Reputational concerns, not altruism, motivate restraint when gambling with other people's money

It has been argued that improving one's reputation is a particularly important way that prosocial behavior can ultimately serve self-interest (Fehr & Fischbacher, 2003). Experimental work is consistent with this view. For example, Sylwester and Roberts (2010) had subjects play two rounds of an economic game. Subjects who were more generous in the first round were chosen by partners who were more generous in the second round. By this means, subjects who were more generous initially earned more overall. Sylwester and Roberts reasoned that more generous subjects had invested in their reputation.

Bereczkei, Birkas, and Kerekes (2007) attempted to distinguish the direct effect of reputational concerns on generosity from altruistic motives. Students were asked in the presence of their classmates to volunteer for charities. When commitments to volunteer were made publicly, so that classmates could observe them, the proportion of subjects who volunteered jumped from 25% to 50%. Thus, Bereczkei et al. inferred that the observed difference in generosity was due to reputational concerns. Similar findings of audience-dependent generosity have been obtained by other investigators, particularly among subjects who are especially self-interested ("egoistic"; Simpson & Willer, 2008) or socially manipulative ("Machiavellian"; Bereczkei, Birkas, & Kerekes, 2010).

There is also evidence that people will take opportunities to appear prosocial without actually being so. In the ultimatum games of Kagel, Kim, and Moser (1996), subjects who knew their own payoff from a deal would be tripled but who knew their opponent did not know this typically offered an even split. Thus, subjects behaved in a way that appeared fair (perhaps to increase the odds the opponent would accept the split) but was actually self-interested. In the dictator games of Dana, Weber, and Kuang (2007), subjects were more likely to choose a selfish split when the opponent did not know the payoffs. Hence, subjects acted less prosocially when their opponents could not evaluate how prosocial they were being. However, these findings are difficult to interpret with respect to reputation management, because in both studies, it is not clear that subjects could identify each other, and thus, reputational consequences were not clearly involved.

In the bulk of the experimental tasks described above, there was no uncertainty other than subjects' ignorance of other people's thoughts and plans. Real-life social dilemmas, on the other hand, are often uncertain. Taking advantage of other parties may or may not ultimately harm them. One class of social dilemma which captures this kind of risk is moral hazard (more precisely, "indirect moral hazard", Crosby, 1905, or "morale hazard", McLeman & Smit, 2006). A person is said to be under moral hazard when potential negative consequences of her decisions will be borne by a third party, regardless of whether the situation would be seen as morally charged according to psychological or philosophical notions of morality. Moral hazard has primarily been studied in the context of insurance (e.g., Abraham, DeLeire, & Royalty, 2010; Grossman, 1992; Quiggin, Karagiannis, & Stanton, 1993). The provision of insurance shifts losses from an individual to her insurance agency, increasing her incentives for risk-taking and therefore possibly making her less sensitive to risk. For example, people who are insured against floods are under moral hazard because they face less potential loss by purchasing property in a flood-prone area (e.g., Bagstad, Stapleton, & D'Agostino, 2007; Burby, 2001; Huber, 2004). However, moral hazard also arises in everyday situations. A person may be less careful to lock her friend's door than her own, because she does not experience the negative consequences of a robbery. Similarly, an HIV-positive person having sex with an HIV-negative partner has less to lose by neglecting to use a condom. More broadly, moral hazard is related to other means of exploiting the contributions of others, such as social loafing (Karau & Williams, 1993), as well as to differences in risk-taking between individuals and groups (Zajonc, Wolosin, Wolosin, & Sherman, 1969; Burnstein, Vinokur, & Trope, 1973; Gardner & Steinberg, 2005).

