Against All Odds: Tentative Steps Toward Efficient Information Sharing in Groups

When groups face difficult problems, the voice of experts may be lost in the noise of others' contributions. We present results from a 'naturally noisy' setting, a large first-year undergraduate class, in which the 'expert's voice is lost' to such a degree that it is in fact optimal for all non-experts to contribute their bits of information. A single individual has little chance to improve the outcome and coordinating with the whole group is impossible. In this setting, we examine the change in behavior before and after people can talk to their neighbors. We find that the number of people who reduce noise by holding back their information strongly and significantly increases.


Introduction
People have varying degrees of expertise and views of the facts. Nevertheless, they often need to come to a decision without the help of some leader who imposes her view. Is the developed product ready to be launched? Should the UK leave the EU? While people are free to discuss facts and what to make of them with a colleague or a friend, the joint time of all involved is a scarce resource, especially if the group is large. Not every team member is asked to elaborate on the chances of a successful launch. Asking every UK citizen to describe the perceived benets and costs of leaving the EU and then agreeing on what these contributions imply seems absurd. Instead, communication is restrictedin the extreme case to an expression of opinion by a vote on a proposal that reduces potentially rich information to the rather coarse`yea,' nay,' or`nil.' When group communication is restricted, holding back less valuable information can become optimala notion that has been formalized by Feddersen and Pesendorfer (1998) as the`swing voter's curse'. Talking about a minor detail in a team meeting may bury potentially important contributions of others.
Voting on whether to leave the EU based on one's gut feeling dilutes the voices of those who correctly foresee the implications and consequences of such a step.
Individuals do not always realize that withholding information may improve the group's decision. In practice, people frequently talk or vote without much to go on. For such groups to improve information aggregation, individuals would have to learn about the benets of restraining themselves and let the expert be heard.
While direct communication can help small groups coordinating on better outcomes, (for examples, see Cooper et al., 1992;Charness, 2000;Duy and Feltovich, 2002;Blume and Ortmann, 2007) such coordination may be ham-while not being able to reach out to one another and coordinate because they are randomly dispersed across the auditorium. In an economic laboratory that seats 20-30 people, this would have been very hard to do. Second, while subjects in the laboratory seem to be aware of the benets of holding back to such a degree that Morton and Tyran (2011) speak of a norm of`letting the experts decide', our experience suggested that participants in a large lecture might not. Third, the classroom setting allows us to have`rich' local communication, while that on the group level is restricted to voting: students sit next to those whom they are likely to know and trust, but they decide together with other students who are dispersed throughout the room.
One out of the 36 group members receives expert knowledge: she is perfectly informed about the most frequent color in an urn with blue and green balls. For the other 35, we draw balls from the urn without replacement and show them the color of their individual draw on a mobile device. In order to introduce some uncertainty about the value of one's own information, one of the 36 individuals is randomly excluded from the group. The perfectly informed individual is thus not certainly but very likely remaining in the group. Every remaining individual can then vote for or against the color that they have seen or abstain.
If the majority votes for the right color, all in the group receive 10 e. Otherwise, they get nothing.
This game features an all-vote equilibrium, where everybody votes for their information, but also an only-expert-votes equilibrium, in which everyone but the expert abstains. We parameterize it so that the only-expert-votes equilibrium is substantially more ecient: the probability of identifying the correct state is more than 30 percentage points higher than when all vote. Our game is similar to that by Morton and Tyran (2011), so that our experiment can be regarded as an extension and robustness check bringing in the above features.
As we expected and hoped for, participants initially overwhelmingly vote for their own color (78.1%). This is not only a coordination problem. In an post-experimental questionnaire, we ask participants the hypothetical question: If you would interact with a group of robots, how would you program them?
More than half of the people program the robots to always vote and only a quarter to play the ecient strategy. This suggests that they are unaware of the ecient strategy. Moreover, participants are more likely to vote against their color (15.5%) than to abstain (6.4%). Given this behavior, voting one's own color is optimal.
Taking inspiration from Cason and Plott (2014), who nd that local communication among students reduces misconceptions in an individual decision task, we give students ve minutes to talk to their immediate neighbors before repeating our interactive decision task with a newly lled urn and in new teams.
While individual changes in voting behavior have little impact, a coordinated response by the group can improve its performance by 5 percentage points if the few people who vote against their color and abstain can be convinced to vote. Alternatively, the group can obtain a gain of over 40 percentage points if almost everybody in the group but the expert abstains. Given that communication is only local and incomplete because members are scattered around the 3 auditorium, that many do not follow and do not even know the sophisticated strategy, and that voting is the best reply if others are unaware of it, an increase in voting seems likely.
Still, we nd that abstentions increase in the second run from 6% to 13%.
This increase by more than 100% is not only economically but also highly statistically signicant. It documents, to our knowledge for the rst time, that self-governed groups are not trapped but make a rst tentative step toward the more ecient use of information. Put dierently, we maybe observing the birth of the 'norm of abstention' seen in laboratory experiments. The increase is all the more surprising because it is (predictably) too small to improve information aggregation: if anything, the majority is now less likely to identify the right color.
