Exploiting the guilt aversion of others: do agents do it and is it effective?

Abstract

The general idea of guilt aversion is that agents may be motivated to avoid letting others down, even at the expense of their own material payoff. Several experimental studies have documented behavior that is consistent with agents exhibiting guilt averse motivations in social interactions. However, there are strategic implications of guilt aversion, which can impact economic outcomes in important ways, that have yet to be explored. I introduce a game that admits the possibility for agents to induce guilt upon others in a manner consistent with the method posited by Baumeister et al. (Psychol Bull 115:243–267, 1994). This game enables me to experimentally test whether agents attempt to exploit the guilt aversion of others by inducing guilt upon them, and whether agents are actually susceptible to this exploitation. Additionally, the design enables me to test whether agents exhibit higher degrees of trust when they are given such an opportunity to exploit the guilt aversion of others. The data suggest that agents do not attempt to fully exploit the guilt aversion of other agents by inducing guilt upon them; however, the data suggest that agents would have been susceptible to guilt induction.

1 Introduction

The results from a growing body of experimental literature suggest that economic agents may not be solely motivated to maximize their own material payoffs. One example of such a “non-selfish” behavioral motivation is guilt aversion. The general idea of guilt aversion is that an agent would suffer disutility, in the form of guilt, from hurting or letting down another agent, relative to that other agent’s expectations;¹ thus, a guilt averse agent may be motivated to avoid hurting or letting down that other agent, even at the expense of his/her own material payoff, to assuage the guilt feeling. Behavior consistent with agents exhibiting guilt aversion has been documented in several experimental studies. For example, Dufwenberg and Gneezy (2000), Charness and Dufwenberg (2006, 2010), Bacharach et al. (2007), Reuben et al. (2009), and Attanasi et al. (2013) find evidence of guilt aversion using variations of two-player experimental “trust” games (Berg et al. 1995).² Specifically, these papers show that the amount of money the trustee (second mover) returns to the trustor (first mover) is positively correlated with how much the trustor expects to get back (or how much the trustee thinks the trustor expects back). The idea being that guilt averse trustees give back more (forgo their own material payoff) the more their trustor expects back to avoid the guilt that would result from failing to meet the trustor’s expectations.

In light of the experimental evidence of exhibited guilt aversion, it is important to consider the richer set of interpersonal strategic implications that can arise when agents are motivated by guilt aversion. In particular, the guilt aversion of one agent can influence the behavior of other agents in important ways. In certain social interactions, the possibility may arise for agents to behave opportunistically and exploit the guilt aversion of others. Charness and Dufwenberg (2006), in their concluding remarks, point to such a possibility by raising the question, “do people manipulate the guilt aversion of others in self-serving ways?” (p. 1595). Given the opportunity and incentive, agents could attempt to influence the behavior of guilt averse others by strategically inducing guilt upon these agents, i.e., increasing the amount of guilt they would feel. Consequently, guilt averse agents may be more motivated to respond in kind to avoid hurting or letting down those agents that had induced guilt upon them. These interpersonal implications of guilt aversion can impact strategic decision making and, consequently, economic outcomes in important ways that have yet to be explored.

The goal of this study is to explore some of these interpersonal implications of guilt aversion. Specifically, this paper introduces an experimental design aimed at shedding light on the following three questions: First, do economic agents attempt to exploit the guilt aversion of other agents in self-serving ways by strategically inducing guilt upon those other agents? Second, is strategic guilt induction effective at influencing the behavior of other agents, i.e., are agents susceptible to this exploitation? Third, does having the opportunity to induce guilt upon other agents impact strategic behavior?

Although previously unexplored in the economics literature, the interpersonal implications of guilt aversion have been studied and documented in the psychology literature (Vangelisti et al. 1991; Baumeister et al. (BSH henceforth) 1994, 1995; Tangney and Fischer 1995; De Hooge et al. 2011). In particular, BSH (1994) argue that one of the primary functions of guilt is to motivate others to behave in a more desirable way. In their study, BSH (1994) note that “we observed ample evidence of the hypothesized function of guilt as an interpersonal influence technique: People induced guilt to get another person to comply with their wishes.” (p. 249) Similarly, Vangelisti et al. (1991) argue that people induce guilt “primarily to achieve their own end-to persuade their listeners to do or not to do something.” (p. 33) These psychology studies provide foundational insights regarding the interpersonal functions of guilt in social relationships by drawing conclusions from non-incentivized personal narratives and surveys. However, these functions of guilt may not be restricted to social interactions; guilt may also function as an “interpersonal influence technique” in strategic economic interactions. An incentivized experimental game provides a suitable platform for investigating these interpersonal implications of guilt aversion in economic settings.

Before I proceed, I pause to highlight some of the economic settings where strategic guilt induction could be relevant in terms of influencing the behavior of others and, consequently, impact outcomes. In contracting environments, guilt induction may allow a disadvantaged party to influence the behavior of an advantaged counterparty. For instance, a contracted firm that had made relationship-specific investments could possibly thwart opportunist re-contracting and hold-up by conveying to the counterparty firm the loss in profits associated with such a hold-up. In the workplace, managers could induce guilt upon employees to mitigate shirking by conveying to employees how their sub-standard effort adversely affects other employees.³ In academia, assistant professors could possibly induce guilt upon journal editors to get a more timely review decision on a submitted paper by gently informing the editor, at the time of submission, that his/her tenure review is rapidly approaching and a lengthy review period could hinder his/her tenure prospects.⁴ Many economic settings, like these mentioned, permit the possibility to induce guilt upon others. Hence, a deeper understanding of the strategic interpersonal implications of guilt aversion is required to better ascertain how the guilt aversion of agents will impact outcomes in such settings; the insights gleaned from this paper are intended to help with this understanding.

To investigate whether agents strategically induce guilt upon others and its potential effectiveness, it is crucial to first identify how agents can attempt to induce guilt upon others. For this, I draw foundational insights from BSH (1994), who posit the following method for how people induce guilt in others: “If Person A wants Person B to do something, A may induce [more] guilt in B by conveying how A suffers [how much A is let down] over B’s failure to act in the desired fashion” (p. 247). In regards to strategic economic settings, this proposed method by BSH (1994) would correspond to Person A conveying how low his/her payoff would be, i.e., how he/she suffers, as a result of Player B choosing an undesirable action toward Player A.⁵ Note that this method for inducing guilt implicitly requires that (i) Person A has private information about his/her own payoff and the degree to which he/she may suffer from Player B’s action, and (ii) Person A has the possibility to convey such private information to Person B.⁶ Previous studies that have investigated guilt aversion in strategic settings mostly consider variations of 2-player “trust” games that do not feature either of these properties.⁷ Hence, a new game is warranted that provides a rich enough strategic structure to allow agents the opportunity to induce guilt upon others.

In this paper, I employ a novel experimental design that uses a 2-player, binary choice trust game featuring both private payoff information and an opportunity to convey this private information. In the game, the privately informed first mover (Player A) is effectively given an opportunity to convey to the second mover (Player B) how low his/her payoff would be if Player B fails to act in the desired fashion.⁸ This game allows me to then derive testable hypotheses regarding whether Player As attempt to strategically induce guilt upon Player Bs, and whether Player Bs are susceptible to strategic guilt induction. Additionally, I show that an important artifact of this game is the ability to identify behavior consistent with guilt aversion without having to elicit beliefs. The experimental design also includes a second, related trust game that does not feature an opportunity for Player As to induce guilt upon Player Bs. This second game provides a baseline trust measure for Player As, which then allows me to derive a testable hypothesis regarding whether Player As are more trusting of Player Bs when Player As have an opportunity to induce guilt upon Player Bs.

The experimental data seem to suggest that Player Bs are susceptible to the guilt induction of Player As. However, the data reveal little evidence that Player As are attempting to induce guilt upon Player Bs, à la BSH (1994). Interpreted differently, the data reveal that Player As are not attempting to fully exploit the guilt aversion of Player Bs by inducing guilt upon them; although, had they done so, it would have been effective at increasing the likelihood of Player B choosing the desired action. Furthermore, I find evidence of marginally significantly higher trust rates by Player As in the trust game where Player As have the opportunity to induce guilt upon Player B, compared to a trust game with no such opportunity.

Furthermore, I show that the derived hypotheses are consistent with predictions of the formal model of “simple” guilt developed by Battigalli and Dufwenberg (2007) (B&D henceforth).⁹ I also show that in the game considered, effective guilt induction can be supported in equilibrium under the B&D framework. The paper proceeds with the experimental design and hypothesis development in Sect. 2. I outline the experimental procedure in Sect. 3. I present and discuss the results in Sect. 4, and Sect. 5 concludes.

