Truth Telling as Justification for Selfish Action

Transcription

1 Truth Telling as Justification for Selfish Action January 2013 Abstract In this paper, we examine how interrogators can get potential sources to provide information which entails defecting from their group. In our experiment, subjects are faced with an interrogator either using coercive techniques or offering rewards. We argue that coercion and reward affect individuals who are conditional defectors differently. These individuals will defect only when they can justify that selfish action as either fair or truth telling. For subjects who possess the information the interrogator desires, these conditional defectors will provide that information in both treatments because they are simply telling the truth. For ignorant subjects, conditional defectors provide bad information under coercion because honestly stating ignorance leads to unequal outcomes. In the reward treatment, truthfully saying I don t know leads to a more equal outcome. This means that interrogators receive more information under coercion, but that information is of lower quality. 1 Introduction Seminal work on group behavior asks this question: how can we encourage individuals to cooperate when it is in their self-interest to defect from the group? For example, one could model sharing of common property as a Prisoner s Dilemma individuals have incentives to take all of the public good they can which hurts all members of the group including themselves if everybody acts selfishly (Ostrom, 1990). Hence, the key research questions centers on the types of institutions that encourage cooperation and allow the group to enjoy larger benefits. If we consider the Prisoner s Dilemma framed around actual prisoners, the normative consequences are murky. The prisoners are unable to cooperate and as a result are punished because they defect. This is bad for the prisoners, but good for the society as a whole as it keeps prisoners off of the street. The concentration on how to encourage cooperation has ignored the fact that sometimes people or society in general would prefer to encourage defection. This is especially true in the case where people want individuals to provide information on members of their group. Examples of this range from a teacher who wants to know what student threw a paper airplane to a police officer attempting to locate where a gang stashes drugs. Interrogators can incentivize group members to provide information through two primary means (1) coercion and (2) rewards 1. While the use of coercive techniques in interrogation is controversial topic, there is little empirical research on its effectiveness (Conrad 1 It is also obviously possible for the interrogator to use some mixture of the two. 1

2 and Moore, 2010). The difficulties with studying this issue while upholding standards of ethical research, has led to a reliance on anecdotal and often contradictory evidence (Armshaw, 2011). Further, even when scientists involve themselves in the issue, they often abandon rigor and nuance in the discussion in favor of moral absolutes (Suedfeld, 2007). This paper serves as an attempt to determine experimentally (1) how individuals behave under rewards and punishment and (2) the effectiveness of the two techniques. Ethical concerns obviously prevent research on the harshest methods, but even evaluating the basic hypothesis is difficult if one observes actual interrogation even in a setting as inconsequential as a high school principal s office. The researcher must know a priori the information the source possesses to evaluate the truthfulness of what the source says which renders the necessity of the interrogation pointless. Therefore, we examine interrogation with an incentivized, group-based experiment. This particular research method is best served at answering a very limited set of questions relating to the effects of the type of interrogator (one that is rewarding or coercive), the potential source s knowledge state, the source s desire for a fair outcome and his or her desire to tell the truth. Our research is not designed to examine how affect toward the interrogator can hinder (or assist) the interrogator. As such, we use a computer interrogator to avoid any potential interactions between treatments and complicated emotional ties toward the interrogator. Although clearly interesting such inquiries lie outside the scope of the paper. We test the roles of lie aversion and preferences of fairness using two economic experiments. In the first experiment, subjects in groups are assigned a playing card, but only two of the subjects know what that card is. Subjects participate in these groups for multiple periods. Each period, the computer asks subjects to reveal the card. If a subject reveals the correct card (regardless of their information), the other group members receive no payment. In the coercion treatment, subjects receive less money than their other group members if they fail to reveal the correct card. In the reward treatment, the computer provides an additional payoff in exchange for correct information. In the second experiment, we remove the framing and have subjects pick their favored income distribution. The possible income distributions are identical to those in the first experiment and are probabilistically assigned in manner that is also identical to the first experiment. The only element that is absent in the second experiment is the possibility of telling a lie. In all treatments, the individual s dominant strategy is to defect from the group i.e., to provide information even if the information is potentially inaccurate or in other words to select the most monetarily selfish outcome. In experiment 1, we argue subjects will defect when they can justify an action as being something other than selfish. This motivation combined with the treatment results in specific predictions. Ignorant subjects defect more frequently in coercion treatment because it is both the selfish action and results in a fairer outcome. Knowledgeable subjects reveal the card in coercion and reward because it is the selfish thing to do, but truth telling is also the honest thing to do. Our hypotheses are supported in the second experiment. We find subjects with income distribution choices equivalent to knowledgeable subjects in the first experiment are significantly more likely to select the fair outcome. Specifically, the ability to tell the truth (even to a computer) acts as a means to reduce self-serving cognitive dissonance. 2

3 2 The Truth and What Is Fair Before we begin in great detail, we attempt to remove the contentious and emotional aspects relating to the sensitive nature of topic. Admittedly, this a tall task. First, in the interest of helping the reader understand the argument, we use contextualized language that is not at any point used in the experiment. Second, we define an interrogator as any person or group of people who wishes to acquire information from members of a group. A detainee is the target of the interrogator. Although this language sounds harsh, it is general enough that it can be applicable to multiple environments, ranging from a mother trying to find out which of her beloved children broke a dish to a police officer questioning a gang member about the location of a stash of drugs or weapons. Of course, in each of these cases, there are varying levels of affect between the interrogator and the detainee. One would hope the mother loves her children while in the other example we would expect a degree of mutual animosity between the criminal and the police officer. Clearly, there are potential interaction effects between the type of interrogator (we will discuss later) and their relationship with the detainee. Such interactions are outside the scope of this paper. 2.1 Hypotheses In determining how individuals will behave under different forms of interrogation, there are at least three individual level characteristics that must be considered: (1) the individual s knowledge state 2 i.e., do they know the information the interrogator desires or not, (2) the individual s regard for a fair outcome relative to his or her group, and (3) the individual s aversion to lying. Outside of the lab, an interrogator does not perfectly know the knowledge state of the detainee, and, under normal circumstances, only has a rough idea of the detainee s knowledge based upon priors that may be improved after interactions with the detainee and other group members (who may or may not be detained), and often after the desired information has been recovered. Testing the effectiveness of various tactics with perfect information outside the lab is a thorny issue as it would require all parties to go through needless and costly interrogations. The other two factors are underlying personality characteristics and potentially context dependent. For example, when a teacher asks a student who pulled a prank the student may have no trouble lying to the teacher, but may reveal the truth when his or her mother serves as the interrogator or vice versa depending on the student s relationship with the two interrogators. We can control for, to some extent, these individual characteristics. However, cases involving affect toward the interrogator are outside the scope of this paper but are of great interest for future work. The nature of the two forms of interrogation we investigate (the within session treatments COERCION or REWARD) means that we are manipulating the fairness of the outcomes between the detainee and the other group members. Clearly, individuals may experience disutility when they receive a smaller payout than other members of their group, 2 I.e., do they know the information the interrogator desires or not. 3