One possible explanation why abstentions increase is that participants teach each other about the sophisticated strategy. Consistent with this notion, people who program the sophisticated strategy are signicantly more likely to start abstaining in the second round (86%) than those who do not (33%). Moreover, abstention spreads locally. With a neighbor who abstained in the rst round the probability that a voter abstains in the second round increases by 13%.
In our setting communication occurs naturally, without interference, structure or observation. The downside of this is that we are unable to use the content of conversations to decide whether people really learn about the sophisticated strategy or whether they already know this strategy and only learn to trust that coordination can be successful. It is even possible that the increase in abstention had nothing to do with communication at all but instead resulted from locally correlated independent learning. More able participants are more likely to nd the sophisticated strategy either in the rst or the second round than less able ones. If people of similar ability are more likely to sit next to each other, having a neighbor who abstained in the last round is associated with a higher probability to switch to abstention in the second roundeven without communication. We cannot exclude this explanation because participants were not seated randomly but (as typical for lectures) chose their seats themselves. 1 Our design can also not exclude that people learn to trust in successful coordination. We can, however, examine whether these alternative explanations are consistent with the data.
Both alternative explanations, local but independent learning as well as learning to trust in the success of coordination, imply that more abstaining neighbors in the rst round are associated with a stronger increase in abstentions in the second round. The more neighbors abstain in the previous round and communicate this to others, the larger should be an individual's assessment that coordination can be successful and hence her willingness to abstain. A participant with more able independent learners around is more likely to discover the sophisticated strategy and abstain.
These alternative explanations, however, are not borne out by the data. How 1 Seating this amount of people randomly and checking that they sit in the designated seats would not have been possible in the allocated time. many neighbors previously abstained is not associated with a higher probability to abstain. The explanations of local independent learning and learning to trust in the success of coordination are thus not consistent with what we observe. Our tentative conclusion is that people teach each other the sophisticated behavior although this is lowering the group's probability of success.
The remainder of the paper is organized as follows. The next section describes our contribution to the literature. Section 2 derives the best-reply and and some equilibria of the game. Section 3 explains the design of the study.
Section 4 presents our ndings and Section 5 concludes.
1 Contribution to the literature Our paper relates to three streams of the literature: (i) optimal extraction of information in groups, (ii) choice of equilibrium when there is communication and (iii) learning in experiments.
The proposition that more people know more has fascinated social philosophers at least since Galton conducted his famous analysis showing that the average guesses of visitors at a fair was closer to the true weight of an ox than that of an expert (Galton, 1907). De Condorcet (1785) asserts that more (at least partially) independent and informative viewpoints lead to better decisions.
Here, we check whether crowds are not only wise in the sense of available information but also whether they wisely aggregate this information.
How groups aggregate information has received quite a bit of attention in the literature. One of the best established results in this regard is that markets work extremely well in extracting and pooling private information from individual traders. This power of markets to collect and disseminate information has been demonstrated by Plott and Sunder (1988). In the presence of complete markets, 2 the price swiftly converges to the rational expectation's benchmark. Plott and Sunder (1982) and Camerer and Weigelt (1991) examine the role of`expertise' (or perfectly informed insider) in experimental asset markets. They show that the experts' information quickly takes a hold and the prices converge to the underlying fundamental value. 3 In contrast, information is shared in our study by voting.
There is a sizable literature examining information aggregation when voting.
For instance, Guarnaschelli et al. (2000) examine the strategic voting incentives in ad hoc committees of various sizes under the majority and unanimity rule. Ali et al. (2008) further extend this work to standing committees that interact repeatedly. In line with assertions of Austen-Smith and Banks (1996) and Feddersen and Pesendorfer (1998), these studies nd that strategic voting is prevalent. Under unanimity rule, as predicted, a substantial fraction of subjects vote against their signal. 4 These ndings demonstrate that in laboratory voting games, subjects are capable of acting with a high degree of sophistication. 5 In our case, there are no incentives to vote against one's own signal.
On the other hand, strategic abstention is central to the ecient information aggregation strategy that we focus on.
Unlike in our study, a high degree of coordination on abstention by nonexperts is observed in laboratory experiments. Exploring Feddersen and Pesendorfer (1996,1999) in a series of experiments, Battaglini et al. (2008Battaglini et al. ( , 2010 and Morton and Tyran (2011) nd that a large fraction of subjects withhold their information if there is a high chance of a better informed expert. In Morton and Tyran (2011) some learning of abstention seems to be going on, which suggests that we might see the emergence of a`norm of abstention'. This, however, is by no means certain. In all these experiments, standard procedures ensure common knowledge of the game and the problem that we are interested in studying, i.e., whether abstention can propagate in a group where many are unaware of the value of abstention, does not arise.
The sophistication of individuals when extracting information in common value environments can reach its limits. If public and private information is available, members of a group can, for example, be prone to updating biases.