2 Experimental design

I begin this section by first introducing the two trust games around which the research hypotheses are developed and the experimental design is based. I refer to both games as trust games because they feature a payoff structure indicative of the broader class of trust games.¹⁰ Trust games, in general, allow for the possibility of guilty feelings, which make them a suitable platform for developing and testing the hypotheses of this study relating to the strategic implications of guilt aversion.

2.1 Experimental trust games

2.1.1 Uncertain payoff trust game: $\Gamma _{\mathrm{UPT}}$

$\Gamma _{\mathrm{UPT}}$ is a 2-player, sequential move game. $\Gamma _{\mathrm{UPT}}$ begins with the first mover, Player A, choosing between In or Out. If Player A chooses Out, then the game ends; Player A receives a payoff of 6, and Player B receives a payoff of 2. If Player A chooses In, then Player B is called upon to move. Player B must choose between Left or Right. If Player B chooses Right, then the game ends; Player A receives a payoff of 10, and Player B receives a payoff of 4. If Player B chooses Left, then the game ends; Player A receives a payoff of X,  and Player B receives a payoff of 6. X is a random variable where $\hbox {prob}(X=0)=1/2$ and the $\hbox {prob}(X=6)=1/2$. At the start of the game, the distribution of X is known to both players. The extensive form of $\Gamma _{\mathrm{UPT}}$ is depicted below in Fig. 1.

Fig. 1

Extensive form of $\Gamma _{\mathrm{UPT}}$

2.1.2 Private payoff trust game: $\Gamma _{\mathrm{PPT}}$

$\Gamma _{\mathrm{PPT}}$ features a similar strategic structure and payoff structure to those of $\Gamma _{\mathrm{UPT}},$ with two important differences. First, $\Gamma _{\mathrm{PPT}}$ features an opportunity for Player A to become privately informed about the value of X. Second, $\Gamma _{\mathrm{PPT}}$ features an additional stage where Player A has the opportunity to credibly convey his/her private information about the value of X to Player B. $\Gamma _{\mathrm{PPT}}$ begins analogously to $\Gamma _{\mathrm{UPT}}$ with Player A first choosing between In or Out. If Player A chooses Out, the game ends; Player A receives a payoff of 6, and Player B receives a payoff of 2. If Player A chooses In, Nature then decides whether Player A becomes privately informed about the value of X. With $\hbox {prob}=4/5,$ Nature Reveals (Rev) the value of X to Player A, and with $\hbox {prob}=1/5,$ Nature does Not Reveal (Not Rev) the value of X to Player A.

If the value of X is revealed to Player A, then an additional stage arises where Player A must decide whether to credibly Convey (C) on Not Convey (NC) the value of X to Player B before Player B gets the move. Upon getting the move, Player B must then decide between Left or Right. Analogous to $\Gamma _{\mathrm{UPT}}$, if Player B chooses Right, then the game ends; Player A receives a payoff of 10, and Player B receives a payoff of 4. If Player B chooses Left, the game ends; Player A receives a payoff of X,  and Player B receives a payoff of 6, where, again, $\hbox {prob}(X=0)=1/2$ and the $\hbox {prob}(X=6)=1/2$. The extensive form of $\Gamma _{\mathrm{PPT}}$ is depicted in Fig. 2. To simplify the extensive form, the two moves by Nature—determining the value of X and determining whether the value of X is revealed to Player A—have been combined into one move.

Fig. 2

Extensive form of $\Gamma _{\mathrm{PPT}}$

If players are “selfish”, i.e., act to maximize their own material payoff, then the unique equilibrium outcome in both $\Gamma _{\mathrm{UPT}}$ and $\Gamma _{\mathrm{PPT}}$ is Player A chooses Out.¹¹ The inclusion of private information and the additional conveyance stage for Player A in $\Gamma _{\mathrm{PPT}}$ has no impact on the equilibrium outcome assuming selfish players. However, it is exactly these two features of $\Gamma _{\mathrm{PPT}}$ that will allow me to derive testable hypotheses regarding whether agents attempt to induce guilt and its subsequent effectiveness.

Before I proceed, I first highlight two important features of $\Gamma _{\mathrm{PPT}}$, and discuss the motivation for including these features, which will be relevant for the upcoming derivation of the research hypotheses and the application of the B&D model of guilt. First, Player A must choose between In or Out before possibly becoming informed about the value of X (and Player B is informed of this timing). Not informing Player A of the value of X prior to the In/Out decision eliminates any possible signaling value, from the perspective of Player B, regarding the value of X inferred from Player A’s In/Out decision. The second feature, is that the value of X is only revealed to a Player A who chooses In with $\hbox {prob}=4/5$. The motivation for including this uncertainty regarding the revelation of X to Player A is the following: If the value of X is not conveyed to Player B, then Player B is unable to perfectly distinguish between whether: (i) the value of X was not revealed to Player A, or (ii) the value of X was revealed to Player A, and Player A chose to Not Convey. The resulting implication is that if the value of X is not conveyed to Player B, then Player B’s expectation of the value of X will be strictly greater than zero and strictly less than six; the strict inequalities play an important role in the hypothesis development, discussed in Sect. 2.2 below.

More precisely, ${\widehat{m}}_{A}\in [1,5]$ where ${\widehat{m}}_{A}$ denotes Player B’s expectation of X,  conditional on the value of X not being conveyed.¹² In calculating ${\widehat{m}}_{A},$ Player B must think about the relative probabilities that: (i) Player A learned $X=0$ and chose Not Convey, (ii) Player A learned $X=6$ and chose Not Convey, and (iii) Player A did not learn the value of X. Although these probabilities are unobservable to the researcher and, thus, the exact value of ${\widehat{m}}_{A}$ is unobservable, it is possible to derive bounds on ${\widehat{m}}_{A}.$ Specifically, the largest expectation that Player B could hold regarding the value of X occurs when he/she thinks that only a Player A who learned that $X=6$ would choose to Not Convey. In this case, ${\widehat{m}}_{A}$ is bounded above by $\frac{1}{3}\cdot E[X]+\frac{2}{3}\cdot 6=5.$ Here, $\frac{1}{3}$ and $\frac{2}{3}$ represent the updated probabilities, via Bayes’ rule, that Player A did not learn the value of X,  and Player A learned that the value of $X=6$, respectively. By a similar argument, the smallest expectation that Player B could hold regarding the value of X occurs when he/she thinks only a Player A who learned $X=0$ would choose to Not Convey. In this case, ${\widehat{m}}_{A}$ is bounded below by $\frac{1}{3}\cdot E[X]+\frac{2}{3}\cdot 0=1.$ Therefore, regardless of Player B’s beliefs at his/her information set where the value of X is not conveyed, it must be that ${\widehat{m}}_{A}\in [1,5].$

2.2 Research hypotheses

The first motivation of this study is to investigate whether agents attempt to exploit the guilt aversion of others by inducing guilt upon them. Recall that BSH (1994) posit that a person can induce guilt upon another by conveying to that person how one suffers over that person’s failure to act in the desired fashion. Let us consider how this method applies to $\Gamma _{\mathrm{PPT}}.$ Conditional on choosing In, Player A would “desire” Player B to choose Right, as it yields him/her a payoff of 10 compared to a payoff of $X<10$ if Player B were to choose Left. Hence, X measures the extent to which Player A would “suffer” from Player B’s failure to choose Right. Given the opportunity, Player A could attempt to induce guilt upon Player B by “conveying” to Player B a low value of X, i.e., by conveying to Player B how much he/she would suffer if Player B were to choose Left.

Conditional on having chosen In in $\Gamma _{\mathrm{PPT}}$ and having the value of X revealed, Player A must decide whether to credibly convey the value of X to Player B (i.e., Player A cannot lie about the value of X). In the case where $X=0$ is revealed, if Player A chooses to Convey $X=0,$ then Player B will know that if he/she chooses Left, Player A will receive a payoff of $X=0$. Whereas, if Player A chooses to Not Convey $X=0$, then Player B will think that if he/she chooses Left, Player A will receive a payoff of ${\widehat{m}}_{A}\in [1,5].$ Analogously, in the case where the value of $X=6$ is revealed, if Player A chooses to Convey $X=6$, then Player B will know that if he/she choose Left, Player A will receive a payoff of $X=6$. Whereas, if Player A chooses to Not Convey $X=6$, Player B will think that if he/she chooses Left, Player A will receive a payoff of ${\widehat{m}}_{A}\in [1,5]$. Hence, from the perspective of Player B, Player A suffers strictly more from Player B’s choice of Left when $X=0$ is conveyed, compared to when the value of X is not conveyed. Similarly, from the perspective of Player B, Player A suffers strictly more from Player B’s choice of Left if the value of X is not conveyed, compared to if $X=6$ is conveyed.¹³ Therefore, a Player A who is attempting to induce guilt upon Player B would Convey $X=0,$ and Not Convey $X=6$. This leads to the first testable hypothesis:

H1::