4 but, at the same time, their utility may decrease when they receive a greater payout than other members of the group (Fehr and Schmidt, 1999) 3. Detainees generally fall into two categories (knowledgeable and ignorant). Obviously these types are going to have different decision spaces and therefore it is prudent to consider their each type s incentives separately. For individuals who know the information, telling the truth increases the inequality of payoffs. This occurs regardless of the interrogation treatment and to the benefit of sources providing the information they either avoid the punishment themselves or they receive a reward. Telling a lie reduces inequality (in comparison to telling the truth) but does so at a cost to the knowledgeable individual. At the same time, a lie told by a knowledgeable group member is unlikely to harm the other group members. Because the absolute value of inequality in payoffs, from the standpoint of the knowledgeable individual is the same across the within session treatments, we would not expect the behavior of knowledgeable individuals to change across treatments. Some subjects who are relatively more inequality adverse will lie in both treatments while subjects who are less inequality adverse will give accurate information in COERCION and REWARD both treatments. For ignorant individuals, telling the truth and revealing their ignorance has different effects on the equality of payoffs. Under COERCION stating ignorance results in punishment for only the individual; the rest of the group will remains unharmed. From the perspective of the ignorant, this is an unfair outcome benefiting the other group members. The ignorant individual should reveal false information in the hopes the interrogator will spare the full punishment. Under REWARD, the ignorant individual could guess at some information in an attempt to receive at least part of the reward. Note that in the case of reward revealing false information does not result in a loss but rather can lead to gains that negatively impacting the earnings of the other group members; in the best case scenario, this ends in an unfair outcome. If the ignorant individual remains silent, he (or she) insures all group members receive the same payoff and does not risk the group members endowments. As such, the fairer outcome is for the individual to truthfully state that they do not know. Hence, if an individual does not possess information, we should expect them to reveal information (i.e., lie) more frequently under coercion than under reward. This should be true regardless of their level of lie aversion. This leads to two seemingly contradictory hypotheses: Hypothesis 1 Individuals will be more likely to reveal information under coercion than reward. Hypothesis 2 The quality of information will be lower under coercion than reward. 3 Bolton and Ockenfels (2000) propose an alternative model to Fehr and Schmidt (1999) in that individuals maximize a motivation function; holding individual earnings fixed, utility is maximized when the average payoff is relatively close to their own. Although these theories generally result in similar predications, certain payoff configurations could lead to differences (Engelmann and Strobel, 2004). For example, in Bolton and Ockenfels (2000) a subject is indifferent between all group members receiving equal payoffs and the subject earning the average amount in which some subjects receive a large payoff and others a small payoff. In Fehr and Schmidt (1999), the subject would prefer that all group members receive the same payoff. Our argument is motivated by Fehr and Schmidt s conception of inequality aversion, but there is no reason to believe that switching to Bolton and Ockenfels model would result in different expectations. 4

5 The seemingly paradoxical nature of these two hypotheses is the result of more ignorant individuals revealing bad information under coercion. They provide information to the interrogator, but often the information is inaccurate (Schiemann, 2012). In some interrogation situations, the source will not want to lie to the interrogator either because of some institution which punishes lying specifically (Lupia and McCubbins, 1998), because of the relationship between the interrogator and the detainee, or because the individual is averse to lying in general or is concerned about the harm the lie will do to the other parties (Gneezy, 2005; Dreber and Johannesson, 2008; Lundquist et al., 2009; Childs, 2012). Hurkens and Kartik (2009) argues that there are two types of individuals: (1) Frist, there are people who will never lie and (2) second there are people who will lie whenever it would be beneficial to do so and regardless of how the effects other group members. Hence, there are some individuals who will not even tell Pareto white lies or a lie that will benefit everyone (Erat and Gneezy, 2012). In the case of knowledgeable subjects under interrogation, telling the truth and revealing information is a selfish action; it increases their payoffs while lowering the payoffs of other group members. This is true under both interrogation techniques. Hence, sources could be of two types similar to the types identified by Hurkens and Kartik (2009): (1) subjects who always tell the truth; (2) subjects who will tell the truth when doing so benefits them 4. These individuals reveal information because truth telling leads to higher payoffs. Hypothesis 3 Knowledgeable subjects will take the higher payoff when they can t say that they are just telling the truth. The corollary of hypothesis 3 is that subjects will be more likely to take the fair outcome in the unframed experiment. For ignorant and knowledgeable individuals, the same logic can explain their actions. There is a set of individuals who are conditional defectors who reveal information depending on the context. What is required to get these individuals to speak is a justification for their actions other than simply wanting the highest possible payoff. For ignorant subjects, this comes under coercive interrogation because seeking the highest payoff is also the fair action but at the same time is dishonest. In reward, on contrary, truthfully stating ignorance is both honest and fair. For knowledgeable subjects, as long as the interrogation is about revealing factual information, the source can take the higher payoff. Our predictions are consistent with cognitive dissonance. Festinger (1962) suggests that when humans experience something psychologically uncomfortable they will either seek information that will reduce the dissonance or avoid information that will increase it. Some individuals who defect from their group and take a selfish payoff would experience dissonance. They are able to reduce the dissonance by justifying their action as the honest thing to do. However, if the frame is absent, dissonance cannot be reduced. Konow (2000) models dictator game behavior in a manner that takes into account cognitive dissonance. This clever model is depicted in the equation below. In it, the dictator 4 One can see this behavior manifest itself in the behavior of whistleblowers or people who write tell-all books. It is possible these people are telling the truth because they believe telling the truth is the right thing to do. It is also possible these people are telling the truth to put themselves in the limelight. In the former, it appears they are doing something altruistic, but, in fact, they are doing so for selfish reasons. 5

6 maximizes her utility as function of her earnings (y) where y [0, ȳ] and ȳ is the size of the endowment, the amount she believes to be fair or how much she thinks she deserves (φ) where φ [0, ȳ], α is a parameter accounting for cognitive dissonance 5 η is their entitlement, β indicates the costliness of self-deception is. Max y,φ u(y, φ, η, α, β) v(y) f (y φ, α) c (φ η, β) s.t.y ȳ, φ ȳ The right hand side functions, from left to right, take into account the dictator s value for their earnings (or v(.)), level of cognitive dissonance (or f(.)), and the cost of self-deception (or c(.)). Following the usual assumptions relating to monetary earnings, v(.) is increasing and concave in y; that, is utility is increasing in earnings but does so at a decreasing rate. The disutility caused by cognitive dissonance f(.) is convex in y φ and α. Cognitive dissonance costs more the greater weight that is placed on it (i.e. larger α) and the difference between the amount the dictator takes for themselves and what they think they deserve. With respect to the arguments within it c(.) has similar assumptions to that of f(.). 6 That is the cost of self-deception is increasing in both the difference between what the dictator thinks they deserve (φ) and what they actually deserve (η). Further the amount that self-deception matters is increasing in β. For knowledgeable subjects, framing the decision to defect as truth telling reduces the cognitive dissonance by acting on α. Without that frame, subjects are less able to justify their actions as something other than selfish. 7 Therefore, the setting of an interrogation can reduce the role of cognitive dissonance and thereby encourage the more selfish outcome. This leads to predictions identical to what would be predicted via standard lie aversion. That is, lies are costly in the non-monetary sense. It follows, if one removes the ability to lie, ceteris paribus, one would expect to observe a movement in the direction in what was formerly dishonest behavior. 3 Experimental Design and Procedures 3.1 Comparison to Previous Experiments We call our experiment The Interrogation Game. In the experiment, subjects are placed in groups and play against an outside party requiring information to take an action. Only some of the group members have the information. To ascertain the piece of information, the outside party (the computer) interrogates one randomly chosen group member, who may or may not know the required information. The outside party takes an action based upon what that source says. The computer may punish sources for not revealing the information or reward them for revealing accurate information. That is, unlike most models of group behavior in which 3rd parties punish or reward to encourage cooperation (Fehr and Fischbacher, 2004), the 3rd party in this experiment encourages defection among 5 One can think of α as the strength of the dissonance reducing mechanism. 6 For a more detailed description of cognitive dissonance see Akerlof and Dickens (1982), Rabin (1994), and Konow (2000). 7 We are not suggesting truth telling is the only way cognitive dissonance may be reduced in dictator-like games. Other examples include earned endowments, group identities, and levels of anonymity. 6

