Partner selection supported by opaque reputation promotes cooperative behavior

Abstract

Reputation plays a major role in human societies, and it has been proposed as an explanation for the evolution of cooperation. While the majority of previous studies equates reputation with a transparent and complete history of players’ past decisions, reputations in real life are often ambiguous and opaque. Using web-based experiments, we explore the extent to which opaque reputation works in isolating defectors, with and without partner selection opportunities. We found that low reputation works as a signal of untrustworthiness, whereas medium or high reputations are not taken into account by subjects for orienting their choices. Reputation without partner selection does not promote cooperative behavior; that is, defectors do not turn into cooperators only for the sake of getting a positive reputation. Finally, in a third study, when reputation is pivotal to selection, then a substantial proportion of would-be-defectors turn into cooperators. Taken together, these results provide insights about the characteristics of reputation and about the way in which humans make use of it when selecting partners, and also when knowing that they will be selected.

1 Introduction

Partner selection, that is the ability to spot and preferentially interact with better social partners, is proposed to be a major factor in maintaining costly cooperation between individuals (Noë & Hammerstein, 1994). Theories on the evolution of cooperation via indirect reciprocity (Nowak & Sigmund, 1998; Panchanathan & Boyd, 2004) emphasize the role of reputation in avoiding cheaters and supporting cooperation, and experimental studies using dynamic networks indeed suggest that human subjects tend to break links with defectors and form new links with cooperators (Rand, Arbesman & Christakis, 2011; Wang, Suri & Watts, 2012).

Reputation-based partner selection requires the ability to evaluate others and to take into account third parties’ evaluations, i.e., the result of being evaluated by others. A sensitivity towards others’ presence and related evaluations is suggested by several studies. When subtle cues of being watched are present, such as a picture of watching eyes (Bateson, Nettle & Roberts, 2005), or even three dots in a face-like configuration (Rigdon et al., 2009), the probability of donating something significantly increases, both in laboratory experiments (Burnham & Hare, 2007; Haley & Fessler, 2005) and in field studies (Ernest-Jones, Nettle & Bateson, 2011; Yoeli et al., 2013). In economic games, cooperation increases when subjects are informed about others’ actions in a transparent and reliable manner, e.g., when they receive information about the amount of other players’ past contributions in a public goods game (Sommerfeld et al., 2007; Sommerfeld, Krambeck and Milinski, 2008), or in a two-players donation game (d’Adda, Capraro & Tavoni, 2015; Wedekind & Milinski, 2000).

Although interesting, these studies present what Granovetter (1985) calls an “undersocialized” notion that equates reputation with a transparent and complete history of players’ past decisions. In real life, reputations are based on personal evaluations, and they are often ambiguous, opaque and ephemeral. Nonetheless, humans strive to acquire positive reputations, and they select partners and make decisions on the basis of partners’ reputations. The fragility of reputation is even more evident if we take into account digital reputation, a widely used tool in online transactions and services. According to Randy Farmer (2011, p. 16): “Every top website is using reputation to improve its products and services, even if only internally to mitigate abuse. In short, reputation systems create real-world value.” Reputation systems are designed to mediate and facilitate the process of assessing reputations within specific communities (Dellarocas, 2011), and they are built upon users’ evaluations (Dellarocas, 2012).

This kind of systems is pivotal to the establishment of online transactions among distant strangers characterized by asymmetric information, in which buyers have little or no information about the goods they are going to buy, as in electronic markets like eBay. In these systems, comments or feedbacks provided by previous users are essential to promote trust among parties, to overcome information asymmetries, and to minimize fraud (Diekman et al., 2014). However, in online reputation, evaluations are largely opaque in many ways. Sources are unknown, as well as their metrics, meaning that what someone rates as good can be below average for someone else, and this is especially true for ratings, like stars. Comments can be misleading too, or, even worse, fraudulent, because interested targets or their competitors can artificially manipulate reviews, and thus portray a very different situation (Matzat & Snjders, 2012).

