Empathy

Looking inside the black box of social preferences, one component that has been hypothesized to play a central role is empathy, the capacity to share the feelings of others. Adam Smith (1789, p. 10) already highlighted in The Theory of Moral Sentiments the potential role of empathy in pro-social behavior: “Empathy is the source of our fellow-feeling for the misery of others, that is by changing places in fancy with the sufferer, (it is) that we come either to conceive or to be affected by what he feels” [13].

More recently, a very similar stance has been taken in the perception-action model of empathy [14]. The model suggests that it is sufficient to observe or imagine someone else in an emotional state to trigger an empathic response. In a review article on the role of empathy and ToM in economics, the authors suggest that if empathy implies the shared experience of emotions, this can undermine the idea that choices are based solely on self-interest [15]. This is captured, for instance, by the inequity aversion model by Fehr and Schmidt (1999), in which an agent can experience not only positive but also negative utility if the agent's own and others' monetary outcomes differ [2]. In line with this reasoning, empathy has been listed among a number of pro-social emotions, such as guilt and shame, which underpin pro-social behavior in human decision making [16].

Individual differences in the capacity for empathy reflect differences in pro-social behavior in domains such as volunteering and donating [17]. In a group of young adults, measures of pro-social dispositions have been found to be stable across a period of five years and relate to ratings of empathy [18]. According to the empathy-altruism hypothesis, empathy is even regarded as the exclusive source of genuine altruism [19]. Empathic feelings are classically associated with helping someone in need [20], [21]. In the context of the Prisoner's Dilemma Game, inducing empathy via portraying the other person as needy resulted in participants choosing to cooperate more often than in a neutral control condition even when participants knew beforehand that the other person would defect on them [22].

Neuroscientific experiments have indicated that people differ in their empathic capacity and that this might relate to individual differences in social preferences [23]. Inviting couples into the laboratory, participants received a painful stimulus and the resulting brain signal was compared to that obtained when they were merely informed that their loved one had received such a stimulus [24]. The same affective brain circuits were active both when receiving pain and when being informed about the beloved one experiencing the same painful stimulus. Importantly, the higher the activation in the pain circuits, the higher participants scored on a psychometric measure of empathy, the Interpersonal Reactivity Index, which is also employed in the current investigation [25]. In a further study participants were initially exposed to a situation where they were treated either fairly or unfairly by a matched partner [26]. In the second part of the experiment, the matched partner received a painful stimulus. For men, empathy-related brain activity was found only in those who faced a fair partner and not in those who faced an unfair partner. For women, such a difference was not observed. The authors link these results on empathy to social preference theories, where people value others' gains positively if they are fair, but the gains of unfair partners are negatively valued.

A number of researchers have hypothesized about the importance of individual differences in empathy for fairness as observed in neutrally framed laboratory games [15], [27], [28]. If allocators were explicitly asked to put themselves in the shoes of the recipient in the DG, offers increased as compared to individuals in a control group that were not given any specific instructions [28]. This illustrates that empathic behavior can be induced by changing the social context. In this study, we specifically focus on individual differences in empathy in a given context and how these relate to social preferences. Singer (2008, p. 264) reemphasizes a point made earlier by Singer and Fehr (2005) that “one prediction that can easily be made is that people with a greater ability to empathize should display more other-regarding behavior” [4], [27]. However, a clear-cut demonstration that individual differences in empathy underlie social preferences as classically measured in neutrally framed laboratory games is still missing.

Hypothesis 1: The greater the individual's capacity of empathy the higher the offer in the DG.

Theory of mind

Empathy relates to the capacity to share social emotions. In contrast, ToM allows an agent to understand the social reasoning and social emotions of others. Utilizing ToM, the decision maker constructs the mental states of others making inferences about beliefs, intentions, and emotions [29]. The concept of ToM was initially proposed by primatologists who suggested that it emerged as a result of the social challenges of living in larger coalitions [30], [31].

In economics, Smith (1776) noted in The Wealth of Nations that understanding the goals and beliefs of one's trading partner facilitates business [32]. The centrality of ToM to strategic interaction implies that it closely relates to issues addressed in game theory. One basic assumption is that of “common knowledge,” which implies that interacting agents reflect on the action of the other and know that the other does the same [33]. More recently, the assumption that all people make the same inferences as others do has been challenged, and individual differences between agents in their thinking steps and hence in their capacity for ToM have been considered central to explaining a number of empirical phenomena [34], [35]. This echoes the hypothesis that Singer and Fehr (2005) postulated, by which individuals who have a higher capacity for ToM can better predict others' motives and actions [27].