7 the group members 8. If the source provides accurate information, the other group members receive a payoff of zero. We use a computer player for two reasons. First, often in the outside world, the interrogator has priors based upon the detainee s individual history (e.g., the principal has some knowledge of the student s credibility and the police interrogator is aware of the detainee s criminal background). Our computer interrogator acts upon a commonly known rule thus removing a strategic element of the game and allowing us to investigate truth telling without confounding effects from strategic interaction that would be present in the examples above. Second, the use of a computer interrogator has the added benefit in that the lie subjects can tell is not to a human player. In effect, a lie told in our experiment does not have the added disutility (or utility) of hurting the interrogator. Note this places this experiment purposefully between the two extremes of telling a lie to a cop (as a gang member) or telling a lie to a parent (as a loving child). The experiment we present shares elements similar to two other experiments. First, there is the Volunteer s Dilemma (Diekmann, 1985). In the Volunteer s dilemma, N players are grouped together and given the choice to cooperate at a cost of K or defect which is costless. If at least one group member selects to cooperate, the subject(s) who choose to cooperate earn U-K, while the group members who defect earn U. If they all defect, the subjects earn nothing. Hence, the dilemma is whether someone is willing to pay a cost for all members to receive a benefit. Rapoport (1988) Volunteer s Dilemma is framed specifically around a coercive interrogation. Subjects are told to imagine they are in a prison or boarding school and the authorities want to find out who committed some transgression. If no one confesses all subjects are punished severely, but if one subject confesses the other group members would not be punished while the one who confessed would only be mildly punished. Rapoport finds that between 40 and 50 percent of subjects cooperate that is, they are willing to bear the cost on behalf of the group regardless of whether or not the nature of the social dilemma has been explained to the subjects. Our experiment differs in that subjects do not choose whether or not to volunteer. The chosen source has been volunteered and in the coercion treatment must decide whether they are willing to pay a cost to ensure the best payoff for other group members. For this reason, the experiment also bears similarity to the Dictator Game 9 where one player is chosen to divide an endowment between him or herself and another player. Like the Dictator Game, what we call the Interrogation Game is not a game in the strict sense because there is no direct intersection between player behaviors and payoffs. The potential source in the Interrogation Game is similar to the dictator in that he or she sets the payoffs for all players. In many cases, dictators choose to keep the entire endowment, but substantial positive offers are also common There are exceptions. Apesteguia et al. (2007) examine corporate leniency where an outside party encourages defection from a cartel. Our experiment differs in that we introduce players who do not have the information the outside party requires and have no strategic interaction with other players. 9 For a review see List (2007). 10 This depends greatly on the conditions of the experiment. Hoffman et al. (1996) find that as anonymity is reduced, zero offers become less common. However, it is possible that the steps taken to insure anonymity also result in subjects becoming skeptical that their offers will actually reach their partners prompting more selfish behavior (Bolton et al., 1998). 7

8 Unlike a single shot dictator game there is the possibility of reciprocal behavior in that subjects observe that the computer guessed the card correctly or incorrectly. Subjects may be reciprocal in the sense that if they kept their endowment in a previous period they may be more likely to take one for the team and truthfully state the don t know or deliberately lie in a future period. However, subjects are not told the number of their group mates who behaved selfishly. As such, this would be a strong assumption as subjects losing their endowment may be due to a simple case of the computer getting lucky (i.e. the ignorant subject guessing the correct card) rather than systemic selfish behavior in the group Subjects and Groups Subjects in our experiment attend a large public university in the southern United States. For each session, we recruit sixteen student subjects using the ORSEE recruitment system (Greiner, 2004). Upon being randomly seated, subjects are given instructions and anonymously assigned to groups of four 12. Subjects then either participate in a coercion treatment or a reward treatment for ten periods. We will explain the treatments in the following section. Following ten periods, subjects are assigned to new groups and participate in the other treatment for ten periods. After completing the main experimental treatments subjects participate in a modified dictator Game and a Holt and Laury (2002) risk aversion task. These final two experiments are primarily used to increase earnings which is necessary to prevent some subjects from leaving the lab with nothing other than their show-up fee. Each session lasts approximately 60 minutes. Subjects in the first experiment earn $6.79 on average in the main experimental treatments in addition to the $10 show up fee and whatever they earned in the dictator game and Holt Laury risk aversion task. Subjects in the second experiment did a little better, making on average $7.15 in the main treatments. The experiment s breakdown of session can be found in Table 1. In all, there are 160 in experiment 1 sessions. In half of the sessions for experiment 1, subjects are randomly assigned to groups of four. After the first ten periods, they are re-matched so that they do not play with the same subjects as they did in the previous 10 periods. In the other sessions of experiment 1, subjects are assigned to groups using minimal group paradigm (MGP) procedure (Tajfel, 1970). Subjects are grouped based on their preferences for different paintings 13. We originally wanted to test the role group identity played in influencing pro-social behavior but were unable to find a significant results. Chen and Li (2009) find when 11 We test for an effect of previous period s outcome on current period s outcome. We find no such effect. 12 The exact instructions can be found in the Supporting Information. 13 Prior to the first treatment, subjects were divided into groups based on their preferences between paintings by Paul Klee and Wassily Kandinsky. For the second treatment, subjects were divided into groups based on their preferences between paintings by Koko the gorilla and Congo the chimpanzee. Like subjects preferences for artwork by Klee and Kandinsky, we have no reason to believe preferences for either primate is correlated with other subject characteristics. Subjects were not told who the painters were and their groups were named simply Group A, Group B, Group C, or Group D. Subjects were told they were placed into those groups because they liked the same paintings. 8