In Mengel and Rivas (2017), ineciencies arise because subjects trust available expert information too much and then vote too often against their signal. In Kawamura and Vlaseros (2017), subjects are locked in a situation where they vote too often for their own uninformative signal rather than abstain. Following voting game. They nd that subjects not only truthfully and publicly reveal their private information to each other but also discuss what to do with that information (i.e., how to vote) to achieve the best outcome. 7 This decentral-5 A somewhat weaker evidence for fully rational benchmark is found in experiments with costly information acquisition, e.g., Elbittar et al. (2016) or Bhattacharya et al. (2014), where subjects are typically found to over-invest in information. 6 The same reason is examined in the context of markets by Ngangoue and Weizsacker (2015). 7 When interests conict, communication has been shown to aect the willingness to give in a dictator game (Andreoni and Rao, 2011), to increase the power of a proposer in multilateral bargaining (Baranski and Kagel, 2015), but also to positively aect investments before sharing the gains from team production (Baranski and Cox, 2019). ized process uniformly improves group decision making and diminishes the role of strategic incentives presented by dierent voting institutions. Palfrey and Pogorelskiy (2017) nd a signicant positive impact of intra-party communication on turnout in laboratory elections. Contrary to Nash equilibrium predictions, the higher turnout primarily benets the majority party. 8 In both these contributions, communication is public (within the relevant group of voters) and can thus be used to establish common knowledge. In our study, we were interested in a situation where communication does not extend to the whole group, so that common knowledge cannot be achieved.
Finally, our paper oers a contribution on learning in experiments. Various experiments examine how subjects learn individually in interactive situations, e.g., in Cournot or Bertrand competition (see, e.g. Huck et al., 1999;Oerman et al., 2002), social-dilemma games (see, e.g. Duy and Kim, 2005;Chaudhuri et al., 2006;Kirchkamp and Nagel, 2007) and coordination games (Schotter and Sopher, 2003;Chaudhuri et al., 2009). In other contributions, subjects learn from others in individual decision problems (Çelen et al., 2010;Cooper and Rege, 2011;Cason and Plott, 2014). We are looking at a situation, where individuals can learn from others in an interactive situation.

Model and Analysis
The game that is played by participants involves 36 players i ∈ {1, . . . , 36}. Each player receives a signal s i ∈ {blue, green} about an unknown state of nature θ ∈ {blue, green}, which represents whether the majority of balls in an urn are blue or green. Signals are of dierent quality t i . One of the signals is perfect t i = 1 and reveals the true state P (s i = θ|t i = 1) = 1. All others are relatively noisy because they are drawn from the urn which contains 99 balls of which 50 have one color and 49 the other color. One randomly drawn player is excluded, so that 35 players remain and it is very likely that one of these is perfectly informed ( 35 36 ≈ 97.2%).
The remaining 35 players can vote blue, green or abstain depending on the content s and type t of the signal. Once signals are realized, strategies result in a voting outcome v, where v measures how many more people voted for the true than the other state v ∈ {−35, . . . , 0, . . . 35}. Let us standardize the payo in the case that the majority voted for the actual state, v > 0, to one and for v ≤ 0 to zero. Player i's vote is decisive in two situations: in the case of a tie, v −i = 0, or in the case of a one-vote lead for the true state, v −i = 1, where v −i is the voting outcome without i's vote. In the rst case, a vote for the true state generates a majority for the true state and gains of one. In the second case, a vote against the true state destroys the majority and leads to losses of one. These situations may never occur. If, for example, all other players always vote blue irrespective 8 The role of local communication in an electoral process has recently been highlighted by Pons (2018). The eld experiment involving French presidential elections shows that a short conversation may aect voting behavior. of their signal, player i's vote will never make a dierence. In our setting of a large class, however, such a coordinated response is very unlikely. This is why we assume the following.
Assumption 1 (Strategic uncertainty assumption). From player i's perspective, it cannot be excluded that her vote aects results: This strategic uncertainty means that player i's behavior matters and gives her a reason to contemplate her choice in the rst place. The assumption has several consequences. First, voting for one's own signal strictly dominates voting against it (see Lemma 1). The reason is that the signal is informative and voting against it is hence more likely to cause harm whenever the player can make a dierence. Moreover, for perfectly informed players (t i = 1) voting strictly dominates abstaining (see Lemma 2). Since they are perfectly informed, they will never destroy a one-vote lead for the true state by voting for their signal but they may resolve a tie in the right way.
For imperfectly informed players (t i = 0), the situation is more interesting.
On the one hand, voting for one's own signal can resolve a tie. On the other hand, it may destroy a one-vote lead. The gains from voting one's signal thus depend on the likelihood that these pivotal situations occur. The top panel in Figure 1 shows the gains from voting relative to abstaining under the assumptions that players with perfect information vote for their signal and given a probability q that other imperfectly informed players (non-experts) vote. The gure actually contains both extreme cases, i.e., when non-experts vote for and 8 against their signal. The dierence, however, is so small that they do not show in the graph. The reason is that the expert is very likely to be present. Hence, the question whether others abstain or vote is more important than the relatively small dierence in getting it right when voting.
