Abstract
It has been proposed that somatosensory reaction to varied social circumstances results in feelings (i.e., conscious emotional experiences). Here, we present two preregistered studies in which we examined the topographical maps of somatosensory reactions associated with violations of different moral concerns. Specifically, participants in Study 1 (N = 596) were randomly assigned to respond to scenarios involving various moral violations and were asked to draw key aspects of their subjective somatosensory experience on two 48,954-pixel silhouettes. Our results show that body patterns corresponding to different moral violations are felt in different regions of the body depending on whether individuals are classified as liberals or conservatives. We also investigated how individual differences in moral concerns relate to body maps of moral violations. Finally, we used natural-language processing to predict activation in body parts on the basis of the semantic representation of textual stimuli. We replicated these findings in a nationally representative sample in Study 2 (N = 300). Overall, our findings shed light on the complex relationships between moral processes and somatosensory experiences.
Keywords
Whether moral judgment is a product of reason or emotion has been an ongoing debate among philosophers and psychologists for decades. When moral psychology separated itself from moral philosophy, it almost exclusively focused on reasoning rather than on affective aspects of morality. The first empirical attempts in moral psychology started by examining cognitive-developmental components of understanding fairness and rules (Kohlberg, 1971; Piaget, 1948). But subsequently, as the field expanded, there was an increasing interest in the affective components of morality. Accumulating evidence suggests that emotion can ensue from, amplify, or directly cause moral judgment (Avramova & Inbar, 2013). Irrespective of the exact nature of the relationship between emotion and morality, distinct emotions are known to be associated with specific moral concerns as well as moral violations (Haidt, 2003).
Emotions are neural and somatic events that have the evolutionary function of preparing an organism to respond adaptively to a change in social or physical circumstances (Darwin, 1872). Once emotions are induced, individuals can consciously experience them by constructing a feeling, that is, generating a conscious mental experience (Damasio, 1999). Constructing feelings of emotions depends on brain systems that map and regulate body responses (Damasio & Carvalho, 2013). Both classic and modern theories of emotion postulate that interoception—the sensing of physiological feedback from the body and its visceral organs—is essential for emotional experience (Damasio, 1999; James, 1994; Schachter & Singer, 1962). The link between interoception and emotion continues to be supported by various studies. For example, Barrett, Quigley, Bliss-Moreau, and Aronson (2004) found that arousal focus, the extent to which individuals emphasize the changes of feelings in their verbal reports of experienced emotion, is related to interoceptive sensitivity. Individuals who were sensitive to their heartbeat change in response to emotion-arousal images reported more intense emotional experiences compared with less sensitive individuals (Barrett et al., 2004), supporting the association between body feedback and emotional states.
Even though emotions have been studied for a long time in psychology, the topographical distribution of the emotion-related body sensations has been identified only recently (Nummenmaa, Glerean, Hari, & Hietanen, 2014). Nummenmaa and colleagues (2014) mapped the “feeling space” for different emotions, demonstrating that consciously felt emotions are represented in the human body by topographically distinct maps with partial overlaps. Particularly, some emotions are associated with “activations” in certain body parts, whereas other body regions might be “deactivated” in the same emotional experience. For example, feelings of fear are paired with activations in the chest and head area, whereas feelings of sadness are represented by exorbitant deactivations in lower limbs and slight activations in the chest.
To our knowledge, topographical representations of moral emotions ensuing moral violations have not yet been studied. As mentioned, distinct emotions are known to be associated with various violations of moral norms (Haidt, 2003). Two decades ago, Rozin, Lowery, Imada, and Haidt (1999) proposed the contempt, anger, and disgust (CAD) hypothesis, indicating that the “other-condemning” moral emotions of contempt, anger, and disgust correspond to violations of the moral codes of community, autonomy, and divinity. Consistent with the CAD hypothesis and based on the intuitionist perspective on moral judgment, the moral-foundations theory (MFT; Graham et al., 2013; Haidt & Joseph, 2004) was developed by searching for the best links between anthropological and evolutionary accounts of moral intuitions across cultures. This framework suggests that moral intuitions derive from innate psychological mechanisms that coevolved with cultural institutions. Each moral system produces fast, automatic, gut-level reactions of like or dislike when certain phenomena are perceived in the social world, which in turn guide moral judgments of right and wrong. These systems, according to the MFT, have evolutionarily adaptive underpinnings present in individuals across cultural norms: care, fairness, loyalty, authority, and purity (Graham et al., 2013).
