C
OUNT
TO
TEN
FIRST
ON PUNISHMENT AND THE HOT
-
COLD EFFECTby Cora Hollander, 10215115 // second version: October 28, 2013
MSc Behavioral Economics and Game Theory // Supervisor: prof. dr. F.A.A.M. van Winden
“When angry, count to ten before you speak. If very angry, count to one hundred.” Thomas Jefferson
1. INTRODUCTION
“When you are upset, count to ten before you react.” This piece of folk wisdom captures the idea that how we react in the heat of a debate or dispute might not be the reaction that we prefer when we are in a more thoughtful state. Moreover, the heated reaction may be costly to oneself and something that we later regret. Does this advice contains a truth? Do you behave differently when you would 'count to ten first'?
This is not only a relevant question in social life, but for economists as well. Each economic transaction takes place in a relationship between two or more parties, for example an employer and an employee or a customer and a supplier. Trust is a vital element of such a relationship. In the absence of complete contract enforcement, parties rely on each other to stick to their part of the deal. When trust is violated, there often exists the opportunity for retaliation or punishment. This can take many forms. Think for example of a misled customer that starts to spread negative rumors about a certain supplier, or of an employer leaving bad references about a former employee. Retaliating behavior often destroys wealth, meaning that both (or even more) parties suffer from it. However, its existence is not all bad. The threat of punishment can help deter exploitative behavior. Understanding when and why people punish, helps to explain this inefficient behavior and may help designing mechanisms that prevent it or enforce its power toward stimulate cooperation.
This study investigates one particular feature of punishment, that is the timing of the punishment decision. Provided that there occurs no new information during the delay, there seems to be no reason why a difference in timing influences the decision. This point of view disregards the impact of the emotional state the decision maker is in. Emotions, or more generally visceral factors, are known for
their ability to override deliberative reasoning and have shown to be very influential in decision making (Zajonc, 1984; Damasio, 1994; LeDoux, 1996; Loewenstein, 1996). When in a state of high emotional intensity or 'hot' state, people often act in ways they themselves view as against their self-interest. When in a less intense or 'cold' state, it is hard to imagine how one would behave in hot state and vice versa. This phenomenon has been dubbed the 'hot-cold empathy gap' and is thought to be particularly present for negative emotions (Loewenstein, 2000). The motivation for punishment rises from a negative experience, when people feel that someone misbehaved. This can come with feelings of disappointment, anger or wrath and therefore induce the urge to punish even when this is costly to oneself. When the flood of emotions has waned, so may the urge to punish. Investigating that suggestion is the focus of this study. I aim to answer the question: Does costly punishment occur less frequently when a time lag is imposed on the punishment decision?
To answer this question properly we have to establish first that people engage in a trust relationship, second that violation of trust gives a distinct emotional response, and third that this emotional response is a decisive factor in the consideration of punishment. In order to study this, I conducted a behavioral laboratory experiment. How trust and its violation are exactly measured will be explained in Section 2 (p. 4-5). To reflect on the impact of emotions, I need a situation in which people can feel angry or disappointed about another persons behavior and subsequently have the option to act on these feelings in a costly manner. The trust game or investment game (Kreps, 1990; Berg et al. 1995) is a very useful vehicle to realize this. The game allows for engaging in a relationship of trust, and the violation of this trust. The main outline of the game applied in this study is as follows (see Section 3 for more details). Two players are anonymously coupled in a pair. One of them – called the Investor – is endowed with an amount of money and has to decide how much to transfer to the other participant – called the Trustee. The transferred amount is multiplied by three. The Trustee then has to decide how much to transfer back to the Investor. A variant of the game adds a third stage in which the Investor has the possibility to punish the Trustee by giving up some of its own money in order to reduce the other person's earnings.
When the decision to punish has to be made directly after observing the Trustee's transfer, the reaction to this transfer and the decision to punish almost coincide. In other words, when contemplating punishment the Investor is still in the emotional hot state that might arise from observing its counterpart's choice. In the control condition, the game is played exactly as just described. In the treatment
condition, participants perform a intermediate task before Investors have the opportunity to punish. This task consists of three reasoning questions and takes approximately five minutes. Its main purpose is to cause a small delay between observing the outcome of the mutual transfers and the punishment decision. Specifically, the intermediate task serves to make time for any emotions to ware off and bring the Investors in a cold state before they decide on punishment. If punishment occurs significantly less often in the treatment condition, the folk wisdom to count to ten first indeed contains a truth.
This study is a pilot study, which means that there are limited resources available in terms of location, money and time. I will discuss the consequences of these practical restrictions in Section 7. I originally planned to collect 48 observations (48 people in 24 pairs playing the game twice). If Trustees would send back zero about half the time (as is common), there would occur ten situations per treatment that are likely to provoke punishment. This is still a very small sample, but could be enough to offer an indication on the hypothesized treatment effect. While performing the sessions, I noticed that Trustees almost always choose the equal split, leaving no motivation for punishment. Since it was clear that in this pattern I would never observe enough punishment situations to be able to do statistical testing, I terminated the experiment. Hence the amount of data collected is not sufficient to answer the research question. To answer the question whether there is a difference in frequency of punishment between treatments, a first requirement is that punishment is observed at all. I observed zero instances of punishment, so I cannot tell whether there is a systematic difference between the control and treatment condition. The presentation of the results will therefore focus more on how the observed behavior can be explained rather than delivering an answer to the research question.
The remainder of this paper is organized as follows. Section 2 presents an overview of the existing literature on trust games, punishments, and hot-versus-cold studies. Section 3 describes the experimental design and derives hypotheses. Section 4 provides a theoretical analysis. Section 5 proposes a set-up for empirical analysis. Section 6 presents the results. Section 7 discusses the results and procedures of this study and provides suggestions for improvements in future research. Section 8 concludes.
2. LITERATURE REVIEW
Traditionally economics has assumed that people are rational decision makers with perfect foresight who pursue only their self-interest. It is now widely recognized that other than selfish considerations play an essential part in both social and business life. A large body of experimental evidence established that people do take factors of trust, reciprocity, and fairness into account (see Camerer, 2003 for a review). Several simple games have been developed that mimic the key features of economic interactions and that can measure social preferences (i.e. displaying behavior that deviates from rational and selfish choice).
