Information exchange in laboratory markets: competition, transfer costs, and the emergence of reputation

Public reputation mechanisms are an effective means to limit opportunistic behavior in markets suffering from moral hazard problems. While previous research was mostly concerned with the influence of exogenous feedback mechanisms, this study considers the endogenous emergence of reputation through deliberate information sharing among actors and the role of barriers in hindering information exchange. Using a repeated investment game, we analyze the effects of competition and transfer costs on players’ willingness to share information with each other. While transfer costs are a direct cost of the information exchange, competition costs represent an indirect cost that arises when the transfer of valuable information to competitors comes at the loss of a competitive advantage. We show that barriers to information exchange not only affect the behavior of the senders of information, but also affect the ones about whom the information is shared. While the possibility of sharing information about others significantly improves trust and market efficiency, both competition and direct transfer costs diminish the positive effect by substantially reducing the level of information exchange. Players about whom the information is shared anticipate and react to the changes in the costs by behaving more or less cooperatively. For reputation building, an environment is needed that fosters the sharing of information. Reciprocity is key to understanding information exchange. Even when it is costly, information sharing is used as a way to sanction others. Electronic supplementary material The online version of this article (10.1007/s10683-020-09652-0) contains supplementary material, which is available to authorized users.


Supplementary Online Material
Information Exchange in Laboratory Markets: Competition, Transfer Costs, and the Emergence of Reputation S1. Literature Overview

S1.1 Reputation Building and Information Sharing
Functioning markets require some level of trust and trustworthiness if contracts are incomplete (Akerlof 1970). The risk of moral hazard and the ensuing low levels of trust, for example in sequential exchanges, can lead to substantial welfare losses. As argued by Schelling (1960) and Kreps and Wilson (1982), economic agents may cooperate -even if opportunities to defect existif others can observe or are informed about their behavior. In such situations, players may have an incentive to build a reputation for trustworthiness if future transactions are valuable and sufficiently likely to occur. This may result in cooperation among interaction partners with information about previous interactions serving as a credible signal.
Information about other actors' past behavior can be acquired in different ways: It can be collected in repeated interactions allowing to privately learn about the type of the transaction partner. In other situations, agents can rely on publicly available information allowing them to learn from others (Bolton et al. 2004;Huck et al. 2012). The possibility to build a public reputation may serve as a particularly strong disciplining device for market participants who may fear sanctions by others and the exclusion from the market.
Compared to privately learned information, public reputation mechanisms offer greater economic potential because they allow cooperation to be sustained in larger groups. Relatively large-scale exchange systems based on reputation systems existed before formal state specification and enforcement of contracts were in place in many parts of the world. Notable historical examples include trade between northern Africa and Europe in the Mediterranean area (Greif 2006), exchange networks in Paris before the French revolution (Hoffman et al. 1999), in the Italian Alps (Casari 2007), and in Mexican California (Clay 1997).
Market participants can benefit from a good reputation as it allows them, for instance, to charge higher prices (Deephouse 2000;Fombrun & Shanley 1990;Rindova et al. 2005), to attract more productive employees, investors, and customers (Turban & Greening 1997;Fombrun 1996), and to raise competitive barriers (Milgrom & Roberts 1982;Fombrun 1996;Abimbola & Vallaster 2007). The social benefits of reputation have also been shown in several laboratory experiments confirming that reputation mechanisms positively influence trust and efficiency in markets prone to moral hazard (Huck et al. 2012;Keser 2002;Bolton et al. 2004;Bohnet & Huck 2004;Bohnet et al. 2005).

S1.2 Reciprocity
Information sharing is an interaction, which is strongly influenced by reciprocity (Fehr & Gächter 1998;Falk & Fischbacher 2006). As shown in previous research, humans are willing to reward kind actions and to punish unkind ones, even if this is costly for them (Güth et al. 1982;Güth 1995;Camerer & Thaler 1995;Chaudhuri 2011). Accordingly, individuals may sanction defectors by informing others about their misbehavior. In the literature on information sharing in credit markets, this form of negative reciprocity is referred to as black listing (Pagano & Jappelli 1993;Brown & Zehnder 2007). White listing, on the other hand, is the sharing of information about good transaction partners with the intention to reward their good conduct (positive reciprocity). Both forms allow other market participants to choose their transaction partners more carefully. Naturally, also the sharing of information by others can evoke reciprocation. Bolton et al. (2014) document reciprocity in feedback giving on eBay (see also Bolton et al. 2011 andDiekmann et al. 2014). Abraham et al. (2016) find that trustors are more likely to transfer information about trustees who returned only small amounts (negative reciprocity, black listing). At the same time, trustors who received information from other trustors in the preceding period are significantly more likely to reward this cooperation by sharing their information in the next period. On the other hand, Gërxhani et al. (2013) find no evidence for an effect of the identifiability of the source of information (and hence the possibility to benefit from reciprocal information sharing in the following rounds) on the trustors' willingness to share information. They argue that this may be due to indirect reciprocity, that is, the general willingness to share information with others, as long as some others have shared their information in the past (Anderhub et al. 2002).