Recently, we (Bixter & Luhmann, 2014) experimentally demonstrated increased risk-taking under moral hazard. Subjects were asked to choose between a mixed gamble (e.g., a 70% chance of gaining $40 and a 30% chance of losing $40) and a sure gain (e.g., $15). Each trial had a different gamble; gambles came in various types. For standard gambles, subjects had to bear the full loss themselves if they chose the mixed gamble and lost. For shared gambles, half of any loss (e.g., $20 of the $40) would be borne by a third party, the subject's partner. Thus, shared gambles were more beneficial to the subject, but more harmful to the partner, than standard gambles. As in any moral-hazard situation, taking a shared gamble, compared to an equivalent standard gamble, meant reducing one's own potential loss by an amount equal to the potential loss inflicted upon someone else. This situation is similar to a dictator game, where one can increase one's own gain by reducing another person's gain by the same amount (Forsythe, Horowitz, Savin, & Sefton, 1994). Altruistic subjects, by refraining from shared gambles, could thus benefit their partner (by protecting her from loss) while forgoing extra gain to themselves. Other trials in Bixter and Luhmann (2014) involved matched gambles, which involved the same loss to the subject as the corresponding shared gamble (e.g., a 70% chance of gaining $40 and a 30% chance of losing $20) but no loss to the partner. A subject entirely insensitive to her partner's welfare should then treat matched gambles identically to shared gambles. The results of the study indicated that subjects were more likely to take shared gambles and matched gambles than standard gambles; in fact, no significant difference was found between the former two. That is, subjects' behavior was consistent with a total disregard for the welfare of others.

In the second experiment presented in Bixter and Luhmann (2014), we tried to increase subjects' prosocial behavior by decreasing the social distance between subject and partner—that is, by making subjects feel personally closer to their partner. Previous research (e.g., Jones & Rachlin, 2006) had found that people are more generous toward others to whom they feel closer. Though there was a suggestion that shared gambles were less attractive than matched gambles, this difference was not significant. Overall, the results were similar to that of the first experiment.

The present studies are another attempt to induce prosocial behavior in this kind of moral-hazard gambling task. We investigated if reputational concerns can reduce subjects' willingness to take shared gambles.

For both Study 1 and Study 2, our research protocol was approved by the Committees on Research Involving Human Subjects of Stony Brook University's Office of Research Compliance, and all subjects provided informed consent.

See http://arfer.net/projects/hazard for raw data, task code, and analysis code for both Study 1 and Study 2.

Four subjects accepted 2 or more catch gambles. These subjects were excluded from further analysis, as were all catch trials. The remaining sample was also evenly split by condition, with 17 subjects in each.

Overall, Anonymous subjects took 28% of standard gambles and 41% of shared gambles, whereas Visible subjects took 36% of standard gambles and 35% of shared gambles. See Figure 2 for a complete breakdown of choices collapsed across subjects.

We analyzed the effects of gamble type and experimental condition on gambling behavior with Bayesian mixed-effects logistic regression. We chose regression over ANOVA or t-tests to avoid treating numerical predictors, such as gamble probability, as categorical, and we chose logistic regression over linear regression because our dependent variable, whether the subject took the gamble, was dichotomous. We used Bayesian methods rather than null-hypothesis significance testing chiefly so that we could describe precisely what could be inferred about model parameters, properly accounting for all sources of uncertainty tracked by the model. Doing Bayesian data analysis is also consistent with growing dissatisfaction with standard uses of significance testing (e.g., Cohen, 1994; Gelman & Stern, 2006; Kruschke, 2010; Wagenmakers, Wetzels, Borsboom, & van der Maas, 2011; Cumming, 2014).

For the dependent variable, we coded taking the gamble as 1 and taking the sure gain of $15 as 0. The model had a single per-subject random effect drawn from a normal distribution with mean 0 and standard deviation σ, where the hyperparameter σ was given a prior density uniform on [1/10, 100]. Fixed effects are listed in Table 2. All fixed effects were assigned improper uniform priors. Notice that the model does not include parameters for loss (or expected value). This means that the values of b_g_shared_anon and b_g_shared_visible will include the effect of reduced loss to the subject as well as increased loss to the partner, so our analyses will not answer the question of whether subjects would make prosocial choices at no cost to themselves. (The effect of reduced loss alone is examined with matched gambles in Study 2.)