Consider for example the ultimatum game in which one player, the proposer, offers a split of a certain amount of money between himself and a second player, the responder. The responder can then accept this split or reject it such that both players receive nothing. Although accepting any offer is always the best response in the traditional economic sense, people often reject offers in which they would get only a small part of the pie. Rejection in the ultimatum game can be viewed as punishment, since the responder pays a cost (foregoing a positive amount of money) such that the other player is left with no earnings as well. Many proposers seem to anticipate this and offer (close to) a fair split. Proposers offer substantial amounts even in dictator games, when the responder does not have the option to reject (see again Camerer, 2003). In gift-exchange games, two players sequentially choose how large a gift that want to send to each other. Second movers often reimburse large amounts sent by first movers (e.g. Charness, 2004; Fehr and Gächter, 2000).
Fehr, Gächter and Kirchsteiger (1997) added a third stage to the game in which first movers can reward or punish the other party. First movers utilized this option frequently, and the amount sent by second-movers increased dramatically. The availability of punishment has shown to sustain cooperation also in public good games (Fehr and Gächter, 2000). Egas and Riedl (2008) find similar results, but stress that this is only the case if the cost-impact ratio (the cost born by the punisher and the impact on the player being punished) is low enough. Taken together, these results clearly demonstrate that people have a taste for equality or fairness at least to some extent. This taste is such an important motivation that people are willing to incur losses (i.e. pay for punishment) to express it.
Since this study in particularly concerned with trust and punishing a violation of trust, I will look at these concepts more closely. Recall that the trust game allows to measure trust and trustworthiness. In defining trust, I follow here Lukas and Walgenbach (2010): “Trust in persons means accepting that they may take
advantage of a trustor although the trustor expects them not to do so.” This definition captures two distinctive features of trust. First, to trust means to recognize that there is a chance at a loss and still show the “willingness to be vulnerable” (Mayer et al., 1993) by engaging in the act. Second, a trustor can only form an expectation about the trustworthiness of the trustee and never be sure of what the trustee will do. In other words, there exists “irreducible uncertainty” (Nootenboom, 2002). An illustration: if you would know that it is in the other party's own interest to repay your 'trust', your act is merely calculated cooperation rather than trust. Knowing that repayment is in the other's self-interest reduces the uncertainty characterizing trust. Recognize that trust in this definition needs the repayment of trust not only to be uncertain, but also needs it to be an act that deviates from selfish maximization of monetary payoff. For an act to truly be of a trusting nature then, it requires the actor to regard its riskiness and nevertheless take the leap of faith in the other party's willingness to share the benefit of the exchange. In the trust game, the Investor faces a chance of a gain (receiving more than his endowment) and a chance of a loss (losing his endowment) if he decides to transfer the money. It is clear that it is mutually beneficial to engage in a relationship, but there is no explicit rule that obliges either one party to share the money. Therefore, this game is suitable to measure trust.
Some models (see Watson (1999; 2002) and Sobel (1985)) explain cooperation in trust settings as the equilibrium behavior of rational self-regarding actors, because sustained exchange of trust and trustworthiness is beneficial to each individual actor as well. However, it is again questionable whether this can be called trust and trustworthiness if it is motivated by self-interest. Such a theory of conditional cooperation (“I trust you as long as you are trustworthy” and vice versa) cannot explain trust in one-shot interactions when there is no 'sustained exchange'. The Trustee faces a social dilemma: s/he can either be fully self-regarding and display opportunistic behavior, or s/he can display trustworthiness and share the proceeds of the investment with the Investor. Note again that the Trustee has no monetary incentive to be trustworthy. This is a classical example of moral hazard, because there is no formal enforcement mechanism. I will call a Trustee trustworthy if his/her action realizes a gain for the Investor, and untrustworthy if it realizes a loss. In case the Trustee pays back exactly the amount invested by the Investor, I call this mildly trustworthy since it does not make him/her worse off then s/he was.
Numerous studies employing the trust game typically observe the following pattern (see again Camerer, 2003 for an extensive review). The large majority of investors
send a positive amount of money, on average half of their endowment. Trustees display a widely dispersed pattern, with substantial numbers of people acting both purely selfish and trustworthy. On average they send back the original investment, which means that trust does not pay in most cases. Recall that the rational and selfish choice is to send no money at all in both roles. That theory can clearly not explain the general findings. Why do people trust and repay trust? Engaging in a trust relationship is mutually beneficial. Under the right circumstances, actors with a tendency to be trusting and trustworthy also in one-shot interactions have an evolutionary advantage (see for a recent argument Manapata et al., 2013). Even when the circumstances in the laboratory are different, participants bring such a mechanism with them. The evolutionary explanation is strengthened by the finding that there exists a biological basis for trust. Kosfeld et al. (2005) found that Investors who inhaled oxytocin, a hormone that is implicated in social bonding, transfered larger proportions of their endowment. Trustee behavior was unaffected by oxytocin. Trustworthiness in trust games is often explained in terms of altruism, social norm adherence, or reciprocity. Direct reciprocity seems to explain only a small part, since the amount Trustees send back is only slightly bigger than the amounts they transfer as Dictators (Cox, 2009).
The availability of punishment enhances both trust and trustworthiness significantly (at least when the punishment is severe enough), and this option is commonly used (see e.g. Andreoni et al. 2003; Rigdon, 2009). Where does this demand for punishment come from? In repeated interactions, punishment can be used for strategic reasons. If the punishment is severe enough, a displayed willingness to punish can convince another party to cooperate with you in the future. Punishment in a one-shot interaction has no such function. It is a wealth-destroying act that serves the punisher no benefit other than perhaps the gratification of releasing its anger, and still it is frequently observed.
There is evidence that punishing is indeed used for such reasons. Based on self-reports, unfair decisions are punished often because they lead to an angry reaction (Camerer, 2003). Anger is found to be the predominant driver of punishment in social dilemma games, with the intensity of anger related to fairness perceptions (Hopfensitz and Reuben, 2009). In support of this view, responders in an ultimatum game rejected significantly less when they had another way to display their emotions by sending a free message along with their decision (Xiao and Houser, 2005). These messages were indeed used to express negative emotions following an unfair offer. Further evidence stems from Sanfey et al. (2003) who looked at brain activity during the decision to reject or accept an offer in again the
ultimatum game. Of particular interest here is the pattern found in the anterior insula, a brain region known to be involved in processing negative emotions in particular. Anterior insula activation was responsive to the degree of unfairness, with more unfair offers leading to more activation. Moreover, activity was predictive for rejection rates: subjects with higher anterior insula activation to unfair offers rejected those offers more often. This suggests that an unfair outcome triggers an emotional reaction, and that this reaction indeed influences the decision to punish. Quervain et al. (2004) found that people derive satisfaction from punishing others who violated their trust. Investors punishing untrustworthy Trustees showed elevated activation in the dorsal striatum, a brain area implicated in the processing of rewards. Also, Investors with higher activations were willing to pay more for punishment.
