stimuli (Table 2, Fig 4A); individuals with low trait empathic concern tended not to differ in terms of valence when observing stimuli of different emotional conditions.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Two-way interactions between empathic concern and emotional conditions on valence, arousal, and corrugator supercilii reactions.

The lines indicate the values predicted according to the trait scores. The semi-transparent band indicates the 95% confidence interval. A: The two-way interaction between the score on the empathic concern subscale of the Interpersonal Reactivity Index (IRIEC) and emotional conditions, as revealed by the valence ratings. B: The two-way interaction between the IRIEC and the emotional conditions, as revealed by the arousal ratings. C: The two-way interaction effects between the IRIEC and emotional conditions in terms of the CS responses.

https://doi.org/10.1371/journal.pone.0290765.g004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Simple-effect contrasts of the IRIEC-emotional condition interactions in terms of the valence ratings.

https://doi.org/10.1371/journal.pone.0290765.t002

Regarding data from the 50 participants of the behavioral study, robust estimation also showed a significant interaction between the IRIEC and the emotional conditions, but no interaction between emotional and presentation conditions (S2 Table). Robust estimation showed a significant interaction between the AQ and the presentation conditions, which was not present in the non-robust estimation. Still, the non-robust estimation after model diagnostics showed no such effect (p = 0.116), and the lower bound of the effect size confidence interval included 0.

Arousal ratings.

The final LME model for arousal ratings exhibited the same fixed effects as the valence rating model (Table 3). The random effects included the random intercept and random slopes for the emotional and presentation conditions and their interactions by the random factor “subject.” The random factor “Type” could not be included because it invariantly caused singular fits. We noted a significant interaction between the emotional conditions and presentation conditions. Simple-effect analysis according to the emotional condition showed that the interaction was attributable to a significant difference between the positive-live (mean = 6.66, SE = 0.115, 95% confidence interval [CI] = [6.44, 6.89]) and positive-video (mean = 6.10, SE = 0.122, 95% CI = [5.85, 6.34]; difference = 0.569, SE = 0.114, df = 91, t = 4.987, p < 0.0001*) conditions; there was a smaller difference between the negative-live (mean = 4.80, SE = 0.137, 95% CI = [4.53, 5.07]) and negative-video (mean = 4.54, SE = 0.122, 95% CI = [4.30, 4.79]; difference = 0.261, SE = 0.092, df = 91, t = 2.833, p = 0.0057*) conditions. Simple-effect analysis according to the presentation condition showed that the differences between both the positive-live and negative-live conditions (difference = 1.86, SE = 0.193, df = 91, t = 9.656, p < 0.0001) and the positive-video and negative-video conditions (estimate = 1.55, SE = 0.178, df = 91, t = 8.729, p < 0.0001) were significant.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Statistical summary of the arousal ratings.

https://doi.org/10.1371/journal.pone.0290765.t003

A significant two-way interaction was apparent between IRIEC scores and emotional conditions. Simple-slope analysis revealed that the 95% CI of the IRIEC slope was above zero under the positive condition (slope = 0.087, SE = 0.021, df = 91, 95% CI = [0.047, 0.128]) but included zero under the negative (slope = -0.023, SE = 0.024, df = 91, 95% CI = [-0.070, 0.025]; slope difference = 0.11, SE = 0.035, df = 91, t = 3.153, p = 0.0022*) condition. Simple-effect contrast analysis was performed comparing the positive and negative conditions at mean-centered IRIEC levels from -3 SD to 3 SD in steps of 1 SD, using the Šidák method to correct for multiple comparisons. Individuals with a middle-to-high trait emotional empathy (IRIEC levels of the mean minus one SD and above) tended to feel more aroused when observing positive stimuli than when observing negative stimuli (Table 4, Fig 4B); individuals with low trait empathic concern tended not to differ in terms of arousal when observing stimuli of different emotional conditions. Regarding data from the 50 participants of the behavioral study, robust estimation also showed a significant interaction between the IRIEC and the emotional conditions, but no interaction between emotional and presentation conditions (S4 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Simple effect contrasts of IRIEC-emotional condition interaction in arousal ratings.

https://doi.org/10.1371/journal.pone.0290765.t004

Zygomaticus major.