The proportion of Player As who Convey $X=0$ in $\Gamma _{\mathrm{PPT}}$ is larger than the proportion of Player As who Convey $X=6.$

The second motivation of this study is to investigate whether agents are susceptible to guilt induction. That is, are agents more motivated to respond kindly after guilt has been induced upon them? BSH (1994) posit that after Person A has induced guilt upon Person B, “Person B finds the guilt aversive and, to escape from guilt, complies with A’s wishes” (p. 247). It is also possible, however, that a guilt averse Player B will recognize that Player A is trying to manipulate his/her behavior by “guilting” him/her, which can result in Player B being more motivated to choose the unkind action of Left in response to Player A’s attempted guilt induction. BSH (1994, 1995) document this potential “cost” of guilt induction by arguing that the target of guilt induction (Player B) might feel resentment and be motivated to respond negatively toward the guilt inducer (Player A). Attempted guilt induction by Player A may actually be counterproductive as it may foster more selfish behavior and motivate Player B to choose Left, contrary to Player A’s intended motivation. Hence, the susceptibility of agents to the exploitation of their guilt aversion, through guilt induction by others, is an open empirical question. If guilt induction by Player A is an effective influence mechanism, then Player B would be more motivated to choose Right after Player A induces guilt by choosing to Convey $X=0$ and Not Convey $X=6$. This leads to the following testable hypothesis:

H2::

The proportion of Player Bs choosing Right in $\Gamma _{\mathrm{PPT}}$ after $X=0$ is conveyed is larger than when the value of X is not conveyed, which is larger than when $X=6$ is conveyed.¹⁴

The third motivation of this study is to investigate whether having the opportunity to induce guilt fosters more trusting behavior. Comparing $\Gamma _{\mathrm{UPT}}$ and $\Gamma _{\mathrm{PPT}},$ we can see that the differences between $\Gamma _{\mathrm{UPT}}$ and $\Gamma _{\mathrm{PPT}}$ are the possibility for Player A to become privately informed about the value of X,  and the ability to convey the learned value of X to Player B. As I have shown, it is these features of $\Gamma _{\mathrm{PPT}}$ that provide an opportunity for Player A to induce guilt upon Player B. Therefore, if having an opportunity to induce guilt fosters more trusting behavior, then Player As would be more motivated to choose In in $\Gamma _{\mathrm{PPT}},$ compared to $\Gamma _{\mathrm{UPT}}$. This leads to the following testable hypothesis:

H3::

The proportion of Player As who choose In when playing $\Gamma _{\mathrm{PPT}}$ is larger than the proportion of Player As who choose In when playing $\Gamma _{\mathrm{UPT}}$.

2.3 Consistency with the B&D model of guilt

Derived from how BSH (1994) posit that people induce guilt upon others, the way in which Player A would attempt to induce guilt upon Player B in $\Gamma _{\mathrm{PPT}}$ is by choosing to Convey $X=0$ and Not Convey $X=6$. I take this time to discuss how this manner of inducing guilt upon others is consistent with the B&D model of simple guilt. Namely, the B&D model would predict that Player B would suffer more guilt from choosing Left when $X=0$ was conveyed, compared to when the value of X was not conveyed, compared to when $X=6$ was conveyed. In what follows, I provide a brief explanation of why this is the case, and I refer readers to Appendix A, where a more formal derivation is provided including the application of the B&D model to $\Gamma _{\mathrm{PPT}}$.

Recall that the B&D model posits that agents suffer disutility from guilt when they let down another relative to that other agent’s expectations. Hence, in $\Gamma _{\mathrm{PPT}}$, the amount of guilt that Player B will suffer from choosing Left will be proportional to the difference between the payoff that Player A was expecting to receive and the payoff that Player A actually receives as a result of Player B choosing Left. Conditional on Player A’s payoff expectation when choosing In, the lower the actual payoff Player A will receive from Player B choosing Left, the more guilt Player B will suffer from choosing Left. By the structure of $\Gamma _{\mathrm{PPT}}$, the lower the value of X the lower is Player A’s actual payoff when Player B chooses Left. It follows that the B&D model would predict that Player B would feel more guilt from choosing Left when $X=0$ was conveyed, compared to when the value of X was not conveyed, compared to when $X=6$ was conveyed. Therefore, Player A can attempt to induce guilt upon Player B in $\Gamma _{\mathrm{PPT}}$ by choosing to Convey $X=0$ and Not Convey $X=6$. The idea is that because of the structure of $\Gamma _{\mathrm{PPT}}$ Player A can manipulate the guilt that Player B would feel via strategic revelation of private information to Player B about the actual payoff that Player A would receive if Player B chose Left.

The primary motivation of this study is to experimentally investigate whether agents attempt to exploit the guilt aversion of others, and whether agents are susceptible to such exploitation. These are questions related to behavioral motivations in games that do not depend upon any equilibrium supposition. Nevertheless, it is important to think about whether such behavior can be supported in equilibrium, as this is informative for determining whether such behavior is sustainable. It is the case that the two possible strategy profiles corresponding to guilt induction by Player A and a kind response to guilt induction by Player B can be supported as sequential equilibria of $\Gamma _{\mathrm{PPT}}$ under the guilt framework of B&D and the assumption of complete information about guilt sensitivities (refer to Appendix A for a more formal derivation of this). I acknowledge that assuming equilibrium play, especially when a game features multiple equilibria, is a rather strong notion; however, an equilibrium supposition is sufficient, and not necessary, for H1 and H2 to be consistent with predictions of the B&D model of guilt.

2.4 Possible alternative motivations

As shown above, $\Gamma _{\mathrm{PPT}}$ provides Player A an opportunity to induce guilt upon Player B and, thus, a way of exploring whether Player As attempt to do so and whether Player Bs are susceptible to such an attempt by Player As. In general, however, the behavior of Player Bs may be influenced by other factors besides guilt aversion. I conclude this section by briefly discussing some alternative motivations that could possibly be impacting the behavior of Player Bs in $\Gamma _{\mathrm{PPT}}$ and how these relate to the research hypotheses.

One possible motivation is inequality aversion—the idea that agents are averse to unequal outcomes, both advantageous and disadvantageous (e.g., Fehr and Schmidt 1999; Bolton and Ockenfels 2000 for seminal models). Applied to trust games, inequality aversion is often cited as a possible explanation for why second movers may by motivated to choose the more kind action (Right), as it generally results in a more equal division of payoffs. However, the payoffs in $\Gamma _{\mathrm{PPT}}$ were designed to attempt to control for possible inequality aversion of Player Bs. Specifically the payoffs are structured such that, from the perspective of Player B, a choice of Left in $\Gamma _{\mathrm{PPT}},$ compared to a choice of Right, yields: (i) a higher material payoff, (ii) weakly less inequality in the payoff distribution between Player A and B (in terms of the absolute difference in payoffs), and (iii) favorable rather than unfavorable payoff inequality for all values of $X\in [0,6].$ Given these three properties, the Fehr and Schmidt model, and the nonlinear extension proposed by Bellemare et al. (2008), would predict that inequality averse Player Bs would always be motivated to choose Left, regardless of the value of X.¹⁵ Thus, as specified by the linear and nonlinear versions of the Fehr and Schmidt model, behavior by Player B in the direction of H2 cannot be explained by possible inequality aversion of Player B; hence, variation in Player A’s conveyance decision regarding the value of X in $\Gamma _{\mathrm{PPT}}$ cannot be explained by any strategic considerations regarding inequality aversion of Player B as characterized by the Fehr and Schmidt model.