The literature discussed in the previous paragraph seems to agree that feelings of anger and satisfaction are important determinants of the decision to punish. How does the timing of the punishment decision come into play? From a fully rational point of view, postponement as such (without new information) should not influence a decision. However, it has been shown that timing of a decision does make a difference. For example, the willingness to invest in risky gambles increases when the decision is made for multiple periods in the future compared to each period in turn (Gneezy and Potter, 2003). Evidence from several social dilemma games show contributions are higher when people decide simultaneously rather than sequentially even without information on the prior moves (e.g. Rapoport, 1997; Abele and Ehrhart, 2005).
The hot-cold state paradigm relates timing to emotional state. The influence of emotional state is formally included in dual-system theories of decision making that distinguish between two types of cognitive processes (Epstein, 1994; Stanovich and West, 2000; Kahneman, 2011). The first type of reasoning is fast, fairly effortless, and automatic - coined “System 1” processes, while the second type is slow, intentional, and requires conscious attention - coined “System 2” processes. System 1 is primarily driven by intuition and is prone to heuristics and emotions, where as System 2 is driven by rules and logic and moderates deliberative reasoning. Timing can influence which system has the upper hand. Imposing high time pressure, and thereby suppressing System 2, resulted in participants being more likely to reject offers in an ultimatum game (Cappelletti et al., 2008). Delaying a decision is thought to mitigate the influence of System 1 and allow more time for System 2 to kick in, thus leading to more rational decisions (e.g. Adler, Rosen, and Silverstein, 1998).
The distinction between hot and cold states is generally created by means of one of two methods. Some studies define as a hot state the situation in which the the decision maker moves after observing the actual action by the first mover. Those studies apply the strategy method to observe cold state behavior, i.e. asking decision makers to state their action in response to hypothetical first mover's actions. Other studies aim to induce a cold state by imposing a time lag between the observation of the first mover's action and the final response to that action. Without a time lag decision makers are said to be still in a hot state. Results on the effect of 'temperature' (hot versus cold) on decision making are mixed, both between and within studies using the methods just explained. The effect seems to depend on other experimental parameters as well.
Employing the strategy method, Charness and Brandts (1998) studied two sequential games that can measure the extent to which players give a positive (cooperative) or negative (selfish) response to a positive or negative action. They found no systematic difference in behavior between states. In a study with comparable games and also applying the strategy method, Brosig and Weimann (2003) do observe more punishment (negative responses) with decision makers in a hot state, but only when the cost of punishment was low enough. In my opinion the strategy method might fail to create a distinction between a hot and a cold state. Participants using the strategy method have to decide immediately after realizing a scenario. If they consider the hypothetical cases by vividly imagining themselves in such a position, they could still be in a hot state. The time lag method better reflects the notion that emotions arise quickly and then wear off over time. Still, the results here are mixed as well.
Oechssler et al. (2008) found that the rejection rate of unequal splits in an ultimatum game dropped by 25 percent when responders were given the unanticipated opportunity to revise their choice 24 hours later. This effect was present only when the stakes were high-prize lottery tickets, and not when they were small amounts of cash.1 In a similar vein, Grimm and Mengel (2011) observed
rejection rates going down by almost half with a cooling-off period of ten minutes. This suggests a cooling-off period induces more rational behavior, at least when the incentive is high enough.
On the contrary, Bosman et al. (2001) do not observe a cooling-off effect. Here, half of the responders in an ultimatum game was asked to make their decision one hour after reviewing the proposer´s offer. Neither behavior nor self-reported emotions upon deciding differed from those who responded directly. This suggests that the impact of emotions on decision making can show up again by the
same magnitude when one has actually has to make a decision. In line with this finding, a ten-minute cooling-off period did not make players respond closer to the profit-maximizing choice in a Stackelberg competition game (Cardella and Chi, 2012). The latter pair of authors suggest themselves that the intended cooling-off period may have acted as a heating-up period instead, by fostering the emotional response rather than damming it. Neo et al. (2013) do find an effect of imposing time delay on responder behavior in ultimatum games, but they do not observe such an effect in trust games.
Which mechanism is precisely at work during an intended cooling-off period is not well articulated in most studies. The general idea rests on the so-called “primacy of affect”; the notion that people can identify an affective reaction more rapidly then they can produce a deliberate response (Zajonc, 1984). Inserting a cooling-off period is one of the most commonly applied methods to reduce anger, in research and in practice (Adler et al., 1998). Taking a break allows deliberation to catch up with affect. This will not work if the time is spent reliving the event that triggered the emotion (Goleman, 1995). VanOyen Witvliet et al. (2001) examined the emotional and physiological effects of rehearsing hurtful memories and harboring grudges (i.e. unforgiving thoughts) compared to taking perspective and granting forgiveness. They found that unforgiving thoughts induced more aversive emotions, as revealed by significantly higher corrugate (brow) electromyography (EMG), skin conductance, heart rate, and blood pressure. This affirms that active remembrance has an emotional impact.
For a cooling-off period to have the intended effect, actors need to devote their attention to other matters, to reason away the urge of an outburst, or to focus on calming their nervous system by recalling pleasant experiences (Doner, 1996). One explanation for the mixed results on the effect of a time lag described above (cooling off versus heating up) could then be that some studies did not succeed in diverting people's attention away from the Trustee's behavior. A second explanation could be that the impact of emotions on decision making is more robust than their fast and flaring nature suggests. This is consistent with “the law of conservation of emotional momentum of emotions” by Frijda (1988) which states that “emotional events retain their power to elicit emotions emotions indefinitely, unless counteracted by repetitive exposures that permit extinction or habituation to the extent that these are possible.” This means that emotional impact cannot be
mitigated by time as such, but that re-evaluation of the eliciting event can.2
The literature reviewed here shows that the hot-cold effect is not well understood
yet. It is not clear in which circumstances a cooling-off period is effective. Several questions ask for investigation. How long a time lag is sufficient? Is the activity during the break of importance? Does it matter whether the decision is delayed or that the decision can be revised? Is certain type of decision making (for example competition, bargaining) perhaps less susceptible to a cooling-off effect? The current research wants to contribute to this quest, by focusing on punishment and the effect of delaying the decision with a short reasoning task.
3. EXPERIMENTAL DESIGN AND HYPOTHESES
In this section I will describe the experimental design and give a detailed account of the recruiting procedures, the setting, and the communication during the sessions. All original statements, both written and verbal, that were communicated to the participants at any time during the experiment are available to the reader in the Appendices (in Dutch). These include the instructions, answer sheets, and my spoken notifications.
