A defining aspect of human cooperation is the use of sophisticated indirect reciprocity. We observe others, talk about others, and act accordingly. We help those who help others, and we cooperate expecting that others will cooperate in return. Indirect reciprocity is based on reputation, which spreads by communication. A crucial aspect of indirect reciprocity is observability: reputation effects can support cooperation as long as peoples’ actions can be observed by others. In evolutionary models of indirect reciprocity, natural selection favors cooperation when observability is sufficiently high. Complementing this theoretical work are experiments where observability promotes cooperation among small groups playing games in the laboratory. Until now, however, there has been little evidence of observability’s power to promote large-scale cooperation in real-world settings. Here we provide such evidence using a field study involving 2,413 subjects. We collaborated with a utility company to study participation in a program designed to prevent blackouts. We show that observability triples participation in this public goods game. The effect is nearly seven times larger than offering a $25 monetary incentive, the com-
*Federal Trade Commission, Washington, DC 20580; †Department of Computer Science and Engineering, University of California at San Diego, La Jolla, CA 92093; ‡Program for Evolutionary Dynamics, Harvard University, Cambridge, MA 02138; §Department of Psychology, Yale University, New Haven, CT 06520; and Departments of ∥ Mathematics and #Organismic and Evolutionary Biology, Harvard University, Cambridge, MA 02138. **To whom correspondence should be addressed. E-mail: email@example.com.
pany’s previous policy. Furthermore, as predicted by indirect reciprocity, we provide evidence that reputational concerns are driving our observability effect. In sum, we show how indirect reciprocity can be harnessed to increase cooperation in a relevant, real-world public goods game.
Cooperation occurs when we take on costs to benefit the greater good. By this definition, everyone is better off when everyone cooperates, but self-interest undermines cooperation and leads to free-riding. Promoting cooperation is a central challenge for human societies, both today and over our evolutionary history (Hardin, 1968; Axelrod, 1984; Ostrom, 1990; Levin, 2006; Helbing and Yu, 2009; Walker et al., 2009; Sigmund, 2010; Nowak and Highfield, 2011; Oakley et al., 2011; Churchland, 2012). There are five mechanisms for the evolution of cooperation (Nowak, 2006b): direct and indirect reciprocity, spatial selection, group selection, and kin selection. Each of these mechanisms is an interaction structure that can lead cooperators to outperform noncooperators, and therefore be favored by selection.
Direct and indirect reciprocity involve repeated interactions, creating future consequences for one’s actions: it can pay to cooperate today to receive cooperation from others tomorrow. Spatial selection occurs when players’ interactions are structured rather than occurring at random. As a result, cooperators may be more likely to interact with other cooperators and thus preferentially receive the benefits of cooperation. Spatial selection operates when cooperaters cluster in physical space, on social networks, in sets, or in phenotype space (Nowak et al., 2010). Group selection (or multilevel selection) occurs when competition and reproduction happen at multiple levels: not only do players compete with others in their group, but groups compete with each other. If cooperative groups outcompete noncooperative groups, then group-level selection can favor the evolution of cooperation. Finally, kin selection may be defined as preferring to cooperate with those who are closely related. Kin recognition can allow players to cooperate with close genetic relatives and defect otherwise.
Most of the literature on the evolution of cooperation uses the Prisoner’s Dilemma and related frameworks: players can pay a cost to give a greater benefit to one or more others. Thus, within the context of these games, cooperation is good for everyone. However, cooperation need not be good for everyone more generally (Oakley et al., 2011; Dasgupta, 2012). There are situations in which cooperating may give a benefit to some, but impose costs on others. For example, in intergroup conflict and war, people cooperate with members of their own group in an attempt to harm members of other groups (Choi and Bowles, 2007).
Or in the context of markets, companies may collude to keep prices high, benefiting each other but harming consumers (Carlton and Perloff, 1994). The five mechanisms for the evolution of cooperation may promote both total welfare-enhancing cooperation, as well as these more pernicious forms of cooperation.