Player i's best reply in the lower panel shows that the likelihood that other imperfectly informed players vote is crucial. If it is unlikely, abstention is better to not dilute the expert's vote. If it is likely, the expert's vote is already so diluted that voting is the best reply.
The share for which it becomes optimal to vote rather abstain depends on the behavior of the other non-experts: if they all vote against their signal, voting for the signal becomes more valuable in comparison to abstaining and the share drops. The best reply function reveals the following equilibria of the game (the proof is in Appendix A). Proposition 1. Under the strategic uncertainty assumption, there are two Nash equilibria in pure strategies: • In the only-expert-votes equilibrium, only the perfectly informed expert votes and all others abstain.
• In the all-vote equilibrium, everyone votes. In addition, there is a Nash equilibrium in mixed strategies, where the expert votes and non-experts vote with some probability (around 70%). Figure 1 shows that apart from extreme situations where nearly everyone either abstains or votes, the individual decision to vote or abstain has very little impact on the group's probability of identifying the right color. The non-expert's joint decisions, however, can have a sizable impact (see Figure 2), so that it matters, which equilibrium will be played (the proof is in Appendix A).
Proposition 2. The only-expert-votes equilibrium Pareto-dominates the allvote and mixed-strategy equilibrium by more than 30 percentage points.
Since perfectly informed players vote in both equilibria, which of them is reached depends on the behavior of imperfectly informed players. This is why we will focus on their behavior when discussing results, later.

Experimental Design
When introducing the game in the lecture, we used physical props to render it easy for subjects to comprehend the problem. Before the lecture, we lled two urns with 49 blue and 49 green balls, added another (blue or green) ball to generate a majority as well as a dice of the same color (to represent the perfect information). We then drew the dice and 35 balls from the urn, numbered them and placed them into a small cardboard box that was sealed and put on display during the whole lecture. After the complete experiment, the box was opened and students were allowed to inspect them. i.e., a mobile phone, tablet or laptop. The ball (imperfect information) was represented on the screen by a circle of the appropriate color; the dice (perfect information) was represented by a square of the respective color. For logging in, subjects had an access code taped to their desk, which could only be used once.
Subjects were told that they would be matched in groups of 35 members who were dispersed throughout the auditorium. Each member of the group had a unique piece of information corresponding to the color of one of the 36 objects from the cardboard box. Subjects were assigned randomly rather than with their immediate neighbors in order to prevent direct communication within the voting group. For practical reasons, the matching of participants into groups was carried out after all decisions were made. One group was selected for payout and each member of this group received 10 e in case that the color with most votes was actually the majority color in the urn.
With 35 group members and 36 objects in the box, one object was left unassigned. This physically conveys the idea that it is unlikely (but possible) that none of the group members might have received the perfect information without having to rely on probabilities.
These rules were explained with a short animated presentationsee online Appendix. After the presentation of the rules, we told subjects that we would 9 The software operated independent of specic platforms using the Internet protocol and was derived from PINGO, a classroom communication software (Kundisch et al., 2013), by adding the functionality of sending information to the students. 10 now proceed under an exam protocol and that anyone caught talking would be excluded.
Before sending out information about colors, we asked subjects several control questions in order to get some indication of their comprehension. We did not provide feedback about correct answers in order to maintain the`natural' heterogeneity. At the end of the round, subjects received information on their mobile device about the voting outcome of their group, whether this outcome was correct, and whether their group was selected for payout. Then, we surprised subjects with the announcement that the voting will be repeated with a new urn and a new draw of 36 objects. A second sealed cardboard box containing the 36 draws from the second urn was presented and placed on the desk next to the box from the rst round.
Before starting the second round, we gave subjects 5 minutes to freely discuss with their neighbors. After the ve minutes, subjects were again put under the exam protocol, i.e., no talking or looking around. They then received their signals and cast their votes.
At the very end, we sent subjects a brief questionnaire on their devices in which we inquired about their gender and age. In addition, we wanted to assess whether they were aware of the ecient strategy. We did so by asking subjects to imagine that they would play with robots that could be programmed to follow a certain behavior. We then wanted to know how they would program these robots. They could specify whether robots should vote for blue or green or abstain depending on the information received by the robot.

Results
The participants in this study were almost 600 students from a large rst-year class (`Introduction to Business Administration') at Paderborn University. It was run on two consecutive days and the overwhelming majority of those present participated. Participants were able to`leave' and`enter' the study at any time by disconnecting from or reconnecting to the server. Still, only 3% were lost during the actual experiment, which might be expected for purely technical reasons (network interruptions, low battery). Attrition during the ex-post questionnaire was larger, particularly on the rst day, where many participants had to move on to the next lecture. Our analysis will focus on the around 590 participants with imperfect information whose behavior determines whether or not the ecient equilibrium is played. (Numbers for perfectly informed subjects are too low for a meaningful analysis.) The whole experiment took between 20-27 minutes and average earnings among subjects in groups selected for payout were 7.50 e.

