Team versus Individual Behavior with Non-Binding Commitments Puja Bhattacharya, John Kagel, Kirby Nielsen, and Arjun Sengupta Draft: June 1, 2016 Abstract We study the behavior of teams versus individuals in a setting where non-binding communication may serve to increase cooperation. We use a hidden-action trust game with pre-play communication (Charness and Dufwenberg, 2006) to analyze the role of cheap-talk in promoting cooperation. We replicate the findings of Charness and Dufwenberg for individuals, and see that these are not generalizable to teams. We find that teams and individuals respond to non-binding communication in the same way, but teams are much less likely than individuals to follow through after making a commitment. We try to understand the decision making processes of teams through analysis of team discussions. We find no mention of expectation-based guilt or preference for commitment underlying teams’ decision to carry through on their promises. Instead, we find that teams decide on their actions first and choose strategic messages to support their predetermined action. A majority of the team decisions can be explained by the first suggested action within the team. Teams’ willingness to trust is largely driven by a team’s willingness to take a risk in order to gain a higher payoff.

PRELIMINARY AND INCOMPLETE. PLEASE DO NOT CIRCULATE.

1

1

Introduction

Economic transactions are often characterized by actions which are imperfectly observable. These unverifiable actions and the lack of trust surrounding them can prevent potentially profitable partnerships from being formed. In environments with hidden action, one party’s actions are not perfectly observable and therefore are not enforceable by contracts. While complete contracts are not possible, non-binding commitments can still lead to more efficient outcomes. Even in the absence of enforceability, individuals can be motivated to fulfill their contracts due to moral obligations. A budding literature on individual decision making has shown that non-binding communication, a form of non-enforceable contracts, can increase cooperation in environments with hidden actions. However, there is no evidence to date on how communication affects cooperation with team decision makers. We use the hidden-action trust game of Charness and Dufwenberg (2006) (hereafter CD) to compare the behavior of two-person teams and individuals in a trust game with pre-play communication. The motivation for this is twofold. First, we want to directly compare the behavior of teams and individuals in this trust environment. The relevant economic decisionmaker in many partnerships and contracts is a group of people, rather than an individual, so it is important to understand the behavioral differences between individuals and teams. Second, our team design allows us to observe and record within-team conversations. We will analyze these intrateam discussions to gain direct insight into the teams’ decision-making processes. An overarching theme from social psychology and economic experiments comparing teams to individuals is that teams tend to be more strategic and less cooperative (see Charness and Sutter (2012) and Kugler et al. (2012) for review). Schopler et al. (2001) review the domain of the “discontinuity effect,” a term used to describe the greater competitiveness and reduced cooperation in intergroup interactions compared to interindividual interactions. This discontinuity effect between teams and individuals has been shown in a variety of environments, for teams of various sizes, and with varying degree of in-group interaction. Many explanations for the discontinuity effect have been proposed, and we review these in

2

Section 2. Our results are consistent with this choice discontinuity. In the baseline trust game without pre-play communication, we find teams to be less trusting than individuals, but just as trustworthy. With the addition of communication, however, we find that this relationship no longer holds. Under pre-play communication, our results show that teams are equally trusting but less trustworthy than individuals. Specifically, we see that teams make just as many promises as individuals do but are much less likely to keep their promises. However, teams and individuals respond to various commitments in the same way, so teams are no more or less likely to trust a cheap-talk message. Our results suggest caution toward intergroup commitments. Talk is much “cheaper” coming from teams than from individuals, and teams on the receiving end are unable to anticipate this. Analysis of the teams’ chat discussions gives some insight into the forces driving these relationships. The literature on pre-play communication in trust games has focused on how promises, or specific statements of intent, increase cooperation. In particular, a strand of literature seeks to determine whether promises are kept due to expectation-based guilt aversion, or due to an internal preference for commitment. The expectation-based guilt aversion model was formalized in CD, and proposes that an agent who sends a promise (the promisor) experiences disutility proportional to the extent to which he fails to meet the promise-recipient’s (promisee’s) expectations. The theory suggests that sending a promise to choose a particular action raises the promisor’s belief of the promisee’s expectations that the promisor will, in fact, choose that action. So a guilt-averse promisor will be more likely to choose the promised action after sending a message indicating that he will, as the promisee’s expectations are now higher and the promisor would experience greater disutility should he let down these expectations. On the other hand, the preference for commitment theory proposes that a promisor’s decision to keep his promise is irrelevant of the promisee’s expectations. Instead, he may keep his promise simply to avoid inconsistent behavior or incurring an internal fixed moral cost of lying.1 We look to see whether teams make any mention of these forces in their chat conversations. 1 Falk

and Zimmermann (2011)

3

Intrateam communication allow us to see teams’ decision making processes and how these vary across treatments. We find no explicit mention of expectation-based guilt or preference for commitment. The only teams to ever mention “feeling bad” in any way still choose to send and break promises. Similarly, we find no discussion of a preference to act consistently or avoid breaking a promise. Instead, we find that teams decide ex-ante whether they want to be “cooperative” or not, and then choose a strategic message to send, conditional on that predetermined action. The type of message sent is the same regardless of predetermined cooperation or non-cooperation, but teams that have decided to be cooperative keep their word and therefore appear trustworthy, and those that have decided to be non-cooperative don’t. Team behavior can be predominantly explained through the first suggested action within the team, and the action does not seem to be conditional on message sent. The absence of any communication regarding guilt or commitment is of course not conclusive evidence that it does not exist within team decision making. However, for these considerations to go completely unspoken, it must be that they are so universally understood that there is no need to speak them (even to an anonymous partner who may disagree) or they are stifled by other concerns.2 In Section 7, we include a discussion of possible explanations for why we do not see teams mention guilt or commitment concerns. The remainder of the paper is organized as follows: Section 2 summarizes the relevant existing literature in team decision-making and in trust games. Section 3 explains our experimental design, while Section 4 details our predictions. We present our results in Sections 5 and 6, and conclude in Section 7.

2 2.1

Related Literature Teams versus Individuals

The literature on team decision making gives reason to believe that teams and individuals will behave differently in a non-contractible environment. Charness and Sutter (2012) review 2 It could be that team decision makers are motivated by guilt, but these concerns are overshadowed by diffusion of responsibility, image concerns, etc.

4

the findings of team decision making to three broad conclusions: first, that teams are more cognitively sophisticated than individuals; second, that team decision making can help control against individual self-control and productivity problems; and third, that teams may reduce welfare compared to individuals due to stronger self-interested preferences. Given these broad predictions, we expect to see teams closer to Nash equilibrium predictions than individuals. In the environment of our paper, this predicts teams will be less trusting and less trustworthy than individuals. Kugler et al. (2012) also review the literature on team decision drawing the same conclusions, and include an additional discussion on the “risky shift,” or the difference between team and individual risk preferences. The risky shift is relevant to our analysis, as “trust” is inherently risky in our environment. It’s still an open question whether teams choose more or less risky outcomes, and the previous results to-date are context-specific.3 Several studies have found teams to be more risk-averse than individuals, others have found teams to be more risk-seeking, while still others find no difference between teams and individuals. Thus, the literature makes no clear predictions on the risk-taking or risk-avoiding behavior of teams in our environment, although contextual factors suggest it will be a relevant consideration in team decision making. A few papers have looked at the behavior of teams versus individuals specifically in a trust game environment, though without pre-play communication. Song (2006) looks at the behavior of group-representatives who privately and independently make decisions on behalf of three-player teams. She finds that group-representatives are less trusting and less trustworthy than individuals. Perhaps the closest paper to ours in terms of game structure is Kugler et al. (2007), which compares the behavior of teams and individuals in a standard trust game. In their game, the “sender,” either a three-person team or an individual, can send any portion of his endowment to the responder team or individual, after which this amount is tripled and the responder can return any portion of this tripled amount. Their main finding is that groups are less trusting than individuals, in that teams in the sender role send less money than individuals do. However, they find that teams are just as 3 See

Kerr et al. (1996) for an early review of the social psychology literature.

5

trustworthy as individuals, given by the fact that teams and individuals in the responder role return, on average, the same fraction of the amount sent. Our paper differs from that of Kugler et al. (2007) in several important ways. Kugler et al. specifically aim to test the trust and trustworthiness of teams and individuals, where our motivation is to study the effect of non-binding communication in enhancing trust and trustworthiness. Thus, the biggest structural difference between our paper and theirs is the introduction of pre-play communication, which stems from our very different motivations. In addition, Kugler et al. use three-person teams with face-to-face discussions, whereas we use two-player teams and anonymous computer interaction.4 Our paper is also related to the literature on deception and the differences in deception between teams and individuals. Gino et al. (2009a) report evidence that the immorality of one individual can influence the decisions of others. They find that one person’s dishonesty increases the dishonesty of those around him, especially when the dishonest individual is an in-group member. While Gino et al. don’t study team decision making per-se, their findings are relevant for our comparison of teams and individuals. Their results suggest that one team member’s endorsement of an immoral action will influence his partner’s decisions in that direction, as well, so the team decision may tend toward immorality. Gino et al. (2009b) show, similarly, that a team member’s unethical actions are more likely to be supported by his partners when there are no out-group observers, and alternatively are more likely to be compensated for in the presence of observers. In a later study, Gino and Galinsky (2012) show that psychological closeness to a dishonest individual leads to lower perceptions of dishonesty and immorality, which can be particularly relevant for team decision making if partners build a rapport. Finally, Gino et al. (2013) and Wiltermuth (2011) give evidence that dishonesty increases when individuals can justify their dishonesty as benefiting others. In our study, team dishonesty necessarily benefits another teammate, so the results from Gino et al. and Wiltermuth would suggest this as an explanation for higher team dishonesty compared to individual decisions. 4 This computerized interaction also allows us to study within-team communication transcripts, which Kugler et al. (2007) cannot do with face-to-face interactions.

6

2.2

Pre-Play Communication in Trust Games

In addition to the literature on team decision making, our paper also builds on the literature of non-enforceable pre-play communication, or “cheap talk” games. Spurred by CD’s original paper, an emergent strand of literature on trust games with pre-play communication seeks to identify the exact mechanisms behind promise-keeping and trustworthy behavior, and in particular whether promises are kept due to expectation-based guilt aversion or fixed cost preferences for commitment. CD give evidence for expectation-based guilt aversion, but the evidence is not conclusive. They observe that trustworthiness is correlated with higher second-order beliefs for the promisors, which suggests that trustworthy actions are taken as a result of guilt aversion.5 However, their results could be driven by a self-similarity effect, where promisors who are more likely to behave in a trustworthy manner are exactly those agents who believe promisees expect them to do so.6 The CD results could also be consistent with a fixed cost preference for commitment. If promisors believe that the recipients can accurately predict promisor behavior, then promisors who choose to behave in a trustworthy manner due to fixed cost preference for commitment will also report higher second-order beliefs. Both of these confounds are due to a similar “consensus effect” phenomenon, where agents expect others to behave and think like themselves. Vanberg (2008) employs a clever experimental design to conclude that expectation-based guilt aversion does not drive promise keeping. He uses a dictator game with a pre-play communication stage, during which the participants are unaware which of them will be the dictator. Therefore, they both tend to use this communication stage to send promises that they will behave fairly, should they be chosen as dictator. Vanberg exogenously manipulates expectations by switching some of the partnerships, where now the dictator may be partnered with the participant he sent his promise to, but might instead be partnered with another participant. Dictators are aware of whether their partner has been switched, but recipients are not. Importantly, if the dictator faces a different participant, he gets to see the 5 In the CD game, there are two players, A and B. A’s first-order beliefs represent the probability with which he believes B will choose the trustworthy action. B’s second-order beliefs are B’s beliefs over A’s first-order belief. 6 Ellingsen et al. (2010)