Violating the norms of care, fairness, loyalty, authority, and purity does not necessarily produce a uniform emotional response. Research suggests that specific predictions can be made regarding the types of response that violations of distinct moral norms may elicit. For instance, the specific-correspondence model (Chapman & Anderson, 2013; Kemper & Newheiser, 2018) posits that specific moral emotions map onto specific violation types. Therefore, any action in which an entity was harmed should reliably elicit anger, and any action in which a bodily norm was violated should elicit disgust (Rozin & Haidt, 2013; Russell & Giner-Sorolla, 2013). The specific-correspondence model thus suggests that witnessing harmful actions should uniquely activate a desire to confront the violator, whereas witnessing impure actions should uniquely elicit avoidance. By contrast, constructionist models of moral emotion posit that there are no exclusive links between moral-content domains and elicited emotions (Cameron, Lindquist, & Gray, 2015). Rather, contextual cues (e.g., framing language) and conceptual knowledge (e.g., who was harmed; Gray & Schein, 2012) inform the interpretation of moral violation, and no specific stimulus would predictably elicit the same emotion across all moral contexts (Cameron et al., 2015). Recent research suggests that individuals express distinctively high levels of desire to avoid (vs. confront) violators of purity norms (e.g., Dehghani et al., 2016). Violations of other moral norms, however, do not similarly elicit unique patterns of avoidance or confrontation. Thus, behavioral responses to moral violations depend in part on the norm that was violated, with impure acts eliciting a uniquely strong avoidance response (Kemper & Newheiser, 2018). Therefore, it stands to reason that the topography of different moral violations would reveal different maps on the human body (see Nummenmaa, Hari, Hietanen, & Glerean, 2018).
Here, we aimed to examine how emotions associated with violations of moral concerns are topographically represented in the body. We preregistered our specific research questions, all of which were written in the descriptive mode, an approach termed “informed curiosity” by Rozin (2001, p. 2). Specifically, we addressed four preregistered questions in Study 1. First, are body-sensation maps distinct for the five moral concerns posited by MFT? Second, can machine-learning techniques be used to distinguish body-sensation maps associated with different moral violations? Third, can we differentiate the body-sensation maps of people with different political ideologies (e.g., do liberals feel “purity” in the same regions of their bodies as conservatives) and individual differences in moral concerns (e.g., do people who score high on purity concerns feel “purity” in the same parts of their body as those who score low on purity concerns)? Finally, can the topographical body-sensation maps be predicted from the textual descriptions of moral-violation vignettes? In Study 2, we replicated our findings in a nationally stratified sample in the United States.
Study 1
Method
Participants
Nummenmaa et al. (2014) suggested using at least 40 participants in each group for studies examining self-reported perceptions of feelings. Following our preregistered intention as well as Nummenmaa et al. (2014), we aimed to recruit 600 participants in total, with 120 participants in each condition (i.e., each moral foundation). We requested 300 liberals and 300 conservatives from TurkPrime (Litman, Robinson, & Abberbock, 2017), and to compensate for potential attention-check failures, we recruited 630 participants overall. Following our preregistration, we excluded participants who failed attention checks, resulting in a total of 596 individuals (age: M = 36.5 years, SD = 11.8 years; gender: female = 355, male = 237, other = 3, unknown = 1). Participants were randomly assigned to five experimental conditions and completed the target task and several individual-differences measures, explained below.
