Empathy

At large, empathy refers to the thoughts and feelings of one individual in response to the observed (emotional) experiences of another individual (e.g., Davis, 1983; cf. Woltin et al., 2011). However, similar to most psychological constructs, the specific definition of empathy is debatable. Batson (2009) identifies eight definitions of empathy that are currently in use; of these eight, we will expand on two that are commonly applied. The first definition refers to knowing the internal state of another person and has been referred to as cognitive empathy, perspective taking, empathic accuracy, and theory of mind (Batson's Concept 1). This type of empathy, referred to in this paper as Cognitive Empathy, is related to better negotiation outcomes and satisfaction with the negotiation process (e.g., Bazerman and Neale, 1982; Galinsky et al., 2008), reduced stereotypes (e.g., Todd et al., 2011), and enhanced intergroup contact (e.g., Falk and Johnson, 1977). The second definition proposed by Batson (2009) refers to feeling the “same emotion that another person feels” (page 5) and has been referred to as emotional contagion, affective empathy, and emotional empathy (Batson's Concept 3). This type of empathy, referred to in this paper as Affective Empathy, refers to an affective connection with another person's emotional state, and closely relates to an other-oriented reaction regarding the perceived welfare of the other person (cf. Batson et al., 1997). Affective empathy is related to behavioral mimicry (Dimberg et al., 2011) and the motor neuron system facilitating behavioral mimicry (Nummenmaa et al., 2008).

Although both types of empathy imply the processing of others' emotional states, emotion-specificity in both the concept as well as the assessment of cognitive and affective empathy has not yet been considered. In the following we will elaborate on why considering emotion specificity in empathy is worthwhile as it presents an advancement in the theorizing and measurement of the concept of empathy.

Emotion specific empathy

In the following section, we present several arguments that suggest researchers may need to consider empathy from an emotion specific perspective. First, we refer to a prominent theoretical model of empathy, the Perception-Action Model (PAM) of empathy (cf. Preston and de Waal, 2002; Preston and Stansfield, 2008; Preston and Hofelich, 2012), which (implicitly) assumes emotion specific empathy. PAM assumes that for an empathic reaction to occur, the observer (i.e., the subject) must attend to another individual's (i.e., the object's) emotional state. By attending to the object's state, the subject's neural representations of similar states are automatically activated, thus, other-oriented object-congruent reactions are facilitated. Herein, PAM provides a theoretical framework that relates empathy to concepts such as facial mimicry (e.g., Sonnby-Borgström et al., 2003), emotional contagion (e.g., Blairy et al., 1999), and recognizing emotions in the face (e.g., Dimberg et al., 2011). PAM suggests that highly empathic individuals are those who are more successful and/or more motivated to attend to others' emotional states and thereby often have similar neural representations consequentially activated. This explains why those who are highly empathic show facial mimicry at an earlier stage of information processing (Sonnby-Borgström et al., 2003), are better at detecting others' emotions (Dimberg et al., 2011), and display more emotional contagion (Blairy et al., 1999).

Importantly, the entire process outlined in PAM implies emotion specificity given that the type of emotional state observed in the object by the subject should elicit different neural representations, setting the stage for different empathic responses and consequently leading to different facial mimicry reactions, emotional contagion, and behavioral reactions. Whereas most research has identified the same areas may be implicated for several emotions (cf. Cunningham et al., 2008; Kober et al., 2008; Shackman et al., 2011; Lindquist et al., 2012; Oosterwijk et al., 2012) some findings indicate that feeling different emotions are associated with activity in different brain regions (e.g., Lane et al., 1997; Pelletier et al., 2003; Tettamanti et al., 2012). However, the potential differentiation in brain region and neural networks, even at the level of positive and negative affect, suggests this differentiation might carry through to the level of specific emotions.