All of these mechanisms are relevant for the evolution of human cooperation, but direct reciprocity and indirect reciprocity occupy a central place: most of our key interactions are repeated and reputation is usually at stake. Direct reciprocity is based on repeated encounters between the same two individuals: my behavior toward you depends on what you have done to me. Indirect reciprocity is based on repeated encounters in a group of individuals: my behavior toward you also depends on what you have done to others (Fig. 12.1). We take a keen interest in who does what to whom and why, which requires sophisticated social intelligence. We talk to each other about others. As David Haig said: “For direct reciprocity you need a face, for indirect reciprocity you need a name” (Nowak and Highfield, 2011). The evolution of indirect reciprocity is linked to the evolution of human language. Supported by human language, reputation systems allow us to track the good and bad behavior of others and to use this information to incentivize cooperation. Whatever is specifically human about our mental machinery is derivative of human language, social intelligence, and thus indirect reciprocity (Cela-Conde and Ayala, 2007; Nowak and Highfield, 2011).
The evolution of cooperation via indirect reciprocity has been a topic of great interest in recent years. Mathematical models and computer simulations have demonstrated the power of indirect reciprocity for promoting cooperative behavior (Kandori, 1992; Nowak and Sigmund, 1998a,b, 2005; Nowak, 2000; Leimar and Hammerstein, 2001; Panchanathan and Boyd, 2003, 2004; Ohtsuki and Iwasa, 2004, 2006, 2007; Brandt and Sigmund, 2006; Pacheco et al., 2006; Masuda and Ohtsuki, 2007; Suzuki and Akiyama, 2007a,b; Fu et al., 2008; Ohtsuki et al., 2009; Uchida and Sigmund, 2010; Berger, 2011; Nakamura and Masuda, 2011; Manapat and Rand, 2012; Manapat et al., 2013; Mani et al., 2013). In these models, players typically engage in a series of one-shot interactions with others selected at random from the population. In some of those interactions, players’ previous decisions are observable by their partners. Observability allows players to use conditional strategies that base their actions on the partner’s behavior in the past. When past actions are sufficiently observable, natural selection can favor strategies that cooperate as long as the partner has behaved well in the past. What constitutes “good” behavior worthy of receiving cooperation depends on the social norm. A simple social norm is called “image scoring” and prescribes cooperating with those who have cooperated in sufficiently many previous interactions (Nowak and Sigmund,
FIGURE 12.1 Indirect reciprocity can support contributions to public goods, as depicted here. In indirect reciprocity, my behavior toward you depends on how you have behaved toward both me and others. This process occurs in three stages. First, people engage in a public goods game (C = cooperators and D = defectors). Peoples’ behavior is observed, both by other players and third parties. Second, information about this behavior can spread from person to person. Based on the information received, peoples’ opinions about the players are updated. Third, as a result, public goods cooperators will receive cooperation in future interactions whereas defectors will be denied cooperation, defected upon, or punished. Thus, indirect reciprocity creates an incentive to contribute to public goods and can promote the evolution of cooperation.
1998a). More complicated norms also take into account the behavior of the partner’s previous partners. For example, under the “standing” norm, players can maintain their good reputation by defecting against those with bad reputation (Ohtsuki and Iwasa, 2006). The many models of indirect reciprocity differ in their details, yet across a wide range of assumptions, making previous decisions observable allows cooperators to selectively target their cooperation at other cooperators and withhold cooperation
from defectors. Thus, free-riders are at a disadvantage, and cooperation can spread.
This body of theoretical work is supported by behavioral experiments where subjects play economic games in the laboratory. People are substantially more cooperative when their decisions are observable and when others can respond accordingly (Wedekind and Milinski, 2000; Milinski et al., 2001, 2002a,b; Wedekind and Braithwaite, 2002; Andreoni and Petrie, 2004; List et al., 2004; Rege and Telle, 2004; Bolton et al., 2005; Semmann et al., 2005; Rockenbach and Milinski, 2006; Seinen and Schram, 2006; Tadelis, 2007; Linardi and McConnell 2008; Andreoni and Bernheim, 2009; Rand et al., 2009; Ule et al., 2009; Jacquet et al., 2011; Pfeiffer et al., 2012; Choi and Ahn, 2013). Subjects understand that having a good reputation is valuable in these settings (Pfeiffer et al., 2012) and so are willing to pay the cost of cooperation. Observability particularly increases cooperation when the prosocial nature of the cooperative choice is made salient (Rege and Telle, 2004; Kossmeier et al., 2009). Moreover, experimental evidence indicates that indirect reciprocity is deeply entrenched in human psychology: subtle cues of observability have large effects on cooperation levels (Haley and Fessler, 2005; Burnham and Hare, 2007; Ernest-Jones et al., 2011; Powell et al., 2012), and our initial impulse to cooperate in one-shot anonymous settings (Cornelissen et al., 2011; Rand et al., 2012, 2013; Schulz et al., in press) is likely the result of adaptation in a world dominated by reputational concerns (Rand et al., 2012, 2013).