3.1 What is measured
The experiment is designed to measure money transfers by Investors and Trustees (reflecting trust and trustworthiness respectively) and punishment by Investors, as will be explained in more detail in Section 3.2. Aside from those, I incorporated two additional measurements: the Self-Assessment Manikin (SAM) and the Cognitive Reflection Test (CRT).
The SAM is a very quick and easy method to measure an affective response. It asks respondents to indicate their present state on two scales of pictures, representing valence (the degree to which a respondent is happy) and arousal (the degree to which a respondent is excited).3 The SAM method has shown to be as
good or even better than more extended and time-consuming methods (Bradley and Lang, 1994). The brief and straightforward nature of the SAM makes it a very nice technique to use here for two reasons. First, a short question is less likely to influence the subsequent punishment decision than a longer list of questions on emotional state. Second, it allows me to introduce the SAM only to the Investors at the time they are asked to complete it. This way, it does not draw explicit attention of the Trustees to the potential emotional response of the Investors. Of course Trustees can think of the emotional response of their counterpart themselves, but then it happens in a natural manner. Also, by concealing the SAM for the Responders they do not receive unnecessary information during the instructions
3 The original SAM also includes a third scale for 'dominance'. I left this scale out to keep time to complete the SAM as minimal as possible. Dominance is not relevant in the current context.
which could be confusing. Note that this is not deception since I do not hide information about action sets or payoff schemes. The answers to the SAM have no consequences for the outcome of the interaction at all.
The CRT is a three-item test designed to assess a specific cognitive ability: the ability to suppress an intuitive and spontaneous but wrong answer in favor of a reflective and deliberate right answer (Frederick, 2005). Consider for example the following question: “A bat and a ball cost $1.10 in total. The bat costs $1.00 more than the ball. How much does the ball cost?”. The immediate intuitive answer is 10 cents, but is easily found to be wrong if one would take a moment to check that answer.4 The test aims to measure the degree of cognitive reflectiveness, with low
scores pointing at 'cognitive impulsiveness'. This spectrum from reflectiveness to impulsiveness relates to dual-system theory mentioned in Section 2. High (low) CRT performance corresponds to high (low) ability to control the input from System 1. CRT performance is shown to be correlated with several behavioral traits. For example, people with high scores seem to be more patient (i.e. have lower implied discount rates) and more risk neutral (Frederick, 2005). High-scorers are also less susceptible to a number of behavioral biases such as the reflection effect, the conjunction fallacy, conservatism bias, base rate fallacy, the endowment effect, overconfidence, and anchoring (Kahneman and Frederick, 2006; Hoppe and Kusterer, 2010).
Finally, participants are asked to complete a questionnaire regarding some background information on age, gender, education, familiarity with game theory, and fairness opinions. Also, I asked participants to state their considerations for each decision they made. This turned out to be insightful in understanding the observed behavior, as I will describe in Section 6 .
3.2. Parameters of the game and treatment
The experiment features two conditions: the control condition which I call T0 or 'Hot', and the treatment condition which I call T1 or 'Cold'. Each participant plays either T0 or T1, which entails a between-participants design. A within-participants design would open up the possibility for round effects and learning effects that blur the treatment effect. These could be in principle be checked for by applying different orders, but the modest number of participants makes it hard to reliably test for order effects. Which condition was assigned to a session was simply determined by the schedule. The first session was assigned T0 and thereafter I let conditions alternate.
One session consists of two parts. Participants are informed that there will be a part 2, but not what that part constitutes before Part 1 is over. In fact, Part 2 consists of the same game but with a different partner. At the beginning of Part 1 they are explicitly informed that they will be matched randomly with another person in the room, and that they will encounter this person only once through the entire experiment. This information reveals something about Part 2, namely that participants will be matched with someone. I nevertheless share this information to emphasize the interaction in both parts is one-shot. At the start of Part 2, it is again assured everyone will meet someone else than in Part 1.
All participants are assigned a show-up fee of 2 euros. Investors are endowed with an additional 5 euro 'game-play money'.5 The game is explained in steps. First, I
describe the control condition T0:
In Step 1, the Investor has to decide if s/he wants to transfer the game-play endowment to the Trustee or not. If s/he transfers, this amount is tripled by the experimenter. The Trustee then receives 15 euro. If the Investor does not transfer, the game ends. Both players do still participate in Step 4.
In Step 2, the Trustee is informed about the Investor's decision. If the endowment is transferred, the Trustee has to decide how much money s/he wants to transfer back to the Investor. S/He can choose to send back 0, 5, or 7.50 euro. The Investor is directly informed about the Trustee's decision. In Step 3, the Investor can decide to punish the Trustee provided that s/he
transferred in Step 1.6 First, his/her emotional response to the Trustee's
action is measured by means of a SAM. Then the Investor decides to punish or not. If so, the punishment will lower the Trustee's payoff by 3 euro at the cost of 1 euro to the Investor. The Investor can always do this, even if s/he transferred the endowment and got nothing in return. In that case, the show-up fee prevents him/her from negative earnings. This is carefully explained to the participants. Any proceeds from punishment are paid back to the experimenter.
In Step 4, all participants individually answer three quiz questions (more details will follow in section 3.3).
The action space for all three decisions (transfer, back transfer, punishment) is deliberately restricted to specific amounts instead of letting participants freely
5 In retrospect, the chosen endowment setting creates an unbalanced playing field and is thereby possibly a confounding factor. I will come back to this in Section 7.
6 Note that the rare but theoretical case that a the Investor does not transfer the game-play endowment but nevertheless punishes, is hereby prohibited. This protects the Trustee from generating negative earnings.
choose an amount in an interval. I acknowledge it restricts freedom and might nudge people toward a particular decision by making certain allocations focal (especially for Trustees). The advantage of a restricted action space is that it is easier to interpret what a given action could mean, both for the actor's partner and for me as a researcher. The first transfer is set to be binary (transfer the full endowment or nothing) so it is clear that a transfer is a sign of trust, or at least a sign of willingness to engage in a relationship. Any other positive transfer is multi-interpretable: at one hand it is a sign of trust because money is given out of hands, but at the other hand a certain amount is kept from the other player which can be seen as a sign of distrust. Regarding the second transfer, I choose these amounts because they correspond to different motivations for determining the transfer by the Trustee: not reciprocate/maximize own earnings (0), reimburse the investment (5), and share the profit and/or equate the payoffs (7.5).7 The punishment decision
is again binary. This cannot reveal the strength of the desire to punish, but that is not the goal. The main objective of this research is to see whether the timing of the punishment decision influences the propensity to punish and not whether it influences the degree to which people punish.8 Also, with the chosen punishment
structure it is easy to prevent negative earnings without the need to give Investors a much larger endowment then Trustees.