Prerequisites
If we want to study whether a group can overcome the problems of unawareness about the ecient strategy and evolve to a more sophisticated use of information, we need to check whether the starting point exhibits such unawareness 11 and whether there is still room for sophistication.
Some evidence for unawareness comes from the control questions. About 6% incorrectly believe that green is more likely to be the majority color when their signal was blue. Around 3% claim that voting blue means that blue becomes less likely to be the majority color in their group. 30% say that them voting blue will not aect the outcome of the vote.

10
More importantly, participants do not agree on what is optimal behavior in the game. The majority programs their voting robots to vote regardless of whether they are perfectly or imperfectly informed, which suggests that they are unaware of the ecient only-expert-votes equilibrium. Slightly more than a quarter restricts the robot to vote only when information is perfect and hence likely to be aware of the ecient equilibrium. The remaining quarter's programming is all over the placesee Table 1. This shows disagreement on how to optimally behave even at the end of the experiment, i.e., after people had the chance to talk to each other. It thus seems likely that these disagreements were more pronounced before subjects communicated. *-vote against information, 0 abstain, + vote for information Subjects do not only express disagreement in the un-incentivized question on how to program voting robots, the discrepancy is also reected in actual behavior. As expected and hoped for, rst-round voting is far from the ecient only-expert-votes equilibrium. On both days and in both rounds, almost 80% of non-experts vote for their colorsee Table 2. The group is not capable of reaping the eciency gains from holding back and letting the expert decide. The probability that the majority in a group coincides with the true state in the rst round can be computed to be 51.1% (using the actual behavior in the respective formula in Appendix B) and is hence only slightly better than mere guessing.
The expert-vote and even the all-vote benchmark of 97.2% and 61.0% are thus missed by a wide margin. The bad performance is also due to a considerable number of imperfectly (and even some perfectly) informed subjects who vote against their signal, which (as we argue in Appendix F) can be traced back to people who misunderstood the game.
Summarizing these observations, initial behavior in our setting exhibits heterogeneity and leaves room for better coordination and sophistication.

Changes in voting behavior
The best-reply to the actual rst-round behavior where only around 6% abstain is to vote for one's signalrecall Figure 1. On the one hand, one might thus expect the share of votes to rise. On the other hand, abstaining becomes optimal if participants believe that suciently many of them will start abstaining after communicating (as limited as this communication may be).
From the rst to the second round, the number of abstention increases from 6.4% to 13.8%see Figure 3. The number of participants who switch to abstention is much higher than those who switch away from abstention, an increase that is highly signicant (p-value for McNemar test is below 0.001)although it is much smaller (and not statistically signicant) on the rst day.
The increase in abstentions is not large enough to improve performance.
Holding the shares of participants who vote against their signal constant, the probability that the majority in a group votes for the true state computed from the observed behavior slightly drops to 51.0% (this is true for both days).
The increase simply does not push abstentions to a level where gains can be expectedrecall Figure 2. If participants started abstaining in the hope that suciently many others will also do so, this hope was not fullled.
Result 1. Subjects are more likely to abstain in the second round although this is not the best response to actual behavior.

Explainations for the switch to abstention
There are three reasons why people might switch from voting to abstaining between the rst and second round. First, they learn the ecient strategy during the discussion. Second, they already know about this strategy but do not dare to act on it because they are afraid that others might be voting. The group discussion could then reduce the fear in so far as it becomes clear that If the increase is related to learning about the only-expert votes equilibrium between the rounds either individually or from others, the increase should be associated with knowledge about the ecient strategy. This seems to be the case. People who later program their robot to that equilibrium are more likely to have switched to abstention (86%) than away from it (14%). On the other hand, people who do not program their robots to the equilibrium are more likely to have switched away from abstention (66%) than to it (33%)see Table 3.
The dierence between the two groups is signicant at any conventional level.
That switchers are more likely to later know about the equilibrium suggests that they have learned about it between the rounds.
We can distinguish more systematically between the reasons by examining how the number of abstaining neighbors in the rst round aects abstention in the second round. If participants already know about the strategy but start believing that other members act on it, their individual beliefs should be increasing in the number of neighbors who abstained around them. With respect to individual learning between the rounds, one would a priori perhaps expect 14 that there is no correlation with the number of neighbors who abstained. If people sit together with friends whose abilities are correlated with their own, however, one would also observe a spurious correlation between the number of abstaining neighbors (indicating a higher ability in this local area) and the likelihood to start abstaining in the second round. If participants learn the ecient strategy from their neighbor, one abstaining neighbor who is willing to share the strategy suces. The last two channels assume that abstention in the rst round is an indicator for knowing about the ecient strategy. This assumption is plausible since the two are highly correlatedsee Appendix E.
In order to nd out about the relationship between switches to abstention and the number of neighbors who abstained before, we regress the change to abstention after either voting in line with or against one's own signal on dummies indicating whether one, two or many neighbors abstained in the previous round.