9 subjects are put into groups based upon their preference for paintings, subjects behave more pro-socially toward those within their own group in comparison to randomly assigned groups. As we find no statistically significant differences between the two types of group assignment, we present only the pooled results here 14. The results broken down by group type is available in the Appendix. In all cases, subjects are accurately told that their group members in one treatment are not be their group members in the other treatment. Table 1: Summary of Treatments Experiment 1 Experiment 2 Group Type MGP RANDOM RANDOM COERCION - REWARD REWARD - COERCION Total Three of the sessions with randomly assigned groups began with subjects participating in the coercion treatment for ten periods and then in the reward treatment for ten periods. In the remaining two random group sessions, subjects participated in REWARD for 10 periods before being re-matched to participate in COERCION. Similarly, in the MGP group sessions, three began with COERCION and two began with REWARD. In experiment 2, all subjects are randomly assigned into groups of 4. In 2 sessions, subjects first participate in COERCION for 10 periods and are then assigned into new groups for the REWARD portion of the session. In the remaining 2 sessions, the order of the within session treatments are flipped. 3.3 The Treatments In experiment 1, at the start of each experimental round, each group is randomly assigned a playing card - one of the four aces 15. Subjects are told this is their group s card and that they are playing against the computer that is trying to guess the card. Half of the group members see the playing card while the other half of the group sees the text unknown. Subjects are randomly assigned to see the card or to see the text unknown in each period. Subjects therefore, could be ignorant in one period and knowledgeable about the card in another. Upon receiving their information state, subjects then are prompted by the computer to reveal their group s card. Subjects have the option of revealing a card, with no requirement of truthfully revealing the card, or to indicate that they do not know the card. 14 Why did the MGP treatment fail to generate differences in behavior? One possibility is that the instructions read to the randomly assigned groups already activate the minimal group identity. Specifically, subjects are told that they are playing as a group against the computer. Hence, even though, the assignment to groups differed in the two treatments, both may result in minimal groups. This does not say that group cohesiveness or identity does not play a role in subject decision but rather that MPG was not sufficient in changing subject behavior. Further, unless there is a reason for interaction effects, any sort of group identity or cohesiveness would be an intercept shift and should not impact marginal effects. 15 Screen shots of the experiment are available in the Appendix. The experiment was programmed using ztree - software for running economic experiments (Fischbacher 2007). 9

10 The computer then randomly selects one of the group members to listen to and payoffs are assigned based upon the randomly selected subject s decision. 16 Subjects observing the card have three possible actions: (1) truthfully reveal the card, (2) reveal an incorrect card, or (3) claim to not know the card. If a subject does not know the card, there are only 2 possible actions: (1) truthfully state that they do not know the card or (2) reveal a card at random. If a subject does not know the card and reveals a random card there is a one in four chance the subject will randomly select the correct card. Table 2 displays the payoffs the subjects receive for each of the different possible decisions in the two within session treatments (COERCION and REWARD). Subjects are paid based on two randomly chosen periods one in each treatment. Subjects do not know their actual earnings until after both treatments are complete. In COERCION, the subject chosen to act as the source earns $5 if they reveal the correct card and the other group members earn nothing. The source receives nothing if he or she says, I don t know and only $2.50 if he or she tells the computer the wrong card. In both cases, the other group members receive $5. In REWARD, the source receives $10 if he or she reveals the correct card and $7.50 if he or she reveals an incorrect card 17. The payoffs for the other group members are the same in both treatments they earn nothing if the correct card is revealed and $5 if it is not. Table 2: Payoffs by treatment for the subject the computer chooses and other group members COERCION REWARD Others in Both Accurate Card Inaccurate Card Don t Know Expected Value of Guessing All subjects are asked to make a decision but the computer only selects to use one subject as its source. The source is chosen at random. The computer then guesses a card based on the information obtained from the chosen subject. The computer reveals whether or not it correctly guessed the card. If the chosen subject claims to not know, the computer tells subjects that it guessed incorrectly. Subjects do not know who was chosen or the decisions of their fellow subjects; they only know the outcome of the period. Subjects participate in each treatment for ten periods. All of the analysis conducted use clustered standard errors to account for the repeated observations on the same subjects. In most respects experiment 2 is identical to experiment 1. The only differences being that subjects select among allocations identical to those in experiment 1 and there is no discussion of a computer player but rather the computer randomly selects the player 16 It is important to note that, since the interrogator is a computer, it is not a peer of the subjects. It is clearly an outside party. 17 Subjects in the reward treatment receive $7.50 even for accurate information to keep the two treatments parallel and because the subjects are behaving as if they are trying to defect from the group. In fact, ignorant subjects who provide information are defecting from the group since they run the risk of sticking group members with the zero payoff in order to gain more money for themselves 10

11 whose decision determines the round s earnings. Because the analysis of experiment 2 is separate from experiment 1, we defer a more detailed discussion of experiment 2 to later in the paper. 4 Results 4.1 Receiving Accurate (and Inaccurate) Information We begin the analysis by testing hypotheses related to the information the interrogator (i.e. the computer) will receive. Recall that H1 states the interrogator will receive more information under coercion and H2 states the information will be of lower quality under coercion. Figure 1 displays the proportion of time subjects reveal the correct card, reveal the incorrect card, and reveal no card. Essentially, this figure tells us how frequently the computer would guess the wrong card or been unable to guess a card. It is important to remember that sometimes when subjects revealed the correct card, they were guessing what the card is because they did not know it. They happened to be correct, but they did not have any more information than the computer.subjects provide the correct card 48.4% 18 of the time in COERCION compared to 44.8%. This difference is significant at the 0.1 level in an F-test of equal proportions with clustered standard errors (F = 3:47 with 150 degrees of freedom; p = 0:06). There is much stronger support for hypotheses H2. Subjects are significantly more likely to reveal an inaccurate card in the COERCION (F = 11:80; p = 0:00). Subjects reveal the wrong card 43.8% of the time in the COERCION, but were 7.1 percentage points less likely to reveal an incorrect card in REWARD. The difference is largely a result of how frequently subjects clicked the I don t know button in REWARD (F = 39:58; p = 0:00). In COERCION, subjects said they would take the zero payoffs for claiming not to know the card only 7.8% of the time. In REWARD, subjects chose to take $5 and said that they did not know 18.6% of the time. As we predict, subjects are less likely to reveal an incorrect card in REWARD, but Figure 1 does not explain why we see this result. Knowledgeable subjects in REWARD can receive an extra $2.50 by lying about what the card is. This would increase the inequality in payments, but would not harm the earnings of other group members. Thus, one might expect knowledgeable subjects to reveal incorrect cards with greater frequency in REWARD. On the other hand, ignorant subjects might not want to accidentally harm their fellow group members in REWARD by guessing the correct card when they are only trying to gain an extra $2.50. To explain the results in Figure 1 while checking for period effects, Figure 2 displays the frequency of decisions by period for knowledgeable and ignorant subjects in each treatment. The figure also divides subjects by their within session treatment order. The decisions for subjects who participated in a treatment (COERCION or REWARD) first are seen in periods 1 through 10, while the decisions for subjects who participated in a treatment second are shown in periods 11 through 20. The first thing to notice is that there does not appear to be any period trends in the data. The main factors influencing behavior are the subject s knowledge and the 18 Remember, all subjects made a decision. Only one subject decision was randomly chosen to set the payoffs. The results, therefore, show the decisions all subjects made and not just the chosen subject. 11