Hypothesis 2: The higher the individual's capacity for ToM the more accurate the stated beliefs about the behavior of others.

ToM is thought to serve two functions: (i) It facilitates the pursuit of one's personal gains, a function that in the psychological literature has also been labeled Machiavellian intelligence [30]; and (ii) it facilitates pro-social behavior, as shown, for instance, in non-human primates where the capacity for ToM can lead to acts of spontaneous helping [36]. Surprisingly, to date, among humans there exists evidence only for the second function but not for the first. For instance, people high on the Machiavellian scale score low on the capacity for ToM, that is they do not seem to effectively employ ToM to further their own goals [37], [38].

Using economic games, researchers have studied individual differences in ToM among children but not adults: In the UG, 6- to 10-year-old autistic children, who are impaired in ToM, were more likely to accept low offers and to refuse fair proposals compared with a normally developed cohort [39]. Likewise, preschoolers who had acquired the capacity for ToM made higher mean offers in the UG than those who had not yet developed this capacity. This has led to the proposition that individual differences in ToM are a psychological component that underlie social preference theories in which those with a higher ToM are fairer [40].

Yet, these findings do not square with the proposition made in economics and game theory that those who seek to maximize their own monetary profit can benefit from an accurate estimation concerning others' likely action. The seminal study of Kelley and Stahelski (1970) on the role of beliefs in social dilemmas allows a reconciliation of these contradictory propositions [41]. The authors showed that cooperative individuals are more accurate than selfish individuals in predicting the actions of others in a social dilemma. They argued that this difference is functional in nature: To pursue their strategy successfully and reduce the likelihood that someone else takes advantage of their pro-social tendency, cooperative people need more accurate beliefs than defectors. For defectors it is unnecessary to make accurate inferences in a one-shot social dilemma, as doing so does not affect their strategic choice which is always to defect regardless what the other does. For those that consider cooperating, it is vital to accurately estimate what others do in order to be able to deduce the potential risk that is entailed in choosing to cooperate. A question that follows from this is whether people that tend to favor a strategy with a pro-social outcome are generally more accurate in estimating what others do regardless the specific context of the game. We used the DG in an attempt to distinguish the principal social preferences of agents, that is, whether the agents seek to maximize their own payoff or also care about the payoff of others.

Hypothesis 3: When comparing fair and selfish participants as measured by their offers in the DG, fair participants will have more accurate beliefs about the decisions of other participants than selfish ones.

However, the situation changes when the focus is on those proposers in the UG who seek to maximize their own payoff. The more accurately a proposer judges the likelihood of a certain offer being rejected, the higher the expected payoff. Hence, in this context it becomes functional for the selfish proposers to have accurate beliefs.

Hypothesis 4: Selfish participants, as measured by offers in the DG, are more likely to employ accurate beliefs for their own material self-interest in the UG than fair participants.

Behavioral measures

The behavioral measures of our experiment consist of three separate sections. The first section includes three decisions to measure individuals' concern for fairness. The second section elicits participants' belief about the behavior of other participants in the first section. The third section measures participants' risk attitude using incentivized risk elicitation task. Decisions in all sections are incentivized by randomly selecting one decision from each section for the final payment (please see the Instructions S1 for the Experimental Instructions).

To measure an individual's concern for fairness, we had participants make decisions in three different roles: As a proposer in the UG and in the DG and as a responder in the UG. When participants must indicate their actions in all possible roles this has been labeled the strategy vector method [44]. When the strategy vector method has been compared to situations where participants play only one role in bargaining experiments, no major differences has been found [45]. These previous results lend support to the argument that individuals' concerns for others as measured in laboratory experiments are robust to elicitation through the strategy vector method. In the role of the responder, participants indicated the minimum acceptance level below which they would reject an offer. At the end of the experiment, one of the three decisions was picked randomly to calculate the monetary payoff. Participants were not informed whether they were matched with a different participant in each decision. This feature may in principle affect the behavior of various individuals.

In both games, the proposer was endowed with 90 experimental currency units (ECUs) and could split these in intervals of 10. An equal split is known to be the modal offer both in the UG and the DG. Hence, our research design precluded the possibility of making the modal offer. It has been shown that after removing the equal split, fair offers become less frequent [46] and that responders are less averse to unequal outcomes [47]. This provides a particularly interesting test bed for individual differences in ToM.

