1. Introduction 

Human altruism is considered an evolutionary puzzle (Nowak, 2006).1 While repeated interaction and reputation formation can explain altruistic behavior in small closely knit groups, altruism is often considerable among strangers who interact only once. Here, we investigate the hypothesis that altruism is caused by feelings of shame and pride and that these feelings are accentuated by others' verbal evaluation.

Our evidence comes from a dictator game experiment with recipient feedback. We find that the opportunity for verbal feedback substantially increases donations compared to a control treatment without any feedback.

In a typical dictator game, one person (the divider) is in charge of dividing a resource between herself and another person (the recipient). Usually, the resource is an amount of money, and the divider is free to choose any division. While nothing prevents the divider from taking all the money, a substantial fraction of the dividers leaves some money to the recipient. In laboratory experiments in Western cultures, the equal split is the second most common allocation, with the average donation typically falling in the interval 10–30% (Camerer, 2003). Henrich et al. (2004) report similar results from subject pools that are isolated from Western culture.

The (equal) sharing norm is widespread, and there are several reasons for thinking that generosity in dictator games is driven by a desire for social esteem in relation to this norm. When the divider's choice is observable to the experimenter (Hoffman, McCabe, Shachat, & Smith, 1994) and to the recipient (Andreoni and Bernheim, 2007, Dana et al., 2006 ), the division becomes more generous. Even pictures of eyes, subtly triggering a sense of being watched, has a significant positive impact on generosity in both laboratory experiments (Haley & Fessler, 2005) and field settings (Bateson, Nettle & Roberts, 2006). This evidence is congruent with the long-standing view that prosocial behavior is fuelled by the desire to feel justified pride and to avoid feeling shame (Cooley, 1922, Scheff, 1988, Smith, 1790), and more generally with the desire to signal favorable characteristics (Bliege Bird and Smith, 2005, Henrich and Gil-White, 2001). Importantly, these are quite tight associations. Andreoni and Bernheim (2007) formally demonstrate that a signaling model based on an equal sharing norm predicts the observed spikes in dictator game data [see also Ellingsen and Johannesson (in press) and the references therein].

Our preferred interpretation of the result that anticipated verbal feedback matters for behavior is that feelings of shame and pride in connection to norm compliance are accentuated by emotional communication. Knowing that someone is justifiably angry does not arouse the same level of shame as facing the angry person, or even having to read angry messages. The hypothesis that people are motivated by a desire for praise is not new. It can be tracked back at least to Thomas Hobbes (1651, Chapter XI), who wrote: “Desire of praise disposeth to laudable action, such as pleaseth them whose judgment they value.”2 In the paper's final section we discuss evolutionary foundations for Hobbes's hypothesis in light of recent theories and evidence.

Although the applied psychology literature has long been finding significant effects of symbolic rewards in the workplace and in schools (Henderlong and Lepper, 2002, Stajkovic and Luthans, 2003), such studies suffer from the problem that symbolic rewards may be closely correlated with subsequent material rewards. A similar objection may be directed at the few experimental studies that consider symbolic punishment in repeated public goods games (Gächter and Fehr, 1999, Masclet et al., 2003, Noussair and Tucker, 2005). While behavior does tend to become more prosocial when communication is possible, it is difficult to know whether behavior changes in order to reduce the suffering generated directly by the verbal sanction or to mitigate the effect that verbal messages have on future material payoffs. Relatedly, the fact that people are generally more cooperative when they are allowed to interact face to face or merely to see their opponents, as documented by Ostrom and Walker, 1997, Bohnet and Frey, 1999 respectively, could in principle be due either to increased empathy or to reputational effects. Together with Xiao and Houser (2007) (to be discussed below) ours is the first study to isolate the impact of anticipated feedback on altruistic behavior.

2. Experimental design and procedures 

We conducted a one-shot anonymous dictator game experiment with and without ex post recipient communication in the form of an unrestricted written message. The dictator game was chosen over the more popular ultimatum game in order to avoid confounding altruism with risk aversion or false beliefs. Moreover, we wanted to study the generosity of the divider rather than the recipient's willingness to engage in costly punishment. According to Koenigs et al. (2007), charity and punishment engage different parts of the brain. We choose written messages over other forms of communication both for simplicity and in order to maintain anonymity.