Measures
Moral-violation scenarios
Participants were randomly assigned to five moral-violation conditions based on MFT, and each read a vignette about violation of a particular moral foundation (Clifford, Iyengar, Cabeza, & Sinnott-Armstrong, 2015). For each foundation, we selected vignettes from a larger pool of stimuli provided by Clifford et al. (2015). For each foundation (care, fairness, loyalty, authority, and purity), we included four vignettes, matched on average perceived wrongness, arousal, and frequency. A sample vignette is “You see a woman clearly avoiding sitting next to an obese woman on the bus.” After presenting the scenario, we asked participants to respond to the following questions on a 5-point Likert-type scale: “How morally wrong is the action depicted in this scenario?” (1 = Not at all wrong, 5 = Very wrong; M = 3.30, SD = 1.18) and “How strong was your emotional response to the behavior depicted in the scenario?” (1 = Not at all strong, 5 = Very strong; M = 2.90, SD = 1.12).
Body-sensation task
Bodily topography of feelings was mapped using the emBODY tool (Nummenmaa et al., 2014). The online data-acquisition package is publicly available at https://version.aalto.fi/gitlab/eglerean/embody. Yet some parts of our analytic framework were different from that used by Nummenmaa et al. (2014). Before engaging in the task, participants were shown a brief tutorial video to make sure they fully understood the task. Participants were then asked to color where activations (regions whose activity became stronger or faster) and deactivations (regions whose activity became weaker or slower) were felt in their body. The subtraction yielded a valid operationalization of pixel-level activation to be used in statistical analyses. Specifically, participants viewed two silhouettes, one on the right side and one on the left side of the screen. After reading the moral-violation scenario, participants were asked to draw on the portions of the silhouette where they felt activation (left silhouette) and deactivation (right silhouette). Individual-level bodily topographies were then computed on the basis of the difference between the left and right silhouettes. Final body-sensation maps were represented by 48,954 pixels.
Political orientation
All participants rated their political affiliation with the Republican party or the Democratic party along a 7-point scale ranging from 1 (Strong Democrat) to 7 (Strong Republican). Another item asked participants to rate their political conservatism on a scale ranging from 1 (Very Liberal) to 7 (Very Conservative). We averaged these two items to create a political-orientation score, on which higher scores indicated more conservative political orientation. A similar method was used in previous work for assessment of political ideology (Jost & Thompson, 2000). The internal consistency of these two items was relatively high in the current sample (α = .94). We labeled participants who scored 1 standard deviation (SD = 1.86) lower than the average political conservatism (M = 3.83) as liberal (n = 91) and those who scored 1 standard deviation higher as conservative (n = 126).
Moral concerns
Individual differences in moral concerns were measured using the 30-item Moral Foundations Questionnaire (MFQ; Graham et al., 2011). The MFQ assesses the degree to which participants deem different considerations as relevant when making moral judgments (1 = Not at all relevant, 5 = Extremely relevant) and their agreement with statements germane to morality (1 = Strongly disagree, 5 = Strongly agree). These items were used to create foundation-level scores for care (α = .69), fairness (α = .63), loyalty (α = .76), authority (α = .76), and purity (α = .83). Participants who scored at least 1 standard deviation above the average of the respective MFQ subscale were considered to have high levels of that moral concern, and those who scored at least 1 standard deviation below the average were considered to have low levels of that moral concern.
Procedure
This study was approved by the institutional review board of the University of Southern California (UP-16-00695-AM003). Potential participants were invited to take part in a psychological study on Amazon’s Mechanical Turk (MTurk) for monetary compensation. Participation was on a voluntary basis, and participants were compensated $0.50 for their time. Each participant completed the body-sensation task after reading a vignette about a particular moral violation and then completed a set of self-administered measures of political ideology and moral concerns. This study’s hypotheses, predictions, and analyses were preregistered on the Open Science Framework (OSF; https://osf.io/zbv6e/), and all data and materials have been made publicly available at https://osf.io/4tdx5/.
Data-analysis strategy
As discussed in our preregistration, we examined the topographic representation of each moral scenario across conditions, using the method described in Nummenmaa et al. (2014). Specifically, for each participant, a single map comprising 48,954 pixels was obtained by subtracting the activation and deactivation maps to obtain 48,954 variables per person in a given condition. The differences between activations and deactivations were then assessed using 48,954 univariate t tests: A one-sample t test against zero was performed for each pixel within a condition, resulting in a statistical t map. We then produced effect-size maps based on the t maps and sample sizes in each condition.