In our preferred specication (G2SLS), we allow for local correlation in the dependent variable as well as in the error term. 11 We only use the 525 subjects who did not abstain in the rst round.
The presence of one neighbor who abstained in the rst round is highly signicantly correlated with abstention in the next roundsee Table 4. With such a neighbor, it is 13% more likely that an individual changes from voting to abstaining in the next round.
Result 2. Having a neighbor who previously abstained is associated with a higher probability of abstentions. Having several abstaining neighbors is not associated with a further increase in the likelihood to abstain. This result is consistent with learning the sophisticated strategy from a neighbor but not with learning independently or learning to trust that others abstain. 11 More precisely, we estimate a spatial autoregressive model with autocorrelated errors (SARAR) using a generalized spatial two-stage least squares regression (G2SLS). As a robustness check, we use maximum likelihood (ML) estimates and a plain vanilla regression without local eects (OLS). Unsurprisingly, abstention in the second round is locally correlated: individuals are signicantly more likely to abstain when their neighbors abstain (see λ in the table). The correlation ρ between neighbor's errors, i.e., after taking the eect of observables on abstention into account, is negative and signicant but only using G2SLS. Signicance levels : * : 10% * * : 5% * * * : 1%

Reasons for the dierence between the days
There is a marked dierence between both days. Participants are much less likely to switch to abstention on the rst than on the second dayrecall Figure 3.
Attributing this dierence to a specic feature is dicult because the two days dier in many ways including starting time (whether a lecture is at Monday 7:30 or Tuesday 9:15 is important to many students, leading to potential selection eects) and study programs to which students signed up. There were also substantially less students on the rst day, which in principle can facilitate coordination. On the other hand, students were less densely seated on the rst day: only about a third of seats were lled and participants were widely dispersed throughout the auditorium. Some did not even have the opportunity to discuss with their neighbors because they had no neighborssee Figure 4.
Information can thus travel less well in the room on this day.
While we cannot identify which of these confounding factors is causing the dierence, there is a hint. Examining the regression, we see that switches to abstention are not related to the day in itself. The coecient of the day dummy is small and insignicantsee Table 4. This suggests that we are not dealing with a direct eect of the day. Instead, abstention seems to be less contagious on the rst day. While on the second day, having an abstaining neighbor is associated with an increase of abstention in 13%, the increase is 15 percentage points lower on the rst day; the respective coecient of the interaction term of whether one neighbor abstained and the day 1 dummy is statistically signicant (albeit at the 10% level). This is consistent with people not bothering to explain the sophisticated strategy to their neighbor when information cannot travel well.

5 Conclusion
Organisations, states, or companies can be in turmoil if they lack a functioning culture or understanding of how to act appropriately, i.e., a system of selfenforcing rules. The reason is diverging views of the facts and what to make of them and no central entity to guide the group. Some rule systems are more ecient but require sophisticated reasoning and trust, whereas others are more resilient in that they are robust to errors and confusion. Given the chance to talk but only locally, is there any chance that the more ecient system might emerge? Or will the organization take the path toward the resilient but less ecient equilibrium, which does not require a common understanding, trust and communication?
Here, we studied this question using a voting game in a setting that includes not only diverging views of the facts but also naturally occurring dierences in what to make of them: a large rst-year undergraduate class. The situation is not favorable for the ecient but sophisticated equilibrium to arise. The initial behavior is already very close to the more resilient but less ecient equilibrium.
Surprisingly, we nd that after the opportunity to communicate the group starts moving in the direction of the more sophisticated equilibrium.
Using econometric analysis, we tentatively conclude that the increase comes from students teaching each other about the more sophisticated strategyeven though this strategy is not the best response to what they have experienced.
This is a sign that it is possible for a group without central agency to move in a promising direction. Lemma 2 (Dominant strategy when perfectly informed). Under the strategic uncertainty assumption, voting strictly dominates abstaining for perfectly informed individuals (t i = 1).

Plott, Charles R and Shyam Sunder, Rational
Proof. Recall that the outcome is only aected in case of a one-vote majority for the true state or a tie. Voting and abstaining, both lead to a payo of one in case of a one-vote majority for the true state. In case of a tie, voting shifts the payo from zero to one because the vote is in line with the true state with certainty, while the outcome after abstaining remains zero. Since ties occur with positive probability, the dominance is thus strict.