S2. Additional Descriptive Statistics
This section provides additional information about the key treatment and outcome variables in our analysis. Figure S1 shows the distribution of the direct transfer cost treatment. The variable takes values from 0-1 and was randomly varied each round except for the first two rounds in which the direct costs were fixed at 0 and 1 explaining the peaks at the two extremes of the cost distribution.
Figure S1 -Distribution of within-subject treatment variable: Direct costs of information transmission  Figure S2 shows the overall distribution of two of our central outcome measures, the amount P sent by trustors and the ROI (dashed line indicates betraying threshold). Clearly, the focal point was to return 150% of the sent amount. This results in an equal split of total earnings if the trustor sent 10 token to the trustee. Nevertheless, some variation is visible, allowing us to test for the role of the different treatment conditions in changing both trustees' and trustors' behavior. Interestingly, we find that some trustees sent back 3 times the sent amount. One possible explanation for these extraordinarily high returns may be that trustees wanted to make up for prior defections and to send strong signals of trustworthiness. Figure S2 -Distribution of sent amount P and trustees' return on investment (Q/P) Figure S3 present the distribution of the sent amount P by the different between subject treatments. Overall, we observe less variation and higher sent amounts for the two information sharing treatments (with and without competition). In all four treatments some trustors chose the outside option of sending zero tokens. Similarly, Figure S4 shows the distribution of the ROI (Q/P) for the four treatment arms. Here, in particular, the sharing, no competition treatment stands out with overall substantially higher resent amounts and less variation in the resending behavior.   Figure S5 shows the distribution of the ROI over time for the different treatment conditions. Also in this illustration, it is clearly visible that trustees in the sharing, no competition treatment returned on average the highest percent of the sent amount P. When the game approaches the 24 th round, we observe a slight endgame effect in the resending behavior in the sharing, competition treatment.  Table S1 provides further summary statistics for the outcomes considered in our analysis. Again, we present the statistics separately for the different between treatment arms. Whereas we observe various differences between the different treatment conditions, the second treatment "sharing, no competition" clearly stands out again. Players in this treatment are more willing to share information as compared to "sharing with competition". Furthermore, trustors send on average  higher amounts and trustees are less likely to betray and return overall higher amounts. The greater levels of trust result in an overall higher payoff for the players. Interestingly, while we find substantial differences in the height of the sent amount P, we do not observe any major differences in trustors' probability to send a positive amount. This suggests that in all treatments, even if trust levels are low, trustors are willing to invest a small share of their endowment into the matched trustee potentially to test them or to leave them a window to re-enter the game (which may correspond with the extraordinarily high ROI of 300% reported above).  Table S2 shows two-sample Wilcoxon rank-sum (Mann-Whitney) tests, analyzing differences in the trustworthiness and trust between the different treatment arms. Unlike the test statistics reported in section 4.1, these tests are performed on matching group level. In line with our other findings, they show higher returns on investment and higher sent amounts with information sharing in the no competition treatment arms (1). At the same time, competition significantly reduces the positive effect of voluntary information exchange on trustees' resending behavior (2) and makes trustors more likely to send higher amounts, if there is no information sharing (Prob > |z|with 0.1172 insignificant) (3).