I acknowledge, however, that the parameterization of $\Gamma _{\mathrm{PPT}}$ does not fully control for inequality averse preferences as specified by the Bolton and Ockenfels (2000) model. In their model, inequality is measured based on a player’s relative share of his/her payoff (i.e., a player’s payoff divided by the total payoff across all players); subsequently, players dislike when their payoff deviates away from the equal, average relative share of the payoffs, which in a 2-player game is simply when a player’s relative payoff share is 1/2. In the context of $\Gamma _{\mathrm{PPT}}$, this implies that it is possible for a Player B, with Bolton and Ockenfels type inequality averse preferences, to prefer Right when $X=0.$ Specifically, it is possible (under the very general assumptions of the model) that Player B would prefer choosing Right and sacrificing $2 in own payoff (moving from $6 to $4), to bring his/her relative payoff share closer to 1/2 (moving from 6/6 = 1 to 4/14 = 2/7). That said, such a choice would require Player B sacrificing 33 % of his/her own material payoff, as well as moving to a position of unfavorable/disadvantageous inequality, to move 28 % closer to the relative payoff share of 1/2. Moreover, Engelmann and Strobel (2004) provide experimental evidence that the Fehr and Schmidt model better predicts behavior than does the Bolton and Ockenfels model. In summary, the payoffs in $\Gamma _{\mathrm{PPT}}$ were structured to control for possible inequality aversion of Player Bs as posited by the Fehr and Schmidt model. However, I cannot go as far as saying the design fully controls for all possible characterizations of preferences for inequality aversion. It is possible, although unlikely in my view, that a Player B may be more motivated to choose Right when $X=0$ (the direction of H2) because it moves Player B’s relative payoff share closer to 1/2.¹⁶

Another possible motivation for Player Bs may be positive reciprocity—the idea that agents may be motivated to respond kindly to agents who are kind to them (e.g., Dufwenberg and Kirchsteiger 2004). In the context of $\Gamma _{\mathrm{PPT}},$ Player B may be motivated to positively reciprocate the kind action of Player A choosing In by choosing Right, and this may be more prevalent if Player B knows that $X=0$ (see Cox 2004 for a general discussion of reciprocity motivations in trust games). To control for this possible reciprocity confound, I consider a third game, which is a “dictator” version of $\Gamma _{\mathrm{PPT}}.$ In this modified dictator version of $\Gamma _{\mathrm{PPT}}$, denoted as $\Gamma _{\mathrm{PPD}}$, the initial In/Out decision of Player A is eliminated. $\Gamma _{\mathrm{PPD}}$ begins with Player A’s decision to convey X to Player B, conditional on X being revealed, and proceeds with Player B’s decision of Right or Left. Hence, the extensive form of $\Gamma _{\mathrm{PPD}}$ is simply the subgame of $\Gamma _{\mathrm{PPT}}$ that begins with Nature’s move. The removal of the initial In/Out decision eliminates possible motivations to positively reciprocate Player A’s In decision from Player B’s Right decision. If there are no significant differences between Player B behavior in $\Gamma _{\mathrm{PPD}}$ and $\Gamma _{\mathrm{PPT}},$ then reciprocity motivations are not a salient concern, and H2 can be tested using pooled data from both $\Gamma _{\mathrm{PPT}}$ and $\Gamma _{\mathrm{PPD}}.$ However, if Player B behavior differs across the two games, then reciprocity could be salient, and H2 will be tested using data from $\Gamma _{\mathrm{PPD}}$ only. By comparing the decision making of Player B in $\Gamma _{\mathrm{PPT}}$ with decision making in the dictator version of $\Gamma _{\mathrm{PPT}},$ I am able to test for and, if necessary, control for possible reciprocity motivations of Player B.¹⁷

However, one limitation of the experimental design is its inability to directly control for efficiency concerns or maximin preferences (Charness and Rabin 2002), both of which may be salient motivations (see Engelmann and Strobel 2004). In regards to efficiency, Player B may be motivated to make choices that maximize the sum of total payoffs. Note that in $\Gamma _{\mathrm{PPT}},$ Right is always more efficient, so in the extreme case where Player B was solely motivated, or very strongly motivated by efficiency concerns, then Player B would always be motivated to choose Right. However, under the more reasonable assumption that efficiency concerns are being balanced against one’s own material payoffs, then at the margin Player B may be more motivated to choose Right after $X=0$ had been conveyed. Similarly, if Player B has maximin preferences—a desire to maximize the lowest payoff of a player in the game—then Player B may be more motivated to choose Right after $X=0$ had been conveyed.

It is important to note that from a conceptual standpoint, it may be difficult to eliminate concerns for efficiency and maximin preferences when testing if agents are susceptible to guilt induction. The reason rests in the manner by which agents induce guilt upon others, namely, by conveying that they will receive a low payoff given an undesirable action of the other agent. Therefore, relative to not conveying such information, the undesirable choice will necessarily lead to a lower payoff for the other agent and a lower level of efficiency, as is the case in $\Gamma _{\mathrm{PPT}}.$ As a result, I am unable to separate out efficiency concerns and maximin preferences from guilt aversion when testing whether Player Bs are susceptible to guilt induction by Player As (H2).

That said, $\Gamma _{\mathrm{PPT}}$ allows Player A to induce guilt upon Player B by conveying $X=0$ and not conveying $X=6$, which is consistent with the method posited by BSH (1994). Furthermore, the B&D model predicts that a Player B would suffer more guilt from choosing Left after $X=0$ is conveyed, compared to when the value of X is not conveyed, compared to when $X=6$ is conveyed. Hence, if the experimental data support H1, then I will interpret this as evidence consistent with Player As attempting to exploit the guilt aversion of Player Bs. Subsequently, if the experimental data support H2, then I will interpret this as evidence consistent with Player Bs being susceptible to the exploitation of their guilt aversion. Readers should certainly be mindful of the fact that other possible motivations, e.g., efficiency concerns and/or maximin preferences, could be influencing the behavior of Player B in addition to guilt aversion. However, in the results section, I provide some results from a post-decision questionnaire that are consistent with Player Bs being motivated, at least in part, by guilt aversion. This is consistent with recent findings by Attanasi et al. (2013) who document that “guilt aversion is a prevalent psychological motivation” (p. 3) in a 2-player trust game.

3 Experimental procedure

All experimental sessions were conducted in the Economic Science Laboratory (ESL) at the University of Arizona in April 2011 and October 2011. The sessions were computerized, and the software was programmed using Z-tree (Fischbacher 2007). The subject pool consisted of undergraduates who were recruited via an online database. In total, 22 sessions were conducted using 444 subjects comprising 222 two-player groups.

To test the three main research hypotheses of this paper (H1–H3), I use a between-groups design where all participants are randomly assigned to one of the following three experimental treatments:

• UPT Treatment Subjects played $\Gamma _{\mathrm{UPT}}$, where the payoffs from $\Gamma _{\mathrm{UPT}}$ corresponded 1:1 with the monetary payoffs in the experiment.

• PPT Treatment Subjects played $\Gamma _{\mathrm{PPT}}$, where the payoffs from $\Gamma _{\mathrm{PPT}}$ corresponded 1:1 with the monetary payoffs in the experiment.

• PPD Treatment Subjects played $\Gamma _{\mathrm{PPD}}$, where the payoffs from $\Gamma _{\mathrm{PPD}}$ corresponded 1:1 with the monetary payoffs in the experiment.

Conditional on their random assignment to treatment, all participants were randomly assigned to either the role of Player A or Player B and then randomly and anonymously matched with a participant of the opposite player role. All participants then proceeded to play their designated game one time in their assigned player role. Of the 222 groups, 111 were assigned to the PPT Treatment, 45 were assigned to the UPT Treatment, and 66 were assigned the PPD Treatment. Each session lasted approximately 25 min and subjects earned an average of $9.68 USD (including a $5 show-up payment).

Before starting the game, the experimental instructions were carefully read aloud to all participants in the session to enhance clarity and general understanding of the task among the participants. A copy of the experimental instructions can be found in Appendix B.1, as well as sample screen shots of the computer interface in Appendix B.2. Upon the completion of the game, the decisions of each player, the corresponding outcome, the profit to each player, and the value of X were displayed to both players. All subjects were informed in the instructions that the value of X would be revealed to both players upon completion of the task, irrespective of the decisions made in the task. Revealing the value of X to all players ensures that Player As were not motivated to choose In (or Conveying X) just so Player A (Player B) could learn the value of X. This design feature eliminates any curiosity biases that may arise from the uncertainty regarding the value of X,  and the consequent payoffs to the other player.

Upon completion of the game, subjects were asked to fill out a short questionnaire. In all treatments, the questionnaire contained 8 general demographic questions. In the PPT and PPD treatments, two additional questions were asked that related to guilt feelings and perceptions of guilt feelings in the game. These specific guilt-related questions, the corresponding responses, and discussion of the possible gleaned insights from these questions are presented in Sect. 4.2.

4 Results

I first present the aggregate decision data from each of the three treatments and the corresponding tests of the three research hypotheses. I then present some results from the post-decision questionnaire. I conclude with discussion and speculative remarks about some of the observed patterns in the data.

4.1 Aggregate data and hypothesis testing

I begin by comparing the aggregate Player B data from the PPT and PPD treatments to test for any possible reciprocity motivations (see Sect. 2.4). Specifically, I compare the frequency of Player Bs choosing Right at each of the three possible conveyance states: (i) $X=0$ was conveyed (Right $\vert X=0$), (ii) the value of X was not conveyed (Right $\vert X=NC$), and (iii) $X=6$ was conveyed (Right $\vert X=6$). Figure 3 displays the histogram of the relevant Player B data.