12 If there is a single-vote majority favoring the wrong state, then i's vote can either boost the majority or convert it into a tie. In either case, the payo is one and the same, namely, zero. If the payo in case of a tie had been to determine the decision with a ip of a fair coin this lemma would still go through. 13 Note, the only dierence between an perfectly and imperfectly informed voter is only the strength of the signal. In the imperfect case the signal is only informative 1/2 < P (s i = θ) < 1; in the perfect case, it is truth revealing P (s i = θ) = 1. The argument above only relies on P (s i = θ) > 1/2 and so applies to both types. B Expected payo when q vote for and r against their draw Lemma 3. When q 0 non-experts vote in line with their draw and r 0 against it and the rest abstains, while q 1 experts vote their signal and r 1 against it, the probability that the majority votes for the correct state is: Q 0 (k, l)P 0 (k, l) and Proof. If no expert is in the sample, the probability of q 0 who vote in line and r 0 who vote against their signal, result in a sample of k voting for their signal, l against it, while 35 − k − l abstain. The probability for this event is: Say that x of those who vote in line with their draw receive the true signal, while y of those who vote against their signal receive the wrong signal. Then, the true state receives x + y votes and a majority for the true state results if and only if x + y > k+l 2 . This requires that x + l − y of the M balls of the majority color to be drawn and y + k − x of the N − M balls of the minority color. Moreover, of the total of k + l − y true signals that are drawn, x must come from the k who vote for and l − y from the l who vote against their signal. The probability for such a specic combination of (x, y, k, l) is thus: Summing these probabilities for all cases where x + y > k+l 2 leads to: If an expert is in the sample, the 34 players consist of k individuals who vote in line with and l who vote against their signal, while 34 − k − l abstain with probability: Electronic copy available at: https://ssrn.com/abstract=3665107 In q 1 percent of the cases, the expert votes in line with her signal and a tie among the remaining 34 members is now enough for a majority for the true state, leading to the following version of (4): In r 1 percent of the cases, the expert votes against her signal. A majority for the true state thus requires at least a two-vote lead among non-experts: x + y ≥ k+l 2 + 2 and the probability becomes: In case that the expert abstains, we are back to the same probability as if no expert would be in the sample, with the dierence that there is now one vote less. Putting together the four cases, we get the result. C Individual gains from voting when q vote for and r against their draw Lemma 4. Suppose there is no expert among the members of the group and an even number of its members k + l vote, where k vote in line and l against their signal. Then, the expected gain of a non-expert from voting relative to abstaining for k > l is: And for k < l it is: Lemma 5. Suppose there is no expert among the members of the group and an odd number k + l of non-experts who vote, where k is those who vote for and l those who vote against their signal. Then, the expected gain of a non-expert from voting relative to abstaining for k > l is: And for k < l it is: Lemma 6. Suppose there is an expert among the members of the group who votes for her signal and an even number k + l of non-experts vote, where k is those who vote for and l those who vote against their signal. Then, the expected gain of a non-expert from voting relative to abstaining for k > l is: And for k < l it is: Lemma 7. Suppose there is an expert among the members of the group who votes for her signal and an odd number k + l of non-experts who vote, where k is those who vote for and l those who vote against their signal. Then, the expected gain of a non-expert from voting relative to abstaining for k > l is: And for k < l it is: Lemma 8. If q is the probability that a non-expert votes for her signal and r the probability that she votes against it. Then, the expected gain of a non-expert from voting relative to abstaining is: Proof. Consider a non-expert. The probability that an expert has not been drawn (from the remaining n − 1 candidates) to be a part of the group is ( 1 0 )( n−2 n−2 ) ( n−1 n−2 ) = 1 n−1 . In this case, there are n − 1 other non-experts and the probability that k of them vote for and l against their signal is given by the multinomial in the top line of the expression. With the complementary probability, n−2 n−1 , there is an expert. In this case, there are n − 2 other non-experts and the probability that k of them vote for and l against their signal is given by the multinomial expression in the bottom line of the expression. The respective gains G were computed in the preceding lemmas.

D Session Protocol
Before the lecture, students were instructed by email to bring an Internet capable device such as a laptop or mobile phone. The email also included instructions on how to connect with the wireless network of the university.
The class was taught by one of the authors, 14 focuses on how management decisions might be biased (based on system 1 rather than system 2) in order to motivate that economic theory and econometrics try to overcome those biases; it features a prisoner's dilemma, introduces the notion of externality and brings a`soft' re-statement of the rst fundamental theorem of welfare. It did not cover voting theory or any other aspects of game theory beyond the prisoner's dilemma.
The study was run on the 9th and 10th of November 2015 in a room that seats 620 peoplesee Figure 5. Originally, the study was scheduled two weeks earlier (the 26th and 27th of October 2015). On the 26th of October, the server broke down when subjects were trying to login. The same happened on the 14 In principle, there is a danger that students want to show o in front of their teacher by choosing the clever`abstain' strategy. The sheer size of the class, however, renders it impossible for the lecturer to remember the names or faces of anyone and signaling cleverness is thus of little value. If anything, there is more reason for showing o in the rst than in the second round; such eects thus run against our nding. October 27th, although several servers and a load distributor were used. The problem was eventually solved by hiring fast server capacity. Most subjects thus were exposed to the instructions twice. There is no evidence of any intermediary discussions between students in between. Notice that any such discussion would have weakened our nding that more abstention occurs after communicating in the lecture.
On the rst day (Monday), the lecture took place at 7:30 and was addressed to students of International Business Studies and Economics & Engineering.