In order to classify the condition to which each participant was assigned, we used support-vector machines (SVMs; Cortes & Vapnik, 1995; Hearst, Dumais, Osuna, Platt, & Scholkopf, 1998), which are a class of supervised machine-learning algorithms considered to be robust to overfitting in classification settings. In the SVM training phase, a hyperplane is chosen that maximizes the margin of separation between the classes while also allowing for some data points to be misclassified. On the basis of the model built from training data, a new data point can then be classified using its position relative to the hyperplane. Further, nonlinear classification can be performed using SVMs by projecting the data into high-dimensional space using the so-called kernel trick. SVMs can also be used in high-dimensional settings in which the number of features (in our case, 48,954) is larger than the number of participants. Therefore, we preregistered and used SVMs that perform well under such constraints and are computationally efficient. In all of our analyses, we balanced the data points between the classes and performed 10-fold cross-validations 100 times to obtain estimates of variance in the performance of the models. We then used Fisher’s one-sample permutation test, with 5,000 permutations, to assess whether the classification algorithm performed significantly more accurately than would be expected by chance. Averaged accuracies (across the 100 folds), their bias-corrected, bootstrapped 95% confidence intervals (CIs), and p values associated with permutation tests are reported for each model.
To examine the role of political orientation in expressed mental representation of felt emotion, we divided the data by political orientation and investigated whether the maps in each condition were significantly different for liberals and conservatives (e.g., do liberals feel “purity” in the same part of their body as conservatives do?). Similarly, for each condition, we balanced the data points between the classes (i.e., liberals and conservatives) and, for each concern, performed 10-fold cross-validations 100 times to obtain estimates of variance in the performance of the models. We then compared the average accuracy against chance (50%) using Fisher’s one-sample permutation test, with 5,000 permutations, to investigate the reliability of the differences between the maps generated by liberals and conservatives.
We also broke down the MFQ responses to examine whether having high (vs. low) levels of a particular moral concern had a similar effect on self-reported body sensations in matching conditions. For each condition, we balanced the data points between the classes and performed 10-fold cross-validations 100 times to obtain estimates of variance in the performance of the models to examine whether high scorers and low scorers felt moral violations in different body regions. We then compared the average accuracy in each condition against chance (50%) using Fisher’s one-sample permutation test, with 5,000 permutations.
In a preregistered exploratory analysis, we explored whether the topographical maps for different moral concerns can be predicted from the textual description of that moral violation as stated in the vignette in natural language. Whereas typical approaches for the analysis of moral language (e.g., Garten et al., 2018) estimate the “moral loadings” of a piece of text on the basis of the moral words (as measured by a prespecified dictionary), here the vignettes described events that are potentially, but not necessarily, moral (Clifford et al., 2015). Therefore, we applied the InferSent 1 methodology (Conneau, Kiela, Schwenk, Barrault, & Bordes, 2017) to generate vector representations of each vignette. InferSent learns sentence-level representations by first training a supervised bidirectional recurrent neural network with long short-term memory (LSTM) on the Stanford Natural Language Inference data set (Bowman, Angeli, Potts, & Manning, 2015). The sentence-encoder function provided by InferSent retrieves pretrained word embeddings of each token in an input sentence and feeds them to the pretrained bi-LSTM model, described above, with maximum pooling. The output is a general-purpose sentence embedding that captures generic information useful for a broad set of tasks. These representations have been shown to achieve superior results compared with most available models. Specifically, the sentence-level embeddings produced by InferSent have been evaluated on 12 transfer-learning tasks and outperformed sentence embeddings learned by models trained in unsupervised conditions or on other supervised tasks. Therefore, we used the InferSent-trained model to obtain sentence encodings as the latent representation of the vignettes. What made InferSent appropriate for our study was that (a) it had been proposed for acquiring generic sentence-level semantic representations of natural-language data and (b) had achieved better results compared with available models described in the literature for sentence-level representations.