From Fig. 3, we can see that the relative frequencies of Right choices at each of the three conveyance states are similar across the two treatments. In fact, a two-sided Fisher’s Exact test does not yield a significant difference between the proportion of Player Bs who choose Right $\vert X=0$, Right $\vert X=NC$ , and Right $\vert X=6$ between the two treatments ($p = 1.000, p = 1.000,$ and $p = 0.542,$ respectively). Thus, we can rule out reciprocity as a salient confounding motivation for Player B’s Right/Left decision in $\Gamma _{\mathrm{PPT}},$ conditional on the conveyance state. As a result, I am able to proceed in testing H2 using pooled data from the PPT and PPD treatments, as discussed in Sect. 2.4, which provides the added benefit of a larger sample size and more statistical power for the hypothesis testing. That said, because of differences in the strategic structure of the two games (dictator vs trust), I will also present the results and hypothesis testing separately for the PPT and PPD treatments.

Fig. 3

Comparison of Right rates—PPD and PPT treatments

I test the hypotheses in the order that corresponds to working backwards through the game. Namely, I first test H2, whether Player Bs are susceptible to guilt induction by Player As. In terms of testing H2, I compare the proportion of Player Bs who choose Right $\vert X=0$, Right $\vert X=NC$, and Right $\vert X=6$. Table 1 presents the relevant data from the Player Bs who were called upon to make a decision (i.e., all 66 Player Bs from the PPD treatment and the 58 Player Bs in the PPT treatment whose corresponding Player A chose In). From Table 1, we can see that in the PPT treatment 3/7 (43 %) Player Bs chose Right $\vert X=0$, 5/40 (13 %) chose Right $\vert X=NC$, and 0/11 (0 %) chose Right $\vert X=6$. The corresponding frequencies in the PPD treatment are 9/23 (39 %), 2/21 (11 %), and 2/22 (10 %). A Jonckheere–Terpstra non-parametric test for ordered alternatives rejects the null of equality of these Right rates in favor of the ordered alternative for each treatment ($p = 0.008$ and $p = 0.006,$ respectively), as well as pooled over both treatments $(p<0.001).$ ¹⁸ Thus, the data suggest that Player Bs are susceptible to guilt induction by Player As, in the sense that Player Bs are more likely to choose Right $\vert X=0,$ compared to Right $\vert X=NC,$ compared to Right $\vert X=6$, which supports H2. An alternative interpretation is that guilt induction by Player A would be effective at increasing the likelihood that Player B chooses Right; this is especially true when explicitly choosing to convey $X=0$, rather than not conveying $X=0.$ The nonlinear pattern in Right rates across these three conveyance states likely results from Player B’s belief about the value of X in the NC state, ${\widehat{m}}_{A},$ not being equal to 3. Specifically, the data suggest that ${\widehat{m}}_{A}$ is close to its upper bound of 5, which corresponds to the case where Player B believes that only those Player As for whom $X=6$ would not convey.

Table 1 Comparison of Player B Right rates by conveyance state

Next, I turn to testing H1. Namely, are Player As attempting to exploit the guilt aversion of Player Bs by choosing to Convey $X=0$ more frequently than they choose to Convey $X=6$? Table 2 presents the relevant conveyance data from the Player As who were called upon to make a conveyance decision (i.e., 56 Player As from the PPD treatment to whom X was revealed, and 42 Player As from the PPT treatment who chose In and to whom X was revealed). From Table 2 we can see that in the PPT treatment 7/16 (44 %) Player As chose to Convey $X=0$ and 11/26 (42 %) Player As chose to Convey $X=6$. The corresponding frequencies in the PPD treatment are 23/27 (85 %) and 22/29 (76 %), respectively. Clearly, the data reveal that in neither treatment did all Player As choose to Convey $X=0$ and Not Convey $X=6.$ Furthermore, while the proportion of Player As who Convey $X=0$ is larger than the proportion who Convey $X=6$ in both treatments, the difference is not significant, using a one-sided Fisher’s exact test, in either treatment ($p = 0.589$ and $p= 0.296$, respectively); likewise, the difference in conveyance rates is not significant for the pooled sample using a one-sided Fisher’s exact test $(p = 0.215)$ or a one-sided t test $(p = 0.161)$. Looking separately at the PPT and PPD treatments, or pooled over both treatments, the Player A conveyance data fail to support H1, which suggests that Player As are not attempting to fully exploit the guilt aversion of Player Bs by inducing guilt upon them.

Table 2 Comparison of Player A conveyance rates by value of X

Before presenting the aggregate trust rates in the PPT and UPT treatments and the corresponding test of H3, I pause to briefly discuss the differences in Player A conveyance rates across the PPT and PPD treatments. Specifically, the overall rate of conveyance is significantly higher in the PPD treatment compared to the PPT treatment, both for $X=0$ and $X=6.$ Given the differences in the strategic structure across the games (i.e., the presence of the outside option in PPT along with the initial In/Out decision for Player A), there is no reason to assume, ex ante, that the conveyance behavior be equal across the two games.¹⁹ That said, the identification of attempted guilt induction by Player A (H1) is based on the relative comparison of the ratio of Player As who convey $X=0$ with the ratio of those who convey $X=6.$ Thus, the level differences in the conveyance rates across the two games do not necessarily invalidate the test of H1 within each treatment, or the overall implication drawn from the data in the previous paragraph that Player As do not appear to be attempting to induce guilt upon Player Bs.

Lastly, I turn to testing H3; namely, does having an opportunity to induce guilt foster more trusting behavior? To test this hypothesis, I compare the aggregate In rate from the 111 Player As in the PPT Treatment with the In rate from the 45 Player As in the UPT Treatment. From Table 3, we can see that 58/111 (52 %) of Player As chose In in the PPT Treatment and 17/45 (38 %) of Player As chose In in the UPT treatment, which is significant using a one-sided Fisher’s Exact test $(p = 0.071)$ and a one-sided t test $(p= 0.051).$ ²⁰ Thus, the aggregate data on Player A In rates across the PPT and UPT treatments marginally support H3, which suggests that strategic settings that provide an opportunity for agents to induce guilt may foster more trusting behavior by those agents.

Before proceeding, I pause to acknowledge that the observed difference in In (trust) rates between the PPT and UPT games could result from alternative motivations besides Player As foreseeing an opportunity to induce guilt in PPT; this is especially so in light of the fact that the data reveal little evidence that Player As attempted to induce guilt. For example, some alternative explanations for why Player As in PPT may be more motivated to choose In could be: (i) they have an opportunity to communicate information, (ii) they can learn the value of X before Player B acts (as opposed to the end of the game in UPT), or (iii) Player A feels they have some increased ability to control the outcome in PPT.²¹ Thus, a more conservative interpretation would be that the marginal support of H3 provides evidence that the inclusion of private payoff information for Player A about their payoff if Player B is not trustworthy and the ability to convey this information to Player B can increase trust by Player A. That said, as I noted in the Introduction, these two features are necessary for providing Player A with the opportunity to induce guilt upon Player B.

Table 3 Comparison of Player A In rates

4.2 Questionnaire results

Next, I present results from two post-decision questionnaire questions asked to both Player As and Player Bs in the PPT and PPD treatments. The motivation of these two questions was to gain additional insights regarding Player B’s feelings of guilt, and Player A’s perceptions of Player B’s feelings of guilt. This questionnaire was not incentivized and did not impact monetary earnings. As a result, the natural amount of discretion must be used in evaluating the resulting data. At the same time, there is really no scope for any type of monetary gains from strategic false reporting and, thus, no obvious material incentive to not report truthfully.²² In addition, all the analyses of the response data are done using matched samples, which control for scaling differences and possible anchoring effects that could exist across unmatched samples.

For Player As, the first question asked how much guilt they thought Player B felt (would have felt) from choosing Left if Player B knew (would have known) the true value of X. The second question asked Player As how much guilt they thought Player B felt (would have felt) from choosing Left if Player B did not know (would not have known) the true value of X. Responses were ranked on a 5-point scale with 5 being a Very High amount of guilt and 1 being a Very Low amount of guilt. For the analysis, I consider only the Player As who had the value of X revealed to them and, thus, had an opportunity to convey the value of X to Player B. Table 4 presents the aggregate response data for these Player As.²³ Table 4 is divided into two panels that correspond to whether $X=0$ (Panel 1) or $X=6$ (Panel 2) was revealed to Player A. Within each panel, the average reported perceptions of Player B’s guilt feelings from choosing Left are separately presented for those Player As who conveyed the corresponding value of X and those who did not.