In spite of these problems, individuals rely heavily on others’ reputations and evaluations, even when these are ambiguous and non transparent, as stars in online reputation systems (such as Tripadvisor). Chevalier and Mayzlin (2006) found that evaluations, both in the form of written reviews and average star rankings, affected book sales by two online booksellers (and that negative and positive reviews had asymmetric effects on consumers’ behavior). Reputation systems are used in a variety of different contexts, from philanthropy to science (Masum & Tovey, 2011), but they are all based on ambiguous, anonymous and usually aggregated evaluations.

The aim of this work is to explore the extent to which opaque reputations in the form of stars might support cooperation in a strategic game. We consider both sides of evaluations, introducing reputation-based partner selection in two variants: weak and strong. Given our interest in understanding ambiguous reputations, pervasive in online environments, we choose to run web-based experiments using the online labor market Amazon Mechanical Turk. To avoid confounding factors related to online transactions (prices, goods, sellers’ features), to measure individuals’ cooperative attitudes we decide to use a standard one-shot Prisoner’s Dilemma (Nowak, 2006; Perc & Szolnoki, 2010; Capraro, 2013; Rand & Nowak, 2013).

In order to single out the effects of evaluations and partner selection on individuals’ behaviors, we designed three different studies. In the first experiment, we implemented a between-subjects design in order to understand how ambiguous evaluations (in the form of a grade obtained in a previous non-specified test) are taken into account, and how subjects use such an ambiguous evaluation system when assessing their partners’ behaviors. In the second and in the third study, we investigated the effect of knowing that one will be evaluated on one’s own behavior, by increasing the consequences of being evaluated, ranging from none to the possibility of being selected for another round of the game by a third-party knowing only the person’s reputation (see Methods for more details).

Our results provide evidence of the importance of reputation even when it is opaque. More specifically, we report four major results: (i) people cooperate much less with low-reputation partners than with medium or high reputation partners, even if they do not know how that reputation was acquired, (ii) when given the opportunity to select a partner knowing only his or her reputation, people tend to select partners with high reputation; (iii) individuals use their own behavior as a baseline for evaluation, disregarding absolute value of actions; (iv) knowing that they will be evaluated by their partner and that a third party will have the opportunity to select them as future partners has the effect to turn a substantial proportion of defectors into cooperators.

In sum, even when opaque and uncertain, reputation affects decision making: bad reputation works as a signal of anti-social behavior and, in combination with partner selection, promotes cooperative choices in digital environments.

2 Methods

We conducted a series of three studies recruiting subjects through the online labour market Amazon Mechanical Turk (Mason & Suri, 2014; Paolacci & Chandler, 2014; Rand, 2012). Here we report the experimental design of each of the three studies. Full instructions are reported in the Appendix.

2.1 Study 1

In this study we were interested in exploring whether opaque reputation is taken into account when interacting with someone, and how people apply an ambiguous reputation system when assessing their partner’s behavior. After entering their TurkID, subjects were randomly assigned to one of seven conditions. In the baseline condition, subjects were randomly matched to play a standard Prisoner’s dilemma (PD) game. Specifically, each subject was given $0.10 and had to decide whether to keep it (i.e., defect) or give it to the other player (i.e., cooperate). In the latter case, the $0.10 would be multiplied by 2 and earned by the other player. After reading the instructions, subjects were asked four general comprehension questions in random order. Subjects failing any of the comprehension questions were automatically excluded from the survey. Those who answered all comprehension questions were directed to the “decision screen”, in which they could select either “keep” or “give”, by means of appropriate buttons.

Subjects in the low reputation condition played the same PD game as in the baseline condition, but, before making their choice, they were told that the person they were matched with had participated in a previous test (without receiving any information about the kind of test), in which he or she was rated 1 out of 5 stars. This information was initially presented in the “instructions screen” and then made salient in the “decision screen”. In reality, there was no previous test and, to compute the payoffs, we paired subjects at random. The neutral evaluation condition was similar to the low reputation condition, with the only difference that subjects were told that they were matched with a person who was rated 3 out of 5 stars in a previous test. In the high reputation condition subjects were told that they were matched with partners who was rated 5 out of 5 stars in a previous test.

Subjects in the evaluated condition played the same PD as in the baseline condition, but, before making their choice, they were informed that their choice would be communicated to the other subject, who would be asked to rate it from 1 to 5 stars. This procedure was real, as cooperators were paired with subjects in the evaluate cooperator condition below, and defectors were paired with subjects in the evaluate defector condition below.