Using the best-performing predictive model, we encoded each moral vignette into a vector length of 4,096. We then ran 48,954 ridge regressions using InferSent vectors as predictors of activation of each pixel and then calculated R2 for each model to build an R2 map to visualize which parts of the body’s activation or deactivation can be explained by the semantic representation of textual stimuli.
We ran all analyses in the R (Version 3.4.1; R Core Team, 2017), Python (Version 3.6; Python Core Team, 2015), and Octave (Version 4.4.1; Eaton, Bateman, Hauberg, & Wehbring, 2019) programming languages.
Results
Bodily sensation of moral violations
Figure 1 displays the body-sensation maps associated with each moral violation. For each condition, we normalized the difference between activation pixels and deactivation pixels subject-wise. Then, we performed 48,954 one-sample t tests against zero to get the resulting t maps for each moral-violation condition. Next, we transformed these t maps into effect-size maps by dividing each pixel’s t value by the square root of sample size in that condition (Cohen’s d). Effect-size maps are visualized in Figure 1. As can be seen, in each condition, the moral violations are associated with slightly distinct distributions of body areas where activation was felt to either increase (or speed up) or decrease (or slow down).
Body-sensation maps of moral violations in Study 1. Each map shows regions where activation increased (warm colors; indexed by positive values) or decreased (cool colors; indexed by negative values) in a particular moral-violation condition. Results are shown separately for liberals and conservatives in each condition. The color bar indicates the effect sizes.
Classification of moral violations
As mentioned previously, we ran binary SVM classifiers with 10-fold cross-validation 100 times for each condition; in each fold, the target concern and a randomly selected subset of data from the other conditions (equal in size) were classified. For care, the average classification accuracy was 49.0% (95% CI = [47.96, 49.83]), which was not significantly higher than chance (p = .991). For fairness, the average classification accuracy was 51.5% (95% CI = [50.55, 52.45]), which was slightly greater than chance (p = .002). For loyalty, the average accuracy was 49.0% (95% CI = [47.95, 49.82]), which was not significantly higher than chance (p = .983). For authority, the average accuracy was 49.0% (95% CI = [48.04, 49.79]), which was not significantly higher than chance (p = .993). Finally, the mean classification accuracy for purity was 51.0% (95% CI = [50.14, 51.68]), which was slightly higher than chance (p = .009).
Individual differences and moral violations
Body-sensation maps of moral violations for liberals and conservatives are presented in Figure 1. We ran binary SVM classifiers with 10-fold cross-validation 100 times for each condition; in each fold, the political orientation of the left-out participants was predicted. The average classification accuracies for care (n = 45, M = 66.7%, 95% CI = [66.67, 66.67], p < .001), fairness (n = 36, M = 62.3%, 95% CI = [59.33, 64.00], p < .001), loyalty (n = 49, M = 74.0%, 95% CI = [72.00, 74.50], p < .001), authority (n = 48, M = 66.7%, 95% CI = [66.67, 66.67], p < .001), and purity (n = 39, M = 72.0%, 95% CI = [69.33, 74.67], p < .001) were significantly greater than chance. Therefore, for all moral concerns, we could classify political ideology on the basis of body-sensation maps, indicating that liberals and conservatives feel moral violations, especially perceptions of feelings of loyalty and purity, in different parts of their body. Although we preregistered our analysis, these findings should be interpreted with caution because the numbers of liberals and conservatives included in the study were small (36 ≤ n ≤ 48 in each group), and middle-of-the-road individuals were not included.
We also examined classification of high (vs. low) moral concerns based on body-sensation maps. We ran binary SVM classifiers with 10-fold cross-validation 100 times for each condition; in each fold, high scorers and low scorers on the MFQ subscales were classified. The average classification accuracies for care (n = 39, M = 66.3%, 95% CI = [64.33, 66.67], p < .001), fairness (n = 37, M = 61.0%, 95% CI = [56.67, 63.00], p < .001), loyalty (n = 37, M = 50.5%, 95% CI = [45.50, 54.50], p = .504), authority (n = 42, M = 49.8%, 95% CI = [48.25, 50.00], p = .999), and purity (n = 43, M = 58.5%, 95% CI = [53.25, 63.25], p < .001) were relatively high, except for loyalty and authority. Therefore, individual differences in moral concerns were associated with where in the body individuals report perceptions of feeling associated with moral violations. Body-sensation maps of moral violations for low scorers and high scorers are presented in the Supplemental Material available online. Of note, high scorers and low scorers formed small subsets in each group (37 ≤ n ≤ 43 per group), and these findings should be viewed with caution until further replicated.