The treatment and control condition differ in one respect: the order of the last two steps. In T0, the punishment decision is carried out directly after Step 2 as described above. In T1, the punishment decision is delayed until after the answering of the three questions, i.e. Step 3 and 4 change places. Hence, the steps in T1 look like this:
In Step 1, the Investor has to decide if s/he wants to transfer the game-play endowment or not. If s/he transfers, the Trustee receives the tripled amount of 15 euro. If the endowment is not transferred, the game ends. Both players do still participate in Step 3.
In Step 2, the Trustee is informed about the Investor's decision and decides how much to transfer back to him/her. The Investor is directly informed about the Trustee's decision.
In Step 3, all participants individually answer three quiz questions.
In Step 4, the Investor completes the SAM and can decide to punish the Trustee.
7 See Section 7 for comments on the possible effects of this action space and its alternatives. 8 This is an interesting follow-up question for future research.
The time devoted to the three individual questions serves in T1 as a cooling-off period. Solving the questions is a short but engaging task, which is meant to pass some time and take especially the Investor's mind of the game for a while. the Trustees also perform this task although they are not the ones who need to cool off. They cannot leave the room since that would reveal information about who played which role. Doing nothing might be boring, and the task will provide them with something to do. In T0, the three questions are merely added for two reasons: 1) to keep the duration of the session and the maximum earning possibilities equal between the control and treatment group, and 2) to have data on CRT performance in the control group too.
3.3. The quiz questions
For the intended cooling off-effect to take place in T1 we need an intermediate task that takes a couple of minutes and that is engaging enough to spend effort on. This quizzes are therefore incentivized. Subject can earn 0.50 euro per correct answer to each quiz question, that is 1.50 euro in total for this task per part. I do not provide feedback on performance.9 Because I let participants play the game twice
(part 1 and 2), I need two different but comparable assignments to serve as the intermediate task. In Part 1, I use the Cognitive Reflection Test (please see above). In Part 2, I collected another set of questions that also appeals to reasoning, and is also famous for pointing at common failures in logical reasoning (adopted from Kahneman, 2011).
3.4. Recruitment and procedures
Participants were recruited through my own social network. I aimed to recruit people that were only indirectly related to me, to prevent potential experimenter effects and a too informal atmosphere as much as possible. Most participants were second-degree connections, i.e. colleges, friends or acquaintances of my friends and family members. As as consequence, participants within (but not across) each session knew each other (which will later show to be potentially problematic, see Section 7). People were invited by e-mail and could sign up through a website I moderated. All sessions were carried out between July 30 and August 13, 2013 and were held on location close to the participants.
Before each session, I requested an empty room with desks and chairs (a large office or a meeting room) to be available for me to create a temporary laboratory. I arranged the seats such that participants could not observe each others instructions and actions. As I prepared the room, participants gathered in
9 Subjects were very eager to hear the correct answers to the questions. This is a clear sign of engagement with the task. With those interested, I discussed the correct answers afterwards.
another room or in the hallway. Then I started the experiment in a formal manner, as is common in research labs. All participants adhered to the instructions very carefully, and maintained silent and concentrated throughout the whole experiment. Upon entering the lab, participants are randomly assigned to a seat by blindly drawing a card with a table number from a stack. For each seat I predetermined the role and matching protocol. Participants are privately informed about their role: half play the part of Investor and half play the part of Trustee.10 All
keep their role throughout the whole experiment. Investors are matched with Trustees into pairs. I use a strangers matching protocol, which means that players are anonymously matched with a different counterpart in each part of the experiment. Each pair therefore has only a singe-shot interaction although the game itself is repeated. Such a matching protocol rules out any strategic motives such as reputation building to play a role.
The experiment was conducted manually; participants indicated their answers and decisions on paper answer sheets. When applicable, participants were privately informed by me about their counterpart's actions. I transferred the information by ticking a box on the answer sheets.11 Trustees received no feedback
about the punishment decision.
The instructions were read out loud in order to create common knowledge and to urge participants to take the time to go through the instructions.10
Participants were invited to ask questions for clarification. Any questions were answered privately.12 In the questionnaire at the end of each session, I explicitly
asked if any part of the instructions or answer sheets were unclear. Nobody indicated so.13
After the instructions, each participant receives answer sheets for the part at hand. A set of answer sheets consists of three to four pages stapled together. The first page displays Step 1 and 2, the second and third page display Step 3 and 4 in the applicable order. In both the control and the treatment version, participants are able to review the history of the game in that part if they want to by flipping back to previous pages. This possibility is indicated on the answer sheet for the punishment decision. The answers of Part 1 are collected before the start of Part 2 so that participants cannot compare both parts on paper, only naturally from memory. The interested reader is again referred to the Appendix to review the
10 To keep the language used in the experiment as neutral as possible, the roles of Investor and Trustee were labeled Player A and Player B respectively.
11 The interested reader is again referred to the appendices for the instructions and sample answer sheets.
12 I received only one question: 'What goal should I pursue? Making the most money, or …'. I answered: 'That is up to you. You may choose a goal yourself.'
13 There were two comments posted here: 1) one participant found it unclear what his objective had to be (this was not the same participant that asked me about this), 2) one participant wondered why I measured 'arousal' in the SAM.
original answer sheets.
At the end of the experiment I randomly selected one pair in one part of the experiment for payment according to their earnings in that part. Normally one prefers to pay all participants but my budget did not allow for that. Theoretically the incentive of acting according to ones true preferences stays in place since the relative outcomes of the game are not altered by the random payment scheme. However, I acknowledge that the earning prospect is heavily diluted by this additional lottery. Payment was executed privately. All participants received an envelope which contained a receipt stating their payment, and (if applicable) their payment in cash. Participants were asked to check their payment and sign the receipts. This concluded the session.
3.5. Hypotheses
Investors who have trusted their endowment have the possibility to punish the Trustee. Even though punishment is costly, Investors may have a desire to punish when they feel Trustees have mistreated them. That is, punishment is assumed to be mainly an emotional reaction to the violation of the Investor's norm or expectation. Delaying the decision is thought to bring the Investor in a cold state and thereby mitigate the influence of emotions. Therefore I expect to observe less punishment when a time lag is imposed on the punishment decision.
Hypothesis I: Punishment occurs less frequently in T1 than in T0.