These laboratory experiments are extremely valuable. They generate powerful insights into human psychology and provide clear evidence for the importance of indirect reciprocity. To do so, however, they typically use abstract economic games and involve the interaction of only a handful of subjects. Thus, the question of whether observability affects large-scale cooperation in real-world settings outside of the laboratory remains largely unexplored [exceptions include Lacetera and Macis (2010), Ashraf et al. (2012), and Karlan and McConnell (2012)]. The extent to which findings from theory and the laboratory generalize to natural field settings is of great importance, both for scientific understanding and for public policy (Levitt and List, 2007).
Here, we address this question by running a large-scale field experiment on the effect of observability in a public goods game (PGG). We collaborated with a major electric utility company to enroll consumers in a “demand response” program. This program is designed to help prevent blackouts by reducing excessive use of air conditioning during periods of high electricity demand. The cost of electricity production can spike hundredsfold during demand peaks. However, the price consumers pay is typically constant across time. Thus, during peak periods, there is a dramatic mismatch between price and actual cost, leading to excessive
energy use. This mismatch reduces grid reliability, drives up energy costs, increases the risk of blackouts, and harms the environment. In recent years, reducing excessive peak energy use has become a target of regulatory efforts to increase efficiency in the electricity industry.
Encouraging participation in demand response programs such as the one used in this study is the primary policy tool available for reducing peak energy use (U.S. Committee on Technology, 2011). Demand response programs are voluntary programs in which people allow their utility to remotely restrict their energy consumption during peak hours. To do so, the utility usually installs a remote switch in-line with the circuitry of an appliance such as a hot water heater or air conditioner. Estimates suggest that these voluntary programs could reduce the need to invest in additional generation capacity by at least 38 percent over the next two decades, generating cost savings of at least $129 billion (Chupka et al., 2008). Voluntary energy efficiency and demand response programs have been widely available for years, but participation is frustratingly low (Federal Energy Regulatory Commission, 2011). Demand response programs exemplify the public goods dilemma: participation helps reduce on-peak demand, benefiting all energy grid users, but energy consumers find participating inconvenient. Participation is socially optimal because the inconvenience is minimal for most individuals relative to the societal costs of a blackout.
To explore the effect of observability on this real-world public goods problem, we solicited residents of 15 homeowners associations (HOAs) to participate in a demand response program. Residents who volunteered for this program allowed the utility to install a device that remotely curbs their central air conditioners when necessary: on days with unusually high demand or in the case of an unexpected plant or transmission failure. Residents who volunteered therefore contributed to a public good by improving the stability of the electrical grid in all of California, at the cost of some personal inconvenience. We solicited volunteers by delivering mailers to residents and asking them to participate. Sign-up sheets were posted in a communal area near their home, usually by a shared mailbox kiosk. In our primary manipulation, we varied whether residents’ neighbors could tell who had signed up for the program. We did so by varying whether the publicly posted sheets required residents to print their name and unit number (observable treatment) or only a code that does not reveal their identity (anonymous treatment).