Semantic representation of moral violations
As mentioned in our data-analysis strategy, we represented each vignette onto a 4,096-dimensional sentence embedding. We used these vectors to predict activation or deactivation of each pixel using a ridge regression with 10-fold cross-validation. We then computed an R2 map, indicating the explained variance in each pixel on the basis of the semantic representation of the textual stimuli. Results of these ridge regressions are presented in Figure 2. It can be seen that semantic representation of the texts used as experimental stimuli can predict larger proportions of variance in activation of the gut area.
Percentage of variance explained in the activation of different body parts in Study 1. Results are based on the semantic representation of textual stimuli.
In the second study, we addressed two limitations of Study 1. First, we collected a nationally representative sample to generalize our findings to American citizens in all demographic layers of the society. Second, we used a longer and more reliable measure to assess political orientation of the participants.
Study 2
Method
We recruited a nationally stratified sample (in terms of age, ethnicity, gender, and political orientation) consisting of 300 participants (age: M = 46.1 years, SD = 16.7; gender: female = 152, male = 148) through Qualtrics Panels (https://www.qualtrics.com/online-sample/). Our sample size was preregistered and included more participants per condition than Nummenmaa et al. (2014) suggested. Participants were randomly assigned to the five experimental conditions and completed the emBODY task (Nummenmaa et al., 2014). In addition to completing all individual-differences measures used in Study 1, participants filled out the 12-item Social and Economic Conservatism Scale (Everett, 2013; α = .86). In addition, internal consistency coefficients were acceptable for care (α = .75), fairness (α = .69), loyalty (α = .74), authority (α = .75), and purity (α = .81). Our data-analysis strategy and procedures were the same as those in Study 1. Study 2’s hypotheses, predictions, and analyses were preregistered on the OSF (https://osf.io/5hfcs/), and all data and materials are publicly available at https://osf.io/4tdx5/.
Results
In each condition, moral violations were associated with slightly distinct distributions of body areas, mirroring our findings in Study 1. The visualizations can be found in the Supplemental Material. For care, the average classification accuracy was 53.3% (95% CI = [50.25, 56.00]), which was slightly higher than chance (p = .011). For fairness, the average classification accuracy was 51.33% (95% CI = [49.46, 53.17]), which was was not different from chance (p = .067). For loyalty, the average accuracy was 49.1% (95% CI = [46.00, 51.80]), which was not significantly higher than chance (p = .747). For authority, the average accuracy was 50.3% (95% CI = [47.33, 52.92]), which was not significantly higher than chance (p = .414). Finally, the mean classification accuracy for purity was 48.3% (95% CI = [45.20, 51.10]), which was not higher than chance (p = .883).
We labeled participants who scored 1 standard deviation (SD = 19.03) lower than the average political conservatism (M = 64.99) as liberal (n = 50) and those who scored 1 standard deviation higher as conservative (n = 55). Body-sensation maps of moral violations for liberals and conservatives are presented in the Supplemental Material. The average classification accuracies for care (n = 26, M = 75.0%, 95% CI = [75.00, 75.00], p < .001), fairness (n = 20, M = 66.7%, 95% CI = [66.67, 66.67], p < .001), loyalty (n = 22, M = 65.3%, 95% CI = [62.00, 66.33], p < .001), authority (n = 19, M = 66.7%, 95% CI = [66.67, 66.67], p < .001), and purity (n = 18, M = 66.7%, 95% CI = [66.67, 66.67], p < .001) were significantly greater than chance. Therefore, for all moral concerns, we could classify political ideology on the basis of body-sensation maps, indicating that liberals and conservatives feel moral violations in different parts of their body. These results fully replicate those of Study 1.