12 Figure 1: Proportion of subject decisions by treatment. Error bars represent 95 confidence intervals treatment. The top graphs display the results for knowledgeable subjects, first in the coercion treatment then in reward treatment. The figure shows knowledgeable subjects do not alter their behavior on the basis of the treatment. In both within session treatments, subjects revealed the correct card about 75% of the time, the wrong card about 22% of the time, and claimed to not know about 3% of the time. It would appear on the basis of these results that subjects took on one of two types when they were knowledgeable. Most maximized their payoffs, but some took the greatest benefit they could without harming the other group members. When subjects are ignorant, however, the treatment had an effect in support of H2. In the COERCION, ignorant subjects truthfully stated that they did not know the card about 12.5% of the time. In REWARD, ignorant subjects honestly state they do not know the card 32.5% of the time. By revealing a card, ignorant subjects are running the risk of harming other group members. In COERCION, however, it is the fair thing to do. If an ignorant subject does reveal a card, they receive nothing in COERCION and the group members receive $5 with certainty. If they do reveal a card, their expected payoff is $3.13 and the expected payoff for the other group members is $3.75. Since the payoffs are more equal if ignorant subjects reveal a card in COERCION, ignorant subjects should reveal a card regardless of their level of inequality aversion. In REWARD, all players will have equal payoffs if the ignorant subject does not reveal a card. Revealing a card both increases inequality and risks the other group members earnings. Thus, in REWARD, ignorant subjects who care about equality should honestly state that they do not know anything. There does appear to be an order effect occurring among the ignorant subjects. Ignorant subjects who participate in the reward treatment first state that they did not know what the card is 40% of the time in that treatment. If the reward treatment is second, then they select don t know only 27% of the time. Table 3 presents a logit model of the decision making of ignorant subjects with fixed effects to account for the repeated 12

13 Figure 2: Subject decisions by period for ignorant and knowledgeable subjects measures on subjects. In the model, the dependent variable is coded one if the subject truthfully indicates they don t know and zero if the subject guessed a card. The table shows that ignorant subjects chose I don t know more frequently in reward condition regardless of which treatment they participated in first, but that the effect was larger if they participated in the reward treatment first. We test for equality of coefficients and find evidence in support of the alternative hypotheses of unequal coefficients 19. These results suggest subjects participating in COERCION first got into the habit of revealing a card potentially lowering the payoffs of their fellow group members. Because of this order effect, we are potentially underestimating the difference in how much inaccurate information the third party receives between the treatments. If we only look at the first treatment, subjects reveal the wrong card 44.1% of the time in COERCION and 35.6% in REWARD which is greater than what is presented in Figure χ 2 =7:52 with 1 degree of freedom and p=

14 Table 3: Did the order of the treatments affect the behavior of ignorant subjects? Coef. Z-Value REWARD First REWARD Second N (Subjects) 1253 (126) AIC Who Tells the Truth? The preferred outcome for an interrogator is for the source to tell the truth. That is, the interrogator wants a source that has the information to reveal it and wants a source who does not know to say, I don t know. The first probit in Table 4 estimates the probability of truth telling as function of the treatment and the subject s knowledge state. The second model includes an interaction between these two variables. We also present a model with four control variables. First, we include a period counter to account for potential period effects. We also include one period lags on the subject s behavior and whether the computer guessed the card correctly. Last, we estimate the marginal effect of the previous period s payoff on the subject s behavior in the current period. These variables control for, respectively, the potential stickiness of subject behavior and for whether subjects change their behavior based on how they believe other group members are behaving 20. Table 4: Predicting subject truth telling Coef. Z-Value Coef. Z-Value Coef. Z-Value COERCION Ignorant COERCION*Ignorant Previous Truth Teller Previous Correct Card Previous Period Pay Period Constant N (Subjects) 3200 (160) 3200 (160) 2880 (160) A.I.C Probit models with standard errors adjusted for clustering on subjects. The dependent variable in the probit models in Table 4 is coded one if the subject tells the truth and zero if the subject lied. Knowledgeable subjects are truthful if they reveal the correct card and ignorant subjects are truthful if they clicked the don t know button. The first model includes only the treatment and knowledge state variables. Both variables are statistically significant and in the expected direction. Subjects tell the truth less frequently in COERCION and ignorant subjects are less likely to reveal their ignorance. 20 The Appendix includes a logit model which replicates the first and second models with fixed effects for subjects. These models more directly control for subjects who tend to exhibit the same behavior. The substantive results are the same as in the probit models in Table 3. 14

15 When we include the interaction effect in the second model, we see that the effect of the treatment on truth telling only holds for ignorant subjects which is consistent to what is seen in Figure 2. This result holds when we include other control variables in the third model 21. The positive, statistically significant coefficient for Previous Truth Teller suggests that subjects behavior is sticky, though, the statistically significant effect of Period shows subjects tell the truth less frequently in later periods. The lack of a statistically significant effect for Previous Correct Card demonstrates that subjects do not react to the outcome. It does not appear that subjects reciprocate or coordinate their behavior; the previous period s earnings have no significant impact on truth telling behavior. The results from the models confirm the problem with COERCION that we observe in the previous analysis. Even if we ignore all ethical concerns with interrogators using coercion, coercion is suboptimal because it increases the risk of receiving false information from sources with no actual information. False statements from ignorant sources are often suspected of making coercion an ineffective mechanism for educing information (Blakeley, 2007). This is clearly the case in our laboratory experiment Self-Serving Lie Aversion We now move to experiment 2, Up to this point, we have shown ignorant subjects are more likely to tell the truth in REWARD possibly because revealing ignorance is unfair to them. We have yet to demonstrate that knowledgeable subjects use the fact that they are telling the truth as a means to justify their selfish behavior. For this reason, we replicate the experiment removing the possibility for subjects to lie by removing the experiment s frame. Specifically, half of the group members had three choices (BLUE, TEAL, and GREEN) and other half had two (PURPLE, and GREEN). Each of these choices has equivalent payoffs to subjects participating in the original framed experiment. Specifically, GREEN s payoff is identical to that of a subject selecting Don t Know ; BLUE is equivalent to a subject truthfully revealing the card; TEAL is equivalent to a subject lying about the card. Last, PURPLE pays BLUE 25 % of the time and TEAL the other 75 % of the time. Figure 3 compares the behavior of subjects in the framed and unframed versions of the experiment. The top row are subjects who can set the payoffs (Knowledgeable subjects in the framed version) while the bottom row are subjects with uncertain payoffs (Ignorant subjects in the framed version). The left side of Figure 3 shows the decisions made by subjects in the framed experiment by treatment and knowledge state. The right side displays the decisions made by subjects in the unframed version of the experiment. As before, subjects who can set payoffs with certainty (i.e. the subjects who knew the card in the framed treatment) do not alter their behavior by the within session treatment. They make the same decisions regardless of whether they are choosing within the coercion payoff scheme or the reward payoff scheme. Subjects who have uncertain payoffs are affected by the treatment and in the same manner as in the framed experiment. It is unfair of them to 21 The N size is smaller in the third model because we drop the first period of each treatment due to the fact that there is no lag for the first period. 22 Of course, we cannot state precisely how the results would change if coercion was the use of force rather than a monetary loss, but there is anecdotal evidence that force often induces false statements from ignorant sources (Armshaw, 2011). 15