Nevertheless, the extent to which one can empathize with another individual may vary depending on the empathized emotion as the PAM suggests. Correspondingly, in other avenues, research suggests emotion specificity may exist in empathy related abilities, such as mimicry, emotion memory, and emotion recognition (e.g., Oberman et al., 2007; Ponari et al., 2012; Wilhelm et al., in press). An emotion specific conceptualization and measurement of empathy will instigate future research that disentangles the relation between, for example, empathizing with someone's sadness and mimicking his/her sadness, happiness, or fear. An emotion specific perspective on empathy appears more appropriate to us, not only because—as elaborated above—empathic reactions to sadness should imply different processes compared to empathic reactions to joy, anger, or disgust (cf. Lane et al., 1997; Preston and de Waal, 2002; Pelletier et al., 2003; Tettamanti et al., 2012), but also because being able to empathize with another's sadness (in contrast to another emotion such as anger) should be related to different behavioral outcomes and help us further understand individual differences in empathy.

Finally, on a behavioral level, we argue individuals will produce behavioral reactions that differ depending on the emotion of interest. For example, perceiving sadness in others increases affiliative tendencies and helping behavior while this is not the case for perceiving anger (Hendriks and Vingerhoets, 2006; Van Doorn et al., 2012). While affectively connecting (i.e., showing affective empathy) with a sad person increases one's willingness to help that person (see Batson et al., 1989, 1991, 1997), it is debatable whether affectively connecting with an angry person results in the same increased willingness to help. This is because anger and sadness differ in their underlying appraisals (cf. Frijda, 1986), which in turn affects behavioral consequences of these emotions (cf. Horstmann, 2003; Seidel et al., 2010). Anger, for example, is an emotion that arises when a sudden obstacle appears in the pursuit of one's goals, leading the subject to perceive unfairness and wish to fight the obstacle (Haidt, 2003). Consequentially, anger is associated with approach-related, aggressive behavioral tendencies (e.g., Carver and Harmon-Jones, 2009). Hence, it appears plausible that an individual who is sensitive to others' anger (i.e., a highly anger-empathic individual) might be more likely to engage in aggressive behavioral acts than if he or she was highly sadness- or fear-empathic.

Taken together, we argue that at least three main arguments speak to the fact that a differentiated perspective on empathy is well indicated. First, theoretical models of empathy suggest emotion specificity (cf. Preston and de Waal, 2002). Second, considering empathy as emotion specific yields new research questions and potentially important insights on the relation of empathy and related concepts and abilities such as mimicry, emotion recognition, and emotion memory, because different emotions are associated with different appraisals and information processing (cf. Frijda, 1986). Finally, given specific appraisals are associated with different emotions, different consequences might be associated with empathy when considered emotion specific.

Subscale statistics

Next, we computed the mean values of the subscales. On all subscales women self-reported higher scores compared with men, which is consistent with other self-report empathy scales (e.g., Jolliffe and Farrington, 2006; Carré et al., 2013). These differences were significant for all subscales, with the exception of Anger Cognitive Empathy and Surprise Cognitive Empathy, however the effect sizes, assessed with Cohen's d (Cohen, 1992), were all small to medium (see Table ​Table11).

The mean values of the subscales were correlated with each other to test the convergent and discriminant validity between the subscales (see Table ​Table3;3; Campbell and Fiske, 1959). The items were derived from the same stem by instantiating emotion specific content so any reduction in the correlation from 1.0 would indicate a discrimination between the subscales that could be attributable to the emotion content differences. The correlations between the ESE affective subscales range from moderately to strongly related (r's range from 0.43 to 0.77; average r = 0.63; Cohen, 1992), with the average shared variance amongst the scales at 40%, suggesting that while the subscales are similar to one another they are not perfectly related and are instead distinct. Correlations between the cognitive subscales were stronger compared with the strength of the relations between the affective subscales (r's range from 0.64 to 0.79; average r = 0.70; average R² = 0.49); likewise, the strength of these correlations and the average shared variance suggest the emotion specific cognitive empathy subscales are also distinct. Finally, correlations between the emotion congruent subscales ranged from moderate to strongly related (r's range from 0.35 to 0.60; average r = 0.45; average R² = 0.20) suggesting that while distinct, there is convergence (indicated by moderate to strong relations between the subscales and an average shared variance above null) between the emotion congruent empathy subscales. These correlations were in general larger than the correlations between the affective and cognitive empathy subscales that were emotion incongruent (r's range from 0.14 to 0.48; average r = 0.31; average R² = 0.10) suggesting discriminant validity between the affective and cognitive empathy subscales and between emotions.