We also examined classification of high (vs. low) moral concerns based on body-sensation maps. The average classification accuracies for care (n = 15, M = 66.3%, 95% CI = [64.33, 66.67], p < .001), fairness (n = 27, M = 85.7%, 95% CI = [85.71, 85.71], p < .001), loyalty (n = 26, M = 85.7%, 95% CI = [85.71, 85.71], p < .001), authority (n = 17, M = 100%, 95% CI = [100.00, 100.00], p < .001), and purity (n = 32, M = 91.7%, 95% CI = [91.67, 91.67], p < .001) were high. Therefore, individual differences in moral concerns are associated with where in the body individuals feel moral violations, replicating our findings in Study 1. Finally, results of these ridge regressions based on sentence embeddings are presented in the Supplemental Material.
General Discussion
The social-intuitionist model of moral cognition (Haidt, 2001; Haidt & Bjorklund, 2008) suggests that moral judgments are caused by emotional responses to a person, an action, or a violation. Drawing on intuitionist models of human morality (Haidt, 2001; Haidt & Joseph, 2004) and recent research on body maps of emotions (Hietanen, Glerean, Hari, & Nummenmaa, 2016; Nummenmaa et al., 2014; Nummenmaa et al., 2018), we conducted two preregistered examinations of body sensations associated with violations of different moral concerns.
In Studies 1 and 2, we demonstrated that moral-violation scenarios are associated with changes in activation and deactivation of specific body regions (Nummenmaa et al., 2018). The topographic maps associated with moral violations manifested more commonalities than differences (see Fig. 1). These similarities are consistent with the findings of Kemper and Newheiser (2018), who did not uncover evidence for a specific-correspondence mapping of behavioral response to moral violations based on MFT. Hence, our results are more in line with the constructionist model of moral emotions, indicating that there is no clear correspondence between foundation-level moral violation and self-reported perceptions of feelings associated with those moral foundations. It can be seen that across moral violations, the head and face area was highly activated, paired with varying levels of activation in the chest. The consistent activation observed in the head area suggests that people subjectively associate moral violations with high-level cognitive processing “in their head.” This is not surprising, because the moral-violation scenarios require a high level of cognitive and emotional processing, as well as an evaluation of personal states in relation to standards of right and wrong. The patterns of activation and deactivation observed for the head area consistently occurred across moral violations, and they may reflect subjective changes not only in the brain but also in the face. On the basis of the current data, we cannot infer whether these activations reflect perceived mental processing, facial blushing, or a combination of both. Deactivation of the limbs represented a consistent pattern across all moral scenarios. Of note, the chest area was activated in response to violations of care, fairness, loyalty, and authority, but less so in response to violations of purity. Instead, people reported higher activations in the abdomen area in purity.
Over the past decade, a body of research has shown that liberals and conservatives rely on different moral foundations and react differently to different moral violations (for a review, see Graham et al., 2013). Accordingly, we trained classifiers that could reliably predict political ideology on the basis of body maps. This is the first work indicating that political orientation influences where and how moral violations are felt in the body. These findings contrast null effects of political orientation on the link between moral transgressions and moral emotions (Landmann & Hess, 2018). Therefore, it is possible that, as we found, liberals and conservatives feel moral violations in different body regions, interpret them as distinct complex feelings (Landmann & Hess, 2018), and subsequently make different moral and political judgments. This seems to be a robust effect, because it was fully replicated in our nationally representative sample, which had the same breakdown of political ideology as in the United States.
We also trained classifiers that were able to reliably predict people’s moral concerns (high vs. low in moral foundations) in both MTurk and representative samples. It stands to reason that participants who score high on purity, for example, may have different reactions to impure actions than their counterparts who do not endorse purity values as strongly (e.g., Heerdink, Koning, van Doorn, & van Kleef, 2019). Those who are more concerned with purity are more sensitive to cues of degradation and purity violations; thus, such individuals are more likely to report stronger subjective activations or deactivations (or both) when primed with stimuli that violate relevant norms. As mentioned, these results are promising but should be viewed with caution until further replicated in future studies. Together, these results suggest that moral violations may evoke a “moral upset” that cannot be differentiated across moral foundations but can differ between groups (e.g., liberals vs. conservatives).