16 take the lowest payoff under the coercion scheme. Hence, they only take the lowest payoff 16% of the time. In the reward scheme, taking the lowest payoff keeps the payoffs equal while inequality increases if they attempt to earn more money. Figure 3: Comparing behaviors in the framed and unframed version of the experiment. Error bars represent 95 percent confidence intervals. The frame impacts the decisions of only those people who can set payoffs. In the framed condition, subjects take the highest payoff leaving the other group members with the lowest possible payoff nearly 75% of the time. They can justify this action as not solely selfish, but acting truthfully. That is, responding with a different card or I don t know would be a lie. In the unframed version, this justification is removed. Now, taking the 16

17 largest payoff is only a selfish decision. Accordingly, subjects who can set the payoffs are significantly less likely to take the highest payoff in the unframed version. Instead more subjects take the TEAL option in the unframed version. This is the equivalent of revealing the wrong card in the framed version. The subject takes a lower individual payoff, but the group ends up with the social optimal outcome. In fact, the percentage of subjects taking the lowest payoff doubles in the unframed version. This is very strong evidence that in the unframed version subjects were unwilling to take an action that would clearly be labeled as selfish. 5 Discussion We have argued a combination of concerns for fairness and selfish truth telling explain the pattern of behavior observed by subjects in our experiment. A potential alternative explanation for the decisions would rely on prospect theory (Kahneman and Tversky, 1979) or simply loss aversion for the knowledgeable subjects since there is no uncertainty for those subjects. In this case, the argument would state that the subjects react differently in COERCION because they do not want to take a loss in payoffs. If this is the case, then we would not need to consider subject concerns for fairness of the outcome for themselves or other group members. There are several problems with this explanation. First, it relies on an assumption regarding the reference point with which subjects compare their final payoffs. Subjects may believe that $5 is the reference point because that is the payoff everyone receives if the card is not revealed. In that case, receiving $2.50 or nothing under coercion is a loss. Alternatively, subjects never leave the lab with less money than they entered. Further, the instructions never state that the subjects begin with an endowment of $5. In this case, the subjects would view no money as the reference point. Hence, they would never be on the loss side of the ledger. They would still have to cope with missing out on potential gains. These foregone gains do not seem to have the same psychological burden as actual losses (Kahneman et al., 1986). More importantly, loss aversion alone cannot explain the actions of the knowledgeable subjects or the subjects who can act as a dictator in the unframed experiment. There is no treatment effect for these subjects at all. If they simply wanted to avoid losses, they would take the higher payoffs in the unframed version of the coercion treatment, but in the reward treatment. The results suggest there are two key types of subjects. First, some subset of subjects will take the largest payoff regardless of fairness or truth telling. The other subset will sometimes take the larger payoff, but only when they can find a way to justify it. This typology of subjects informs how one should think about the external validity of the experiment. Much like the conditional cooperation observed in public goods games (Fischbacher et al., 2001), the subjects who are affected by the treatment were conditional defectors. Hence, the proportion of conditional defectors in the population outside the lab will affect the proportions of behaviors observed. If a group is made up almost entirely of the type that will never take the selfish payoff (e.g. a committed terrorist cell) then neither technique would reveal information. Alternatively, if the group is made up of a significant number of selfish types (e.g. an elementary school class) then both techniques 17

18 will reveal information. It remains the case however, that rewarding is the weakly dominant strategy because it does not result in the ignorant selfish type revealing false information while coercion does. Still, we must proceed with caution concerning the external validity of all results. In interrogation settings outside the lab, whether in a principal s office or a police station, the incentives are often much larger. The disincentives to remain quiet when facing a coercive interrogator are much higher than the risk of losing five dollars. However, the rewards that an interrogator can offer are similarly much larger and go beyond simple monetary compensation for information 23. While everything in actual interrogations is scaled up, the comparison between the two types of incentives should be similar. Behavior in Dictator Games and Ultimate Games, for example, is remarkably stable even when the monetary stakes are raised to levels greater than weekly wages (Slonim and Roth, 1998). Further experiments with higher stakes are necessary to see whether the size of the treatment effect is conditional on the size of the stakes. Furthermore, our results may also unreliable in cases where there is strong affect (both positive and negative) toward the interrogator. It is easy to imagine that punishment by a hated interrogator may serve as a mechanism that strengthens group identity and consequently lead to significantly more instances of pro-social behavior. Finally, this study probably cannot shed light on interrogations involving physical coercion. In part, this is because it is unclear how physical trauma would affect the behavior of a target. In the extremes, it may even be the case that victims of especially harsh interrogation tactics may be physically unable to respond to the interrogator s questions. Further, our experiment assumes the state uses coercion only to obtain information. These results do not speak at all to states using torture to stifle opposition or during ethnic strife (Poe and Tate, 1994; Wisnewski, 2010). In those cases, the visible effects of torture are required to demonstrate ruthlessness (Rejali, 2009). 6 Conclusion Obviously, the stakes are higher in the world outside the lab, but the comparison between the treatments should provide some insights. Real world interrogators in many settings offer harsher punishments but they can also offer greater rewards. We would not suggest that the size of the lab treatment effect is the size of the effect in actual interrogations nor can we speak to effectiveness of extreme coercive techniques that many would consider to be torture. We do believe, however, that we demonstrate coercive interrogation does not need to be the inevitable behavior of rational interrogators dealing with rational sources as previous models suggest (Wantchekon and Healy, 1999; Schiemann, 2012). Moreover, this paper argues certain individuals will take a selfish action only when they can justify the action as being the right thing to do. In the case of our experiment, that justification comes from subjects selfishly telling the truth if they are knowledgeable. Further, ignorant subjects lie because lying decreases inequality in the coercion condition even though it is a selfish act. 23 In an extreme example, the state can offer to protect the family members of a potential source from a war torn country. 18

19 These behaviors suggest that reward is a better interrogation technique than coercion. This is because conditional defectors behave exactly the same way in COERCION and REWARD when they are knowledgeable, but there is a significant treatment effect when they are ignorant. When they are ignorant they provide poor information to avoid a low payoff under coercion. In REWARD, they truthfully state that they do not know because any other action would be purely selfish. Our experiment provides evidence of the common suspicion that coercion is an ineffective means of interrogation because of false statements from ignorant sources (Blakeley, 2007). This paper also contributes to the literature on how subjects behave in experiments. The extant research has shown that punishment is an effective mechanism to increase cooperation (Fehr and Gächter, 2002). Our experiment centers on instances in which society wants encourage individuals to defect from a group and we show reasons why rewards might be a preferable treatment. Additionally, we put a spin on the previous literature concerning lie aversion. That literature argues many individuals will lie only when doing so benefits them (Hurkens and Kartik, 2009). In our experiment, we show that you can incentivize people to tell the truth. Instead of the selfish lies observed in previous studies (Erat and Gneezy, 2012), our subjects participate in selfish truth telling. References Akerlof, George A, and William T Dickens (1982) The economic consequences of cognitive dissonance. The American Economic Review 72(3), Apesteguia, Jose, Martin Dufwenberg, and Reinhard Selten (2007) Blowing the whistle. Economic Theory 31(1), Armshaw, Patrick (2011) The torturer s dilemma: Analyzing the logic of torture for information. PhD dissertation, Florida State University Blakeley, Ruth (2007) Why torture? Review of International Studies 33(3), 373 Bolton, Gary E, and Axel Ockenfels (2000) Erc: A theory of equity, reciprocity, and competition. American economic review pp Bolton, Gary E, Elena Katok, and Rami Zwick (1998) Dictator game giving: Rules of fairness versus acts of kindness. International Journal of Game Theory 27(2), Chen, Yan, and Sherry Xin Li (2009) Group identity and social preferences. The American Economic Review pp Childs, Jason (2012) Gender differences in lying. Economics Letters 114(2), Conrad, Courtenay Ryals, and Will H. Moore (2010) What stops the torture? American Journal of Political Science 54(1), Diekmann, Andreas (1985) Volunteer s dilemma. Journal of Conflict Resolution pp