Table 4 Player A’s perceptions of Player B’s guilt feelings from choosing Left

From Panel 1 of Table 4, we can see that the Player As who chose to Convey $X=0$ perceived that Player B would have felt significantly more guilt from choosing Left $\vert X=0$ compared to Left $\vert X=NC$ $(p = 0.008)$. However, the Player As who chose to Not Convey $X=0$ did not perceive that Player B would have felt significantly more guilt from choosing Left $\vert X=0$ compared to Left $\vert X=0$ $(p = 0.524)$. Similarly, from Panel 2, we see that the Player As who chose to Not Convey $X=6$ perceived that Player B would have felt significantly more guilt from choosing Left $\vert X=NC$ compared to Left $\vert X=6$ $(p = 0.002)$. Yet, Player As who chose to Convey $X=6$ perceived that Player B would have felt only marginally more guilt from choosing Left $\vert X=NC$ compared to Left $\vert X=6$ $(p = 0.061).$ ²⁴

One of the primary things revealed from the Player A response data is that the Player As whose conveyance decisions were consistent with attempted guilt induction (those Player As who conveyed $X=0$ and did not convey $X=6$) seemed to think that doing so would induce more guilt upon Player B if Player B were to choose Left. However, the Player As whose conveyance decisions where not consistent with attempted guilt induction (those Player As who did not convey $X=0$ and did convey $X=6$) seemed to think that doing so would not have induced as much guilt upon Player B if Player B were to choose Left.

I turn now to the Player B questionnaire data. For Player Bs, the first question asked how much guilt he/she felt (would have felt) from choosing Left (if he/she had chosen Left) if he/she did not know the value of X. The second question asked Player B how much guilt he/she felt (would have felt) from choosing Left (if he/she had chosen Left) if he/she knew the value of X. Again, responses were ranked on a 5-point scale with 5 being a Very High amount of guilt and 1 being a Very Low amount of guilt. For the analysis, I consider the Player Bs who actually made a Left/Right decision in either $\Gamma _{\mathrm{PPT}}$ or $\Gamma _{\mathrm{PPD}}$. Table 5 presents the aggregate response data for these Player Bs.²⁵ Table 5 is divided into two panels that correspond to the actual value X, and each panel shows the average reported Player B guilt feelings from choosing Left.

Table 5 Player B’s reported guilt feelings from choosing Left

From Panel 1 of Table 5 we see that when $X=0,$ Player Bs reported that they would have felt significantly more guilt from choosing Left $\vert X=0$ compared to Left $\vert X=NC$ $(p < 0.001).$ Similarly, from Panel 2 we see that when $X=6$, Player Bs reported that they would have felt significantly more guilt from choosing Left $\vert X=NC$ compared to Left $\vert X=6$ $(p<0.001).$ These reported perceived guilt feelings by Player Bs are consistent with guilt aversion acting as a predominant motivation in the observed aggregate behavior of Player Bs being more likely to choose Right $\vert X=0,$ compared to Right $\vert X=NC,$ compared to Right $\vert X=6$. This reinforces the idea that Player Bs are susceptible to the exploitation of their guilt aversion by Player A.

4.3 Discussion

Although the motivation of this study is to test H1–H3, I take this time to make a few speculative remarks regarding some of the observed patterns in the data. I begin by proposing some plausible explanations of why Player As seem to not be fully exploiting the guilt aversion of Player Bs (i.e., failure to support H1). Failure by Player A in attempting to induce guilt upon Player B in $\Gamma _{\mathrm{PPT}}$ (and $\Gamma _{\mathrm{PPD}}$) corresponds to Player As either (i) not conveying $X=0$, and/or (ii) conveying $X=6$. The data revealed a non-trivial percentage of Player As did not convey $X=0$ (30 %) and did convey $X=6$ (60 %), which suggests that failure to support H1 seems to be a result of a combination of both (i) and (ii).

One possibility is that some Player As just do not realize or think that a sufficient amount of guilt could have been induced upon Player B by conveying $X=0$ and not conveying $X=6$. Consequently, these types of Player As would not necessarily be motivated to convey $X=0$ and/or not convey $X=6$ as a means of inducing guilt. The questionnaire data presented in the previous section provided some evidence consistent with this possibility; namely, we saw that the Player As who made conveyance decisions that were inconsistent with attempted guilt induction did not seem to think as much guilt could have been induced by conveying $X=0$ and not conveying $X=6$, compared to those Player As whose conveyance decisions were consistent with attempted guilt induction. A second possibility is that some Player As feel guilt over inducing guilt, an idea that BSH (1994) refer to as metaguilt. Player As who are averse enough to these metaguilty feelings would then be motivated to not induce guilt by not conveying $X=0$ and conveying $X=6$. A third possibility is that some Player As were attempting to “signal” honesty to Player B by not attempting to manipulate Player B’s behavior by letting them think that $X=0$, when in fact $X=6$; said differently, some Player As who learned $X=6$ might not have wanted to deceive Player B into thinking $X=0.$ The motivation for doing this is that Player As may have strategically reasoned that this display of honesty (as opposed to deception) could, in turn, increase B’s trustworthiness and increase the likelihood of Player B choosing Right.²⁶ In fact, Wang et al. (2009) and Wang and Leung (2010) document recent experimental evidence of this effect where honesty (in the form of information revelation) is rewarded, while deception is punished; relatedly, Brandts and Charness (2003) document experimental evidence that agents are punished less frequently when honest.

Regarding the Player A conveyance data, it is important to note that the observed failure to support H1 is not evidence that agents universally would not attempt to induce guilt upon others across all strategic environments, and I would caution readers from interpreting it as such. Rather, the data from this study provide evidence that, at least in the specific strategic settings considered (the PPT and PPD games), participants acting in the role of Player A did not fully capitalize on their opportunity to induce guilt upon Player Bs; this is especially true with regard to conveying $X=0$, where they could have increased the likelihood of Player B choosing Right by about 30 % (in expectation). Amongst the set of Player As who thought guilt could be induced, their behavior seems to be in line with attempted guilt induction (Table 4). However, there is also evidence of a non-trivial fraction of Player As exhibiting behavior inconsistent with inducing guilt, which could be for a myriad of reasons; a non-exhaustive list of some plausible reasons is listed above. Future research is warranted that aims at exploring the possible mediating and moderating factors of an agent’s decision to attempt to induce guilt upon others in strategic decision making environments.

With regard to Player B behavior, a pattern that emerges from the data is Player Bs’ propensity to choose Right is not linearly increasing across the conveyance states. Player Bs are much more likely to choose Right $\vert X=0$ (40 %) than Right $\vert X=NC$ (11 %), as compared to Right $\vert X=6$ (6 %). Recall, BSH (1994) posit that guilt is induced by conveying how one suffers if another fails to act in a desired fashion. You can think of Player A choosing to convey $X=0$ as an explicit attempt to induce guilt because Player A is actually conveying how he/she suffers. On the other hand, you can think of Player A choosing to not convey $X=6$ as an implicit attempt to induce guilt because Player A is conveying how he/she suffers by not conveying how he/she does not suffer. This distinction might be important in terms of the effectiveness of guilt induction, in light of the fact that Player Bs were much more likely to choose Right after Player A had explicitly induced guilt by conveying $X=0.$ This suggests that guilt induction may be the most effective when an agent knows he/she will suffer over another’s failure to act in a desired fashion and is able to explicitly convey this degree of suffering. However, if an agent knows he/she is not going to suffer, then not conveying that information may be less effective.

5 Conclusion

The main motivations of this study were to (i) experimentally test whether agents attempted to exploit the guilt aversion of others in self-serving ways by inducing guilt upon them, (ii) whether agents where susceptible to this type of exploitation, and (iii) whether having such an opportunity fostered more trusting behavior. The experimental data are consistent with Player Bs being susceptible to guilt induction by Player As. This susceptibility of Player Bs to guilt induction is reinforced by data from the post-decision questionnaire where Player Bs’ self-reported feelings of guilt where consistent with guilt aversion. However, the data reveal that Player As did not attempt to fully exploit the guilt aversion of Player Bs in the setting considered, despite the fact that it would have increased the likelihood that Player B would have chosen the kind action. Although, the data from the post-decision questionnaire reveal that there may be two types of Player As: (i) Player As who think inducing guilt would be effective and, thus, attempt to do it, and (ii) Player As who do not think inducing guilt would be effective and, therefore, do not attempt to do it. Lastly, the data reveal evidence that Player As may be marginally more trusting when playing a trust game where the strategic structure is rich enough for the opportunity to induce guilt upon Player Bs.

The susceptibility of Player Bs to guilt induction that is observed in the data can be viewed as additional experimental evidence consistent with the hypothesis that agents are motivated by guilt aversion. Hence, the experimental design provides an alternative approach for investigating guilt aversion from the previously implemented belief elicitation based approaches, both of which present previously established limitations.²⁷ The ability to test for guilt aversion without eliciting beliefs is particularly relevant in light of the recent studies by Reuben et al. (2009) and Ellingsen et al. (2010), which both test for the presence of guilt aversion using similar experimental designs that feature conveyance of elicited beliefs, yet reach opposing conclusions. This paper joins Nelissen et al. (2011) and Charness and Dufwenberg (2011), in its ability to identify behavior consistent with belief-dependent models of utility without having to elicit or convey elicited beliefs.