In both our treatments, one subject (the divider) decided how to allocate SEK 120 between herself and another subject in another room (the recipient) (SEK=Swedish Kronor; US$1≈SEK 7.5 at the time of the experiment). In the feedback treatment, the recipient had the opportunity to send a message after learning the divider's allocation; in the control treatment, the recipient had no communication option.

The subjects were recently enrolled undergraduate business and economics students at the Stockholm School of Economics. The experiment was conducted in early September 2006. Subjects were randomly allocated between the two treatments and we carried out five sessions (three with the feedback treatment and two with the control treatment). A total of 276 subjects participated in the experiment yielding 134 pairs of observations (85 in the feedback group and 53 in the control group). Subjects in a pair were anonymous with respect to each other, and the decision of a specific subject could not be observed by other subjects or the experimenters. The two treatments are further described below (the complete experimental instructions are available in Appendix B in the electronic supplements).

In the feedback group, subjects were recruited to two separate rooms called Room A and Room B. Dividers were in Room A and recipients were in Room B. The subjects were welcomed and told not to talk to each other. Subjects in both rooms received numbered instruction sheets. Subjects in Room A also received an envelope marked with the same number that contained six SEK 20 bills. The subjects read the instructions, and thereafter they were allowed to ask questions individually and privately (after first raising their hands).

The experimenter in Room A called one person at a time and the subject took her envelope and went behind a screen. In private behind the screen, the subject decided how many SEK 20 bills to leave in the envelope and how many to keep for her own use. The subject then sealed the envelope and dropped it in a box marked “Mail.” The subject thereafter returned to their seat. When all subjects in Room A had made their decisions, the experimenter gave the box marked “Mail” to an assistant that was waiting outside of Room A. The assistant took the box with the envelopes to Room B and distributed the envelopes to the respective recipients. Each recipient opened the envelope and pocketed any money in the envelope. The recipients then wrote down the amount in the envelope on a form marked “Result/Message.” The recipients were told that they had the opportunity to write a message on the form and to thereafter put the form in the numbered envelope. The assistant then collected the envelopes and gave them to the experimenter in Room A (the assistant thereafter left the room). The experimenter then distributed the envelopes to the dividers in Room A. The dividers were told to open the envelope and read the message and thereafter to fold the “Result/Message” form, without putting it back into the envelope. The experimenter then passed around the box marked “Mail” and the dividers were told to put the folded forms into the box (the forms now lacked any identification numbers and a specific form/decision could therefore not be linked with any specific divider in Room A). The experiment was then over.

In the control group, the experiment proceeded in exactly the same way until the recipients had written down the result on the results form. The recipients were told to fold the form after they had written down the amount in the envelope, and the assistant then passed around the box marked “Mail” and the recipients put the forms in the box. The experiment was then over.

We also carried out a classification of the messages in the feedback group. The messages were classified according to whether they expressed disapproval, approval or a neutral evaluation of the divider. To classify the messages we used the approach of Xiao and Houser (2005). We recruited 11 message evaluators among the undergraduate students at Stockholm School of Economics (who did not participate in the experiment or any previous dictator game experiment). The evaluators were asked to classify each message into one of the following three preset categories: positive feedback/approval, negative feedback/disapproval, neutral feedback. The evaluators were given a copy of the experimental instructions and after reading the instructions they classified all the messages in the study (the messages were provided in a random ordering). The evaluators were not given any information about the decision of the divider. Evaluators were paid SEK 150 for participating. To motivate the evaluators to attend carefully to their task, three messages were also randomly chosen for real payments. If the evaluator's evaluation matched the most popular evaluation for this message they earned an additional SEK 50 (the maximum total earnings were thus SEK 300). Messages were classified according to the most popular classification chosen by the evaluators. There were only two ties, which were broken by our own evaluations.

3. Results 

Fig. 1 shows the distribution of donations in the two treatments. On average, subjects in the control group donated 24.84% of the endowment to the recipients. In the feedback group the average donation was 34.12% of the endowment, an increase of almost 40%. The difference between the groups is statistically significant at the 5% level (p=.023 according to a nonparametric Mann–Whitney test; two-sided p value).3 The fraction of subjects who donate zero decreases substantially with feedback, from 42% to 21%. With feedback there is also a sizeable increase in the fraction of subjects that divide the endowment equally between the subjects. This fraction increases from 30% to 48%.

Fig. 1. The distribution of donations in the experiment.