The final LME model of the ZM responses exhibited the same fixed effects as the rating models. The random effects included the random intercept and random slopes for emotional and presentation conditions over the random factor “subject,” as well as the random intercept over the random factor “Type” (Table 5). Adding the random intercept over the factor “Type” did not improve the model fitting (model comparison p = 0.9242), but we included it to account for possible influence by the study type. We found significant interactions between the emotional conditions and presentation conditions. Simple-effect analysis according to the emotional condition showed that ZM reactions under the positive-live condition (mean = 0.070, SE = 0.022, df = 28.26, 95% CI = [0.025, 0.115]) were stronger than ZM reactions under the positive-video condition (mean = 0.053, SE = 0.020, df = 21.87, 95% CI = [0.011, 0.095]); mean difference = 0.018, SE = 0.007, df = 185, t = 2.581, p = 0.0106*); ZM reactions under the negative-live condition (mean = -0.004, SE = 0.009, df = 1.10, 95% CI = [-0.096, 0.088]) was no different from ZM reactions under the negative-video condition (mean = 0.006, SE = 0.009, df = 1.42, 95% CI = [-0.055, 0.068]; mean difference = 0.011 SE = 0.007, df = 190, t = 1.568, p = 0.1186) conditions. Simple-effect analysis according to presentation condition showed that the ZM reaction under the positive-live condition was stronger than the ZM reaction under the negative-live condition (mean difference = 0.074, SE = 0.020, df = 93.1, t = 3.698, p = 0.0004*); the ZM reaction under the positive-video condition was stronger than the ZM reaction under the negative-video condition (mean difference = 0.047, SE = 0.020, df = 93.0, t = 2.312, p = 0.0230*). Regarding data from the 50 participants of the behavioral study, robust estimation also showed a significant interaction between the emotional and presentation conditions (S6 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Statistical summary of zygomaticus major (ZM) responses.

https://doi.org/10.1371/journal.pone.0290765.t005

Corrugator supercilia.

The final LME model of the CS responses exhibited the same fixed effects as the rating model; the random effects included the random intercept and random slopes for emotional and presentation conditions over the factor “subject,” as well as the random intercept over the factor “Type” (Table 6). Adding the random intercept over the factor “Type” did not improve the model fitting (model comparison p = 1), but we included it to account for possible influence by the study type. After model diagnostics, 8,862 data points from 84 participants remained in the model. We found significant interactions between the emotional conditions and the presentation conditions. Simple-effect analysis according to the emotional condition showed that although the CS was more relaxed under positive-live conditions (mean = -0.029, SE = 0.008, df = 1.55, 95% CI = [-0.078, 0.019]) than under positive-video conditions (mean = -0.017, SE = 0.008, df = 1.34, 95% CI = [-0.075, 0.041]; difference = -0.013, SE = 0.004, df = 181, t = -3.420, p = 0.0008*), there was no difference between the negative-live (mean = 0.0029, SE = 0.008, df = 1.27, 95% CI = [-0.060, 0.065]) and negative-video (mean = -0.001, SE = 0.008, df = 1.13, 95% CI = [-0.078, 0.075]; difference = 0.0039, SE = 0.004, df = 193, t = 1.077, p = 0.2828) conditions. Simple-effect analysis according to the presentation condition revealed stronger CS contraction under the negative-live condition than under the positive-live condition (mean difference = 0.0321, SE = 0.005, df = 103, t = 6.523, p < 0.0001*), and under the negative-video condition than under the positive-video condition (mean difference = 0.0162, SE = 0.005, df = 104, t = 3.279, p = 0.0014*).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. Statistical summary of corrugator responses.

https://doi.org/10.1371/journal.pone.0290765.t006

A significant two-way interaction was apparent between IRIEC scores and emotional conditions. Simple-slope analysis revealed that the 95% CI of the IRIEC slope was more negative under the positive condition (slope = -0.0011, SE = 0.0009, df = 78.3, 95% CI = [-0.003, 0.0006]) than under the negative (slope = 0.0010, SE = 0.0007, df = 87.0, 95% CI = [-0.0004, 0.002]; slope difference = 0.0021, SE = 0.0009, df = 76, t = 2.196, p = 0.0311*) condition. Simple-effect contrast analysis was performed comparing the positive and negative conditions at mean-centered IRIEC levels from -3 SD to 3 SD in steps of 1 SD, using the Šidák method to correct for multiple comparisons. Individuals with a middle-to-high trait emotional empathy (IRIEC levels of the mean minus SD and above) tended to relax CS more when observing positive stimuli than when observing negative stimuli (Table 7, Fig 4C); individuals with low trait empathic concern tended not to differ in terms of CS responses when observing stimuli of different emotional conditions. The IRIEC-emotional condition interaction was not significant in the absence of model diagnostics. However, the robust estimation showed the same significant result pattern as model diagnostics (S7 Table). Regarding data from the 50 participants of the behavioral study, robust estimation also showed a significant interaction between the emotional and presentation conditions, but not between IRIEC and the emotional conditions (S8 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 7. Simple effect contrasts of IRIEC-emotional condition interaction in corrugator responses.

https://doi.org/10.1371/journal.pone.0290765.t007