In the evaluate cooperator condition subjects first played the PD and then were informed that their partner had cooperated. At this stage, players were given the opportunity to rate the other subject’s action from 1 to 5 stars. Finally, the evaluate defector condition was similar to the evaluate cooperator condition, with the only difference that subjects were required to rate the behavior of an opponent who defected.

2.2 Study 2

As it will be shown in the Results section, the average cooperation in the evaluated condition of Study 1 is statistically the same as the average cooperation in the baseline condition, suggesting that the opportunity of being evaluated does not affect subjects’ choices in one-shot PD games, when the resulting reputation has no real consequences. We designed Study 2 in order to understand why the reputation threat had had no effect on individuals’ behaviors in Study 1. We introduced a light manipulation to the setting used in the previous study, in which we informed players that their decisions (to cooperate or to defect) would be communicated to the other player who could give a rating going from 1 to 5 stars. Here, subjects were also told that ratings were collected with the purpose of creating a rank of Turkers among which to select players with the highest ranks for further participation in a particularly rewarding task. To build a reputation we used the data collected in the evaluate cooperator and evaluate defector conditions of Study 1. Specifically, each cooperator in this condition was randomly assigned to a subject in the evaluate cooperator condition and was assigned the evaluation given by this particular subject. Similarly with defectors.

2.3 Study 3

As it will be shown in the Results section, the light increase in the consequences of the evaluation experimented in Study 2 does not lead to an increase in cooperative behavior. The aim of Study 3 is to test whether a stronger form of partner selection would increase cooperative choices. To this end, we employed a two-stage game in order to test for differences in cooperation levels between the first and the second game. After entering their TurkID, subjects were randomly assigned to either of two conditions. In the Random+Evaluated condition, the first stage consisted of a standard PD (as in the baseline condition in Study 1) played with a randomly selected partner. The following stage, instead, was divided in three parts: a game part, an evaluation part, and a selection part. In the game part, subjects played another PD, neutrally framed, with a randomly selected person, denoted Person A, but they were told that the experimenter, in the next part of the stage (i.e., the evaluation part), will communicate their decision to another person, Person B (different from Person A), who was in charge of assigning subject’s behavior a grade ranging from 1 to 5 stars. Subjects were also told that, in the third part (i.e., the selection part), another player (Person C) was given a list of 5 subjects (including themselves), each characterized by a different grade, from which they could choose their partner for playing a PD. Subjects were told that, in case they were selected by Person C, they would be playing another round of the PD with Person C. In reality, there was no other round. So, the total payoff of a subject was given by the sum of the payoffs obtained in the two PDs. Complete information about the three parts of Stage 2 was given all together at the beginning of the stage itself, and two comprehension questions (in addition to the four comprehension questions asked in Stage 1) were asked before subjects were allowed to make their decision.

The other treatment, Choose+Evaluated, was again in two stages. In the first stage, after reading the instructions of the PD, subjects were told that they were grouped with five other subjects, each of whom was characterized by a different number of stars obtained in a previous unspecified test. Subjects were asked to select the subject with whom they wanted to play. In reality, this selection procedure was fictitious, and subjects played with a randomly selected subject playing in the same condition, regardless of their selection. After the choices were made, subjects enter the second stage, which was exactly the same as in the Choose+Evaluated condition.

3 Results

A total of 962 subjects, located in the US, passed the comprehension questions and participated in our three studies. This corresponds to about 59% of the total number of subjects: about 41% of subjects failed the attention check, and were automatically excluded from the survey. This is in line with previous studies using similar strategic situations. For instance, Capraro, Jordan and Rand (2014) report 32% of subjects failing a very similar attention test. To avoid multiple observations from the same subject, each time we found a subject identified with the same IP address and/or the same TurkID, we kept only the first decision and eliminated the rest. As a consequence of this, the 962 subjects that we analyze are distinct in all measurable variables. Subjects failing the comprehension questions in one study were allowed to participate in the subsequent studies.