7

message that his new partner received from another dictator in the communication stage. Vanberg hypothesises that, if dictators were guilt-averse, they would give similar amounts to recipients regardless of whether they had been switched or not, since the recipients’ expectations were raised by receiving a promise. Contrary to this hypothesis, he finds that dictators give significantly more when paired with their original partners, and give much less when paired with new recipients to whom they had not made a promise. This gives evidence that guilt aversion was not driving their decisions, and leans toward a preference for commitment explanation. Ederer and Stremitzer (2014) claim that Vanberg’s results could be explained by the fact that an individual may experience guilt only if he was responsible for raising the promisee’s expectations. Their design exogenously manipulates the promisee’s expectations by changing the state of the world, rather than changing the promisor-promisee relationship. In their experiment, there is a varying exogenous probability that the promisor will be able to keep his promise, even if he intends to. This manipulation exogenously varies the promisor’s second-order beliefs, and their results show that promisors are sensitive to these beliefs. Thus, Ederer and Stremizter give evidence for expectation-based preferences for promise keeping and conclude that promisors are sensitive to promisee’s payoff expectations. A recent paper by Bhattacharya and Sengupta (2016) endogenously varies promisors’ second-order beliefs, and finds evidence of both guilt-aversion and preference for commitment in different portions of their sample. Their design offers promisees the option of buying ‘insurance’ against the trusting decision. This insurance purchase is visible to the promisors, so promisors can take insurance purchase as a signal of promisee’s expectations. The promisee would prefer to buy insurance if she distrusts the promisor and would not buy insurance if she trusts the promisor, so the promisor can directly infer the promisee’s expectations from her insurance purchase decision. They find heterogeneity in their sample. Some promisors are sensitive to the expectations of the promisee and adjust choices based on insurance purchase. However, in a larger portion of the population, promisors keep their promise regardless of promisee expectations. Another recent paper by Ismayilov and Potters (2016) presents a different interpretation

8

of promises. Ismayilov and Potters use the game from CD, but modify it so that messages from promisors were delivered to promisees with probability 0.5. Both players learn whether the message was delivered or not before making their decisions. They find that players who sent a promise were significantly more likely to behave trustworthy than participants who did not send a promise, regardless of whether the message was actually delivered. However, a baseline treatment where only non-promising messages were permitted showed the same level of trustworthiness as in the control treatment. They therefore conclude that the correlation between promises and trustworthiness is due to self-selection. It is not that promises increase cooperation, but that cooperative people are more likely to send promises than non-cooperative people. The sample above is not exhaustive, but shows the various ways in which the previous literature has tried to disentangle expectation-based explanations from pure preferencebased explanations of promise keeping. By exogenously or endogenously varying secondorder beliefs, these papers try to infer promise-keeping behavior in different environments. However, the evidence from previous literature on the channel through which promises work to increase cooperative behavior is still inconclusive.

2.3

Pre-Play Communication with Teams

The reviewed papers on pre-play communication all use individual, rather than team, decision makers. A handful of papers have looked at the behavior of teams in cheap talk environments. Cohen et al. (2009) test whether teams lie more than individuals, and find that teams are more willing than individuals to lie when it is payoff-maximizing. Teams are also more willing to tell the truth when truth-telling can be used strategically, so the results generally suggest that teams communication more strategically than individuals. Sutter (2009) confirms these results, also showing that teams strategically tell the truth when it benefits them. The above papers focus on strategic deception and truthful reporting of a state of nature. We are the first paper that we know of to look at the effect of non-binding commitments on teams, and specifically in a trust game environment.

9

3

Experimental Design

The Game We use the hidden-action trust game of CD, keeping payoffs consistent with their design. The game tree is reproduced below, and the names of players and strategies are the same as those used in the experimental protocol.7 We implement the game sequentially, using z-Tree (Fischbacher, 2007). Participants played 5 periods of the same game with perfect stranger matching. Roles were held constant throughout the session, so each subject only experienced the game as Player A or as Player B.

Figure 1: Game Tree We use a 2x2 experimental design, varying communication, {no communication, communication}, and the type of decision maker, {individual, team}. In the no communication treatment, participants played the game exactly as shown in the game tree. All Player As decided In or Out, then Bs decided Roll or Don’t Roll. In the communication treatment, 7 See Appendix for subjects’ instructions and images from the experimental interface. The given instructions and images are for the team treatments, which were designed based on the original CD instructions. Instructions and images for the individual treatments are analogous and available upon request.

10

B had the opportunity to send a free-form message to A through the computer before A decided In or Out, after which the decisions proceeded as in the no communication treatment. In our individual treatments, each A or B was a single individual, whereas in our team treatments, decisions were made by 2-person teams.8 That is, each A player was a team of two anonymous subjects acting together, as was each B player. The teammates could nearly-continuously communicate with one another through a chat box, and were required to reach agreement on all actions made and messages sent by the team.9 Partners were held constant throughout the session, so the team composition remained consistent with the same two players throughout all 5 periods. Payoffs were given to each player of a team, so monetary incentives were held constant between the two treatments. Feedback We used the strategy method for Bs’ decisions in all treatments, so B did not learn A’s decision before making his choice. We suppressed all choice feedback until the end of the session. After each Period, participants proceeded immediately to the next Period. Neither As nor Bs received any feedback on the decisions made by the other participants, and they did not learn their payoff outcomes in between Periods. At the end of the session, all participants learned the payoff they would have received in each of the 5 Periods, and then learned which randomly selected Period would determine their payment. Participants never learned the actual decisions made by their matched counterparts.10 Team Procedure In order to implement the design using teams, we allowed the teammates to communicate with one another using an on-screen chat box. They were not restricted in their dialogue in any way, except they were told to refrain from using profanity and were told not to 8 In the experiment sessions, we used the word “group” rather than “team” in order to reduce any pure competition effects from being on a team. We also refrained from using the word “partner” whenever lexically convenient. 9 Note, teammates could chat with one another in BOTH the communication and no communication treatments. Our differentiation between communication and no communication refers to between-team communication in the form of messages sent from B to A, not within-team chat communication. 10 Bs receiving a payoff of $5 could easily infer A had chosen Out, and As receiving a payoff of $12 could infer B had chosen Roll. However, upon receiving a payoff of $0, A did not know whether B chose Roll and the Chance move was unsuccessful, or whether B chose Don’t Roll.

11

identify themselves in any way.11 In the communication treatment, the B teams had 2 minutes during which they could chat and come to agreement on a message. In order to give both members input into the message content while still requiring agreement, either member could propose a message to send to the A team. Once he proposed a message, his partner would see the message and choose to accept it or reject it. If he accepted the message, it would be sent to the A team after the 2 minutes expired. If he rejected it, he had the opportunity to rewrite the message. If the teammates did not come to agreement on a message within the 2 minutes, one member was randomly selected and given 30 seconds to write a message on behalf of the group.12 While the B teams decided on messages, the A teams had 2 minutes to freely chat with their partners. After all B teams had written a message, the messages were delivered to their respective matched A teams, and the A teams had 1 minute to decide on their action. The teammates were required to agree on their decision to choose In or Out. If they could not come to agreement within the 1 minute time limit, one teammate was randomly selected to make the decision on behalf of the group.13 While the A teams decided on their actions, the B teams were able to chat. After all A teams had decided on their actions, the B teams had 1 minute to come to agreement on their decision to choose Roll or Don’t Roll. As with the A teams, if they could not reach agreement, one member was chosen to make the decision on behalf of the team.14 In the no communication treatment, the message stage was omitted but decision times were held consistent with the communication treatment. A total of 330 subjects participated across all sessions. Subjects were primarily from the undergraduate student population at the Ohio State University, recruited through ORSEE (Greiner, 2004). We excluded subjects who had participated in a previous trust game 11 Chat analysis indicates subjects generally followed these instructions, with the exception of a few subjects disclosing minor identifying characteristics such as major and class schedule. 12 We did not have any instances where group members disagreed with the message proposed. There were only 5 rejected messages, and these were mostly instances where one teammate accidentally typed a portion of the chat conversation into the message box, so his teammate rejected the message to prevent it from being sent to the A team. We had 17 instances where the teammates ran out of time deciding on the message (typically in the first round), and one member proposed the message on behalf of the group. however, chat conversations indicate that, in most of these situations, the teammates had already agreed on their message but ran out of time in typing it. 13 We only had 5 instances of this across all 275 decisions made by A teams, and all 5 occurred in Round 1. 14 There were only 2 instances where B teams did not reach agreement within the time limit.

12

experiment.15 Sessions lasted under 1 hour, and payments averaged $11.50 in the team treatments and $13 in the individual treatments, including a $5 show-up fee.

4

Predictions

4.1

A and B Decisions

As our experiment features a complete 2x2 design, we can look at the full relationships between teams and individuals, across no communication and communication environments. In comparing teams and individuals without pre-play communication, we will look to confirm the results of Kugler et al. (2007). Kugler et al. find that, in a similar trust game without preplay communication, teams are less trusting than individuals, but are just as trustworthy. Our first hypothesis is that we will see similar patterns in our environment. Hypothesis 1a: In our no communication treatments, teams will be less trusting than individuals, measured by A teams choosing In less often than A individuals. Hypothesis 1b: In our no communication treatments, teams will be as trustworthy as individuals, measured by B teams choosing Roll as often as B individuals. In comparing across no communication and communication environments with individual decision makers, we will look to confirm the CD results. CD find that communication increases cooperation for both A and B players. Our individual treatments are a close replication of the CD design, so we expect to see the same results. Hypothesis 2a: A individuals will be more likely to choose In in the communication treatment than in the no communication treatment. Hypothesis 2b: B individuals will be more likely to choose Roll in the communication treatment than in the no communication treatment. 15 Chat

conversations indicated that two subjects had experience with a similar game, but this experience did not seem to influence their decisions and they did not use it to persuade their teammate one way or another. One subject also indicated that his roommate had participated in the experiment, but again this did not seem to influence his choices.

13

Our results will provide the first evidence comparing team decision makers across environments with and without communication. There is no clear prediction for the effect of communication on team decision makers. The abundance of literature on individual communication would suggest higher cooperation under communication. However, the literature on team decision making would suggest that the higher selfishness and/or strategic reasoning of teams might mitigate the moral force of commitments. The relative magnitude of these effects will determine the results. However, it’s clear that cooperation levels should be at least weakly greater under communication than under no communication. Hypothesis 3a: Communication will weakly increase cooperation for A teams, in that A teams will be no less likely to choose In in the communication treatment than in the no communication treatment. Hypothesis 3b: Communication will weakly increase cooperation for B teams, in that B teams will be no less likely to choose Roll in the communication treatment than in the no communication treatment. Our results will also provide the first evidence comparing teams to individuals in an environment with pre-play communication. Given the heightened rationality of teams, we predict A teams will discount messages and be less likely to choose In after receiving a message. As for Bs, the previous literature suggests that teams are more willing than individuals to engage in strategic deception. In our environment, this implies B teams will be less likely than individuals to choose Roll after sending a promise indicating they will do so. Hypothesis 4a: Teams will be less trusting of messages than individuals, so A teams will choose In less often than A individuals after receiving a message. Hypothesis 4b: The efficiency-enhancing effect of communication will be greater for individuals than for teams. That is, B teams will Roll less often than B individuals after sending a message.