BSH (1994) note that “guilt [induction] does not depend on formal power or influence and may even work best in the absence of such power, because one induces guilt by depicting oneself as the helpless victim of another’s actions” (p. 247). This suggests that guilt induction could be particularly effective in economies with less developed legal systems. In such economies, guilt induction could serve as an informal mechanism for enforcing contracts and mitigating corrupt behavior, which might otherwise transpire in the absence of formal prohibitive legislation (Lee 2010). Guilt induction could also prove to be effective at influencing behavior and impacting outcomes in credence goods markets (Dulleck and Kerschbamer 2006; Dulleck et al. 2011; Beck et al. 2013; Balafoutas et al. 2013). With credence goods (e.g., doctors, mechanics, taxis, or other expert services), the consumer is often the “helpless victim” of the expert’s actions. Guilt induction by the consumer could be implemented to thwart opportunistic behavior by the expert, especially in developing economies where the incentives for opportunistic behavior are likely to be much stronger.²⁸

Partnerships, principle-agent contracting, and employee-employer relationships represent some of the many economic settings where trust is pivotal for successful and efficient relations. There is a growing body of literature that investigates the importance of trust in social and economic settings, and how trust can be fostered (see Fehr 2009; Charness et al. 2011; Sapienza et al. 2013 for reviews). Much of this literature focuses on the effectiveness of reputation building in fostering trust.²⁹ While there is often an incentive to trust in economic settings, this incentive is often offset by exposure to the risk of opportunistic behavior by the trusted agent. However, guilt induction by the trusting agent can serve as a mechanism for thwarting such opportunistic behavior, thus mitigating the risk associated with trusting actions. Therefore, having an opportunity to induce guilt would then lead to more trusting behavior, which is what is observed in this paper. This might help shed light on why trust is so prevalent in many economic interactions in our society today, where the strategic environments are often rich enough to allow the possibility to induce guilt.

In his seminal work, Rabin (1993, p. 1296) raises a concluding question about whether agents can “force” emotions in sequential move games? The answer to this question is of clear importance in determining economic outcomes, given that emotions can impact strategic decision making in very systematic and considerable ways. Along these lines, Gneezy and Imas (2014) provide experimental evidence that agents strategically anger others (i.e., they force anger upon others) in self-serving ways. The results from this study suggest that agents can also force guilt upon others, which can then lead to more kind or favorable behavior by those other agents. While this is suggestive that it may be possible for agents to force or manipulate other emotions in sequential move games in self-serving ways, this remains an interesting and open empirical question for future research.

I conclude by noting that the effectiveness of guilt induction as an influence mechanism in strategic settings may have limitations. In particular, repeated applications of guilt induction may become less effective since the target of the guilt induction will likely become resentful or angered over its repeated application. This could ultimately lead to fewer kind actions in response to guilt induction, which is counter to its intended purpose. BSH (1995) recognize this and argue that “although guilt may often be an effective way of getting one’s way, it appears to be costly and to carry some stigma. This suggests that inducing guilt may be a technique that has to be used with caution and restraint” (p. 184). Perhaps guilt induction in strategic economic settings should be a mechanism that is reserved for instances when there is little scope for reputational effects, and when the payoff and potential risk associated with a trusting action are largest.

Appendices

Appendix A: Application of B&D model of guilt

In this Appendix, I formalize the arguments put forth in Sect. 2.3 regarding the connection between the posited method by BSH (1994) of how guilt is induced and the formal model of guilt developed by B&D. In doing so, I first provide a brief outline of the B&D model of guilt, followed by its application to $\Gamma _{\mathrm{PPT}}$. I then derive conditions under which the BSH (1994) method is consistent with predictions of the B&D model of guilt. I conclude by showing that the BSH (1994) method can be supported as an equilibrium in $\Gamma _{\mathrm{PPT}}$ under the framework of the B&D model. I conclude by briefly discussing the recent extension of the B&D model developed by Attanasi et al. (2015) that incorporates incomplete information about guilt sensitivities, and how incomplete information about guilt sensitivities can impact behavioral predictions in $\Gamma _{\mathrm{PPT}}$.

1.1 B&D model of guilt applied to $\Gamma _{\mathrm{PPT}}$

Before I proceed in formally applying the B&D model of guilt to $\Gamma _{\mathrm{PPT}},$ I first provide a general overview of the B&D model of simple guilt.³⁰ What follows is only a simplified outline of the model. Interested readers should refer to B&D for the full, technical presentation of the model including illustrative examples. Informally, the model posits that agents suffer disutility, in the form of guilt, from failing to live up to others’ expectations. This is captured by modeling an agent’s utility as a function of his/her own material payoffs and the extent to which he/she let other agents down.

Formally, simple guilt is modeled by specifying a utility function for player i given by:

$$\begin{aligned} u_{i}^{\mathrm{SG}}=m_{i}-\underset{j\ne i}{\sum }\theta _{ij}\cdot D_{j} \end{aligned}$$

(Simple Guilt Utility)

In this expression, $m_{i}$ represents player i’s material payoff, and $\underset{j\ne i}{\sum }\theta _{ij}\cdot D_{j}$ represents player i’s disutility from simple guilt. The latter component is composed of two pieces. The first, $\theta _{ij}$, is an exogenously given constant that measures player i’s sensitivity of feeling guilty toward player j. The second, $D_{j}$, represents the amount by which player i lets player j down, as a result of player i’s strategy. $D_{j}=E_{j}-m_{j}$ is expressed as the difference between the material payoff that player j was expecting, $E_{j}$, and the material payoff that player j actually receives, $m_{j}.$ $E_{j}$ itself is a function of player j’s strategy, and player j’s vector of “first-order” beliefs regarding the strategies of the other players. Note, player i does not actually observe $D_{j}$, as it is a function of player j’s first-order beliefs. Therefore, it is assumed that player i maximizes the expected value of $u_{i}^{\mathrm{SG}}$, given player i’s first-order beliefs regarding player j’s strategy, and player i’s “second-order” belief regarding player i’s first-order beliefs.

In proceeding with the application of the B&D model of simple guilt to $\Gamma _{\mathrm{PPT}}$, I derive the guilt that Player B would experience from choosing Left in $\Gamma _{\mathrm{PPT}}$ at each of the three possible conveyance states (histories): (i) $X=0$ was conveyed, which I denote $C^{0}$, (ii) $X=6$ was conveyed, which I denote $C^{6}$, and (iii) the value of X was not conveyed, which I denote $C^{N}$. A strategy for Player B is a probability distribution over Player B’s possible actions, {Left, Right}, at each of the three possible conveyance states. In deriving this guilt that Player B would suffer from choosing Left, it is necessary to first derive Player A’s material expectation. Before Player A makes her initial In or Out decision, Player A forms an initial first-order belief regarding Player B’s strategy, which can be represented as the probabilities that Player B would choose Right at each of the three conveyance states; I denote this vector of probabilities as ${\varvec{\alpha }}_{A}=(\Pr ({ Right }|C^{0}),\Pr ({ Right }|C^{N}),\Pr ({ Right }|C^{6})).$ For a given strategy, Player A forms an initial expectation, which I denote $E_{A},$ weighted over ${\varvec{\alpha }}_{A}$ and moves by Nature, of her material payoff.

Player B experiences disutility from simple guilt when he chooses a strategy that yields a payoff to Player A that is lower than $E_{A}$. However, Player B does not observe $E_{A}.$ Therefore, Player B must form an expectation of $E_{A},$ conditional on the conveyance state. I define this conditional expectation of $E_{A}$ as $E_{B}|h$ where $h\in \{C^{0},C^{N},C^{6}\}$, which is implicitly a function of Player B’s conditional second-order beliefs regarding ${\varvec{\alpha }}_{A}$. The guilt that Player B would suffer from choosing Left is proportional to the difference between $E_{B}|h$ and $m_{A},$ where $m_{A}$ is the material payoff that Player A actually receives as a result of Player B’s Left decision. Let $\theta _{B}\ge 0$ denote Player B’s sensitivity to feeling guilty. Below is the amount of simple guilt that Player B would suffer from choosing Left at each of the three possible conveyance states.