Participants were additionally asked to indicate their beliefs about the likely action of a randomly chosen partner in each of the three decision tasks (DG offer, UG offer, and UG minimum acceptance level). They thereby stated the probability they considered most likely for a given action by answering the question such as the following example for DG offers: “Please indicate the likelihood that a randomly determined person taking part in this experiment has chosen one of the 10 possible divisions”. Beliefs about the actions of others were rewarded based on the Quadratic Scoring Rule (QSR) [48]. It is used to measure and reward the accuracy of predictions. In our case, each participant states a probability vector r = (r₁,…,r_n) where r_i refers to the probability that an event i occurs. The QSR is then used to give a reward of Q(r,j) where j is the event that actually occurs. The functional form of the QSRis strictly concave with respect to r_i in which case the highest possible earning is received by placing 100% on the occurring event and zero probability on any other event. Participants get penalized for placing a positive probability on events that actually do not occur. The penalty is disproportionately increasing in the probabilities placed on those events. We set α = 10 and β = 10 which guarantees a positive payoff unless the participant assigns 100 per cent probability on a single false event in which case the participant receives no payment. An empirical assessment of various proper scoring rules (logarithmic, quadratic and spherical) show that participants are likely to employ various suboptimal reporting strategies when the score contains both positive and negative scores [49]. In our experiment, we have limited the QSR score to positive scores except in case where 100 percent probability is assigned to a single false event to facilitate truthful revelation. Schotter and Trevino [50] present a recent survey of the literature on belief elicitation in laboratory experiments and discusses among other techniques the QSR. Our implementation of the rule closely follows the standards for incentive compatible belief elicitation in economic experiments.

The QSR rule induces truthful revelation of subjective beliefs about the expected behavior of other participants in strategic games and is widely used to elicit beliefs in behavioral experiments [51]. To ensure full comprehension of the payoff mechanism, participants were carefully instructed on the procedure. Each participant had to go through three learning episodes with control questions that became increasingly difficult. A post-hoc analysis of the data shows that 93 per cent of participants provided the right answer to all three questions. At the same time, 96 per cent of participants provided the right answer to the most complex question (Question 3). However, a considerable effort and learning through trial and error was needed to gain full comprehension. Had we allowed only one attempt per question, only 52 per cent of the participants would have been able to find the right answer to all three questions.

For the purpose of paying, the QSR score for each participant in a given task was calculated on the basis of how well stated beliefs predicted the decision of another, randomly selected, participant. In order to make a robust assessment of the accuracy of beliefs and their relation to social preferences in the results section of the paper, the reported belief distribution of each participant was compared to the distribution of choices of all other participants. The belief assessment of every participant was matched with each of the observed outcomes in the study population, resulting in 115 scores, as there were 116 participants in total. We then computed for every participant a mean score for the three separate tasks and an overall score that is aggregated over the three tasks. These scores could range from zero to two. A score of zero indicates that the participant assigns 100 percent probability to a single false outcome. A score of two indicates that the participant assigns 100 percent probability to a single correct outcome.

When beliefs and actions from the same participants are elicited in an incentive-compatible manner, participants might hedge across tasks that are independently incentivized. In other words, participants might take in some task higher risks to increase their chance of higher earnings and compensate for this in other tasks where they take lower risks ensuring that they get a minimum payoff. Previous studies however show that unless hedging opportunities are very prominent, results are unlikely to be confounded due to hedging [52]. Because participants did not know about the belief elicitation until they were actually asked to state their beliefs, the prominence of hedging opportunities between the games and belief elicitation tasks was not an issue here. Eliciting a separate belief distribution in each game may encourage participants to balance their reported beliefs. For example, participants could report systematically wider belief distributions in a certain game to secure a high minimum payment if this task gets chosen for the payment. As we only paid one randomly chosen belief elicitation task, such hedging should be minimized.

A possible remaining confound is that participants adjust their reported beliefs to their risk attitude. Likewise, the practice of paying participants based on the action chosen by one randomly determined person may affect the reported belief distributions. This may invite risk-seeking participants to report narrower distributions centered on the option that they believe to be the most likely event. Risk-averse participants may be inclined to report flatter distributions to secure a high minimum payment. Not only beliefs but also behavior in the games might be shaped by individual differences in risk aversion. Greater risk aversion can, for example, lead to higher offers of proposers in the UG as these are less likely to be rejected.

In order to control for individual differences in risk aversion regarding the stated beliefs and decisions in games we used the Holt-Laury lottery task [53]. This is an incentive-compatible task where participants choose 10 times between two different lotteries that gradually vary the combination of probabilities and monetary outcomes in order to measure each individual's degree of risk aversion.