14

4.2

A and B Chat

A benefit of our experimental design lies in the role of within-team chat. We will analyze the chat conversations in light of the A and B team decisions. In particular, we will look to the chat to see any indicators of trust and trustworthiness. Chat analysis may give insight into the motivations behind trusting and trustworthy behavior and the decision processes involved therein. Additionally, there have been many clever experiments designed to explain the decision processes leading toward greater cooperation under communication. We propose looking at within-team chat to gain insight into team decision-making and the differences in decision-making between communication and no communication treatments.16 Within the communication treatment, the expectation-based guilt aversion hypothesis suggests that sending a promise will raise A’s expectations that B will choose Roll.17 We look at A team chat to see whether A teams actually consider the messages they receive, and whether these messages affect beliefs. Previous studies have shown correlation between promises and higher beliefs, but we want to see whether these considerations emerge organically, without the intrusion of a belief-elicitation which forces promisees to consider the message and their expectations. Given the predictions in the literature, we hypothesize that A team chats will reveal that As consider the messages they receive and different messages will have differing impact on their expectations. The guilt aversion hypothesis also predicts that Bs will react to As’ expectations of Bs in choosing their decisions. That is, Bs consider As’ expectations when making their decisions, and Bs are more likely to choose Roll if they believe As expect them to do so. We want to see whether Bs ever discuss these second-order expectations, without the added mechanism of a belief elicitation forcing them to do so. Both preference for commitment and expectationbased guilt aversion have been hypothesized in the literature, and we look for evidence of these in the recorded team decision-making processes. 16 To be clear, most previous studies have used individual, rather than team, decision makers. Withinteam chat analysis gives us the benefit of direct evidence on team decision-making processes, but all results will not necessarily carry over to individual decision makers. We will use chat analysis to gain insight into the patterns we see in team behavior. 17 Guilt aversion doesn’t necessarily require A’s expectations to change after receiving a message. It requires B to believe A’s expectations will change. However, A’s expectations actually changing is only natural for the consistency of the theory.

15

5

Experimental Results

We analyze data from 10 team sessions (220 subjects) and 6 individual sessions (110 subjects). Of the team sessions, 5 are with communication (112 subjects) and 5 are without communication (108 subjects). Of the individual sessions, 3 are with communication (58 subjects) and 3 are without (52 subjects). Given that each A and B team comprises two subjects, we therefore have observations for 27 of each A and B teams without communication, 28 A and B teams with communication, 29 A and B individuals with communication, and 26 A and B individuals without communication. We randomized our 330 subjects across treatments in order to have balanced observations, but our numbers of independent observations are not quite comparable to the studies we replicate.18 In order to be equivalently powered in comparison to CD, we will collect up to 45 observations in each treatment in future research. In analyzing the experimental results, the unit of observation for the individual sessions will be an individual, and the unit of observation for the team sessions will be a team. The statistics we present will be based on calculations at the individual- or team-level, unless stated otherwise.

5.1

The Effect of Communication

We first look to see whether communication in general increases cooperation rates for teams and individuals. A large literature has shown that the ability to communicate makes individuals more cooperative and more likely to achieve efficient outcomes, even in environments where incentives are misaligned. However, it’s unclear whether this result will remain true with team decision makers. Our individual and no communication treatments are a replication of CD and Kugler et al., and our results serve as confirmatory evidence of the hypotheses in the literature and a baseline for comparison. Our team results present the first evidence of the effect of non-binding contracts on team decision makers.

18 Charness and Dufwenberg have observations for 42 indiviuals in their communication treatment and 45 in the no communication treatment. Kugler et al. have observations for 27 teams and 32 observations for individuals.

16

Figure 2: A comparison of the average In and Roll rates for teams and individuals, between communication and no communication treatments. First, we look to see whether our results replicate the findings from Kugler et al., who show that teams are less trusting than individuals, but are just as trustworthy, in an environment without communication. When communication was not possible, A individuals chose In 50% of the time and A teams chose In 35% of the time. This is a substantial 15% difference, and Wilcoxon rank-sum tests confirm this difference is statistically significant (p=0.086). B individuals choose Roll 28% of the time compared to B teams who choose Roll 19% of the time, but this difference is not statistically significant (p=0.102).19 We therefore replicate the findings of Kugler et al. in our trust game environment. 19 We acknowledge the magnitudes of these two results are not drastically different in declaring one significant and the other not significant. We expect these results to strengthen with more data.

17

Result 1a: Teams are less trusting than individuals in an environment with no communication. A teams choose In less often than A individuals. Result 1b: Teams are as trustworthy as individuals in an environment with no communication. B teams choose Roll as often as B individuals. Next, we look at the difference between communication and no communication, focusing on individual decision makers. CD see that communication increases cooperation for both A and B individuals, so As are more likely to choose In and Bs are more likely to choose Roll in the communication treatments. For A individuals, we see they are 13% more likely to choose In in the communication treatment. Similarly, communication increases the likelihood of B individuals choosing Roll by 14%. We test for the significance of these differences using Wilcoxon rank-sum tests, as well as probit regression analysis. See Table 12 and Table 13 in the Appendix for results. We replicate the magnitude and direction of the effect of communication on individuals as found in CD’s original paper, and the increase in communication is statistically significant for As (p=0.085). However, we have fewer observations than CD, so the effect does not turn up significant for Bs (p=0.163), though the magnitude is of the same order. More sessions are needed to establish the effect of communication on individuals, though directionally the results are confirmed. Result 2a: A individuals are more likely to chose In in the communication treatment than in the no communication treatment. Result 2b: B individuals are more likely to choose Roll in the communication treatment than in the no communication treatment. To test our third set of hypotheses, we look to see the difference between communication and no communication, focusing on team decision makers. As Figure 2 shows, A teams are 21% more likely to choose In under the communication treatment, while B teams are only 5% more likely to choose Roll. Wilcoxon rank-sum tests confirm that the effect of communication is statistically significant for A teams (p=0.024), but the increase is very small and statistically insignificant for B teams (p=0.267). On an aggregate level, we can conclude that the B teams’ messages are influential on A teams’ behavior, but they do not 18

change Bs’ behavior. Therefore, we can see that the CD result does not generalize to teams. Our results suggest that A teams may place value in messages and their content, but B teams do not base their actions on their own messages. Result 3a: Communication increases cooperation for A teams. Result 3b: Communication does not increase cooperation for B teams. Finally, we look to see the differences between teams and individuals in an environment with pre-play communication. In our communication treatments, A individuals chose In 63% of the time, on average, and A teams chose in 56% of the time. This difference is fairly small and is not statistically significant (rank-sum p=0.357). B individuals choose to Roll 42% of the time, compared to B teams who Roll only 24% of the time. This 18% difference is large in magnitude and is statistically significant (p=0.043). We conclude, therefore, that teams and individuals are equally trusting after receiving communication, but teams are less trustworthy after making commitments. Result 4a: Teams and individuals respond to strategic communication in the same way. Conditional on message received, A teams are just as likely to go In as A individuals. Result 4b: B teams are less trustworthy under communication than B individuals. Results 4a and 4b are fairly surprising, and Result 4a is contrary to our Hypothesis 4a. We do not see that teams respond differently to messages than do individuals, as teams are no more distrusting of message content than individuals are. Result 4b is in line with our hypothesis, but both 4a and 4b are contrary to the results found in Kugler et al. These results show that the findings from Kugler et al., as replicated in our no communication baseline, do not extend to environments with pre-play communication. Although teams are less trusting than individuals in the absence of communication, this difference goes away when teams and individuals can receive messages. It seems as though communication equalizes trust across teams and individuals. On the other hand, while teams and individuals are equally trustworthy in the absence of communication, teams are much less trustworthy 19

than individuals after sending messages. The ability to engage in strategic communication seems to make evident the behavioral differences between teams and individuals. After establishing our main results, we look to analyze our communication treatments in more detail. The previous literature suggests that the efficiency-enhancing impact of communication comes from promises, specifically. While it does not appear that communication increases cooperation for B teams overall, it could still be that communication increases cooperation for B teams conditional on sending a promising message.

5.2

Message Coding

In order to study the role of promises in particular, we first code the free-form messages sent from B to A. To remove experimenter bias, we had two outside coders classify the messages into the following categories: Strong Promise, Weak Promise, Empty Talk, and No Message. Coders were undergraduate students from the Ohio State University, neither of whom participated as a subject in the experiment. Given that they are drawn from the same subject pool as our participants, we feel they have a fairly accurate interpretation of their peers’ conversations. Coders were given a description of the message classifications along with a few examples, and were asked to code each message sent from B to A into exactly 1 of the 4 categories.20 Consistent with the classifications and findings from Houser and Xiao (2010), we categorize a message as a Strong Promise if the sender clearly makes a promise to choose the action Roll. We categorize a message as a Weak Promise if it includes any statement of intent or reference to choosing Roll, but is not a clear promise. Empty Talk messages were messages unrelated to the game or messages without any statement of intent, and No Message was reserved for blank messages or messages where the sender wrote “No Message.”21 20 Coders reached an agreement rate of 94% on message categorization. Full classification of message coding available upon request. 21 Subjects were told they had the option to leave the message blank or write “No Message” if they wished not to send a message.

20

Individuals

Teams

Example

Number

Percentage

Number

Percentage

Strong Promise

80

55%

74

53%

“We will choose ROLL”

Weak Promise

24

17%

33

24%

“It would be wise to choose In”

Empty Talk

7

5%

8

6%

“Hi!”

No Message

34

23%

25

18%

Table 1: Message frequencies across treatments While subjects were free to use the message opportunity as they wished and the experimenters never mentioned promises, a majority of our subjects use the opportunity to signal trustworthiness by sending a promise to choose Roll. Out of 145 total messages sent in the individual treatment and 140 total messages sent in the team treatment, 55% of messages sent from individuals were Strong Promises, compared to 53% of messages sent from teams.22 17% of messages in the individual treatment were Weak Promises, compared to 24% in the team treatment. 5% of message in the individual treatment were Empty Talk compared to 6% in the team treatment, and 23% of messages in the individual treatment were No Messages compared to 18% in the team treatment. Fisher exact test confirms these differences are not statistically significant (p=0.40). Thus, it does not appear that teams and individuals use messages differently. On an aggregate level, teams and individuals use messages to the same strategic purposes.