We found that residents in the observable treatment are nearly three times as likely to participate in the demand response program as residents in the anonymous treatment (fraction of residents participating:
FIGURE 12.2 We solicited 1,408 customers of a major electric utility for participation in a program designed to prevent blackouts. Residents signed up for the program on sheets posted in a communal area near their home, usually by a shared mailbox kiosk. We varied whether residents’ neighbors could tell who signed up for the program: publicly posted sheets required residents to print their name and unit number (observable treatment) or only a code that does not reveal their identity (anonymous treatment). Observability tripled participation in the program.
anonymous = 0.030, observable = 0.088, P < 0.01, n = 1,408; Fig. 12.2). All statistics presented are from probit regressions including various controls, with SEs clustered at the HOA level
The effect of the observable treatment was nearly seven times that of offering a $25 incentive (the estimated effect of the incentive is 0.009; a Wald test rejects that the coefficients on observability and the $25 incentive are identical, P = 0.024). This incentive was what the utility had used before the experiment, and they had previously argued the incentive would be far more effective than observability. In fact, this incentive appears to have been too small to be effective, and such small financial incentives are known to sometimes backfire (Gneezy and Rustichini, 2000). For the sake of comparison, we followed convention and estimated how large the financial incentive would have to be to achieve the same results if its effect is linear (Bertrand et al., 2010). We found that the utility would have had to offer an incentive of $174 to increase participation as much as our observable treatment.
We now explore the mechanism through which observability functions to increase participation. Indirect reciprocity theory is based on reputational concerns: when groups of people interact repeatedly and actions are observable, it becomes advantageous to be seen contributing
to public goods. Based on this account, we predict that observability will have a greater effect among populations where ongoing relationships and reputations are expected to play a larger role. We evaluate this prediction in two ways.
First, we tested whether the effect of the observable treatment was greater in apartment buildings compared with row houses and individual homes. In apartment buildings, residents are more likely to interact with their neighbors in public spaces, and signup sheets were typically posted in especially conspicuous locations. Thus, indirect reciprocity theory predicts that observability will have a larger effect in apartment buildings. As shown in Fig. 12.3A, the results confirm this prediction: observability increased participation among those living in apartment buildings (fraction of residents participating: anonymous = 0.048, observable = 0.114, P < 0.01, n = 582), whereas it had little effect on the inhabitants of row houses or individual homes (fraction of residents participating: anonymous = 0.024, observable = 0.038, not significant, n = 826; yielding an estimated interaction of 0.052, P = 0.04).
Second, we tested whether observability had a larger effect among those who own their homes/apartments relative to those who rent. Renters are more transient and therefore likely to be less invested in relationships with their neighbors. Thus, indirect reciprocity theory predicts that observability will have a larger effect among owners. As shown in Fig. 12.3B, the results were again consistent with this prediction: observability dramatically increased participation among owners (fraction of residents participating: anonymous = 0.024, observable = 0.099, P < 0.01, n = 1,015) but had little effect on renters (fraction of residents participating: anonymous = 0.045, observable = 0.059, not significant, n = 393; yielding an estimated interaction of 0.046, P < 0.01).
Residents of apartment buildings and individual homes differ on any number of dimensions, as do those who own versus rent their homes. Thus, although the results are consistent with the predictions of the indirect reciprocity framework, alternative explanations of these results are possible. To partially address this issue, we ran the same analysis with additional controls such as Spanish language preference, ethnicity, missed payments, rebate use, and historical electrical use, and found even stronger results. Therefore, differences on these dimensions do not account for the differential effects of observability seen in Figs. 12.2 and 12.3.
Finally, we provide evidence that the effect of observability is unique to public goods. Not participating in the program should carry the threat of social sanctions only if participation is considered to be a public good. Therefore, indirect reciprocity theory predicts that observability should not increase participation among subjects who do not think of participa-
FIGURE 12.3 Observability increased participation more in settings where reputational concerns matter more. (A) Observability increased participation more in apartment buildings where residents are more likely to interact with their neighbors in public spaces and sign-up sheets were typically posted in especially conspicuous locations, compared with row houses or individual homes, where neighbors are less likely to interact and sign-up sheets were less easily visible by others. (B) Similarly, observability increased participation more among those who own their homes/apartments relative to those who rent because renters are more transient and therefore likely to be less invested in long-term relationships with their neighbors.
tion as a public good. To test this prediction, an additional 1,005 subjects received exactly the same treatment as described above, except that the mailers they received were stripped of any language that framed the demand response program as a public good. Consistent with our hypothesis, Fig. 12.4 shows that the effect of observability was reduced in this cohort (fraction of residents participating: anonymous = 0.061, observable = 0.086, not significant, n = 1,005; estimated interaction
FIGURE 12.4 We solicited an additional 1,005 customers with exactly the same treatment as described earlier, except that the informational materials they received were stripped of any language that framed the blackout prevention program as a public good. The effect of observability was dramatically reduced among subjects who did not receive the public good framing.
between observability and the public good message in a pooled regression is 0.035, P = 0.098).