Table 3

ESE subscale correlations and covariances.

Open in a separate window

^*p < 0.05; In the top half of the table, correlations are presented below the diagonal, variances on the diagonal, and covariances above the diagonal. Correlation coefficients in the top left quadrant indicate the convergent validity between the affective empathy subscales and correlation coefficients in the bottom right quadrant indicate the convergent validity between the cognitive empathy subscales. Correlation coefficients in the bottom left quadrant indicate the discriminant validity between the emotion specific affective and cognitive empathy subscales, with the exception of the those coefficients presented in boldface, which indicate the convergent validity between the emotion specific subscales (i.e., Anger Affective Empathy with Anger Cognitive Empathy). The bottom half of the table presents correlations between the ESE and IRI subscales.

Finally, the ESE subscales were correlated with Empathic Concern and Perspective Taking, from the IRI. When correlated with Empathic Concern, all of the ESE affective subscales show stronger relations compared with their ESE cognitive empathy counterpart. And, when correlated with Perspective Taking, all of the ESE cognitive empathy subscales, with the exception of happy show stronger relations compared with their ESE affective empathy counterpart.

ESE measurement model

Because there is a lot of overlap between the subscales in terms of what they measure (i.e., affective subscales measure affective empathy, however each affective subscale should also measure an emotion specific empathy and thus is conceptually related to its emotion-specific cognitive empathy counterpart), we tested a series of hypothesized measurement model structures for the ESE measure, with the 12 subscales as indicators, that apply constraints to test specific hypotheses about the structure of the ESE measure and the differentiation between emotion specific empathies. Because we hypothesized emotion specific differentiation in empathy, the presented models focus on an emotion specific differentiation structure to the ESE measure. Each model tests specific hypotheses regarding the convergent and discriminate validity of the emotion specific empathies, ranging from complete discrimination between constructs to complete convergence. These models apply various constraints in a Multitrait-Multimethod framework (MTMM; Widaman, 1985; see Figure ​Figure11 for hypothesized models). Because of known convergence problems with the final solution of correlated trait and method factors, this structure was not modeled (Kenny and Kashy, 1992).

Open in a separate window

Figure 1

Hypothesized Measurement Model Structures based on the MTMM framework.

Regarding the modeling of method variance, we opted for the inclusion of two methods factors (i.e., an affective empathy latent variable and a cognitive empathy latent variable) instead of the correlated uniqueness model (Marsh, 1989) or the correlated-trait correlated-method-1 structure (CTCM-1; Eid et al., 2008) in order to identify latent constructs for each empathy construct measured by ESE, so that later these latent constructs are available for testing covariate relations. In addition, the correlated uniqueness model is known to have bias when there are strong loadings on methods factors and strong correlations between method factors, which is a characteristic of our data (see Figure ​Figure22 below for loadings on method factors; Conway et al., 2004). Because the emotion factors are indicated by only two manifest variables, we added an additional constraint to adequately identify the emotion latent variable by constraining the unstandardized loadings from both manifest variables to equality forcing the affective and cognitive emotion subscale to load equally on the corresponding emotion latent factor³. Through an evaluation of model fit indices we can identify the best fitting structure to the data (see Table ​Table44).

Open in a separate window

Figure 2

Final ESE Measurement Model (Model 6).