5.3

Team vs. Individual Behavior

Behavior of As We look at the behavior of As by analyzing the percentage of A teams and individuals who choose to go In. We condition the decision on type of message received – Strong Promise, Weak Promise, Empty Talk, or No Message.23 22 The 145 and 140 messages are the result of 29 individual participants and 28 team participants in role B, sending one messages in each of 5 Periods. 23 Because we have repeated observations, statistics are based on individual- or team-level averages over Periods, so n=29 for individuals and n=28 for teams. For individuals, we calculate the average of each

21

Individuals

Teams

p-value

Strong Promise

71%

60%

(0.26)

Weak Promise

69%

68%

(0.89)

Empty Talk

17%

21%

(0.70)

No Message

36%

39%

(0.84)

No Communication

50%

35%

(0.086)

Table 2: The percentage of A individuals and teams who choose to go In, conditional on type of message received. Reported p-values are from Wilcoxon rank-sum tests on individualor team-level averages. Table 2 shows that individuals are slightly more likely to go In than teams conditional on receiving a Strong or Weak Promise, and are slightly less likely than teams to go In conditional on receiving an Empty Talk or No Message. These differences are small, and Wilcoxon rank-sum tests show that none of the differences are statistically significant. We confirm these results using random effects probit regressions. Our aggregate measures are based on individual- or team-level averages, which give higher weight to some teams or individuals compared to others, depending on types of messages received.24 Regression analysis with errors clustered at the individual or team level allows us to confirm that our results are not an artificial effect of our aggregation procedure. We split the sample by type of message received in a given round, and regress A’s decision to choose In or Out on a team dummy, controlling for period effects. Results can be found in Table 14 in the Appendix, and show that there is no significant difference between team and individual In rates, conditional on any type of message received.25 Thus, in our comparison between individual’s In rate across all 5 Periods. That is, for each individual, we calculate the number of times he went In conditional on receiving each message type. Then, we average across all A individuals. For example, an individual who received 3 promises and went In in 2 of those 3 Periods would have an individual-level In rate of 66%, conditional on receiving a promise. We average across these individual-level In rates for all subjects to obtain the reported averages. We do the analogous calculation for groups, finding the group-level In rate conditional on message type, and then averaging across groups. 24 For example, a team or individual who receives 1 Strong Promise and chooses to go In has a 100% In rate, conditional on a Strong Promise. Another individual who receives 5 Strong Promises and chooses to go In 4 times also has only an 80% In rate, but this average is based on more observations. Similarly, not all As receive all message types, so some As are represented in multiple categories while others are not. 25 In addition to the aggregate measures, we also look at the distribution of In rates conditional on message received for both individuals and teams (see Figure 7 in the Appendix). Kolgomorov-Smirnov tests show

22

teams and individuals, we see that teams and individuals are not just equally trusting under communication in general, but that they’re equally trusting conditional on receiving any type of communication from B players. The results from Table 2 also suggest that the A teams and individuals are more likely to trust B after receiving a Strong or Weak Promise than after receiving an Empty Talk or No Message. For both teams and individuals, As are more likely to choose In after receiving a Promise than after receiving a non-promising message. These results are also confirmed using probit regression analysis. We report random effects probit regressions predicting the decision to choose In, given type of message received and controlling for Period effects.26 The reported effects for different message types are relative to the no communication treatment. We analyze the data separately for teams and individuals, to see if the effect of receiving a Promise is significant for both types of decision makers. Then, we pool the teams and individuals together to look at the effect of message and the effect of decision maker together.27 no differences between the distributions, p=0.38 for Strong & Weak Promise combined, p=0.99 for Empty Talk & No Message combined. 26 Errors are clustered at the individual or team level. 27 Ideally, in this regression as well as the analogous regression for Bs in Table 5, a specification where the team treatment is interacted with each message category would capture the differential impact of message types across teams and individuals. However, we defer this analysis until we have collected more data. Splitting our already small number of observations would not give us enough power to make conclusive statements, so we leave the analysis for when we can make more robust predictions.

23

VARIABLES Strong Promise Weak Promise Empty Talk No Message

Individuals

Teams

Pooled

Choosing In

Choosing In

Choosing In

0.850**

1.020**

0.932***

(0.412)

(0.409)

(0.286)

0.956**

1.275***

1.119***

(0.478)

(0.472)

(0.330)

-0.645

-0.513

-0.623

(0.946)

(0.615)

(0.564)

-0.401

0.164

-0.149

(0.416)

(0.492)

(0.315)

Team

-0.447* (0.244)

Period

-0.179***

-0.124*

-0.154***

(0.0624)

(0.0709)

(0.0478)

0.521*

-0.24

0.375

(0.294)

(0.358)

(0.252)

Observations

275

275

550

Number of Clusters

55

55

110

Constant

Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Table 3: Random effects probit regression for A teams and individuals The results largely present the same picture shown from the aggregate measures. The coefficients on both Strong Promise and Weak Promise are positive and statistically significant for both teams and individuals, reflecting the higher likelihood of A choosing In after receiving a Promise. In the pooled regression, we see a slight negative effect of being in a Team, which is identifying the higher trust of individuals in the no communication baseline. These results conclude that both A teams and individuals are responsive to Promises, and are more likely to choose In after receiving a Promise.28 28 As is common in repeated games, we see a substantial end-game effect, where participants are less cooperative in later rounds. Chat analysis reveals part of this could be due to subjects’ experience in previous experiments, e.g. “haha maybe we can do in for the first two or three rounds. people tend to be more nice the first several rounds.” We run the same regressions on different subsets of our data, to be

24

Result 5: Both A teams and A individuals are more likely to trust Bs and choose In after receiving a Promise than after receiving an Empty Talk or blank message.

Behavior of Bs We look at the behavior of Bs by analyzing the percentage of B teams and individuals who choose to Roll. We condition the decision on type of message sent – Strong Promise, Weak Promise, Empty Talk, or No Message.29

Individuals

Teams

p-value

Strong Promise

56%

27%

(0.03)

Weak Promise

39%

33%

(0.72)

Empty Talk

28%

20%

(0.39)

No Message

43%

18%

(0.19)

No Communication

28%

19%

(0.102)

Table 4: The percentage of B individuals and teams who choose to Roll, conditional on message sent. Reported p-values are from Wilcoxon rank-sum tests on individual- or teamlevel averages. Results from Table 4 show that teams are less likely than individuals to choose Roll, conditional on any message sent. These differences between teams and individuals are statistically significant for Strong Promises (Wilcoxon rank-sum test p=0.03). We again confim these results using random effects probit regressions. We split the sample by type of message received in a given round, and regress B’s decision to choose Roll or Don’t Roll on a team dummy, controlling for Period effects. Results can be found in Table 15 in the Appendix. The coefficient on Team is significantly negative in the Strong Promises regression, indicating that teams are significantly less likely than individuals to Roll after sure the results aren’t driven by early-round confusion or late-round deterioration of cooperation. The same results hold qualitatively if we look at just Periods 2-5, or at just Periods 1-4. 29 Statistics are calculated at the individual- or team-level, so n=29 for individuals and n=28 for teams. These percentages are calculated in the same way as the A In rates. See footnote 23 for reference.

25

sending a Strong Promise. The coefficient on Team is not significant in any other regression, concluding that there is no significant difference between team and individual Roll rates for the other types of messages.30 We see that B teams are less trustworthy than individuals, as teams are significantly less likely to choose Roll after A chooses In, especially after sending a Promise to choose Roll. The difference in trustworthiness between B teams and individuals in the communication treatment is primarily driven by Bs who send a Strong Promise. Result 6: B teams are equally likely to send a promise as B individuals, but are significantly less likely to keep their promises. The results from Table 4 also suggest that both B teams and individuals are somewhat more likely to Roll after sending a Promise than after sending a non-promise. We test this using probit regression analysis. We report random effects probit regressions, predicting the decision to choose Roll given the type of message sent and controlling for Period effects. The reported effects for different message types are relative to the no communication treatment. We analyze the data separately for teams and individuals, to see if the effect of sending a Promise is significant for both types of decision makers. Then, we pool the teams and individuals together to look at the effect of message and the effect of decision maker together.31 30 In addition to the aggregate measures, we also look at the distribution of Roll rates conditional on message sent for both individuals and teams (see Figure 8 in the Appendix. Kolgomorov-Smirnov tests show borderline significant differences between the distributions, p=0.08 for Strong Promise, p=0.11 for Strong & Weak Promise, p=0.37 for Empty Talk & No Message 31 Again, we defer analysis of the interaction effect of messages on teams.

26

VARIABLES Strong Promise Weak Promise Empty Talk No Message

Individuals

Teams

Pooled

Choosing Roll

Choosing Roll

Choosing Roll

1.563**

1.210

1.500**

(0.801)

(0.926)

(0.656)

0.593

2.057*

1.427**

(0.829)

(1.071)

(0.721)

1.316*

-1.737

0.468

(0.781)

(2.315)

(0.791)

-0.571

-1.907

-0.806

(0.798)

(1.420)

(0.710)

Team

-1.566** (0.617)

Period

-0.294***

-0.354**

- 0.297***

(0.062)

(0.143)

(0.0568)

-0.410

-3.657***

-0.585

(0.494)

(0.543)

(0.555)

Observations

275

275

550

Number of Clusters

55

55

110

Constant

Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Table 5: Random effects probit regression for B teams and individuals Again, the probit regression results draw the same conclusions as the aggregate statistic measures. The coefficient on Strong Promise is positive and significant for individuals, indicating that individuals are more likely to Roll after sending a Promise than in the no communication treatment.32 For teams, the effect of a Weak Promise is marginally significant, but overall teams are not significantly more likely to Roll after sending any message than they were when communication was not possible. If we look only at Periods 25, the results hold similarly for Individuals, but the entire effect of communication disappears 32 While we need more data to confirm the CD result of higher B cooperation under communication in general, we do replicate the efficiency-enhancing effect of promises for B individuals.

27

for teams.33 In the pooled regression, the coefficient on Team is significantly negative and large in magnitude, mitigating any effect of sending a Promise. Our results suggest that the ability to communicate has a strong effect in increasing the cooperation of B individuals, but has a negligible effect on the cooperation of B teams. Result 7: Compared to the no communication baseline, B individuals are more likely to Roll after sending a Promise, but B teams are not. To summarize, we gain important insights by looking at specific types of messages within the communication treatment. First, we see that A teams are no less likely than A individuals to go In, conditional on any type of message received. Both A teams and A individuals are more likely to trust Bs and go In after receiving a Promise, but fixing message content, A teams and individuals respond similarly to communication. This is in contrast to the differences in trust we see between A teams and individuals in the no communication treatments. For Bs, we see a significant difference in trustworthiness between teams and individuals in the communication treatments. B teams and individuals are equally likely to send a Promise, but B teams are less likely to keep this Promise and choose Roll. We see that communication has an efficiency-enhancing effect for B individuals, as they are more likely to Roll after sending a Promise than they are in the no communication treatment. However, for B teams, sending a Promise does not result in significantly higher Roll rates.

Welfare Implications The observed In rates and Roll rates are far from the SPNE (Out, Don’t Roll) predictions. We test whether this trusting behavior pays off for As, given B Roll decisions. The SPNE predicted payoff for both A and B is $5, and the SPNE expected payoff to As from choosing In is $0. However, this SPNE payoff is conditional on B choosing Don’t Roll. If B Roll rates are high enough, it is possible for In to be a best response for As. This is the case when As’ expected payoff from choosing In, given Bs’ observed likelihood of choosing Roll, 33 One motivation for looking at Periods 2-5 is that chat analysis indicates some teams were confused about the repeated nature of the game. Although the instructions were clear that no teams would receive feedback on decisions, some teams still chose to Roll in Round 1 to establish a “reputation” and encourage As to go In. Once they realized that As would not learn their decisions, they quickly adjusted their decisions toward Don’t Roll.

28

is greater than $5.34 Results from Table 6 show that choosing to go In is never an empirical best response for A teams, regardless of message received, but is an empirical best response for A individuals after receiving a Strong Promise.

Individuals

Teams

Strong Promise

$5.60

$2.70

Weak Promise

$3.90

$3.30

Empty Talk

$3.80

$2.00

No Message

$4.30

$1.80

Table 6: A’s expected payoff from choosing In conditional on receiving a given type of message, given B observed Roll rates. A teams’ trust doesn’t pay off for them, but it certainly raises the expected payoff of B teams. Bs’ payoff depends on whether B chooses Roll or Don’t Roll, but A choosing In unambiguously raises B’s expected payoff. The expected payoffs for B, given A In rates, are given in Table 7, conditional on B choosing Roll or Don’t Roll.35

Conditional on Roll

Conditional on Don’t Roll

Individuals

Teams

Individuals

Teams

Strong Promise

$8.55

$8.00

$11.39

$10.40

Weak Promise

$8.45

$8.40

$11.21

$11.12

Empty Talk

$5.85

$6.05

$6.53

$6.89

No Message

$6.80

$6.95

$8.24

$8.51

Table 7: B’s expected payoff, given observed A In rates. 34 We calculate this simply by taking A’s expected payoff from going In given Chance, and then multiplying it by the probability he receives that expected payoff, which is $10*pr(Roll). These expected payoffs are determined by B roll rates, so any significant differences in B roll rates conditional on different message types will carry over into significance in A expected payoffs. 35 We calculate the expected payoff conditional on Roll as $10*pr(In) + $5*pr(Out), and we calculate the expected payoff conditional on Don’t Roll as $14*pr(In) + $5*pr(Out).