The messages themselves also provide interesting information. A complete translation of all messages is provided in Appendix A in the electronic supplements. All 18 recipients who received no money sent messages to the dividers, and 17 of them expressed disapproval. Many of these messages communicated strongly negative views, frequently in foul language. Of the 41 recipients who received half the endowment, 38 sent a message to the divider. All of these messages except two praised the divider. For the recipients who received amounts of 20% or 40%, reactions were more mixed with about half the messages expressing approval and about half expressing disapproval or being neutral.

Common definitions of shame involve a concern for what others think about one's behavior or character, whereas guilt is predominantly a matter of self-assessment, involving remorse and empathy (Lindsay-Hartz et al., 1995, Tangney and Dearing, 2002, Fessler, 2004). Since the negative messages are predominantly angry, blaming the divider's bad behavior and assaulting his or her character, they appear designed to induce shame rather than guilt.

An alternative explanation to our pride and shame hypothesis is that the receiver uses emotional contagion (Hatfield, Cacioppo, & Rapson, 1993) as a reward mechanism. Emotional contagion describes the tendency of one person's emotions to rub off on others. While some of the positive messages may possibly be understood as an attempt to let the divider share the responder's feeling of happiness, the negative messages are typically angry. It appears unlikely that the purpose of these messages is to induce anger in the divider.

Message writers sometimes try to remove the divider's sense of anonymity. For example, one message includes the threat “I'll come looking for you” and another message states, “I know who you are” — despite the fact that subjects are located in different rooms and cannot identify their partner. Another recipient writes: “The person to the right of you (nr 26) will get a long and happy life.” Presumably, the writer here has seen that recipient 26 has gotten some money, or at least wants to convey this impression, and passes on this information to one's own less generous divider. While the statement about the location of divider 26 is of course pure speculation, these are the kinds of details that could potentially make the message's reader feel “observed” and thus more uneasy about one's own lack of generosity. Two subjects apparently even revealed their identity (either their name or their student number) in the message, although there is no way of knowing whether these revelations were truthful without contacting the students with these identities.

As we did not measure dividers' emotions, we cannot attribute states of shame or pride to them. However, our findings are consistent with our hypothesis that people feel shame and pride when considering others' opinion of them, and that others can magnify the feelings through emotional verbal feedback. Given this interpretation pride and shame in turn promote altruistic behavior.

4. Discussion 

Our findings demonstrate that anticipated verbal feedback in the form of anonymous written messages induces a substantial increase in altruistic behavior towards otherwise defenseless opponents. Since an anonymous written message is a mild form of feedback compared to naturally occurring personal communication, the identified effect is likely to underestimate the behavioral impact of anticipated emotional feedback in more realistic settings.

Comparing the donation level in our feedback condition with ultimatum game proposals reported by Ellingsen and Johannesson (2005) for an almost identical subject pool, we find that they are both about 35%. With the caveat that the comparison is not based on strict randomization, and that it neglects differential framing effects, the impact on divider behavior of anticipated verbal feedback is thus of similar average size as the effect of the recipient's punishment power in ultimatum games. Moreover, since material sanctions are themselves expressive of anger, it seems likely that some of the disciplining effect of rejections in the ultimatum game is due to the divider's desire to avoid the feeling of shame rather than just the anticipated material loss.

Symbolic rewards and punishments have important advantages compared to material ditto. Material punishments often use up resources, and material rewards necessitate a transfer of resources. Symbolic rewards and punishments neither use nor transfer resources.

Independently of us, Xiao and Houser (2007) have conducted a similar study. Like us, they study a dictator game in which the recipient could send an anonymous message to the divider after the allocation decision. However, there are several design differences. Notably, Xiao and Houser restrict the set of offers to seven discrete levels, ruling out the possibility that the divider takes more than 90% of the available amount. Broadly, the findings coincide, but our effects are larger in magnitude and statistical significance, and our increase in equal splits is not matched by their data. There are several possible explanations for these differences. Xiao and Houser's offer restriction and the fact that they ask their subjects not to use foul or threatening language imply that verbal feedback is milder, and the fact that they do not insist that subjects read the messages means that the subjects can protect themselves from strongly negative feedback. (We did not force our subjects, but we believe that all dividers read the recipients' messages.) The differences in findings between the experiments are thus well in line with our hypothesis.

The same authors, Xiao and Houser (2005), also studied how the possibility of feedback messages affects recipients in ultimatum games. They found that ex post verbal feedback among recipients decreased their likelihood of rejecting an unfair offer, suggesting that expression of disapproval is a substitute for monetary punishment.