Psychometric measures

Individual differences in the capacity for ToM and empathy have traditionally been assessed using psychometric tests. We used two tests that both measure empathy and ToM. The two empathy measures are convergent, whereas the two ToM measures complement each other: (i) “Cold” ToM is about inferences regarding the epistemic state of others and refers to the knowledge, beliefs, and intentions that someone else holds; (ii) “hot” ToM is about inferences about others' emotions [54].

To measure cold ToM and empathy, we employed the Interpersonal Reactivity Index (IRI),which is the most widely used psychometric test to evaluate both empathy and ToM. The test has been extensively investigated and validated [25]. The German translation of the IRI, the Saarbrücker Persönlichkeits-Fragebogen [55] was used in the experiment. The IRI is a self-report questionnaire using abstract descriptions of social interaction which participants respond to. The modified version we used had four dimensions, each containing four statements. Two dimensions were used for the final analysis: Cold ToM was measured using the perspective-taking scale where participants responded to statements such as, “I try to look at everybody's side of a disagreement before I make a decision.” The capacity for empathy was measured with the scale for “empathic concern”. Participants had to respond to statements such as, “I often have tender, concerned feelings for people less fortunate than I.” Individual differences in the empathic concern scale correlate positively with brain activity associated with empathy [24]. Participants indicated the extent to which a statement described them on a 5-point scale, from does not describe me well to describes me very well. Cronbach's α was .73 for the ToM measure and .80 for the empathy measure, which correspond to the values found by Davis (1980) [25].

A potential problem with psychometric tests, in particular regarding participants that are primarily motivated to increase their own earnings, is that responses are not incentivized. Specifically, Lyons, Caldwell, and Shultz (2010) point out that this might be problematic when it comes to assessing the relationship between ToM and Machiavellianism [38]. In their experiment, participants high on the Machiavellian scale showed a lack of effort in the ToM task, indicating that they might not have been sufficiently motivated by non-incentivized tests to reveal their ToM level. Note that measuring how accurate a participant's believes are what others will do in the games should be equivalent to how accurate a person is on inferring the epistemic state of someone else. The former is assessed using the QSR and is incentive compatible, the latter is assessed using the psychometric scales for ToM.

To measure hot ToM and empathy, we used the Multifaceted Empathy Test (MET), which employs 40 realistic photographs of faces expressing positive or negative emotions as stimuli [56]. The MET was originally developed to measure individual differences of empathy and ToM for people with autism as these have difficulties with abstract descriptions employed for instance by the IRI. The test is also suitable to measure individual differences in a normally developed cohort. It uses pictures as stimuli. Participants answered for each picture three types of questions reflecting three subscales. The subscale emotion recognition measured hot ToM by assessing to what extent participants could correctly infer the emotional state of others as depicted in the photographs. Participants answered the question, “What does this person feel?” by selecting one of four possible options, where only one was correct. A similar test for hot ToM, relying on emotion recognition in faces, was used by Paal and Bereczkei (2007) and Lyons, Caldwell, and Shultz (2010), who found a positive relationship between cooperativeness and the capacity for ToM [37], [38]. To measure empathy, the Multifaceted Empathy Test provides two subscales: The subscale direct empathy asked participants to answer the question, “How strongly do you feel with this person?” The subscale indirect empathy asked the question, “How aroused are you by the picture?” Direct and indirect empathy were measured with a 9-point scale ranging from not at all to a lot. Cronbach's α was .95 for direct empathy, .96 for indirect empathy, and .70 for hot ToM, which correspond to the values found by Dziobek et al. (2007) [56].

Experimental procedure

Participants did not know about the content of the separate sections of the experiment at the beginning of the experiment but were informed about the subsequent tasks after the completion of each section. However, participants were aware of the content of the individual tasks within each section before submitting their decisions. For example, when reporting their beliefs about the expected behavior in the DG, participants knew that they are also requested to report their beliefs concerning UG behavior.

Half of the participants played the DG, then the UG, followed by the psychometric tests; the remaining participants completed the psychometric tests first and then played the games. We did not find any order effects between the psychometric tests and games. We therefore have pooled the data. Subsequently, participants indicated their beliefs about the likely actions of others in the games. This was followed by measuring participants' risk attitude. Questions on the demographic background of each participant ended the experiment. Participants received their earnings in cash immediately after the completion of the experiment.

When analyzing the data we exclude from the dataset four non-native German speakers, who indicated limited ability to fully respond to the verbal descriptions of emotional states used in one of the psychometric tests. Furthermore, the data reveal that 13 participants reported non-monotonic risk preferences. These individuals are excluded from the analysis when analyzing the impact of risk aversion on participants' decisions and belief formation.