We have shown that indirect reciprocity promotes cooperation in a real-world public goods game affecting thousands of people. Making participation in the public good observable substantially increased sign-ups and did so significantly more than offering a cash incentive. Moreover, the effect of observability was larger in settings where individuals were more likely to have future interactions with those who observed them, and when participation was framed as a public good. These results provide evidence that reputational concerns were the driving force behind the effect of observability in our study.
Our study is part of a nascent literature exploring reputation and prosociality using field experiments. Consistent with our findings in the domain of energy efficiency, there is evidence that publicizing the names of donors increases the frequency of blood donation (Lacetera and Macis, 2010) as well as the level of giving to a college charity (Karlan and McConnell, 2012). Nonfinancial incentives involving reputation have also been shown to outperform monetary incentives in motivating the sale of condoms on behalf of a health organization in Namibia (Ashraf et al., 2012). Our work adds to these studies by directly manipulating observability, allowing a comparison with monetary incentives while avoiding
other potential confounds present in previous experiments. We also test specific hypotheses generated by indirect reciprocity theory regarding when observability will and will not increase cooperation. Taken together, this body of work provides clear evidence that reputational incentives can be a powerful force for increasing cooperation in the field. Our paper in particular adds to efforts aimed at promoting energy conservation via nonfinancial incentives, such as providing people information about their own energy use and how it compares with the energy use of their neighbors (Minosi et al., 2003; Thompson, 2007; Ayres et al., 2009; Costa and Kahn, 2010).
A question arising from our study is the extent to which our subjects were conscious of their indirect reciprocity motives. One possibility is that they explicitly considered the reputational costs of not participating in the observable public goods treatment. Alternatively, they may have learned or evolved sensitivity to subtle cues that subconsciously increased their desire to participate when their decisions were observable, as has been shown in other settings (Haley and Fessler, 2005; Bateson et al., 2006; Burnham and Hare, 2007; Ernest-Jones et al., 2011). Perhaps the degree of “warm glow” they feel is sensitive to the degree of observability in their environment and the likelihood of interacting with observers in the future. Subsequent studies should further investigate this issue.
A related issue is the universality of reputation concerns. Observability can promote cooperation, but only in populations where the proper social norms are in place. For example, in a laboratory experiment in the United States, making public goods contributions observable by linking the PGG to a set of pairwise Prisoner’s Dilemma games led to high contributions (Rand et al., 2009). However, when the same experiment was run using students in Romania, no such positive effects were observed because the Romanians did not sanction bad behavior in the PGG (Ellingsen et al., 2012). Similarly, providing feedback on how one’s energy use compares with one’s neighbors had reduced consumption among American liberals but may have had the opposite effect among conservatives (Costa and Kahn, 2010). Studying the interaction between norms and institutional policies is an important direction for future research.
In our experiment, the observability mechanism was designed so that participation was automatically displayed to all: because sign-up sheets were posted in public areas, no special effort was required by individuals to spread reputational information. Most indirect reciprocity models, however, rely on individuals communicating information about the observed actions of others (Nowak and Sigmund, 2005). Fortunately, we are more than happy to talk about how others have behaved: gossip is a central element of human communication (Dunbar et al., 1997;
Sommerfeld et al., 2007). However, why did we come to have this predilection for gossiping about the previous behavior of others? Why spend time and effort on evaluating others, and why give honest evaluations of competitors? Indirect reciprocity itself offers a potential answer: providing honest information or not is another game of cooperation and defection, which is also linked to reputation. Your reputation can be damaged not just by defection in the primary public goods game, but also by the distribution of incomplete or false information. Another important question involves large-scale reputation systems such as those used by the online market eBay (Resnick et al., 2006) or the business rating website Yelp (Luca, 2011): to what extent does our intrinsic desire to gossip extend to these more distributed settings? Why do people bother to leave evaluations, and how can secondary reputation systems be designed to encourage honest feedback? Exploring these issues is an important direction for further study.