A proximate explanation for emotional sensitivity to communication runs as follows. Shame is associated with a desire to flee and hide from others, presumably to avoid punishment. Being exposed to explicit disapproval naturally exacerbates the sensation of threat. Indeed, the visible displays of inferiority — averted gaze, shrinking posture — associated with shame only make sense in the presence of others. Likewise, feelings of justified pride are accentuated by laudatory speeches, even if the speeches contain no new information. If so, altruistic behavior ought to increase when recipients can provide verbal feedback. The recipient's approval is a symbolic reward; the recipient's disapproval is a symbolic punishment. Consistent with our hypothesis, work in social psychology indicates that feedback in the form of praise raises the level of pride and satisfaction in performing a task (Webster et al., 2003, Gaines et al., 2005). Interestingly, studies of computer–human interaction similarly document a significant impact of the affective content of computer messages on user satisfaction and behavior [see, for example, Partala and Surakka (2004)].

If the feelings of pride and shame, and the associated search for praise and avoidance of blame, constitutes the proximate mechanism behind our findings, what is the (ultimate) evolutionary mechanism? In Chapter 5 of Descent of Man,Darwin (1871) proposes that natural selection is at work and proposed an explicit evolutionary theory for people's sensitivity to others' approval and disapproval. Darwin offered a group selection argument. More recently, several authors have instead proposed individual selection arguments for emotions such as shame and pride [see e.g., Keltner, 1995, Weisfeld, 1997, Gilbert and McGuire, 1998, Fessler, 1999, Fessler, 2004)]. One key idea is that desirable behavior is rewarded by high in-group rank (or at least group inclusion) — which in turn produces survival and/or better mating opportunities. The association of high (low) in-group rank with pleasant (unpleasant) feelings is therefore adaptive.

Our view that people are aware of their propensity to feel shame and pride, and are unable to fully control these feelings, has additional testable implications. For example, if dictator game dividers are given freedom to choose whether to read the feedback messages, generous dividers ought to be more prone to read. Likewise, the theory explains why people prefer not to obtain information about whether a self-interested action will cause harm to others, as documented by Dana, Weber, and Kuang (2007). Finally, the reputational theory has the potential to explain field evidence on cooperation in common pools. According to Ostrom (1990), such cooperation tends to be sustained primarily through the use of small symbolic fines.

The reputational theory of altruism does not logically preclude other theories, such as strong reciprocity (Gintis, 2000, Fehr et al., 2002) and empathy (Batson et al., 1988, Batson and Shaw, 1991, Singer et al., 2004, Fehr and Singer, 2005). It seems that some people behave altruistically even when all reputational cues are absent (Johannesson & Persson, 2000). Quite possibly, altruistic behavior is caused by several proximate mechanisms, just as the evolution of these proximate mechanisms may have been caused by several ultimate mechanisms (Nowak, 2006).

Disentangling the relative importance of different mechanisms requires empirical ingenuity. Since it is difficult to discriminate between evolutionary theories based on observed actions alone, a potential avenue for future work is to study how activity in the brain is associated with anticipated and actual experiences of approval and disapproval. Brain imaging has usefully located brain activity associated with humans' expression of moralistic aggression (de Quervain et al, 2004). We are unaware of corresponding studies concerning the brain activity of the targets of moralistic aggression. Recently, several studies have used brain image comparisons of people with and without certain lesions to identify the neural causes of moral behavior. For example, Koenigs et al. (2007) find that the ventromedial prefrontal cortex (VMPC), which is associated with social emotions, affects the pattern of moral judgments when self-interest is pitted against social interests. They argue that the VMPC is essential for the generation of normal moral judgments. However, these studies do not address specifically the emotions of pride or shame. Berthoz, Grezes, Armony, Passingham, and Dolan (2006) document that an old part of the brain, the amygdala, is involved in affective responses to one's own moral violations. The amygdala has previously been found to be central in the processing of fear (e.g., Miller, Taber, Gabbard, & Hurley, 2005, and the references therein). If the amygdala also plays a major role in generating emotional responses to blame, that would seem consistent with the evolutionary theories described above, linking shame to subordination. If the amygdala plays a similar role in the connection to praise, that would corroborate the view that shame and pride are opposite poles of the same underlying emotion; if separate parts of the brain process blame and praise, that would be an indicator that the feelings of shame and pride may not have the same evolutionary origin.