Model parameters were estimated with the Markov chain Monte Carlo sampler Stan (http://mc-stan.org). Markov chain Monte Carlo is a numerical method for estimating the parameters of a Bayesian model. In accordance with the recommendations of Gelman, Carlin, Stern, and Rubin (2004), eight chains were run with random initial values, and for each chain, 250 adaptive burn-in iterations were discarded and 250 non-adaptive sampling iterations were kept, for a total of 2,000 samples per parameter. This resulted in a Gelman-Rubin diagnostic (Gelman & Rubin, 1992) of less than 1.05 for all parameters.

Parameter estimates are shown in Figure 3. With posterior samples in hand, to answer the question of how the experimental condition affected subjects' willingness to take shared gambles, we computed the posterior probability that b_g_shared_anon is greater than b_g_shared_visible. This inequality was satisfied by every sample, so the posterior probability that it is true exceeds 99%. We conclude that subjects were indeed less willing to take shared gambles, exposing their partner to risk, when their reputation was at stake. In fact, in the Visible condition, it is more likely than not that subjects preferred standard gambles to shared gambles (specifically, the posterior probability of b_g_shared_visible being less than 0 is 68%), despite how a standard gamble has twice the potential loss to the subject as a shared gamble, suggesting subjects preferred increased monetary losses to themselves over reputational fallout.

Two subjects accepted 2 or more catch gambles. These subjects were excluded from further analysis, as were all catch trials.

Overall, subjects took 28% of standard gambles, but 42% of matched gambles, 41% of local shared gambles, 46% of remote shared gambles, and 48% of deceit gambles. See Figure 4 for a complete breakdown of choices collapsed across subjects.

The median difference between social-distance ranks of the local and remote partners was 3, with the local partner ranked closer. This difference, being on a 12-point scale, was smaller than might have been expected; however, the rated social distance of the remote partner appeared to suffer from a strong ceiling effect, with 84% of subjects selecting ranks 10, 11, or 12.

We analyzed the effects of gamble type and rated social distance on gambling behavior with a Bayesian mixed-effects logistic regression, as in Study 1. The fixed effects for the Study 2 model are shown in Table 3. Since social-distance rank of the remote partner appeared limited by a ceiling effect, this variable was not included as a predictor. We generated posterior samples as in Study 1.

Our subjects' greater willingness in Study 2 to take remote shared gambles than local shared gambles is consistent with past work on social distance. As the social distance between a subject and a third party decreases, generosity increases (Jones & Rachlin, 2006). Manipulations to decrease social distance that have increased generosity in economic games include allowing subjects to see each others' faces (Bohnet & Frey, 1999) and revealing last names (Charness & Gneezy, 2008). By contrast, our earlier study using a similar moral-hazard paradigm (Bixter & Luhmann, 2014) found no significant effect of whether subjects had personally met their partner. Our strengthened design, which involved an opportunity to become socially closer to the local partner (the RCIT) and an especially distant remote partner for contrast, found an effect of social distance that Bixter and Luhmann (2014) did not. The importance of social distance for economic decision-making is further suggested by its effects on variables other than generosity. Ziegler and Tunney (2012), for example, found that subjects made more patient choices on the behalf of third parties the less genetically related the third party was to the subject.

We found that most kinds of shared gambles were preferred to standard gambles. This is the basic moral-hazard effect of people weighting losses to others less than losses to themselves, replicating Bixter and Luhmann (2014). On the other hand, Visible subjects in Study 1, for whom reputational concerns were particularly salient, may have preferred standard gambles to shared gambles, suggesting that a personal loss may be preferred to facing blame for causing a loss to somebody else.

Our most surprising finding was subjects' preference in Study 2 for deceit gambles over matched gambles (and their less likely preference for remote shared gambles over matched gambles). Effectively, these shared losses were treated as gains. Possible explanations include misery loving company (e.g., Cooper & Rege, 2011); a dissonance effect creating antipathy toward partners (Glass, 1964); an antisocial desire to punish the remote partner (Brañas-Garza, Espin, Exadaktylos, & Herrmann, 2014) by harming her reputation, on the presumption that the remote partner has taken shared gambles harming the local partner; or a view of the deceit gambles as special opportunities that should be taken advantage of.