S3. Model Variations and Extensions
In this section, we vary the main model specifications to explore further the underlying mechanisms influencing players' decision-making and to test for the robustness of our findings. In particular, extending our main models, we more directly control for direct reciprocity in influencing the outcome of the game (S3.1), estimate models without any additional controls (S3.2), separate our main models for the competition and non-competition treatments (S3.3), provide further evidence on the competition treatment (S3.4), present models with alternative outcome indicators (S3.5). and estimate additional models, which further explore the role of reciprocity in information sharing (S3.6). Table S3 and Table S4 extend the main models in the paper ( Table 2 and 4) by additionally including a dummy indicating whether the trustor was betrayed by the trustee in the past interaction prior to the current round. Since the betrayal variable is itself an outcome of our treatments, it was not included it in our main models. Trustor's information sharing is not significantly influenced by the past history of interactions with the trustee, but only by trustee's behavior in the current round. Trustworthiness and the willingness to trust (Table S3) significantly depend on trustee's behavior in the past interaction. A trustee who betrayed a trustor in the past interaction sends back on average smaller returns on investment (if given a chance), which points towards a persistence of defective behavior in the game. At the same time, trustors send on average 4.755 tokens less to trustees who have betrayed them in the past round, resulting in lower average payoffs for both player types, trustees and trustors. Importantly, even under control for this measure of direct reciprocity, the treatment effects are robust. While direct reciprocity is important in our game, it has not affected (i) the positive impact of the possibility to share information and to build a public reputation, and (ii) the negative impact of direct and indirect costs on information sharing and the outcomes of the market.  Table S5 and S6 re-estimate the main models excluding any control variables from the specification. The aim here is to estimate models which do not control for any game dynamics, which might potentially confound the effect estimates. Even if estimated without additional controls, all models remain perfectly robust.

S3.3 Separate Models for Competition vs. Non-Competition Treatments
Aside of introducing competition, the tournament mechanism also leads to an increase in income for trustors. Although, it is unlikely that the observed treatment effects are merely driven by the artificially induced jump in the incomes, we re-estimate the main models in this section separately for treatments with and without competition to show that the results for endogenous information sharing and the influence of direct costs hold for both arms of the competition treatment. Table  S7 shows the results for the main game outcomes (information sharing, ROI, send amount P, trustee payoffs, and trustor payoffs) separated for competition vs. non-competition treatments.

S3.4 Further Evidence on Competition Treatment
In this sub-section we test how different positions in the ranking affected trustor's behavior. The models in Table S8 analyze how the ranking position of trustors who participated in the tournament (competition treatment) affected their willingness to share information with others, their willingness to trust reflected in the amount sent to trustees, and their game payoffs.
We do not observe any statistically significant effect of the ranking position in the last round on information sharing behavior, i.e. all trustors in the competition treatment shared similarly little information. Trustors with a higher ranking, on the other hand, were more likely to send higher amounts P to trustees (p<0.05) and benefited of a higher overall payoff from the game.  Table S9 presents the findings of the treatment effects on trustees' and trustors' behavior in the investment game for two alternative binary trustworthiness and trust indicators: Whether the trustee betrayed the trustor by returning less than the sent amount P and whether the trustor decided to send a positive amount to the trustee in the first place (P>0). All findings remain qualitatively robust with the use of those alternative outcomes.

S3.5 Models with Alternative Outcome Indicators
Compared to a baseline without information sharing, the possibility to exchange information between trustors reduced the probability for betrayal by 18.3% and for trustors to engage in the exchange by 10.4% (not significant). This positive effect is reduced once competition is introduced leading to a significant reduction in trustworthiness and trust in the market. Likewise, an increase in direct information transfer costs leads to an increase in the betrayal risk and a reduced willingness to send a positive amount, but these effects are not statistically significant.

S3.6 Further Tests on the Role of Reciprocity in Information Sharing
Table S10 shows logit models in which the binary information sharing outcome is regressed on the different treatment conditions. Mainly, we are interested here whether costs diminish the effects of negative reciprocity on trustors' willingness to share information with each other. For this, we interact the dummy measuring if a trustor was betrayed in the previous round with the competition treatment (model b) and the direct transfer cost treatment (model c). Model a shows the baseline result from Table 3 in the main text.
While, we observe that both direct and indirect costs reduce information sharing in the baseline model (a), betraying has a positive effect. The betraying effect remains significant in the second model. The interaction with competition, on the other hand, is insignificant suggesting that players are also willing to sanction betrayers about others if this is costly for them. Also in the final model (c) the interaction between betraying and direct costs is insignificant. However, in this model the main effect of betraying is also considerably reduced, even though it still points in the expected direction.
In the model displayed in Table S11, the binary information sharing outcome is regressed on a variable which takes the value one if none of the other trustors has shared information with the trustor in any of the previous rounds. Here, we are interested how the peers' refusal to share information, even if ego shared information in the past, affects ego's willingness to share another time information. For this, we condition our sample on those cases for which the trustor has sent some information in the past, but for which this positive sharing behavior was not reciprocated by the fellow trustors. As this condition significantly reduces the sample size, we focus on simple correlations between the two main variables of interest here. 516.869 Note: Random effects (RE) logit models accounting for the hierarchical clustering of the data. Coefficients displayed as marginal probability changes calculated at the mean of all covariates. Clustered standard errors in brackets (unit of clustering: matching groups). Models control for session fixed effects. Sample restricted to treatments with information sharing. P-values: * p≤0.1, ** p≤0.05, *** p≤0.01 The results clearly confirm the findings from the main analysis. If a trustor has shared her experiences with other trustors in the past, but these did not reciprocate this behavior, then a trustor's willingness to share information is reduced by a significant 36.8%. This suggests that negative reciprocity as a behavioral motive does not only play a role for interactions between trustors and trustees, but also for the interactions among the trustors. Information sharing mechanisms are hence only functional if all members of the information sharing network are equally willing to contribute by sharing their previous experiences in the market.