Students were asked to stay for the study after the lecture at around 8:40. The number of logged in participants uctuated by 4 people around 203. This was the overwhelming majority of those present at the lecture. It dropped to about 120 during the ex-post questionnaire, i.e., when the actual study was over at around 9:00 and when it was time to go to the next class.
On the second day (Tuesday), the lecture started at 9:15 and was aimed at students with a major in Economics. The study took place before the lecture and the number of logged in participants uctuated by 16 people around 400.
Again, this was the overwhelming majority of students present. Apart from the drop in ex-post questionnaire answers on the rst day, there is no noticeable attrition.
The study lasted 20 min on the rst day and seven minutes longer on the second day. The largest chunk of the time was spent on the instructions (6 min).
The shortest part was the actual voting which took only 2 min on the rst day and 4 min on the second day.
Given the number of participants, the probability of being selected for payout was around 35 200 = 17% in each of the two rounds of the rst day and 35 400 = 8.75% on the second day.

29
E Abstentions in the rst round stongly associated with programming of ecient strategy Abstention in the rst round is an indication of knowing about the ecient only-expert equilibrium. Participants who program their robots in accordance with the ecient strategy are much more likely to abstain. Among those who vote only 23%-26% program the ecient strategy, whereas the share is between 56% to 75% among those who abstain. The positive relationship between programming the ecient strategy and abstaining is highly signicant on both days.  Table 2 indicate that a non-negligable number of participants vote against their own signal. Given our observation that this strategy is dominated, this behavior is unexpected. 15 This section inquires into possible reasons. The analysis here is retrospective. It is an exploratory exercise which means the results should be taken with a grain of salt.
One reason for voting against one's own signal might be cheating. We can distinguish between two types of cheating. First, participants may be sitting next to an expert and see the color of the majority of balls. If they have a dierent color, it would be optimal for them to vote against their signal.
This type of cheating relies on having a suciently good view on the neighbors' phone to distinguish between the shape being a circle or a square and identify the color. Second, participants may not be seeing the exact shape of the information but only get an impression of whether there is a dominant color in their local neighborhood. Again, they might vote for this color rather than their own color.
While there are a few isolated cases consistent with the rst kind of cheating, there is no evidence consistent with the second kind of cheatingsee Figure 6.  For a more systematic analysis on whether participants cheat by copying 15 Voting against one's own signal seems to be a more widespread phenomenon that is also observed, e.g., by Bouton et al. (2016) or Herrera et al. (2016). 31 an expert, we run a regression that includes a variable whether an expert with opposing signal is nearby. We control for the second explanation by including a variable that measures the percentage of visible neighbors with an opposing color. Moreover, cheating should be easier if people sit further at the back or closer together, which is why we include the row number and a dummy variable for the rst day.
Before we discuss the results, let us outline two other possible explanations for voting against one's own information. One of them is that some subjects may have been confused or not paid attention. In this case, we speculate that participants might simply vote for whichever choice is presented to them rst on the decision screen. In our case, this is the color`green'. According to this explanation, participants with blue signals would be more likely to vote against their signal (perhaps less so for the blue colored experts).
Lastly, subjects may have suered from some misconceptions regarding the nature of the game. We can learn about these from the control questions.
About 6% incorrectly believe that green is more likely to be the majority color when their signal was blue. Around 3% claim that voting blue means that blue becomes less likely to be the majority color in their group. 30% say that them voting blue will not aect the outcome of the vote. (This answer might actually not be wrong if subjects interpreted the question to mean that they are unlikely to be pivotal.) For all three cases, we include respective indicator variables.
We ran an ordinary least squares regression in which we explain rst round behavior using the above control variables. We took the rst round because participants have not yet talked to each other. (When allowing for spatial autocorrelation in the dependent variable and the error term or using both rounds, results remain essentially the same. 16 ) The proxy variables indicating cheating are not signicant and in the case of opposite expert information even have the wrong sign (having an expert with opposing information renders it less likely that I vote against my signal)see Table 6. The key variables are related to misconceptions about the nature of the game. Participants who believe that having received a green signal goes along with a higher likelihood of the majority of balls in the urn being blue are 30% more likely to vote against their signal. Similarly, participants who believe that voting blue renders it less likely to be the elected color are 27% more likely to vote against their signal.
Votes against one's signal thus seem to be driven by misconceptions about the nature of the game. This nding is in line with Cason and Plott (2014).
Just as on any exam, it is not uncommon to see some students misunderstand a question. In our case, having some fraction of participants misunderstand parts is a natural feature that makes it even less likely that participants abstain initially but gives them a chance to learn during communication. A sizable share of subjects stops voting against their information in the second round (44% on day 1 and 57% on day 2). 17 16 When estimating both rounds with an OLS (errors clustered at individual level) being 32 Electronic copy available at: https://ssrn.com/abstract=3665107 expert reduces contrarian votes at the 10% level. 17 There are also subjects who start voting against their information. This concerns 14% on day 1 and 13% on day 2. Still, we do not observe an overall reduction in voting against one's information simply because the 13% and 14% relate to the much larger base of subjects voting for their information or abstaining.