3.1 The effect of partner’s opaque reputation on cooperative behavior

We begin by analyzing how information about the other person’s (opaque) reputation is taken into account when interacting with them. To this end, we analyze the data of the baseline (N = 96), the low reputation (N = 91), the neutral reputation (N = 87), and the high reputation (N = 82) conditions of Study 1. Results, summarized in Figure 1, show that subjects cooperated much less with partners with low reputation (one star out of five) than with the others (Rank sum test. Low reputation vs baseline: Z = −3.41, p = 0.0006, effect size = 29%; low reputation vs neutral reputation: Z = −3.30, p = 0.0010, effect size = 29%; low reputation vs high reputation: Z = −3.22, p = 0.0013, effect size = 28%;), even if they had no information about the way in which this reputation was acquired. However, there is no statistically significant difference between the rate of cooperation in the baseline condition and that in the “neutral” (Rank sum test: Z = 0, p = 1) and “high” (Rank sum test: Z = 0.05, p = 0.9601) reputation conditions. Thus, low reputation is a signal of anti-sociality, but high reputation is not a signal of pro-sociality.

Figure 1: Comparison among levels of cooperation when there is no information about the partner (baseline), or when they are ranked as having a high reputation (5 stars out of 5), a neutral one (3 stars) or a low reputation (1 stars). Error bars represent the standard error of the mean. Only low reputation seems to be informative for subjects, who do not cooperate with people with low reputation, while no differences in cooperation are found between the three other evaluations.

3.2 The use of opaque reputation to assess others’ behavior

Next, we analyze how subjects use opaque reputation to assess their partner’s behavior. To this end, we analyze the data of the evaluate cooperation (N = 95) and the evaluate defector (N = 93) conditions of Study 1. Results, summarized in Figure 2, show that, when asked to assign a rate ranging from 1 to 5 stars to their partner, subjects rated cooperators overwhelmingly higher than defectors. Specifically, the average grade of a cooperator was 4.91, while the average rate of a defector was 2.14 (Rank sum test: Z=10.76, p < .0001). Both defectors and cooperators gave cooperators very high rates. Indeed, the average grade of a cooperator when rated by another cooperator was 4.94, while the average grade of a cooperator when rated by a defector was 4.86 (Rank sum test: Z = 0.64, p=0.5222). On the other hand, cooperators evaluated defectors significantly worse than other defectors did: the average grade of a defector when rated by another defector was 2.80, while the average grade of a defector when rated by a cooperator was 1.48 (Rank sum test: Z = 4.72, p < .0001). The figure shows that some defectors were rated 4 or 5 stars. Data show that, in these cases, the evaluator was herself a defector. In other words, the maximum grade given to a defector by a cooperator was 3 stars. This means that evaluations were conditional on one’s own behavior, and not based on the absolute positive or negative value of subjects’ choices.

Figure 2: Distributions of evaluations of cooperators and defectors on a scale from 1 to 5 stars. Cooperators received very positive evaluations, in contrast with defectors’ grades.

3.3 The effect of being evaluated and external partner selection on cooperation

Next, we analyze the effect of being evaluated on cooperative behavior in two cases: when the evaluation phase is not followed by real partner selection; and when it is followed by external partner selection, that is, by the possibility of being selected by the experimenter for new studies. To this end, we analyze the data of the evaluated condition of Study 1 (N = 95) and Study 2 (N = 96). We compare the results of Study 2 with those of the evaluated and the baseline conditions in Study 1, although these experiments were conducted in different times, only to understand whether the light increase in the consequences of the evaluation in Study 2 is likely to produce relevant changes in cooperative behavior. Figure 3 summarizes the results and shows that neither treatments had a significant effect on cooperative behavior (Rank sum test. Baseline vs Evaluated: Z = −0.32, p = 0.749; baseline vs study 2: Z = −0.57, p = 0.453).

Figure 3: Percentage of cooperation for the evaluation without partner selection condition (evaluated, Study 1) and for the evaluation with possibility to be selected by the experimenter for another study (Study 2). Error bars represent the standard error of the mean.