Semantic representation of textual stimuli (Clifford et al., 2015) predicted activation or deactivation of different body parts, especially in the abdomen or gut area. Research in cognitive linguistics and social cognition suggests that individuals construe the world in large part through conceptual metaphors, which enable them to understand abstract concepts using knowledge of superficially dissimilar but more concrete phenomena (Lakoff, 2016). Interestingly, the vignettes that we used did not include highly moral words; rather, they simply described a social scenario in which a particular moral norm was violated. These vignettes did differ from each other in their semantic representation in important ways to predict activation or deactivation of different body parts. We found that the semantic representations can locate body representation of moral violations, suggesting that the semantic space of natural language describing moral violations can be coupled with emotional states and their body-sensation maps.
Although we used different populations in Studies 1 and 2, we consistently found that (a) body maps of moral foundations were not substantially different from each other, (b) body maps associated with moral violations were reliably different between liberals and conservatives, (c) body maps associated with moral violations were associated with self-reported moral values, and (d) semantic representation of stimuli could predict activation or deactivation of body regions—especially the abdomen or gut area—in response to moral violations.
These studies have limitations worth noting. First, we collected self-report data regarding where activations and deactivations were felt in the body because our main objective was to examine representation of moral transgressions in the body; however, we did not collect data with regard to change in physiological states. Second, we collected data on two 2-D silhouettes to represent activation and deactivation of different body parts. Future research can use more accurate silhouettes, including 3-D ones, to better disentangle different body parts (e.g., occipital and frontal parts of the head). Third, we used textual stimuli to evoke body sensations; however, more vivid evocative stimuli (e.g., videos of moral violations) might evoke stronger responses. Fourth, we used vignettes that were matched on frequency and wrongness, representing only the moral foundations suggested by MFT. A good next step would be to include other morally controversial scenarios (e.g., transgressions from norms of honesty or humility) with varying levels of wrongness, frequency, and weirdness. Finally, we mention two constraints on generality of these findings for replication and follow-up studies (Simons, Shoda, & Lindsay, 2017). First, our study included only Western participants, and our findings cannot be generalized to other cultures. Second, we used a limited number of moral-violation vignettes. Therefore, in follow-up studies, researchers are encouraged to test boundary conditions of the present findings and use alternative sets of validated stimuli.
Supplemental Material
Dehghani_Open_Practices_Disclosure – Supplemental material for Body Maps of Moral Concerns
Supplemental material, Dehghani_Open_Practices_Disclosure for Body Maps of Moral Concerns by Mohammad Atari, Aida Mostafazadeh Davani and Morteza Dehghani in Psychological Science
Supplemental Material
Dehghani_Supplemental_Material_rev – Supplemental material for Body Maps of Moral Concerns
Supplemental material, Dehghani_Supplemental_Material_rev for Body Maps of Moral Concerns by Mohammad Atari, Aida Mostafazadeh Davani and Morteza Dehghani in Psychological Science
Footnotes
Action Editor
Leaf Van Boven served as action editor for this article.
Author Contributions
M. Dehghani developed the study concept and collected the data. M. Atari and A. Mostafazadeh Davani analyzed and interpreted the data under the supervision of M. Dehghani. M. Atari drafted the manuscript, and M. Dehghani provided critical revisions. All authors approved the final version of the manuscript for submission.
Declaration of Conflicting Interests
The author(s) declared that there were no conflicts of interest with respect to the authorship or the publication of this article.
Open Practices
All data and materials have been made publicly available via the Open Science Framework and can be accessed at https://osf.io/4tdx5/. The design and analysis plans for the studies were preregistered at https://osf.io/zbv6e (Study 1) and https://osf.io/5hfcs (Study 2). The complete Open Practices Disclosure for this article can be found at http://journals.sagepub.com/doi/suppl/10.1177/0956797619895284. This article has received the badges for Open Data, Open Materials, and Preregistration. More information about the Open Practices badges can be found at 
.
Notes
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