Indirect reciprocity offers a powerful tool for promoting cooperation in contexts of great societal importance. Here, we offer quantitative evidence for one example: curbing electricity use during periods of high demand. However, this is just one of many such opportunities (Rand and Nowak, 2009; Lacetera and Macis, 2010; Ashraf et al., 2012; Karlan and McConnell, 2012). For example, people might be induced to drive more efficient cars if all vehicles bore a visible indication of fuel efficiency, perhaps via mandated color coding of license plates for the most efficient and most wasteful vehicles. Or home energy use might be reduced if utility companies made individuals’ power use statistics publicly available. One might even apply this logic to scientific discovery: a measure of “scientific carbon efficiency” could be calculated by dividing an author’s number of citations (or h-index) by the number of miles flown to attend conferences. Of course, privacy is an important issue that must be balanced against the benefits of reputational pressure. However, there are also indirect reciprocity applications that do not infringe on the privacy rights of individuals. For example, businesses might reduce their environmental impact if they were required to disclose the overall carbon footprint of their operations. Reputational concerns might discourage financial institutions from taking excessive risk because of changes in the price at which they have to borrow. Or government agencies might reduce waste if the relevant statistics were readily accessible to the public. Developing interventions that harness indirect reciprocity is a promising direction for future public policy initiatives.
We administered the field experiment with the collaboration of the Pacific Gas and Electric Company (PG&E), a regulated utility that operates the majority of Northern California’s retail residential electricity market. The experiment was incorporated into a routine marketing effort for a demand response program called SmartAC, which is designed to help prevent or shorten power interruptions by curbing demand from central air conditioners on days with unusually high demand, or in the case of an unexpected plant or transmission failure. The program is voluntary; subjects who participate contribute to a public good by contributing to the stability of the electrical grid in all of California, at the cost of some personal inconvenience and possibly some discomfort. The SmartAC switch is installed free-of-charge. At the time of the program, participants received a $25 check for signing up. The SmartAC program is a typical demand-side management, direct load control, or load shedding program.
Subjects in the field experiment were residential customers living in homeowners associations (HOAs) and one rental complex in Santa Clara County. We focused on tenants of HOAs because it was necessary to choose residences with public spaces where sign-up sheets could be posted. We focused on Santa Clara County because PG&E had not marketed in this area before the field experiment. Furthermore, Santa Clara County is hot enough that customers there were likely to have air conditioners, and dense enough to have a sufficient number of HOAs. Finally, we restricted the analysis to HOAs where all residents were known to have central air conditioning because central air conditioning was required to participate in the SmartAC program.
We invited subjects to participate in the program by sliding marketing materials under subjects’ doors, placing them on their doorstep, or mailing materials to subjects. The materials included an informational letter describing the program and an instruction card that directed subjects to sign up for the program on sign-up sheets posted next to their mailboxes or in another central location. We left the sign-up sheets up for 3 to 10 days, depending on managers’ preferences, the weather, and other conditions. After distributing the marketing materials, we removed the sign-up sheets, noted subjects’ participation decisions, and provided the list of participants to PG&E’s contractor for processing and installation. Note that subjects were not aware that they were participating in an experiment. This study is therefore classified as a natural field experiment (Harrison and List, 2004).
In the experiment’s main treatment, we varied observability by varying the design of the sign-up sheets on which subjects register for the program: some sheets were designed so that subjects’ identities were
easily revealed to others who observed the sign-up sheet whereas others were designed to conceal subjects’ identities. In the latter “anonymous’” design, the fields for subjects’ names and apartment numbers were omitted from the sign-up sheet. Instead, subjects were identified only by their randomly generated personal code.
Simultaneously, we varied the design of the marketing materials along two dimensions. First, we varied whether the materials framed the decision to sign up as a contribution to a public good that would benefit others, or just as a new feature being offered by PG&E. Second, we varied whether subjects were offered a $25 incentive for signing up for the program.
This research was partially funded by National Science Foundation Proposal 0752818, Russell Sage Foundation Award 980803, the Stigler Center for the Study of the Economy and the State’s Energy Initiative, the University of Chicago Graduate School of Business, and the John Templeton Foundation. Any opinions expressed in this article are those of the authors and not of the Federal Trade Commission or any individual Commissioner.