20 Dreber, Anna, and Magnus Johannesson (2008) Gender differences in deception. Economics Letters 99(1), Engelmann, Dirk, and Martin Strobel (2004) Inequality aversion, efficiency, and maximin preferences in simple distribution experiments. The American Economic Review 94(4), Erat, Sanjiv, and Uri Gneezy (2012) White lies. Management Science 58(4), Fehr, Ernst, and Klaus M Schmidt (1999) A theory of fairness, competition, and cooperation. The quarterly journal of economics 114(3), Fehr, Ernst, and Simon Gächter (2002) Altruistic punishment in humans. Nature 415(6868), Fehr, Ernst, and Urs Fischbacher (2004) Third-party punishment and social norms. Evolution and human behavior 25(2), Festinger, Leon (1962) A theory of cognitive dissonance, vol. 2 (Stanford university press) Fischbacher, Urs, Simon Gächter, and Ernst Fehr (2001) Are people conditionally cooperative? evidence from a public goods experiment. Economics Letters 71(3), Gneezy, Uri (2005) Deception: The role of consequences. The American Economic Review 95(1), Greiner, Ben (2004) An online recruitment system for economic experiments Hoffman, Elizabeth, Kevin McCabe, and Vernon L Smith (1996) Social distance and otherregarding behavior in dictator games. The American Economic Review 86(3), Holt, Charles A, and Susan K Laury (2002) Risk aversion and incentive effects. American economic review 92(5), Hurkens, Sjaak, and Navin Kartik (2009) Would i lie to you? on social preferences and lying aversion. Experimental Economics 12(2), Kahneman, Daniel, and Amos Tversky (1979) Prospect theory: An analysis of decision under risk. Econometrica: Journal of the Econometric Society pp Kahneman, Daniel, Jack L Knetsch, and Richard Thaler (1986) Fairness as a constraint on profit seeking: Entitlements in the market. The American economic review pp Konow, James (2000) Fair shares: Accountability and cognitive dissonance in allocation decisions. The American Economic Review 90(4), List, John A (2007) On the interpretation of giving in dictator games. Journal of Political Economy 115(3), Lundquist, Tobias, Tore Ellingsen, Erik Gribbe, and Magnus Johannesson (2009) The aversion to lying. Journal of Economic Behavior & Organization 70(1),

21 Lupia, Arthur, and Mathew D McCubbins (1998) Political credibility and economic reform. A Report for the World Bank Ostrom, Elinor (1990) Governing the commons: The evolution of institutions for collective action (Cambridge university press) Poe, Steven C, and C Neal Tate (1994) Repression of human rights to personal integrity in the 1980s: A global analysis. American Political Science Review pp Rabin, Matthew (1994) Cognitive dissonance and social change. Journal of Economic Behavior & Organization 23(2), Rapoport, Anatol (1988) Experiments with n-person social traps i prisoner s dilemma, weak prisoner s dilemma, volunteer s dilemma, and largest number. Journal of Conflict Resolution 32(3), Rejali, Darius (2009) Torture and democracy (Princeton University Press) Schiemann, John W (2012) Interrogational torture or how good guys get bad information with ugly methods. Political Research Quarterly 65(1), 3 19 Slonim, Robert, and Alvin E. Roth (1998) Learning in high stakes ultimatum games: An experiment in the slovak republic. Econometrica pp Suedfeld, Peter (2007) Torture, interrogation, security, and psychology: Absolutistic versus complex thinking. Analyses of Social Issues and Public Policy 7(1), Tajfel, Henri (1970) Experiments in intergroup discrimination. Scientific American 223(5), Wantchekon, Leonard, and Andrew Healy (1999) The game of torture. Journal of Conflict Resolution 43(5), Wisnewski, Jeremy (2010) Understanding torture (Edinburgh University Press) 21

22 A Instructions Instructions to Participants The following instructions were read to the subjects out loud at the beginning of each session. When text is in parentheses subjects received different instructions depending on whether they were randomly assign to groups or if they were participating in the minimal group paradigm sessions. The plain text was read to subjects in the random group treatments the italicized text that follows is what was read to subjects in the minimal group paradigm treatments. Text in brackets was not read aloud but included to help the reader understand what is taking place on the subject s monitors. Thank you for participating in today s experiment. I will be reading from a script to ensure that every session of this experiment receives the same instructions. These instructions explain the nature of today s experiment as well as how to navigate the computer interface you will be working with. We ask that you please refrain from talking or looking at the monitors of other participants during the experiment. If you still have a question or problem, please raise your hand and one of us will come to you. In today s experiment, all monetary amounts, earnings, and costs are denominated in Experimental Currency Units or ECUs. At the end of the experiment, your earnings in ECUs will be translated into dollars at the rate of 1 ECU equals one cent. So, if you end with a balance of 1,500 ECUs, you would be paid 15 dollars plus the10 dollar show up fee for a total of 25 dollars. We will pay you by check at the end of the experiment. Please turn to your computer screens, (we have prepared demonstration screens to help you understand the computer interface you will be working with. / and we will begin the experiment.) There is an index card in front of you. Please, enter the number on the card as your identification number and press OK. In this experiment, there are multiple sections. In the first section of the experiment, you will participate in groups of four and you will participate together for ten periods. After completing the first section of the experiment, you will be regrouped into another group of four and you will participate with that group for another ten periods. Please, press OK. In the first period of a section, you will be placed into a group and you will participate with that group for the entire section. In all subsequent sections, (you will participate with completely different group members. None of the people who were in your group in the first section will be in your group in the second section. / you will participate with different group members. It is possible that the people who were in your group in the first section will be in your group in the second section, but it is also possible that none of your group members in the second section were in your group in the first section.) Please, press OK. A.1 COERCION Instructions To determine who is in your group, we would like you to look at a series of paintings by two different abstract artists. The paintings will be presented in pairs. All of the paintings shown on the left were made by one artist and all of the paintings shown on the right were made by a different artist. For each pair, please indicate which painting you prefer. After you make your choice please press OK. Please press OK now and begin choosing between the paintings. [THEY CHOOSE; THEN THE SCREEN DISPLAYS: YOU ARE IN GROUP 22