3.4 The effect of being evaluated and internal partner selection on cooperation

Finally, we analyze the effect of being evaluated on cooperative behavior, when the evaluation is followed by internal partner selection, that is, the possibility of being selected by another subject for playing another round of the PD. To this end, we analyze the data of Study 3 (Random+Evaluated: N = 104; Choose+Evaluated: N = 123). Figures 4 and 5 show an increase of cooperation in the second PD game with respect to the first one, in both experimental conditions. Cooperation increased both when partners were randomly assigned (Random+Evaluated), and when subjects chose with whom to interact (Choose+Evaluated), suggesting that the opportunity of being selected as a partner in the next game increased cooperative choices.

To understand what drives this increase in cooperative behavior, we do a within-subject analysis looking at those subjects who changed strategy from the first PD to the second PD. In both experimental conditions, we find that virtually all of those subjects who cooperated in Stage 1 remained cooperators in Stage 2, whereas a substantial proportion of subjects who defected in Stage 1 became cooperative in Stage 2. Specifically, in the random+evaluated condition, we find that 55 subjects cooperated and 49 defected in the first PD. Among these cooperators, only 3 of them changed strategy and defected in the second PD. On the other hand, among the defectors, 40% of them (20 out of 49) changed strategy and cooperated in the second PD. Similarly, in the choose+evaluated condition, 69 subjects cooperated and 53 defected in the first PD. Among these cooperators, none of them changed strategy, that is, all of them cooperated also in the next PD. Among the defectors, about 30% (16 out of 53) changed strategy and cooperated in the second PD. Thus, in both cases, we found that the combination of reputation and partner choice was effective in turning a substantial proportion of defectors into cooperators.

Figure 4: Percentage of cooperative choices in the first and second stage of the Random+Evaluated condition. Error bars represent the standard error of the mean. In the second stage, subjects played the PD knowing that their choice would be evaluated by another person and that a third party could select them, for playing another round of the PD, from a list of five subjects, one for each possible grade. This significantly increased cooperative choices, compared to the first, completely neutral, PD.

Figure 5: Percentage of cooperative choices in the first and second stage of the Choose+Evaluated condition. Error bars represent the standard error of the mean. Even when choosing a partner, subjects did not show increases in cooperation until the second game, when they were evaluated and, possibly, selected for another game.

Finally, we investigated people’s preferences when they had the opportunity to choose a partner for playing the first PD (i.e., in the choose+evaluated condition). We find that the majority of people, but not all (75%), preferred to play with a partner with 5 stars. Those who decided to play with an opponent with less than 5 stars were significantly less cooperative than those who decided to play with an opponent with 5 stars (average cooperation: 35% vs 63%, Rank sum, p=0.022).

4 Discussion

In this experimental work, we focused on the role of evaluations on cooperative behaviors, disentangling the effects of evaluating others from those of being evaluated, and assessing the relative importance of partner selection in these processes. The role of reputation on cooperative behaviors in social dilemmas is well-established (Alexander, 1987; Nowak & Sigmund, 2005; Rand & Nowak, 2013; Nax, Perc, Szolnoki & Helbing, 2015), but less is known on the consequences of reputation format and presence or absence of partner selection, especially in the noisy and ambiguous online environment.

We tested whether a reputation, low, neutral or high, coming from a completely unknown source and acquired in an obscure situation (i.e., what we termed opaque reputation) was used in a web-based experiment by subjects who had to decide whether to cooperate or defect in a one-shot Prisoner Dilemma. Cooperation rates are affected by partners’ rankings (expressed as one, three or five stars out of five), but this effect is not symmetrical. Subjects rated with one star received significantly less cooperation than subjects with neutral (three stars) and high (five stars) reputations.

The preminence of bad evaluations over good ones is a distinguishing trait of human psychology (Baumeister et al., 2001), and it seems especially salient in social contexts. Anderson and colleagues (2011) show that negative gossip associated with neutral faces dominates longer in a visual discrimination task. In the online world, Chevalier and Mayzlin (2006) find that a negative review of a book has a stronger influence than a positive one, in online book sellers websites like Amazon. Our results are thus in line with previous findings. But they could also be due to the fact that there was less room for increases on the positive side of neutral, if some subjects were inclined not cooperate regardless of who their partner was.