S4. Translated Experimental Instructions (English) THANK YOU FOR YOUR PARTICIPATION!
Please do not talk to other participants during the experiment! Dear participants, Thank you for your participation in our experiment. In the experiment, we are interested in decisions made in groups. For the success of our study it is very important that you read the following instructions carefully.
The experiment will take about 2 hours of your time. In the experiment you can earn money whereas the total amount of your payoff depends on your decisions and the decisions of other participants. After the experiment, you will be paid individually and anonymously in cash. During the experiment we do not speak of Euro, but of token. At the end of the experiment, the token earned will be converted according to the following exchange rate: 10 token = 1.00 Euro Please take your time when reading the instructions and making your decisions. You cannot influence the duration of the experiment by making quick decisions. Also, remember that the amount of your payout depends on your choices. Before starting the experiment, you will be provided with some test questions to help you check whether you have understood the instructions. For each correctly answered test question you get 1 extra token. At no time during the experiment will the identity of the participants and your payoffs be revealed. Your anonymity towards the researchers and the other participants is guaranteed.
If you have any questions while reading the instructions, please raise your hand. One of the assistants will come to you and answer your question privately. Please do not communicate with other participants and ask your questions quietly. Communication between the participants leads to exclusion from the experiment. After the start of the experiment no further questions may be asked

Study procedure
At the beginning of the experiment, you are randomly assigned to a group of 9 participants (hereafter we speak of players). Each player is given a unique ID and one of two possible roles: Role A or Role B. Each of the groups consists of 3 players with the role A (A-player) and 6 players with the role B (B-player). The group composition, the role distribution and the IDs of the participants do not change throughout the entire experiment: You interact with the same participants and keep your individually assigned role. The experiment lasts at least for 24 rounds. After the 24 th round it ends with a probability of 1/2 (= 50%) in each additional round. Each of the minimum 24 identical rounds consists of different parts, which are presented below Part 1: In each round, each of the 3 A-players is randomly assigned to a B-player, with whom she/he interacts in the round. At the beginning of the round, the A-player is informed about the identity of the assigned B-player (B1, B2, B3, B4, B5 or B6) and the B-player learns the identity of the assigned A-player (A1, A2 or A3). The 3 B-players, who are not assigned to an A-player, pause for one round. Instead of interacting with an A-player, the pausing B-players are shown arithmetic tasks they can solve. To solve the tasks, the B-players have 30 seconds. They receive 3 additional tokens for each solved task. The following parts of the experiment refer only to A and B-players who interact with each other. payoffs in tokens over all previous rounds receives the first rank. The A-player with the second highest sum is ranked second. The A-player with the lowest sum is ranked third. If two A-players have acquired the same sum over all rounds, they are assigned the same rank. After each eight rounds (after the end of the 8th, the 16th and the 24th round), the A-player who is ranked 1st will receive a bonus of 100 tokens (~ 10 €) in addition to his previous payoffs. The player on the second rank receives a bonus of 50 tokens (~ 5 €). The player on the third rank does not receive any bonus. After the 8th, 16th and 24th round, a new ranking list is created for which only the sum of the payoffs from the 8th and 16th round on are considered (without the bonus payments received). The A-players are shown only the sum of the payoffs of the previous rounds since the last bonus payout. For the B-players, no ranking list and bonus payments exist.