There are various ways a person can behave prosocially toward another party. A person can perform an action that incurs a personal cost while simultaneously increasing the welfare of the third party. The real-world action that best represents this form of prosociality is charitable donation. However, people can also behave prosocially toward another party by refraining from committing an action that benefits themselves but decreases the welfare of the third party. Moral hazard is a situation that captures this latter scenario, because increased risk-taking under moral hazard can lead to personal benefits while exposing another party to losses that she is not directly responsible for.

Moral hazards are pervasive throughout the real world and can lead to large societal costs (e.g., Blanchard-Boehm, Berry, & Showalter, 2001; Okamoto, 2009). As a result, it is important to gain a better understanding of the processes that affect people's decision making when they are under moral hazard. Furthermore, research in the laboratory has overwhelmingly focused on the former class of prosocial behavior, such as allotments in ultimatum and dictator games (Camerer & Thaler, 1995; Charness & Gneezy, 2008; Hoffman, McCabe, & Smith, 1996). The experiments of Leliveld, van Beest, van Dijk, and Tenbrunsel (2009), however, suggest that people care more about fairness and less about self-interest when losses rather than gains are at stake, and Zhou and Wu (2011) found a higher rate of altruistic punishment in ultimatum games with losses than in ultimatum games with gains. Clearly, in order to achieve a better understanding of the origins of prosociality, it is necessary to study instances of prosocial behavior across a variety of contexts and situations. The results of our studies suggest that reputational concern is one available method to increase prosocial behavior when a decision-maker is under moral hazard.

As just discussed, there are many kinds of situations to which questions of altruism and prosocial behavior apply. Our studies focused on moral hazard, which is only one such kind of situation, and therefore they do not say much about altruism in other situations. Furthermore, it is believable that people behave differently in different contexts of moral hazard. For example, perhaps people are less prosocial toward insurance agencies than individuals, or perhaps people are more prosocial when exploitation would endanger human health rather than finances. Future research would need to examine such moderating effects directly in order to discover their influence.

In our studies, subjects were placed in an implicit position of power by being offered opportunities to exploit partners. Such an opportunity is part of the definition of moral hazard, but our task differed from most real-life moral-hazard situations in that exploitation was presented as an explicit option. Thus, by a sort of demand characteristic driven by the perceived legitimacy of exploitation, subjects may have been more willing to take shared gambles than they would been in an equivalent real-life situation. Fortunately, any such demand characteristic would apply equally to the two conditions of Study 1 and to all the kinds of shared gambles in Study 2, meaning that the contrasts of chief interest cannot have been affected.

Our studies used an unusual incentive scheme that paid subjects real money if they had reached an unspecified threshold of in-task earnings, as opposed to, for example, paying subjects in direct proportion to in-task earnings, or choosing a random trial to realize. This scheme has the advantage of making every trial count without allowing subjects to exploit dependencies between gambles or "play with the house's money". On the other hand, it complicates comparison to other studies. Future research may benefit from standardizing on compensation.

In Study 2, although we attempted to create social closeness between subjects with the RCIT, and the distance-ranking task confirmed that subjects generally saw the remote partner as more distant than the local partner, subjects still did not feel very close to the local partner. After all, a short in-lab exercise between strangers cannot be expected to forge a social bond comparable in magnitude to long-term friendships or romantic relationships. As mentioned in the introduction, research on social discounting (e.g., Jones & Rachlin, 2006) shows that people are willing to make larger sacrifices to benefit a third party the closer they are to the third party. Thus, our results leave open the possibility that subjects will be more altruistic under moral hazard when the third party whose resources they can gamble with is a true intimate.

In previous studies, subjects could directly, materially gain from maintaining a reputation (Sylwester & Roberts, 2010; Kagel et al., 1996; Simpson & Willer, 2008) or at least expected to repeatedly interact with the audience in the future (Bereczkei et al., 2007; Bereczkei et al., 2010). In our studies, by contrast, the advantage to maintaining a good reputation in the eyes of partners was less clear, since subjects had no reason to believe they would interact with their partners again. We suggest, then, that people inherently value reputation.

We found that subjects were self-interested but not transparently so. This observation is an example of how prosocial behavior is not, in general, incompatible with the notion that organisms are largely self-interested. On the contrary, self-interest can be useful as an explanation for prosocial behavior.