In the second study, subjects were told that their actions would have been evaluated using the same opaque reputation system used in the first study, but our results showed that evaluation alone, without any actual partner selection, was not effective in increasing cooperative choices. Informing subjects that their partners had the opportunity to rate them did not make cooperation increase with respect to a baseline, not even when we told them that a general ranking of players with the possibility of participating in future rewarding tasks would be created. This result allows us to narrow down the effect of partner selection, while stressing the fact that cooperation is not enhanced by the simple awareness that someone evaluates us.

In order to understand better how and under which conditions partner selection is effective in promoting cooperation, we design a third study using a repeated game in which subjects were told that partner selection was real. Our results show an increase in cooperation levels between the first game (without evaluation and partner selection) and the second one, in which players were told that their behavior would have been evaluated and this evaluation transmitted to another potential player. The increase in cooperation was not due to “active partner choice”, because it happened also when partners were randomly assigned, therefore leading us to conclude that the combination of being evaluated and being selected explained the observed increase in cooperation. In the third experiment, we also observed that a subset of subjects selected partners with less than five stars and did not cooperate with them.

Although centered on digital reputation, our work adds to the literature on reputation systems but also to the general literature on cooperation. Studies on web-based reputation systems usually analyze real websites designed for enhancing trust in online transactions, in which real goods or services are exchanged (Dellarocas, 2006; Farmer and Glass, 2010). Our studies use a completely neutral setting, in which motivations related to individual preferences or needs do not enter decision making, but the main elements of reputation systems, i.e., evaluations and partner selection, are present. Showing that an opaque negative reputation can support cooperation complements evolutionary accounts that consider negative reputation as a powerful means for detecting and avoiding cheaters (Barkow, 1992; Giardini & Conte, 2012; Hess & Hagen, 2006; Nowak & Sigmund, 1998), a view supported also by several experimental findings (Anderson et al., 2011; Feinberg et al, 2012; Sommerfeld et al., 2008). Reputation alone, however, seems to have no effect when it is devoid of partner choice, showing that what is called the “threat of gossip” (Beersma & Van Kleef, 2011; Piazza & Bering, 2008) is effective only when consequences of reputation are evident.

Reputation plays a major role in human sociality, and it has been proposed as an explanation for the evolution of costly cooperation. In recent years, reputation has become central also in online systems, even if it is much less controllable and completely opaque. Our findings suggest that reputation, even if opaque, works in isolating defectors, but its value is conditional on subjects’ behaviors. Moreover, when partner selection is not effective, individuals do not become more cooperative only for the sake of getting a positive reputation, at least not in an anonymous online environment. Only when reputation is pivotal to selection does it leads individuals to change their behaviors and to cooperate. The behavioral switch is strong: the mere possibility of being selected for a new interaction turn about 35% of defectors into cooperators. This finding is interesting since it suggests yet another way to promote cooperative behavior in the field (see Kraft-Todd et al., 2015, for a recent review on interventions to promote cooperation in the field).

More generally, our experiments provide insights on the way in which humans use reputational information in uncertain environments such as online interactions. This has implications that exceed online markets and can be applied to several domains. For example, companies or universities, whose success is highly based on cooperation among their employees, might develop a reputational system, according to which colleagues that have been working together on the same project can rate one another.

Our experiments certainly have some limitations. Study 1 uses deception in three out of seven conditions; specifically, those in which subjects are told that their partners obtained a certain grade in a previous unspecified test. In reality there was no previous test. Study 3 uses deception when subjects are told that they could be selected for another round of the PD, depending on how their choice would be evaluated by a third party. In reality, although the evaluation procedure was real, there was no selection for other rounds. In general, the use of deceptive messages leads to a decrease of the effect size (when there is a true effect), driven by a proportion of subjects that may anticipate the fact that the manipulation is not real. Thus, the effect sizes that we found in Study 1 and Study 3 are likely to be a lower bound for the true effect sizes. Understanding the true sizes of these effects is a direction for future work. It is even possible that the asymmetric effect of reputation information on cooperative behavior (Study 1) is an artifact of the use of deceptive messages: subjects paired with high reputation partners may be more skeptical about the reality of the manipulation than those paired with low reputation partners. Although, as discussed above, this asymmetry is in line with previous studies, we cannot exclude that, in our case, it is driven by the use of deception and thus we leave this question for further research.