29

We can see that, regardless of intended choice, B has a strong incentive to send a Strong or Weak Promise. A is much more likely to choose In after receiving a Promise than after receiving any other message, as shown earlier in Table 2, and B benefits from this higher In rate as it results in higher expected payoff. In the same way, sending a Promise is welfareenhancing. Regardless of B’s subsequent decision, A choosing In leads to weakly higher total surplus than A staying Out.

6 6.1

Chat Analysis Selection of Chat Categories

Overall, the results conclude A teams are just as cooperative as A individuals under communication. On the other hand, B teams are generally less cooperative than individuals, and B teams are much less likely to follow through on their promises. We conduct exploratory analysis on within-team chat to further understand the teams’ decision processes. The chat content is very rich, but this poses a challenge for objective quantification. Toward this goal, we identified relevant categories for better understanding behavior, focusing on categories to address the major hypotheses in the literature and observed patterns of behavior. We had two undergraduate coders independently read through and analyze the chat transcripts for both A and B teams.36 All categories were coded at the period-level for each team. That is, the coders determined whether each team satisfied or did not satisfy each of the categories independently across all 5 periods. The coders were instructed to make their coding history-independent, so discussions in earlier periods should not affect satisfying a category in later periods.37 The full list of categories, as well as observation frequencies and 36 Coding

instructions can be found in the Appendix. The chat analysis presented will focus on teams in the communication treatment. Coded chat from teams in the no communication treatment is forthcoming. 37 We do this for a few reasons. First, while discussions are certainly correlated across Periods within a team, history-independent coding should reduce this, especially for A teams who receive new messages each period. Second, history-dependent coding leads to coders making inferences based on teams’ observed decisions. For example, an A team might discuss in Period 1 that they will choose In if the message is believable, and choose Out if the message is not believable. Coders will therefore infer the team’s perception of the message based on the team’s action, even if the team does not chat about the message. We want to focus only on what teams actually say, and we avoid drawing conclusions outside of this.

30

coder agreement rates can be found in Tables 8 and 9.38 A sample of chats in each coding category can be found in the Appendix.39 As our coders didn’t always agree on chat classifications, in our analysis we will use classifications where at least one coder coded the category in question. Our coders reached relatively high agreement on the more objective categories (e.g. A1, B5, and B6), and disagreements arose mostly in the subjective categories. We acknowledge that these agreement rates are fairly low in some categories, but we see this as to be expected given the nature of the topic at hand. Risk preferences and guilt, for example, are rather subtle, and chat classification is necessarily subjective. We use multiple coders from the subject population in order to capture a fairly representative picture of the conversations and their interpretation. Our A team categories primarily focused on the team’s interpretation of messages. An argument in the literature for why individuals keep their promises is because a promise raises the expectations of the recipient, and thus instills expectation-based guilt in the promisor. Our message discussion category captures whether A teams even consider the messages they receive, and the believe/disbelieve message categories capture a team’s perceived credibility of the message.40 Together, these categories roughly capture how A teams interpret the messages they receive. We find that 89% of A teams do discuss the message. While this doesn’t necessarily capture whether As find the message to be valuable in determining their expectations, it does inform us as to whether the groups consider the message in their decision-making processes. We find that 39% of teams give a reason as to why they would believe the content of the message, and 46% of teams give a reason why they should not believe the message. The reasons given for believing the message generally focus on the idea that people think it’s important to keep their word, and only relate to very specific messages. A groups 38 The percentage of teams satisfying a category is calculated as the number of teams either coder coded as satisfying the category in any Period, divided by the total number of teams. Agreement rates were calculated as the number of instances where both coders classified a chat message as satisfying a given category, divided by the number of instances where either coder classified chat in that category. 39 A full report of chats and chat codings is available upon request. 40 Note, some of the categorized discussions surrounding messages revolve around hypothetical messages or counterfactuals. For example, some teams discuss what types of messages they would believe before they receive a message at all, and some teams discuss the significance behind receiving a blank message. We include this as message discussion, as they show the value As place in the communication overall.

31

considering message content seem to strongly value messages with words like ‘promise’ and ‘we swear’ in them, and they point out these indicators as reasons to believe the message. For example, the following group makes a contingent plan to choose their action based on the content of the message, and will only choose to go In after receiving a strongly-worded promise41 –

4: So, should we just go out every time unless they send us a message saying they swear or something? 20: yeah, lets just see based on context On the other hand, many A groups recognize that the messages from B groups might be untruthful. This explicit mention of a reason not to believe any message may negate promissory content in the message. It could be that these teams might not believe any message, regardless of what the B team communicates. For example,

5: I feel like B has the power because of the message because they could say that they will choose roll so we choose in and then totally screw us and choose don’t roll 3: yeah I was just thinking the same 5: lets go out for the rest of the time As trust games are inherently “risky” given that the trustworthiness of the other team is unknown, we code for instances where a team mentions their willingness to take a chance or not. 54% of A teams mention being willing to take a chance and risk going In to get a higher reward, while 43% of teams state that they’re not willing to risk going In. We cannot directly compare to individuals’ risk preferences, but it appears that more teams are willing to take a chance than want to stay “safe.” We also code for strategic reasoning in the instances where a team considers what they would do if they were in the other team’s shoes. This reasoning demonstrates higher strategic thinking, and we predict that it would bring teams closer to Nash equilibrium predictions. The literature on teams has shown teams to be more strategic than individuals, so we want to see whether teams do think strategically, and if this strategy reduces trusting behavior. 41 All chat conversations are transcribed word-for-word. Subject numbers precede text in all transcribed conversations.

32

36% of teams mention what they would do if they were in B’s shoes, most saying that they would choose Don’t Roll to get more money. We predict that this strategic reasoning will be predictive of A groups to choosing Out more often.

Coding Category

Description

Percentage of Teams

A1

Discuss actual or hypothetical message, and how it might influence their decision

89% (0.75)

A2

They give a reason as to why they should believe the content of the message (e.g. people think it’s important to keep their word)

39% (0.29)

A3

They give a reason as to why they should not believe the content of the message (e.g. the other team could just lie)

46% (0.46)

A4

Mention what they would do if they were B group

36% (0.54)

A5

They recognize that going In is risky, but they are willing to take a chance to get a higher reward

54% (0.25)

A6

They recognize that going In is risky, and they are not willing to take a chance to get a higher reward

43% (0.27)

A TEAMS

Table 8: Coding categories for A teams (agreement rates in parentheses). The B team categories focused on the reasons why the team might send and/or keep a promise. Specifically, we focus on the team’s decision to choose Roll or Don’t Roll, and the corresponding decision to send a promise or not. The literature suggests B teams might keep their promises out of guilt or a desire to keep their word, so we also consider a B team’s statements of guilt or remorse for their actions. We code the first two categories, justifying Roll or Don’t Roll, to separate the choice of action from the choice of message. We find that 61% of teams give justification for why they should choose Don’t Roll, and 25% of teams justify choosing Roll. Teams justifying the choice of Don’t Roll are primarily focused on the higher reward that they will receive. They recognize the payoff-maximizing choice and use this monetary reasoning as justification. For example,

20: Do you want to just choose don’t roll?? 33

5: Yes 5: Highest payoff no matter what On the other hand, teams justifying Roll generally focus on other-regarding preferences or fairness considerations, rather than living up to their message. No teams state the promissory content of their message as justification for wanting to choose Roll. Teams justify Roll by indicating a desire to ‘cooperate’ and give money to the other team, so decisions to choose Roll are primarily driven by other-regarding preferences. For example,

13: What are your thoughts? 16: I think we’d better choose to cooperate 13: Yeah, I agree. We get money either way, but it’s important to give back We code instances where teams acknowledge the selfishness or “immorality” of choosing Don’t Roll, as well as instances where they explicitly feel bad about their choices. Since the literature focuses on guilt, we want to see whether any of these teams seem to act in accordance with what guilt would predict.42 54% of teams acknowledge that choosing Don’t Roll is selfish or in some way is an immoral action, and 29% of teams seem to explicitly feel bad about choosing Don’t Roll or about sending an untruthful message. In accordance with the guilt aversion hypothesis, we will look at whether these are teams choosing to Roll in order to avoid suffering this guilt, or if they are choosing Don’t Roll and breaking their promises regardless of the guilt. We also code for the first proposed action among teammates, since the team literature suggests that teams offer diffusion of responsibility, so one member suggesting a morally questionable action might make the other members more willing to follow suit. 64% of teams first propose choosing Don’t Roll and 36% first propose Roll, so Don’t Roll is the predominant choice before any other team discussion occurs.

42 To be clear, “guilt” as defined in the literature is not the same thing as feeling bad. Guilt refers specifically and exclusively to expectation-based preferences, as defined in the Introduction.

34

Coding Category

Description

Percentage of Teams

B1

They give a justification for why they should choose Don’t Roll (e.g. it will make them more money no matter what)

61% (0.36)

B2

They give a justification for why they should choose Roll (e.g. it is more fair to the other team)

25% (0.25)

B3

Acknowledge the selfishness of choosing Don’t Roll

54% (0.48)

B4

Explicitly feel bad about choosing Don’t Roll or sending an untruthful message

29% (0.50)

B5

First proposed action was Roll

39% (0.82)

B6

First proposed action was Don’t Roll

64% (0.56)

B TEAMS

Table 9: Coding categories for B teams (agreement rates in parentheses).

6.2

A Team Chats

Through the A team chats, we look to see how the A teams perceive the messages they receive and how the ‘riskiness’ of the environment affects their decisions. There are only 3 groups who never discuss the message received, so they are irrelevant to the guilt hypothesis. Expectation-based guilt suggests that a message raises or lowers the recipient’s expectations, and these 3 groups make their decisions on something wholly separate from their expectations.

43

43 These groups almost always choose to stay Out. 2 of these 3 groups mention not wanting risk, so their decision to stay Out is primarily driven by unwillingness to take a chance rather than a response to the message.

35

Figure 3: In Rate Conditional on Message Discussion Teams that discuss the message in a given period are much more responsive to the message content. Teams that don’t discuss a message in a given period go In 47% of the time after receiving an Empty Talk message compared to 53% of the time after receiving a Promise, and the difference is not significant (rank-sum p=0.68).44 On the other hand, teams who do discuss a message in a given period are more than twice as likely to go In after receiving a Promise than after receiving an Empty Talk message (p=0.001).45 It is clear that A teams are responsive to the message content in periods where they analyze and consider the message. Furthermore, A teams that discuss the message perceive a significant difference between Promises and Empty Talk, and choose their actions accordingly. For these teams, it seems as though Promises do raise their expectations and increase their subjective expected profit from choosing In. However, some teams don’t discuss the message and therefore this difference between Promise and Empty Talk does not influence their decisions.46 44 We combine Strong and Weak Promises, and Empty Talk and No Messages, for simplicity, but results still hold when messages are separated. 45 These results are not driven by a higher tendency to discuss Promises compared to Empty Talk. 46 Typically, the teams that don’t discuss the message have already decided on what action they will take. Their decisions are usually based on risk preferences or a pre-determined randomization strategy.