First, we modeled a single general empathy factor for all 12 ESE subscales (Model 1), suggesting a single type of empathy was measured by all subscales; however, this model showed poor fit to the data. Next, we modeled emotion specific latent variables, which were either constrained to be orthogonal (Model 2) or allowed to covary (Model 3); both showed poor fit (with Model 3 producing an improper solution, most likely because the emotion factors are indicated by only two variables each, something that Gerbing and Anderson, 1985, 1987 found often results in models with improper solutions) suggesting modeling only differentiation between emotion specific empathies cannot account for the data.

The last three models (Models 4–6) included an affective and cognitive latent variable, indicated by the six affective or cognitive empathy scales, that were constrained to be either orthogonal to one another or covary. Of these three models, Model 6 showed acceptable fit, with the exception of the RMSEA statistic (which will be discussed next), and showed the best fit compared to the other five models, so this model was selected as our final measurement model⁴ (see Figure ​Figure22).

The RMSEA statistic for Model 6 was above the recommended value of 0.06 made by Hu and Bentler (1999); however, the value is also considered mediocre fit by MacCallum et al. (1996) and good fit by Steiger (1989). RMSEA can be easily inflated when a model has the correct number of specified factors but has incorrectly omitted a single covariance between residuals (Savalei, 2012). An examination of the modification indices suggests that the addition of a covariance between the residuals of Happy Affective Empathy and Sad Affective Empathy, which is weak in size (r = 0.12, p < 0.05), results in a lower RMSEA statistic [χ²₍₃₂₎ = 97.50, p < 0.05, SRMR = 0.044, RMESA = 0.065_{(0.051−0.080)}, CFI = 0.987, TLI = 0.973]. However, we refrain from including this covariance in our final measurement model because we did not specify this covariance a priori (Steiger, 1990), the additional covariance will most likely not replicate in future samples (MacCallum et al., 1992), and omitting this covariance will most likely not impact the basic correlational pattern between latent factors (Newcomb and Bentler, 1988).

Model 6 identified orthogonal affective and cognitive factors, which suggest the ESE measure has adequate differentiation between the affective and cognitive subscales, and emotion specific latent factors, suggesting there is emotion specific empathy for the six basic emotions. The emotion latent variables were allowed to correlate with one another, showing strong correlations with one another, suggesting that individuals who self-report high scores on one emotion empathy will also report high scores on the other empathies. This model also supports our earlier hypothesis of emotion specific differentiation in empathy.

Limitations and future directions

The ESE measure is structured based on the six basic emotions (anger, disgust, fear, happy, sad, and surprise). This is just one framework through which to address emotion, however given the emotion bias we found, we suggest that future research should test other proposed emotion structures, such as examining positive and negative affect. While the ESE measure includes more negative affect subscales than positive, we attempted to test this structure with the current data. We remodeled the data and removed the emotion specific latent variables and instead modeled only two latent variables, one for negative empathy (indicated by the ESE items for anger, disgust, fear, and sad empathy) and one for positive empathy (indicated by the ESE items for happy empathy). Because we modeled these structures at the item level, the models were specific to only affective or only cognitive empathy, because otherwise with 50 manifest variables (i.e., 5 items per affective and cognitive empathy subscale), that model would have required us to estimate more parameters than is suggested with our current sample size. Modeling the data in this way presumes that there is only a positive and negative specific empathy, and that the emotion categories are not a proper structure to the data. However, model fit for both affective and cognitive empathy was poor⁶ suggesting that this structure did not fit the data. The measure, however, contains more items assessing emotions considered to have a negative valence, compared with the items considered to have a positive valence, which leads to perhaps a better identification of negative affect than the data allows for positive affect. A proper balance of positive and negative emotions would allow us a better dataset to properly test for higher-order positive and negative affective and/or cognitive empathy factors. This is fruitful direction for future research, however it was not addressed here.

We would like to thank Richard Roberts, Jeremy Burrus, and Mattis Geiger for their help with data collection.