23 A OR GROUP B OR GROUP C OR GROUP D ] As you can see, you ve been placed into groups. You were placed into these groups based on the paintings you just chose. The people in your group liked the same paintings. The people in the other groups preferred different paintings. This is your group for the first section of the experiment. Please press OK and we will walk through several demonstration screens to help you understand the first section of the experiment. We will now talk about what will take place in the first section of the experiment. In each period, you will see one of five possible signals which represent the state of the world. Four of the signals are the playing cards displayed on the screen. Half of the group will see one of those cards. This is your group s card. The fifth signal is the text unknown. If you see unknown, then you don t know what your group s card. Half of the group will see the text and not know what the group s card is. Please, press OK. In this section of the experiment, you are playing as a group against the computer. The computer is attempting to guess your group s card. The computer will ask you and your other group members if you want to reveal what the card is. You can say that you don t know what the card is and the period will end without the computer guessing. You can say that you will reveal a card and then the computer will ask what card you wish to reveal. You can select any of the four possible cards. Please, press OK. All group members are asked if they want to reveal a card, but the computer will only randomly select one of the subjects to listen to. If the computer listens to you, there are different payoffs depending on whether you choose to reveal a card or choose to say you don t know. If you choose to reveal a card, then the group s payoffs also depend on whether or not the computer can guess the card correctly. If you tell the computer the correct card, then you earn 500 ECUs and the other members of your group earn nothing. If the computer guesses incorrectly based on the card you revealed, then you receive 250 ECUs and the rest of your group members earn 500 ECUs. Remember, all group members can reveal a card even those group members who do not know what the card is. There is a 25 percent chance that the group member who does not know what the card is will reveal the correct card. If you decide to say that you don t know, then the computer does not guess a card. You earn 0 ECUs and your group members earn 500 ECUs. Please, press OK. After each period, you will see a table that shows your payoff for the current period, as well as, your payoffs in previous periods. It also displays whether or not the computer guessed correctly. In each section, you will be paid based on a randomly chosen period. After the first two sections of the experiment, you will be told which periods were chosen for payment. Please, press OK. This screen shows you what the screens will look like in this section. Where it says practice, either the card will be displayed or it will say unknown. If you want to remain silent, you can click the silent button. If you want to reveal the card, select a card and then click OK. You will get a chance to confirm your selection. Make a selection and then click OK. We are about to begin the first section of the experiment which will last ten periods. We ask that you follow the rules of the experiment. Anyone who violates the rules may be 23

24 asked to leave the experiment with only the 10 dollar show up fee. Are there any questions before we start? A.2 REWARD Instructions We are now going to place you in different groups. Once again, we would like you to look at a series of paintings by two different abstract artists they are different artists this time. The paintings will be presented in pairs. All of the paintings shown on the left were made by one artist and all of the paintings shown on the right were made by a different artist. For each pair, please indicate which painting you prefer. After you make your choice please press OK. Please press OK now and begin choosing between the paintings. [THEY CHOOSE; THEN THE SCREEN DISPLAYS: YOU ARE IN GROUP A OR GROUP B OR GROUP C OR GROUP D ] As you can see, you ve been placed into groups. You were placed into these groups based on the paintings you just chose. The people in your group liked the same paintings. The people in the other groups preferred different paintings. This is your group for the second section of the experiment. The people in your group now may not have been in your group in the first section. Please press OK and we will explain this section of the experiment. This is the second section of the experiment. This part of the experiment is similar to the previous section of the experiment except that this time the payoffs are changed. If you tell the computer the correct card, then you earn 1,000 ECUs and the other members of your group earn nothing. If the computer guesses incorrectly based on the card you revealed, then you receive 750 ECUs and the rest of your group members earn 500 ECUs. Remember, all group members can reveal a card even those group members who do not know what the card is. There is a 25 percent chance that the group member who does not know what the card is will reveal the correct card. If you decide to say that you don t know, then the computer does not guess a card. You earn 500 ECUs and your group members earn 500 ECUs. We will now begin the second section of the experiment which will last ten periods. Are there any questions before we start? Please, press OK. 24

25 A.3 Results by Group Type In half of the sessions, subjects were randomly assigned to their groups of four. After the first ten sessions, they were re-matched so that they did not play with the same subjects as they did in the previous 10 periods. In the other sessions, subjects were assigned to groups using a standard minimal group paradigm (MGP) procedure. Subjects were placed into groups based on their preferences for different abstract paintings. The results presented in the paper combined both types of groups. The following tables present results split by group type. Frequency of subject decisions by treatment and knowledge state for different group types. Table 5: Knowledgeable Subjects COERCION REWARD All Random Minimal All Random Minimal Right Card 74.40% 71.50% 77.30% 73.30% 72.30% 74.30% Wrong Card 22.60% 26.00% 19.30% 22.40% 23.00% 21.80% Don t Know 3.00% 2.50% 3.50% 4.40% 4.80% 4.00% N Table 6: Ignorant Subjects COERCION REWARD All Random Minimal All Random Minimal Reveal Card 87.50% 86.00% 89.00% 67.30% 67.00% 67.50% Don t Know 12.50% 14.00% 11.00% 32.80% 33.00% 32.50% N Table 7: Replication of third model in Table 4 for different group types, predicting subject truth telling. Probit models with standard errors adjusted for clustering on subjects All Random Groups Minimal Groups Coef. Z-Value Coef. Z-Value Coef. Z-Value COERCION Ignorant COERCION*Ignorant Previous Truth Teller Previous Correct Card Period Constant N (Subjects) 2,880 (160) 1,440 (80) 1,440 (80) A.I.C

26 A.4 Model with Subject Fixed Effects To account for the repeated measures on subjects, the models in the paper adjusted the standard errors for clustering on subjects. These models replicate those results and account for the repeated measures with fixed effects for subjects. Table 8: Replication of first and second models in Table 5 with subject level fixed effects. One subject is dropped due to no variance in the dependent variable. Fixed-effects logit model. Coef. Z-Value Coef. Z-Value COERCION Ignorant COERCION*Ignorant N (Subjects) 3180 (159) 3180 (159) AIC Table 9: Predicting subject truth telling. (MFX) Coef. Z-Value Coef. Z-Value Coef. Z-Value COERCION Ignorant COERCION*Ignorant Previous Truth Teller Previous Correct Card Previous Period Pay Period Table 10: Predicting selfish behavior. Coef. Z-Value Coef. Z-Value Coef. Z-Value COERCION Ignorant COERCION*Ignorant Previous Correct Card Pervious Selfish Previous Payment Period Constant N (Subjects) 3200 (160) 3200 (160) 2880 (160) A.I.C

27 A.5 Decision Screens Decision Screens The following four screen shots are the main decision screen that subjects saw in the experiment. The first two screen shots are the coercion and reward screens from the randomly assigned groups. The next two screen shots are the coercion and reward screens from the minimal group paradigm treatments. Top screen shots are from COERCION. Bottom are REWARD Figure 4: Random Groups 27

28 Figure 5: Minimal Group Paradigm 28