36

Next, we look to the decisions of teams who give a reason to believe or a reason not to believe the messages. Out of our 28 teams, 4 give only a reason to believe the message, 6 give only a reason not to believe the message, and 7 give both a reason to believe and a reason not to believe the message received. For teams that never give a reason to believe the message or not, their likelihood of choosing In is the same regardless of message type received. However, teams that give a reason to believe a message are much more likely to choose In after receiving a promise than after receiving an Empty Talk message. These teams strongly value the strong promissory language in certain messages, and therefore are much more likely to go In after receiving a Promise. This is verified via regression analysis in Table 10. The higher strategic reasoning and greater rationality of teams might suggest that they would be less likely to choose In than individuals, but we find no significant difference between the In rates of teams and individuals. We look to explain this through the A teams’ chat conversations, and we find that decisions to choose In are primarily driven by a team’s willingness to take a chance to make more money. This fits with the “risky shift” hypothesis that, in some situations, teams are more willing to enter into risky situations. Of course we cannot directly test the difference between teams and individuals, since we cannot collect chat data for individuals, but the results are consistent with risky behavior canceling out any mitigating effects from higher strategic reasoning.

37

Figure 4: In Rate Conditional on Willingness to Take a Chance We can see that, in a period where a team does not mention anything about willingness to take a chance, the team is significantly more likely to go In after receiving a Promise (p=0.001). Teams that mention being willing to take a risk to earn a higher reward choose to go In 60% of the time after receiving an Empty Talk message and go In 91% of the time after receiving a Promise. These teams have a much higher baseline rate of choosing In as a result of their willingness to take a chance, but they are still somewhat responsive to message content (p=0.15). On the other hand, teams who mention not being willing to take a risk choose to go In at a much lower rate regardless of message, 50% after receiving an Empty Talk message and 33% after receiving a Promise, and this difference is not statistically significant (p=0.67). Thus, it seems that stated preference for risk increases a team’s underlying propensity to go In, but does not completely eliminate the effect of promises.47 47 In forthcoming results, we will compare these results with coded risk preferences in the no communication treatment. Since teams are significantly less trusting than individuals in our no communication treatment, we expect to see a reduced effect of risk preferences under no communication. We hypothesize that risk preferences are more predictive of behavior in the communication treatments, which can, in part, explain the equal trust levels between teams and individuals.

38

We run random-effects probit regressions, predicting the teams’ decision to choose In. Errors are clustered at the team level, and the message coefficients are relative to receiving No Message. The regression confirms that much of the predictive power of teams choosing In comes from the message content and the team’s perception of the message, as well as team risk preferences. VARIABLES

Choosing In

Strong Promise

0.778 (0.559)

Weak Promise

1.340** (0.597)

Empty Talk

-1.111 (0.803)

Period

-0.138 (0.126)

Discussed Message

0.308 (0.335)

Reason to Believe Message

1.625*** (0.539)

Reason not to Believe Message

-1.472*** (0.515)

Put Self in B’s Shoes

0.119 (0.523)

Willing to Take a Risk

1.568** (0.621)

Unwilling to Take a Risk

-0.614 (0.604)

Constant

-0.253 (0.650)

Observations

140

Number of Clusters

28

Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Table 10: A Probit with Chat Regressors 39

The coefficient on Reason to Believe Message is positive and significant, confirming that teams are significantly more likely to choose In after stating a reason to believe the message, conditional on message content received. The coefficient on Reason not the Believe the Message is negative and significant, confirming that teams are much less likely to go In when they distrust the B teams. Additionally, confirming the above statistics, a team stating their willingness to take a chance makes them much more likely to choose In.

6.3

B Team Chats

One of our observations from the chat is that B teams very rarely ever express guilt, preference for commitment, or consideration of A teams’ expectations.48 Only 8 teams ever explicitly mention feeling bad about choosing Don’t Roll or sending an untruthful message, and only 1 of those 8 teams chooses Roll in that period.49 Additionally, out of the 7 teams who feel bad but still choose Don’t Roll, 5 of them send a promise. Thus, it seems the teams that mention feeling bad aren’t doing so as a justification for keeping their promises. Instead, they mention feeling bad in spite of their decision to break their promise. For example, the only team to ever use the word ‘guilt’ mentions it dismissively, after sending a Promise and choosing Don’t Roll in every Period. 2: i really hope we get groups who want to take risks haha 2: does that make me a bad person? 15: yeah but me too so whatever 2: im buying chipotle with whatever money i get 15: worth the guilt From what we can observe in the team chat conversations, B teams give neither guilt nor preference for commitment as motivations for their decisions. No B teams ever discuss what the A teams expect of them, and they do not take this into account when making their decisions. In fact, the only teams to ever discuss how A will perceive their message 48 We initially coded only for instances of guilt as defined in the literature – a response to second-order expectations. However, when coding in this way, the category is empty. No teams explicitly mention anything resembling expectation-based guilt as the literature suggests, or chat about anything like secondorder beliefs. 49 Teams mention “feeling bad” primarily about choosing Don’t Roll. Once they have decided to choose Don’t Roll, the choice of message is mostly irrelevant to them. Their primary source of “guilt” is in terms of other-regarding preferences.

40

do so with the intention of raising A’s expectations. They have already decided to choose Don’t Roll, and they send their message to persuade A to choose In so they can make more money. For example: 5: lets tell them to go in, than we dont roll again. 7: Sounds good to me Message sent: Go in, and we’ll roll. Id rather have 10 than 5 with you guys going out every time. 7: Well done. I’d buy that 5: i try 7: How about the same thing next time? 5: sounds like a plan In addition to the content of communication, chat analysis reveals an interesting observation about the timing of B teams’ decisions. Recall that in the structure of the experiment, B teams first write and send their message to the A teams, and then make their decision to Roll or Don’t Roll. However, conversations reveal that B teams first decide on the action they will take, Roll or Don’t Roll, and then discuss what message to send.50 So rather than Bs’ decisions being informed by and a response to their messages, their messages are conditional on their decisions. B teams decide FIRST on their action and then on their message, and never change the action they’ve agreed upon after sending the message. Dialogues of the following sort are very common: 2: we should definitely not roll 3: Hello! I agree 2: hello! 2: should we write them a message? 3: however, we should tell the other group that since we have an 80% shot at getting the 12/10 thats what we want For groups that decided to Roll: 13: 16: 13: 16: 13:

Hi Hey What are your thoughts? I think we’d better choose to cooperate Yeah, I agree. We get money either way, but it’s important to give back

50 We

would have included this as a chat category, but every single team with meaningful chat satisfied this timing, so it is not informative.

41

13: What should our message say? 16: A chooses in, B chooses roll, that way we’re better off as a whole Given this timing of decisions, we see little room for decisions to be made based on guilt or preference for commitment as they have been defined in the literature. Guilt would predict that subjects who have decided to choose Don’t Roll would prefer to avoid raising A’s expectations, and therefore would not send promises. Preference for commitment predicts a consistency between actions and messages, so a team that has decided to choose Don’t Roll would prefer to send a message consistent with that decision. We see neither of these explanations holding. Although messages and actions are technically determined endogenously, we can look at the distribution of messages sent, conditional on action chosen, rather than looking at actions conditional on messages as in the previous analysis. Chat data reveals this is how teams actually make their decisions, so we can see the effect of predetermined action on the type of messages sent. Figure 5 shows that the distribution of messages sent, conditional on action, is almost exactly the same across choosing Don’t Roll and choosing Roll.51 Therefore, although teams decide on their actions first, they ultimately send the same messages regardless of their decided action. Thus, it’s not the case that teams who have decided to choose Don’t Roll send fewer Promises to avoid breaking their commitments.52 51 Though we cannot directly test for self-selection as proposed in Ismayilov and Potters (2016), this result suggests self-similarity is not driving our results, as we don’t see cooperative individuals more likely to send promises. 52 One could argue that team members are already anticipating their future promises and incorporating that disutility into their decisions before deciding to Roll or Don’t Roll. However, we don’t see evidence of this. Teams still discuss whether or not to send a message, and what the message would say. It’s not the case that teams are already assuming they will send a promise.

42

Figure 5: Message Conditional on Action We see that B teams are significantly less likely to choose Roll than B individuals in the communication treatment, and we can gain insight into this difference through the B team chat. Previous literature (see Gino et al. (2009a)) suggests that a team member’s immoral suggestion can influence the ultimate decision of the team, so we look to see the influence of the first suggestion made by either teammate. The first proposed action of a team is arguably the closest thing we can observe to individual behavior before team influences take over, so we compare the initial proposed actions within teams to the individuals’ decisions in Period 1. We had our coders document whether each team first proposed choosing Roll or Don’t Roll, and we look to see how the first proposed action affects a team’s subsequent decisions.53 64% of teams first propose choosing Don’t Roll, and 36% first propose choosing Roll in Period 1. We compare this with the B individual choices in Period 1 of the communication treatment. Individuals actually choose Roll 52% of the time, which is significantly higher than the 36% of teams who first propose Roll. 53 We only consider first proposed actions in Period 1. We instructed the coders to identify the first time either team member made a suggestion on what action to take (e.g. “I think we should choose Roll”), and this did not include asking for input ( e.g. “Do you want to choose Roll?”)

43

While we can’t definitively conclude why this difference emerges, the literature suggests a few explanations. It could be that selfish individuals are more likely to speak up in group settings, or that participants are more willing to be selfish when in a group (Song (2006)). Even so, a team that first proposed choosing Roll does not always choose to Roll. However, these teams are much more likely to actually Roll, conditional on sending a promise or an Empty Talk message (p=0.0003 for Promise and p=0.01 for Empty Talk).54 Rank-sum tests indicate that, within teams who first propose choosing Roll, there is no significant difference in Roll rates after sending a Promise versus after sending Empty Talk (p=0.80), and there is also no significant difference for teams who first propose Don’t Roll (p=0.78).

Figure 6: Roll Rate Conditonal on Message Sent, by First Proposed Action We can thus partition B teams into teams who first propose choosing Roll and those who first propose choosing Don’t Roll. This first proposed action is very predictive of the team’s eventual decision. While teams who first propose choosing Roll don’t always Roll in every round, they are much more likely to choose Roll regardless of message sent. On the other hand, teams who first propose choosing Don’t Roll are very unlikely to ever choose Roll. Conditional on first proposed action, a team is no more likely to choose Roll after sending 54 We

combine Strong and Weak Promises, and Empty Talk and No Messages, for simplicity, but results still hold when messages are separated.

44

a Promise compared to after sending an Empty Talk message. We therefore conclude that, conditional on a team’s decided action, the choice of message sent does not affect likelihood of choosing Roll. VARIABLES

Predicting Roll

Strong Promise

2.515 (2.647)

Weak Promise

3.526 (2.758)

Empty Talk

-2.042* (1.082)

Period

-0.244 (0.187)

Justify Roll

4.737* (2.541)

Justify Don’t Roll

-1.314 (0.93)

Acknowledge Selfishness

-0.494 (0.797)

Explicitly Feel Bad

-0.723 (0.815)

First Propose Roll

4.895** (2.288)

Constant

-5.299 (3.875)

Observations

135

Number of Clusters

27

Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Table 11: B Probit with Chat Regressors We run random effects probit regressions, predicting the decision to choose Roll, including all chat categories as regressors. The message coefficients are relative to sending No Message.

45

All of the predictive power in a team choosing Roll comes from a team first proposing to choose Roll, or justifying a reason to do so. Feeling bad has a negative effect on the likelihood of choosing Roll, and sending a Promise has no significant impact. Thus, after controlling for a team’s decided action, sending a promise has no significant impact on the team’s decision to Roll.