The following instructions are provided only in information sharing treatments
Part 5: At the end of each round, A-players can share their experiences from this round with other A-players. The A-players can decide whether they want to share their experiences with none, one or both other A-players. Sharing information with one of the other two A-players costs the communicating A-player an amount C that lies between zero points and one point (0 ≤ C ≤ 1). The amount changes randomly in each round and is communicated to all A-players and B-players at the beginning of the round. If an A-player passes on information to both other A-players, the doubled amount of token (2C) is deducted from his payoff in this round. When information is shared, the receiving A-player(s) receive the following information in the next round: 1. ID of the B-player the information-sharing A-player has played with in the last round (B1-B6) 2. What amount (X) the A-player has sent and the level of the amount after tripling (3X) 3. What amount (Y) the assigned B-player has sent back.
The information shared by other A-players is displayed to the A-players in form of two history tables (a table for each other A-player) in the first part of the next round (see screenshot next page). Each row of the tables represents the transferred information from one of the previous rounds. If no information was shared, this is displayed in the table with the words "no info". Note: Sharing information with another A-Player does not mean that this A-Player also wants to share information with you. An information exchange between B-players is not possible.

Overview of information that was shared by other A-players in the previous rounds
After the 5th part, the subsequent of at least 24 rounds of the experiment starts. During the individual rounds of the experiment, you are repeatedly asked about your expectations regarding the behavior of your interaction partner. If your expectations match the actual behavior of your fellow players, you will receive an additional payout at the end of the experiment.
Take your time to read the instructions again. If you have a question, please raise your hand. On the following page you will find an overview of the instructions. Please keep it in front of you during the experiment.

OVERVIEW OF INSTRUCTIONS
Please keep this overview always in front of you during the entire experiment In total, 9 player per group, 3 A-players and 6 B-players At least 24 rounds. Each round consists of 5 parts After the 24 th round the experiment ends with probability 50% 1. Part -A and B players are randomly matched and are informed with whom they interact in the round -A-players receive information shared by others in previous rounds (only in sharing treatment) -Unassigned B-players solve arithmetic tasks and get 3 tokens per correct solution 2. Part -A-players receive 10 tokens per round as initial endowment -A-players can send amount X to assigned B-players (0≤X≤10).
-The amount X is tripled on the way to B (3X) 3. Part -B-players can send amount Y back to A-player. Y not greater than 3X (0≤Y≤3X) 4. Part -A and B-players learn the game history and their payoffs as well as the sum of the payoffs from all previous rounds -A-players are informed of their ranking and the ranking of the other A-players -The rank depends on the sum of the payoffs from the previous rounds -After the 8 th , 16 th , and 24 th round, the ranking will start again. The first ranked A-player receives a bonus of 100 token, the second ranked a bonus of 50 token (only in competition treatment) 5. Part -A-players can share their experiences from this round with other A-players (only in sharing treatment) -Sharing information costs an amount C, which varies between 0 and 1 token in each round -The selected A-players are informed with who the information sending A-player interacted with in the last round, which amount (X) was sent, and how much was returned (Y) Wenn Sie beim Durchgehen der Anleitung Fragen haben, heben Sie bitte die Hand. Einer der Studienleiter wird dann zu Ihnen kommen und Ihre Frage privat beantworten. Bitte tauschen Sie sich nicht mit anderen TeilnehmerInnen aus und stellen Sie Ihre Fragen leise. Kommunikation zwischen den TeilnehmerInnen führt zum Ausschluss vom Experiment. Nach Beginn des Experiments dürfen keine weiteren Fragen gestellt werden.

S6. Test Questions for Participants
Please answer the following questions. You earn 1 token for each correctly answered question.
Question Answer 1. At the beginning of the experiment all participants are assigned to groups. a) How many A-players does each group consist of? 3 b) How many rounds are at least played in the experiment? 24 c) Does the player composition of the groups change throughout the experiment? No 2. You are playing as an A-player. In this round you send 0 tokens to your assigned B-player. a) What is your payoff in this round? 10 b) What is the payoff of the B-player in this round? 0 3 You are playing as a B-player. You were randomly matched with an A-player in this round. The A-player sends you 8 tokens. The sent amount (X) is increased to 24 token. From these 24 tokens you return 16 tokens to the A-player. a) By which factor is the sent amount multiplied before reaching the B-player? 3 b) What is the payoff of the A-player in this round? 18 c) What is your payoff in this round? 8 d) What would be the payoff of the A-player if you would send back nothing? 2