If punishment is indeed assigned less in T0, the next step is to figure out which mechanism caused this. As explained earlier, the effect of a time lag is assumed to go through emotional state. If this is correct, this pathway can be identified in two ways. First, participants who punish are expected to be more emotional then those who do not punish. This should show up in the SAM-scores regardless of treatment. Second, the participants experiencing the time lag are expected to be in a more neutral emotional state then those who experience no delay. The SAM-scores should reflect this as well regardless of the punishment decision. Since both punishment and emotional response are likely to vary with Trustee behavior, these expectations ought to be tested conditional on the Trustee's transfer.
Hypothesis IIa: SAM-scores of non-punishers are closer to zero than the SAM-scores of punishers, controlling for Trustee behavior.
Hypothesis IIb: SAM-scores in T1 are closer to zero than SAM-scores in T0, controlling for Trustee behavior.
The design allows for one more way to reveal the mechanisms underlying punishment behavior by using the data on CRT performance. Dual-system theory describes the emotional system as an impulsive system. If the decision to punish (importantly, in the absence of strategic motives14) is mainly induced by an intuitive
and emotional System 1-response, participants who are more capable of suppressing impulses are expected to punish less (across treatments).
Hypothesis III: CRT score is negatively correlated with punishment.
I do need to make a precautionary remark concerning the last hypothesis. In the treatment condition T1, the CRT takes place before the punishment decision and hence compromises the purity of the measurement of emotional response. It is possible that the CRT induces emotion in itself and that this subsequently alters punishment behavior. If this is the case, it would show as different correlations of CRT score and punishment between the treatment and the control condition. To fully address this concern, one should separately test whether the CRT on its own induces changes in emotional state or can be regarded as 'neutral' in terms of emotions.
4. THEORETICAL ANALYSIS
Let me present a formal description of the game. Investors and Trustees are called A and B, respectively. Their respective transfers are denoted by X and Y. Punishment is denoted by Z. Qi states the number of quiz questions answered
correctly by player i in Step 3 (in T1) or Step 4 (in T0). Note that a description of one part suffices since each part represents a single-shot interaction and payment is effectuated in only one part. Below I define the set of players N, the set of all possible histories of the game H15, the player functions P that assigns a player to
each subhistory and the payoff functions πi for all N.
Players: N = {A, B}
Histories: H = {(Ø),
(0), (5),
(5,0), (5,5), (5, 7.5),
(5,0,0), (5,0,1), (5,5,0), (5,5,1), (5,7.5,0), (5,7.5,1)}
14 In the presence of strategic motives, for example incentives for reputation building, punishing can also be a sign of advanced forward looking behavior.
Player functions: P(Ø) = P(5,0) = P(5,5) = P(5,7.5) = A P(5) = B
Payoff functions: πA = 2+5–X+Y-Z+0.5QA and πB = 2+3X–Y-3Z+0.5QB
with X = {0,5}, Y = {0,5,7.5}, Z = {0,1}, Q = {0,1,2,3} Since each player answers the quiz questions independently, Q can be viewed as an exogenously determined constant and as having no interaction with X, Y, or Z. However, the possibility that quiz performance somehow influences the behavior in the game or vice versa cannot be excluded. Consider first the possible influence of quiz performance on behavior. It is not clear in which direction such an influence would run. Participants know that the quiz questions are coming, but do not know what kind of questions it will be or how well they performed. On the one hand, good performance can enable punishment. Recall that the show-up fee ensures Proposers have budget for performance at all times. Still, imagine that a participant adopts some kind of threshold strategy (“I want to earn at least C euros”). Quiz performance may determine whether this threshold is met and hence whether this participant has a 'surplus' to spend on punishment. On the other hand, good performance may lessen punishment indirectly. The prospect of earning money in the quiz later may make the threshold easier to reach. Responders may then be inclined to share more money and thereby removing grounds for punishment. Consider second the possible influence of behavior on quiz performance. Participants adopting a threshold strategy that have already met their threshold may be less intrigued by the quiz. This is turn could mitigate the effect of the time lag and hence reduce its effect on the punishment decision. The same type of reasoning goes for participants not having met their threshold by Step 2. Although this reasoning may seem far-fetched, it cannot be ruled out entirely. To keep the analysis simple, I will nonetheless assume that Q is determined outside the game and can be disregarded in setting X, Y, and Z.
Simplified
payoff functions: πA = 2+5–X+Y-Z and πB = 2+3X–Y-3Z
4.1. Four common frameworks
The next step in the analysis requires an understanding of the functional form of a player's utility functions, i.e. of its preferences. Here is where different theoretical frameworks depart from each other. I will discuss four frameworks.
Standard game theory
The traditional game-theoretic perspective (von Neumann and Morgenstern, 1947; Selten, 1965) assumes actors to be rational, selfish, and utility maximizing (with
one's own monetary payoff as the only argument of the utility function). Provided that these preferences are common, a Responder keeps all s/he receives, and anticipating this a Proposer will transfer no positive amount to begin with. Also no Proposer will punish, because it is costly for himself and it serves no purpose later on since players meet each other only once. In the subgame perfect Nash equilibrium (SPNE) then, X=Y=Z=0 for for every player in every part. The corresponding outcome for respectively A and B is (7,2).
Prediction SPNE: X = Y = Z = P = 0
Standard game theory predicts no transfers and no punishment.
The observation that people in fact frequently display non-selfish behavior violates the traditional assumptions. This evidence is too abundant to be explained by trembling hands or insufficiently high stakes (Camerer, 2003). This has led theorists to develop models that account for motivational mechanisms other than pure self-interest, i.e. other-regarding preferences. The way decision makers are assumed to have a regard for others (positively or negatively) differs among approaches. I will briefly describe three frameworks that have shown to be helpful in explaining experimental data (see e.g. Lopez-Perez, 2008; Loewenstein, 2000, for more formal summaries of the modeling).16
Distributional concerns
Models that capture distributional concerns assume that utility is increasing in own income, but is also affected by a player's income compared to others. Utility is often decreasing in the difference between the payoff of oneself and the other. Famous formalizations are the theory of inequality aversion by Fehr and Schmidt (1999) and ERC theory by Bolton and Ockenfels (2000). Those models feature utility functions that balance advantageous and disadvantageous inequalities in income. Inequity is called advantageous when you earn more than the other player, and disadvantageous when you earn less. Usually it is assumed that the weight put on an advantageous difference in payoff is less than the weight put on a disadvantageous difference. Both versions predict that people will sacrifice money to make incomes more equal, at least under certain parametric values. In the present game, ERC theory predicts participants to share the proceeds. The prediction of inequity aversion depends on the exact values of the parameters (i.e. tbe weights in the utility functions mentioned above). It is certainly consistent with trust and trustworthiness given that the parameter values are high enough. With
16 I exclude here a (plain) care for efficiency, altruism, and warm glow since these notions cannot account for the richness of the experimental evidence.