7

Discussion

This paper studies team versus individual play in a trust game with pre-play communication. We replicate previous findings that communication increases cooperation for individuals and that teams are less trusting than individuals absent communication. We provide new evidence on the effect of communication with team decision makers. We find that teams and individuals respond to non-binding communication in the same way, but teams are much less likely than individuals to follow through after making a commitment. In general, non-binding commitments increase cooperation for individual decision makers, but have a negligible effect for teams. Chat analysis allows us to gain insight into the decision-making processes of teams. We see no spoken evidence of expectation-based guilt or preference for commitment as defined in the literature. While this could mean guilt totally dissipates in team settings, we suspect other factors may be in effect. First, it could be that guilt is overshadowed by diffusion of responsibility, image concerns, or other side-effects of being in a team. This would imply that guilt is still a concern for teams, but is not a large enough concern to warrant voicing. Second, it could be that team members avoid speaking of guilt to avoid making immorality salient. Acknowledging the source of disutility only serves to increase its strength. We don’t know of any studies looking at whether individuals prefer to ignore discussing immoral actions as they engage in them. Future work is needed to answer these questions. As with most studies that rely on team chat, our work would greatly benefit from novel experimental methods on individual decision-making. To-date, researchers are unable to view individual decision making processes in a manner analogous to team chat, though

46

progress has been made in this direction (e.g. Burchardi and Penczynski (2014)). Methods that would allow us to analyze individual reasoning as we can analyze team chat would give more insight into the differences between teams and individuals. For example, we see that teams decide on their actions before deciding on their messages, even when decisions are presented in the reverse order. We cannot say whether individuals do the same, and whether an individual’s subsequent decisions are ever influenced by earlier ones. Our design used team and individual decision makers, but we always had teams matched with teams, and individuals matched with individuals. Chat conversations reveal that teams playing against teams do think of the other participants as a “team” rather than as an individual. It’s unclear whether teams would behave the same if matched with individuals, rather than other teams.

47

References Bateson, N. (1966). “Familiarization, Group Discussion, and Risk Taking”. In: Journal of Experimental Social Psychology. Bhattacharya, Puja and Arjun Sengupta (2016). “Promises and Guilt”. In: Working Paper. Burchardi, Konrad and Stefan Penczynski (2014). “Out of Your Mind: Eliciting Individual Reasoning in One Shot Games”. In: Games and Economic Behavior. Charness, Gary and Martin Dufwenberg (2006). “Promises and Partnership”. In: Econometrica. Charness, Gary and Matthias Sutter (2012). “Groups Make Better Self-Interested Decisions”. In: Journal of Economic Perspectives. Charness, Gary, Edi Karni, and Dan Levin (2007). “Individual and Group Decision Making under Risk: An Experimental Study of Bayesian Updating and Violations of First-Order Stochastic Dominance”. In: Journal of Risk and Uncertainty. — (2010). “On the Conjunction Fallacy in Probability Judgment: New Experimental Evidence Regarding Linda”. In: Games and Economic Behavior. Cohen, Taya et al. (2009). “Do Groups Lie More than Individuals? honesty and Deception as a Function of Strategic Self-Interest”. In: Journal of Experimental Social Psychology. Collins, B. and H. Guetzkow (1964). A Social Psychology of Group Processes for DecisionMaking. Ederer, Florian and Alexander Stremitzer (2014). “Promises and Expectations”. In: Working Paper. Ellingsen, Tore et al. (2010). “Testing Guilt Aversion”. In: Games and Economic Behavior. Falk, Armin and Florian Zimmermann (2011). “Preferences for Consistency”. In: IZA Discussion Paper No. 5840. Fischbacher, Urs (2007). “z-Tree: Zurich Toolbox for Ready-Made Economic Experiments”. In: Experimental Economics.

48

Gino, Francesca and Adam Galinsky (2012). “Vicarious Dishonesty: When Psychological Closeness Creates Distance from One’s Moral Compass”. In: Organizational Behavior and Human Decision Processes. Gino, Francesca, Shahar Ayal, and Dan Ariely (2009a). “Contagion and Differentiation in Unethical Behavior: The Effect of One Bad Apple on the Barrel”. In: Psychological Science. Gino, Francesca, Jun Gu, and Chen-Bo Zhong (2009b). “Contagion or Restitution? When Bad Apples can Motivate Ethical Behavior”. In: Journal of Experimental Social Psychology. Gino, Francesca, Shahar Ayal, and Dan Ariely (2013). “Self-Serving Altrusm? The Lure of Unethical Actions that Benefit Others”. In: Journal of Economic Behavior and Organization. Greiner, B (2004). An Online Recruitment System for Economics Experiments. Hernandez-Lagos, Pablo (2015). “The Context Determines Who Leads: Cooperative Initiative through Pre-Play Communication in One-Shot Games”. In: Working Paper. Houser, Daniel and Erte Xiao (2010). “Classification of Natural Language Messages Using a Coordination Game”. In: Experimental Economics. Ismayilov, Huseyn and Jan Potters (2016). “Why do Promises Affect Trustworthiness, or Do They?” In: Experimental Economics. Kerr, Norbert L., Robert J. MacCoun, and Geoffrey P. Kramer (1996). “Bias in Judgment: Comparing Individuals and Groups”. In: Psychological Review. Khalmetski, Kiryl, Axel Ockenfels, and Peter Werner (2015). “Surprising Gifts – Theory and Laboratory Evidence”. In: Journal of Economic Theory. King-Casas, Brooks et al. (2005). “Getting to Know You. Reputation and Trust in a TwoPerson Economic Exchange”. In: Science. Kocher, Martin and Matthias Sutter (2005). “The Decision Maker Matters: Individuals versus Group Behavior in Experimental Beauty-Contest Games”. In: Economic Journal.

49

Kocher, Martin, Sabine Strauss, and Matthias Sutter (2006). “Individual or Team DecisionMaking – Causes and Consequences of Self-Selection”. In: Games and Economic Behavior. Kugler, Tamar et al. (2007). “Trust Between Individuals and Groups: Groups are Less Trusting Than Individuals but Just as Trustworthy”. In: Journal of Economic Psychology. Kugler, Tamar, Edgar Kausel, and Martin Kocher (2012). “Are Groups More Rational than Individuals? A Review of Interactive Decision Making in Groups”. In: CESifo Working Paper No. 3701. Masclet, David et al. (2009). “Group and Individual Risk Preferences: A Lottery-Choice Experiment with Self-Employed and Salaried Workers”. In: Journal of Economic Behavior & Organization. Schopler, John et al. (2001). “When Groups Are More Competitive Than Individuals: The Domain of the Discontinuity Effect”. In: Journal of Personality and Social Psychology. Song, Fei (2006). “Trust and Reciprocity Behavior and Behavioral Forecasts: Individuals Versus Group-Representatives”. In: Games and Economic Behavior. Sutter, Matthias (2009). “Deception Through Telling the Truth?! Experimental Evidence from Individuals and Teams”. In: The Economic Journal. Vanberg, Christoph (2008). “Why Do People Keep Their Promises? An Experimental Test of Two Explanations”. In: Econometrica. Wallach, M., A. Kogan, and D. Bem (1964). “Diffusion of Responsibility and Level of Risk Taking in Groups”. In: Journal of Abnormal and Social Psychology. Wiltermuth, Scott (2011). “Cheating More when the Spoils are Split”. In: Organizational Behavior and Human Decision Processes.

50

Appendix A’s In Rate

B’s Roll Rate

Individuals

Teams

p-value

Individuals

Teams

p-value

No Communication

50%

35%

(0.086)

28%

19%

(0.102)

Communication

63%

56%

(0.357)

42%

24%

(0.043)

(0.085)

(0.027)

(0.163)

(0.379)

p-value

Table 12: Comparison of A and B teams and individuals

A Individuals

A Teams

B Individuals

B Teams

Choosing In

Choosing In

Choosing Roll

Choosing Roll

0.436

0.797**

0.759

0.788

(0.315)

(0.353)

(0.544)

(0.711)

-0.166***

-0.0798

-0.251***

-0.215

(0.060)

(0.0673)

(0.062)

(0.150)

0.485

-0.348

-0.386

-2.968

(0.284)

(0.347)

(0.433)

(4.890)

Observations

275

275

275

275

Number of Clusters

55

55

55

55

Communication

Period

Constant

Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Table 13: Effect of Communication

51

Strong Promise

Weak Promise

Empty Talk

No Message

In

In

In

In

-0.689

-0.290

-0.415

-0.054

(0.634)

(0.723)

(1.781)

(0.713)

-0.125

0.157

-0.714**

-0.432*

(0.129)

(0.225)

(0.305)

(0.226)

1.628**

0.579

0.196

0.629

(0.686)

(0.897) )

(1.967)

(0.683)

Observations

154

57

15

59

Number of GroupIdentifier

39

19

8

23

VARIABLES Team Period Constant

Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Table 14: Probit regressions predicting A’s decision by type of message received

Figure 7: Distribution of In rates

52

Strong Promise

Weak Promise

Empty Talk

No Message

Roll

Roll

Roll

Roll

-3.969***

-0.325

-0.938

-1.694

(1.075)

(0.957)

(0.779)

(1.354)

-0.378*

-0.137

-0.315*

-0.429***

(0.202)

(0.221)

(0.188)

(0.143)

1.001

-0.214

0.411

0.478

(1.167)

(0.925)

(0.885)

(0.875)

Observations

154

57

15

59

Number of GroupIdentifier

39

19

8

23

VARIABLES Team Period Constant

Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Table 15: Probit regressions predicting B’s decision by type of message sent

Figure 8: Distribution of Roll rates

53

Examples of chats in each category –

A1: Discussed message 28: alright, so what’s the plan? 10: But let’s see what they say. If they say they’re going to roll, we go in. if they don’t say anything I think we go out. How does that sound? A2: Reason to believe the message (chat occurred right after receiving a message) 17: not very convincing, dang you think they would try and really sell us 19: i agree 17: better than the first one tho 19: true 17: if they would have typed “we promise” afterwards, i’d give it to em A3: Reason not to believe the message 16: assuming group B sends a positive message i think we should vote in 25: sure 25: they might just send a positive message and choose not roll 16: you think they would do that? 25: yea 25: it makes them 14 A4: What they would do if they were B 12: out the next two? 18: If I were a B, I don’t roll 18: yeah 12: That’s what Im saying A5: Willing to risk 18: one more in? 12: maybe two. you have to risk it for the biscuit sometimes haha A6: Not willing to risk 20: Do we keep going as is? Or take the $5? 20: They would be stupid to roll it. They get a guaranteed $14... 4: I would have to say play it safe than. B1: Justify choosing Don’t Roll 20: Do you want to just choose don’t roll?? 5: Yes 5: Highest payoff no matter what 11: Since A chooses first, I have no problem being selfish for the last two rounds ... B2: Justify choosing Roll 6: I feel like we should roll

54

6: I know if I was in A I would want the other team to roll B3: Acknowledge selfishness 24: I almost feel bad saying we promise 24: almost 7: haha we can be terrible people and start telling groups were gonna pick roll and then pick don’t roll. get that extra 4 bucks B4: Explicitly feel bad 2: i really hope we get groups who want to take risks haha 2: does that make me a bad person? 15: yeah but me too so whatever 2: im buying chipotle with whatever money i get 15: worth the guilt B5: First propose Roll 18: hey 12: hello 18: I think we should choose roll if they go IN. 12: i agree B6: First propose Don’t Roll 12: Hi Partner 21: Hi 12: Since we are in the B group, it is in our best interest to select don’t roll for every round, but we want our A group to select in

55

INSTRUCTIONS Thank you for participating in today’s experiment. Please put away your cell phones and listen carefully as we go through the instructions. This is an experiment in the economics of decision making. The National Science Foundation and the Ohio State University have provided the funds for this research. Feel free to ask questions while we go over the instructions. Please do not speak with any other participants during the experiment and do not take out your cell phones. GROUP FORMATION 1. In this experiment, you will be asked to make decisions over several rounds. Decisions will be made by two person groups. Your group will consist of you and one other person in the room that you have been randomly paired with. The two of you will be in the SAME group throughout today’s session. There are two types of Groups involved in each round – Group “A” and Group “B.” A’s and B’s will be randomly matched together in each round (more on this below). The amount of money you will earn depends on the decisions made by your Group and the Group you’re matched with. 2. Your assignment as an A or B group will be determined randomly at the start of the session, and will be the SAME in all rounds. 3. In each round, A and B Groups will be randomly matched. The computer matches Groups in such a way that no A Group will ever be matched with the same B Group more than once, and vice versa. The Group you are matched with in the current round is called your “matched group.” This means that your matched group will be DIFFERENT in every round. 4. To help each group coordinate their decisions, you will have an on-screen chat box where you can send text back and forth to one another, like an instant messaging system. This chat box is only for you and the other member of your group -- no other participants will be able to see your chat. In communicating with each other we request that you follow two simple rules: (1) Be civil to one another and don’t use profanity, and (2) Do not identify yourself IN ANY WAY. This chat service is intended for discussion and coordination on decisions and should be used as such.

EXPERIMENT OVERVIEW

5. Group A Decision:  In each round, A’s will see a decision screen indicating whether they want to choose IN or OUT. If Group A chooses OUT, they and the B Group they have been matched with will receive $5. If A’s choose IN, the amount of money that Group A and Group B will receive

depends on what Group B chooses. (Note, any payments mentioned will go to EACH member of a group – payments are NOT shared between group members.) 6. Group B Decision: o B’s will see a decision screen where they choose ROLL or DON’T ROLL (referring to the roll of a 6-sided die). o If Group A has chosen IN and Group B chooses DON’T ROLL, Group B receives $14 and Group A receives $0. o If Group A has chosen IN and Group B chooses ROLL, Group B receives $10 and the computer will roll a six-sided die to determine Group A’s payoff.  If the die comes up 2-6, Group A receives $12.  If the die comes up 1, Group A receives $0  Note the computer will be rolling the die – that is using a random number generator to determine which number between 1 and 6 will be drawn. Each number is equally likely to be drawn. o If Group A has chosen OUT, both Group A and Group B will receive $5 regardless of whether Group B chooses ROLL or DON’T ROLL.  When Group B makes their choice, they will not know whether Group A has chosen IN or OUT (this is referred to as the “strategy” method). However, since Group B’s decision is binding only if the A they have been matched with has chosen IN, we ask B’s to presume, for the purpose of making their decision, that A has chosen IN.

Payoffs are summarized in the chart below and will be available on your computer screens for reference throughout the experiment:

7. Prior to Group A making their decision on IN or OUT, the B Group they have been matched with has the option to send a message to Group A via the computer terminal. In these messages, Group B is not allowed to identify themselves IN ANY WAY. Other than this restriction, B may send anything they wish in this message.

TIMING

8. The sequence of the experiment will work as follows: First, B Groups will send a message to the A Group they have been matched with. After receiving B’s message, A Groups will decide whether they are IN or OUT. After A’s decide IN or OUT, B Groups will decide ROLL or DON’T ROLL Now, we will go over the specifics of how you will make these decisions. While we do this, I will show you what your computer screens will look like in these different stages.

MESSAGE STAGE

9. B Groups will have 2 minutes to decide on whether they want to send a message and, if so, what the message will say. As a Group, the two of you must agree on the content of the message. So in order to do that, either group member can propose a message for the other member to review before sending to your matched A Group. [Put up message proposing screen] Both members of the B group will see this screen as they decide on their message. On the top of the screen, you see a reminder that you are a B group. Your Group ID is just a number that identifies your group. If the experimenters ever need to get your attention, we will address you by your Group ID. The top of the screen also reminds you which round it is. There is a timer in the upper right corner which tells you how many seconds are left in the stage. You have 2 minutes to decide on your message, so the timer starts at 120. (point to payoffs) This box will remind you of the payoffs. This table is the same as the one you have on your instructions sheet. (point to chat box) This is the chat box where you can communicate with the other person in your group. To send something, first click your mouse into this darker box at the bottom and then type. You have to click “Enter” on the keyboard for your chat to send. Remember, this chat is only for you and the other member of your group. Members of your matched A group cannot see this chat. (point to message proposing box) You can propose a message by typing it in this box. You must click this “Propose” button for the message to be sent to the other member in your group for review. [Put up message accept/reject screen] If the other member of your group proposes a message, you will see the message here along with these two buttons – “Accept and Send” and “Reject and Rewrite.”

If you are in agreement with the proposed message, click the “accept” button, and it will be sent to your matched A Group at the end of the message stage. If you don’t agree, you can click “reject” and propose another message. Each group member only gets one opportunity to propose a message. If the group members cannot agree on a message within the 2 minutes, one of the members will be randomly selected and he will have 30 seconds to write a message on behalf of the Group. The chat box will be closed during this 30 second period. If B’s do not want to send any message, just type “No Message” or leave the message space blank. Note that messages from B’s will only be delivered after ALL B Groups have written their messages. While B’s decide on their messages, A’s will be able to chat with each other. [Put up A message stage screen] A groups will see this screen while B groups are deciding

on their messages. On the top of the screen, you see a reminder that you are an A group, as well as your Group ID, round number, and timer. (point to payoffs) Here is the payoff table. It is the same as in your instructions. (point to message box) The message from the B group will be displayed in this box once it is sent. For now, it just reminds you that B groups are deciding on their messages. (point to chat box) Here is your chat box where you can communicate with the other person in you’re a group. Remember, you have to click “Enter” on the keyboard for you chat to send. DECISION STAGES 10. Group A: Once A’s receive a message they will have 1 one more minute to decide between IN or OUT. [Put up A decision screen] This is what A groups will see while they make their decisions. (Point to message box) If you matched B group has sent a message, it will show up here. You can use the chat box to coordinate your decision with the other member of your group, and you input your decision here. (Point to decision box) You will indicate your decision where it says “My Decision.” When the other member of your group makes a choice and clicks one of these options, you will see that button fill in under the “Partner’s Decision” column. Once you and the other member of you group have clicked on the same decision, your decision will be automatically accepted.

If you cannot come to agreement during the time allotted, the chat box will close and one member will be randomly chosen to make the decision on behalf of the Group.

Group B: 11. While A’s are deciding on IN or OUT, B’s will be able to chat with each other. Once ALL the A Groups have made their decisions, B’s will have an additional minute to choose whether to ROLL or DON’T ROLL. [Put up B decision screen] This is what B groups will see while they make their decisions. (Point to message box) Whatever message you sent will show up here. Remember, you won’t know whether your matched Group A has chosen IN or OUT, but you should make your decision as if they chose IN. You can use the chat box to coordinate your decision with the other member of your group, and you input your decision here. Like the A groups, you and the other member of your group must select the same decision, and then it will be automatically recorded. Once you have agreed and clicked on the same decision, your decision will automatically be recorded and your decision is binding. If your Group is unable to come to agreement on the decision within the allotted time period, one member will be randomly chosen to make the decision on behalf of the Group.

FEEDBACK & PAYMENT

12. Once all B Groups have made their choices, you will proceed immediately to the next round. You will receive NO feedback regarding the choices made in the previous round. At the end of the session you will get to see a record of your choices and the payoff you would have received in each round. 13. Your payment will be determined by ONE RANDOMLY SELECTED round. Each round you play is equally likely to be the round selected for payment. This means that you should treat each round as if your decisions in that round will directly determine your payment.

SUMMARY

1. You will be making decisions in Groups. Your Group consists of you and one other person. You will be paired with the SAME person in your group for the entire experiment. 2. Your role as either Group A or Group B will be the SAME for the entire experiment. 3. In every round, each Group A is matched with a DIFFERENT Group B (and vice versa). You will never be matched with the same Group more than once.

4. B’s will first have an opportunity to send a message to the A Group they have been matched with. 5. After these messages have been delivered, Group A will decide IN or OUT. 6. After Group A has decided on IN or OUT, Group B will decide ROLL or DON’T ROLL. 7. The experiment will consist of 5 separate rounds, and you will not receive any feedback between rounds. 8. One randomly selected round will determine your payment. In addition to these earnings, everyone will receive a $5 show-up fee.

QUESTIONS 1. Suppose you are in Group A and your group chooses OUT. Your payoff will depend on the decision made by the Group B you have been paired with. (True/False) ____________

2. If you are a Group B member and your group cannot agree on a message to send to your matched Group A, then the computer will randomly select you or your partner to write a message on behalf of your group. (True/False) ____________

3. Are B Groups required to send a message to their matched A Group? (Yes/No) ___________

4. Suppose you are in a B Group. Even though you do not know the choice of your matched A Group, your decision to Roll or Not Roll will be binding if the A Group chooses IN. (True/False) _____________

5. Suppose you are in Group A. You will be matched with the same Group B in every round. (True/False) _______________ The other member of your group will be the same person in every round. (True/False) ______________

6. Suppose you are in Group B and the A Group you have been matched with chooses IN. If your Group chooses ROLL, 5 out of 6 times you would earn $___________ and the A Group you are matched with would earn $___________. In the remaining 1 out of 6 times, you would earn $__________ and the A Group you have been matched with would earn $___________.

7. Suppose you are in a B Group and your matched A Group has chosen IN. If you choose DON’T ROLL, you will earn $__________ and the A Group you have been matched with will earn $___________.

Message Coding Instructions

We categorize a message as one of 4 categories – Strong Promise, Weak Promise, Empty Talk, or No Message. Each message will fall in exactly one of these 4 categories. So it must be either a strong promise, a weak promise, empty talk, or no message, and it cannot be categorized as more than one.

0. No Message No message is when a message is actually left blank, or if they specifically type “No Message.” If you feel that it is a no message, mark a “0” in the PromiseVariable column.

1. Strong Promise We categorize something as a strong promise if the group gives a specific statement of intent to choose the action Roll. Classify a message as a strong promise only if it is clearly a promise to choose Roll. If you feel that it is a strong promise, mark a “1” in the PromiseVariable column. Examples of Strong Promises: “We will choose ROLL” “Choose to go In and we will choose to Roll” “We will choose Roll, so you should choose in as it helps both groups.”

2. Weak Promise Weak Promise is a weaker version of a strong promise. We categorize something as a weak promise if the group perhaps alludes to Rolling or choosing a “nice action” but does not give a specific statement of intent or a strong promise. The line between weak and strong promise is not black and white, so use your best judgment. If you think the group is intending their message to promise they will choose Roll, it should be categorized as a Strong Promise (regardless of whether the message is believable or not). If you feel that it is a weak promise, mark a “2” in the PromiseVariable column. Examples of Weak Promises: “Take the chance and make the big fortune” “It would be wise to choose In”

3. Empty Talk

Empty Talk is any message which doesn’t directly refer to the game or give any degree of intent to choose an action. If you feel that it is empty talk, mark a “3” in the PromiseVariable column. Examples of Empty Talk: “Urban Meyer is bae” “Download my mixtape” “GRONK SPIKE”