very low parameters (i.e. almost no care for inequality), no trust can also be an equilibrium. Both theories too can explain punishment in the case of an unequal split, because punishment reduces the difference between payoffs.17 Under the
assumed preferences, this increases utility. Prediction Inequity Aversion/ERC: i) X=5, Y=7.5, P=1 iff Y=0, 5
provided that weights on advantageous (for decision on X and Y) and disadvantageous (for decision on P) income inequalities are sufficiently high
ii) X=Y=Z=0
in case of sufficiently low weights in both income inequalities
Fairness and reciprocity
Models that capture reciprocity let the regard for the other depend on the other's behavior. Utility functions include terms measuring kindness and fairness of an offer such that people prefer being nice to those that have been nice to them, and act unkindly towards those who acted unkindly before. Rabin (1993) showed that a mutual cooperative 'fairness equilibrium' exists when the care for fairness is sufficiently high. Again, for very low parameter values no exchange is also an equilibrium. This approach has been extended by several authors, for instance to sequential games (Dufwenberg and Kirchsteiger, 2004), by including intentions (Falk and Fischbacher, 1998), and social welfare considerations (Charness and Rabin, 2002). These models explain trust, trustworthiness, and punishment in terms of positive and negative reciprocity.
Prediction theories of reciprocity: i) X=5, Y=7.5, P=1 iff Y=0,5
provided that the care for fairness and the degree of intentionality is high enough
ii) X=Y=Z=0
in case of sufficiently low parameter values
Role of affect
While both models of inequity aversion and reciprocity can explain punishment, both cannot explain a hot-cold effect. Although some authors acknowledge that emotional satisfaction may be the proxy mechanism that guides people to behave reciprocal, all three frameworks do not differentiate between decision-making in
17 A numerical example: the Investor sends the endowment and the Trustee sends back 5 euro. The intermediate outcome is 5,10. When punishment is applied, the final outcome is mere equal namely 4,7. This is explained by sufficient weight put on disadvantageous inequity (under inequity aversion) or having earnings closer to the average (under ERC). The same reasoning applies when the Trustee sends back 0 euro.
various emotional states. That is, all three frameworks discussed above predict no effect of temperature. With the rising interest of economists in the role of affect, several models have been developed that incorporate dual-system theories of behavior. Most of these attempts are concerned with intertemporal choice and/or self control (Shefrin and Thaler, 1988; Metcalfe and Mischel, 1999; Caplin and Leahy, 2001; Bernheim and Rangel, 2003). Loewenstein and O'Donoghue (2007) provide a framework that models the hot-cold concept more generally. Behavior depends on the interplay of a cool deliberative system and a hot affective system. In the domain of social interactions, the regard for the other person is determined by the weighted average of two motivational functions. The deliberative system puts a stable weight on the other person's payoff (on some moral grounds), and the affective system puts a variable weight on the other person's payoff that depends on how sympathetic one feels toward the other. The main implication is that ultimately the affective system is decisive. Little sympathy in the affective system moves behavior closer to self-interest, and high sympathy pushes behavior towards altruism.
Although Loewenstein and O'Donoghue do not discuss the case of negative sympathy (antipathy), I do recognize the potential to extend their framework to antipathy as well. Negative affect would push behavior also away from self-interest, but in the direction of spiteful behavior. According to this reasoning, costly punishment in one-shot interactions can be explained by high intensity of negative affect. Lowering the affective intensity – for instance by a cooling-off period – should then move behavior back towards self-interest (no punishment).
5. EMPIRICAL ANALYSIS: A SETUP
Although I do not estimate an empirical model because of insufficient data, I still like to present my setup for the empirical analysis. As a first concern, I have to account for the fact that each participant plays the game twice. This means that not all observations are independent and that decisions within any one player can be correlated. This causes problems in the estimation of the standard errors and hence in the estimation of the test statistics. To control for correlated errors I form twenty clusters; one for each player. These clusters allow for correlated standard errors within each participant, but not between them.
Second, I have to choose a specification. Recall that the main research question asks whether punishment occurs less frequently when a time lag is imposed on the punishment decision. The assumed underlying force of the treatment effect is emotional state. In other words, the treatment effect is assumed
to be mediated by emotional state. Therefore I would perform mediation analysis (Baron and Kenny, 1986) to investigate whether variation in punishment can be explained by imposing a time lag and if so, whether this effect can be (partially) explained by variation in emotional state. Below I prepare the steps of the proposed analysis.
The first step is to test whether there is an effect at all that may be mediated, i.e. whether there is a relationship between punishment and time lag. The outcome variable is the dummy variable 'Z', which equals 1 if punishment occurred and 0 otherwise. The variable of interest or predictor is the treatment variable 'D' (for delayed), which equals 1 if a time lag was imposed and 0 otherwise. Emotional state is characterized by valence and arousal, which are each measured on a separate nine-point scale. The scales run from respectively very happy or excited (+4) to very unhappy or calm (-4). Let SAM-A denote the score on arousal and SAM-V the score on valence. Notice that a score of zero corresponds to a neutral emotional state in that dimension. I combine the two SAM scores into one measure called 'A' for Affect. I do this by adding SAM-A and SAM-V in absolute value. A high A then indicates a non-neutral emotional state (in one or both dimension(s)), while a low A indicates a state closer to neutrality. The variable A is assumed to be the mediator variable. The transfer by the Trustee, Y, is included as an explanatory variable since punishment is foremost a response to Y. A possible treatment effect is assumed to run through Y, since the time lag is thought to diminish the initial emotional response to the Trustee's choice. To capture this, I include an interaction term of D and Y18:
Specification I Z = β0 + β1 D + β2 Y + β3 D*Y + ε
Since the dependent variable is binary, a probit estimation is appropriate. To test whether there is a treatment effect, I would then test the following hypothesis:
H0: β1 = β3 = 0
H1: β1 v β3 ≠ 0
If the null hypothesis can be rejected and the sum of the coefficients β1 and β3 is
negative, then I would have found statistical support for the notion that imposing a cooling-off period reduces the propensity to punish. If the null hypothesis is rejected but the sum of the coefficients is positive, the time lag had the opposite effect and stirred up the urge to punishment. If the coefficients are not significantly different from zero, I would have found no evidence for a treatment effect and the analysis would stop here.
18 Note that X, the initial transfer by the Investor, need not to be included since punishment is possible only when the endowment is trusted.
If I do find a treatment effect, the next question is why a time lag influences punishment behavior and in particular whether emotional state of the punisher can explain the effect. Therefore, the second step is to see whether there exists a relationship between the assumed mediator A and the predictor D (while again controlling for Y):
Specification II A =
γ
0 +γ
1 D+γ
2 Y +γ
3 D*Y + εThis specification can be estimated by ordinary least squares-regression. A minimum requirement for a mediation effect of emotional state is a significant relationship between A and D. To verify this, I would test the following hypothesis:
H0:
γ
1 =γ
3 = 0H1:
γ
1 vγ
3 ≠ 0The requirement is met if the null hypothesis can be rejected and the sum of the coefficients
γ
1 andγ
3 is non-zero. I would then proceed to the third step. If thisrequirement is not met, I would have found no evidence for a (non-trivial) mediation effect of emotional state. The hypothesis that a time lag serves as a cooling-off period predicts that the sum of
γ
1 andγ
3 is negative, i.e. that a time lag'neutralizes' emotion.
In the third step I test a full model. For a mediation effect to exist, the mediator not only needs to be related to the predictor but also to the outcome variable. Furthermore, the effect of A on Z could depend on Y. This relation could have shown up already as a jointly significant non-zero sum of
γ
2 andγ
3 inSpecification II. To capture this possibility, I add an interaction term of A and Y. The full model then looks as follows19:
Specification III Z = δ0 + δ1 D+ δ2 A + δ3 Y + δ4 D*Y + δ5 A*Y+ ε
Now the final test on the mediation effect can be performed by testing the following hypothesis:
H0: δ2 = δ5 = 0
H1: δ2 v δ5 ≠ 0
Rejection of the null hypothesis would confirm that emotional state mediates the treatment effect. Emotional state completely mediates the effect of the time lag if the coefficients δ1 and δ4 are jointly insignificant and/or their sum is zero. In that
case, the terms involving D can be dropped. Evidence of this mediation effect would support the idea that emotional state of the decision maker is a determinant of punishment. The economic importance of this determinacy can be judged by the
19 Note that an interaction of A and D is not included since their interdependence is investigated in earlier steps of the mediation analysis.
size of the sum of δ2 and δ5. The hypothesis that decision makers in a more neutral
emotional state punish less predicts that the sum of δ2 and δ5 is negative.
The specifications above could be further refined by controlling for additional measures. Differences in punishment rates or emotional state can also stem from differences in personal characteristics, which are now captured by the error term. Treatment is assigned (quasi-)randomly and is therefore expected not to correlate with any personal characteristics. For those characteristics that are measured, auxiliary regressions can be run to check whether those are correlated with the explanatory variables. Those that do correlate should be included in the regression to prevent omitted variable bias.
If the full model tests significant, it would be interesting to look whether arousal, valence, or both are affected by the delay. For this purpose, I would re-run Specification II but now with SAM-A and SAM-V each in turn as the independent variable. This could indicate in which dimension(s) emotional state is altered by a time lag.
6. RESULTS
As explained in Section 1, the number of sessions was less than originally planned in the sight of insufficient variation in behavior. I performed three sessions of each six (2 times) or eight (1 time) participants, twenty people in total. Since each participant plays the game twice (part 1 and 2), this gives twenty observations (though not all unique, see Section 5). A session took approximately forty minutes, with ten minutes spent on reading the instructions. All participants were native Dutch and enjoyed higher education (higher professional education (in Dutch 'HBO') or university). The average age was 34.7 years (sd 12.187)20. Sixty percent was
male. Thirty percent was familiar with game theory. Six out of twenty participants were selected for payment, earning on average 10.42 euros.
Although I cannot answer my main research question, there are some other insights that can be derived. In informal talks with participants after a session was done, I learned that many of them immediately recognized the tension between greed and fairness. To me this is a sign that these kinds of simple laboratory games do succeed in mimicking 'real' situations, also with people who are completely unfamiliar with experimental economics. I also learned that questionnaires in which participants were asked to state their considerations during decision making were very helpful. Participants answered these questions seriously, only one person did
not write something down here.
There is little evidence of subgame-perfect gameplay, as has been replicated in many studies before. Investors transferred their endowment 80 percent of the time (only two unique21 Investors did not). The average transfer by Trustees was 6.56
euros (sd 1.983), meaning that trust was generally repaid. Out of the sixteen occasions that Investors trusted, 75 percent of Trustee decisions showed trustworthiness (nine unique), and 18.75 percent mild trustworthiness (three unique). Untrustworthiness was enacted only once. Punishment is not observed at all. In this data set, this can be explained by the high degree of trustworthiness. When a Trustee transfers back the maximum s/he can, there is no rationale to punish him/her.22 The question then is how the high degrees of trust and
trustworthiness can be explained.
The high return rate of the Trustees can have several explanations. Trustees might anticipate Investors to punish them in case of low transfers, and send a high amount to prevent that. Alternatively they might have some form of other-regarding preferences, which can be a desire to equalize payoffs, to adhere to an implicit norm (social or private), or to reciprocate the Investor's trust. Lastly, the choice set employed might have nudged Trustees toward high trustworthiness since transferring 7.50 euros was the only way to make the exchange pay for both players23.
To understand the high degree of trust and trustworthiness, I looked at the self-stated considerations in the questionnaires. I asked for these in a neutral and open manner, without hinting on any particular motives. The statements align remarkably well with the common frameworks of social games. I categorized the answers as follows. Responses that read “try to earn as much as possible myself” and “maximize profits” were coded as 'maximize own earnings'. Motivations that included “share equally” and “increase joint earnings” were coded as 'share earnings'. Lastly, phrases like “I trust that the other will pay me back” and “return the gesture” were coded as 'reciprocate'. In case participants gave account of different considerations in Part 2 then Part 1 (when they took different actions in both parts), these were recorded separately. If they were the same (this could be indicated by the participants in the questionnaire), I counted that same explanation twice. All but six decisions (three unique) could be coded in this manner. Three people could not be linked to one of the motives coded above. One participant said
21 By 'unique' I mean that one individual participant made such a decision. For example: I observed four times that the initial endowment was kept. This was done by two Investors who kept it in both parts, hence two unique decision makers. I report the number of unique decision makers to put the frequency which a particular choice or answer occurred in the right perspective.
22 Note that punishment after receiving 7.50 euros is allowed. In case of extremely competitive preferences (wanting to be ahead more then wanting to earn the most), people may want to do so. 23 I will come back to this notion in Section 7.