• $X=0$ was conveyed: Because the conveyance is credible, Player B knows that if he chooses Left, Player A will receive a payoff of $m_{A}=0$. Player B’s expectation of Player A’s expectation is $E_{B}|C^{0}.$ By choosing Left, Player B will suffer disutility from guilt equal to $\theta _{B}\cdot (E_{B}|C^{0}-0).$ Thus, Player B’s utility from choosing Left, after $X=0$ was conveyed, is equal to:

  $$\begin{aligned} 6-\theta _{B}\cdot (E_{B}|C^{0}-0) \end{aligned}$$

• $X=6$ was conveyed: Because the conveyance is credible, Player B knows that if he chooses Left, Player A will receive a payoff of $m_{A}=6$. Player B’s expectation of Player A’s expectation is $E_{B}|C^{6}.$ By choosing Left, Player B will suffer disutility from guilt equal to $\theta _{B}\cdot (E_{B}|C^{6}-6).$ Thus, Player B’s utility from choosing Left, after $X=0$ was conveyed, is equal to:

  $$\begin{aligned} 6-\theta _{B}\cdot (E_{B}|C^{6}-6) \end{aligned}$$

• Value of X not conveyed: If the value of X is not conveyed, then Player B must think about the expected material payoff that Player A would receive if he chooses Left. Let ${\widehat{m}}_{A}=E_{B}[m_{A}|C^{N}]$ denote this expectation. As I have previously shown (see Sect. 2.1), ${\widehat{m}}_{A}\in [1,5]$. Player B’s expectation of Player A’s expectation is $E_{B}|C^{N}.$ By choosing Left, Player B will suffer disutility from guilt equal to: $\theta \cdot (E_{B}|C^{N}-{\widehat{m}}_{A})$ where ${\widehat{m}}_{A}\in [1,5].$ Thus, Player B’s utility from choosing Left, after the value of X was not conveyed, is equal to:

  $$\begin{aligned} 6-\theta _{B}\cdot (E_{B}|C^{N}-{\widehat{m}}_{A})\quad \hbox {where } {\widehat{m}}_{A}\in [1,5] \end{aligned}$$

1.2 Guilt induction in $\Gamma _{\mathrm{PPT}}$ via the B&D model of guilt

Based on the BSH (1994) method for how an agent can induce guilt upon another, it was derived that Player A could attempt to induce guilt upon Player B in $\Gamma _{\mathrm{PPT}}$ by choosing to Convey $X=0$ and Not Convey $X=6$. In what follows, I derive conditions under which this is consistent with predictions of the B&D model of simple guilt. From the previous section, we derived that the disutilities, from feeling guilt, that Player B would suffer from choosing Left at each of the three possible conveyance states are:

• $X=0$ was conveyed: $\theta _{B}\cdot (E_{B}|C^{0}-0)$

• $X=6$ was conveyed: $\theta _{B}\cdot (E_{B}|C^{6}-6)$

• Value of X not conveyed: $\theta _{B}\cdot (E_{B}|C^{N}-{\widehat{m}}_{A})$ where ${\widehat{m}}_{A}\in [1,5]$

According to B&D, Player B would suffer more disutility from choosing Left when $X=0$ was conveyed, compared to when the value of X was not conveyed, when:

$$\begin{aligned} \theta _{B}\cdot (E_{B}|C^{0}-0)\ge \theta _{B}\cdot (E_{B}|C^{N}-\widehat{m }_{A})\Longrightarrow E_{B}|C^{0}\ge E_{B}|C^{N}-{\widehat{m}}_{A} \end{aligned}$$

(Condition 1)

Similarly, Player B would suffer more disutility from choosing Left when the value of X was not conveyed, compared to when $X=6$ was conveyed, when:

$$\begin{aligned} \theta _{B}\cdot (E_{B}|C^{N}-{\widehat{m}}_{A})\ge \theta _{B}\cdot (E_{B}|C^{6}-6)\Longrightarrow E_{B}|C^{N}-{\widehat{m}}_{A}\ge E_{B}|C^{6}-6 \end{aligned}$$

(Condition 2)

Essentially, Condition 1 states that Player B’s second-order belief of Player A’s expectation after $X=0$ is conveyed, $E_{B}|C^{0},$ is not too much lower than Player B’s second-order belief of Player A’s expectation after the value of X is not conveyed, $E_{B}|C^{N}.$ Similarly, Condition 2 states that Player B’s second-order belief of Player A’s expectation after the value of X is not conveyed, $E_{B}|C^{N},$ is not too much lower than Player B’s second-order belief of Player A’s expectation after $X=6$ is conveyed, $E_{B}|C^{6}.$ Conditions 1 and 2 would certainly be satisfied if we assumed that Player B does not update his belief of Player A’s expectation (i.e., $E_{B}|C^{0}=E_{B}|C^{N}=E_{B}|C^{6}$), which would be satisfied in an equilibrium. However, making the assumption that $E_{B}|C^{0}=E_{B}|C^{N}=E_{B}|C^{6}$ is clearly stronger than is needed for Condition 1 and 2 to be satisfied.

If Conditions 1 and 2 are satisfied, then the B&D model predicts that Player B would feel more guilt from choosing left after $X=0$ was conveyed and the value of X was not conveyed, compared to when the value of X was not conveyed and $X=6$ was conveyed, respectively. Therefore, if Condition 1 and 2 hold, then according to the B&D model, Player A can induce guilt upon Player B by choosing to Convey $X=0$ and Not Convey $X=6$, which is consistent with the BSH (1994) method for inducing guilt.

1.3 Guilt induction as an equilibrium of $\Gamma _{\mathrm{PPT}}$

Next, I show that effective guilt induction in $\Gamma _{\mathrm{PPT}}$ can be supported as sequential equilibrium (SE) of $\Gamma _{\mathrm{PPT}}$ under the framework of B&D.³¹ The intuition behind this rests in the fact that in an SE, an assessment (profile of behavioral strategies and conditional hierarchical beliefs) will be consistent. Battigalli and Dufwenberg (2009) show that in equilibrium, “players never change their beliefs about the conditional beliefs that the opponents would hold at each h (history)” (pp. 16). Thus, in equilibrium, there is no belief updating. It follows that $E_{B}|C^{0}=E_{B}|C^{N}=E_{B}|C^{6}$, which implies that Conditions 1 and 2 from above would be satisfied in an equilibrium. In this equilibrium analysis, I will assume that there is complete information regarding Player B’s guilt sensitivity, $\theta _{B}$. While this assumption may be unrealistic, and not a feature of the design, it is common in the extant literature related to belief-dependent motivations (e.g., Charness and Dufwenberg 2006, 2011; Dufwenberg et al. 2013); furthermore, it allows me to focus on the motivation of this analysis, which is to establish that guilt induction can be supported as an equilibrium, without the information complexities that would otherwise need to be addressed. I also assume that Player A is not guilt averse, i.e., Player A’s guilt sensitivity is zero.³²

Guilt induction by Player A in $\Gamma _{\mathrm{PPT}}$ is characterized by the choice to Convey $X=0$ and Not Covey $X=6$. Therefore, the strategy (In, Convey $X=0$, Not Convey $X=6$) for Player A is consistent with attempted guilt induction. Effective guilt induction is characterized by Player B choosing Right as a response to the guilt induction by Player A. The following two strategies for Player B are consistent with responding in kind to Player A’s guilt induction: (Right $\vert C^{0}$, Right $\vert C^{N}$, Left $\vert C^{6}$), and (Right $\vert C^{0}$, Left $\vert C^{N}$, Left $\vert C^{6}$). Thus, the strategy profiles ((In, Convey $X=0$, Not Convey $X=6$), (Right $\vert C^{0}$, Right $\vert C^{N}$, Left $\vert C^{6}$)), and ((In, Convey $X=0$, Not Convey $X=6$), (Right $\vert C^{0}$, Left $\vert C^{N}$, Left $\vert C^{6}$)) are the candidate equilibrium profiles for effective guilt induction in $\Gamma _{\mathrm{PPT}}$. In Claims 1 and 2 below, I show that each of these strategy profiles can be supported as an equilibrium of $\Gamma _{\mathrm{PPT}}$.

Claim 1

The strategy profile ((In, Convey $X=0$, Not Convey $X=6$), (Right $\vert C^{0}$, Right $\vert C^{N}$, Left $\vert C^{6}$)) can be supported as a SE of $\Gamma _{\mathrm{PPT}}$ for $\theta _{B}\in [\frac{2}{5}, \frac{1}{2}]$.

To verify that this strategy profile is an equilibrium, we need to check that neither player has a profitable deviation. For Player A, this is rather trivial. By following the equilibrium strategy, Player A earns a payoff of 10, which is the highest payoff of the game. Therefore, Player A has no profitable deviation. For Player B, we need to consider deviations at each of the possible conveyance states. Given consistent beliefs in equilibrium, we have that ${\varvec{\alpha }}_{A}={\varvec{\beta }}_{B}=(1,1,0), E_{A}=E_{B}|h=10\ \forall h\in \{C^{0},C^{N},C^{6}\},$ and ${\widehat{m}}_{A}=5.$ Player B will not deviate to Right $\vert C^{6}$ so long as: $6-\theta _{B}\cdot [10-6]\ge 4\Leftrightarrow \theta _{B}\le \frac{1}{2}.$ Player B will not deviate to Left $\vert C^{N}$ so long as: $4\ge 6-\theta _{B}\cdot [10-5]\Leftrightarrow \theta _{B}\ge \frac{2}{5}.$ Similarly, Player B will not deviate to Left $\vert C^{0}$ so long as: $4\ge 6-\theta _{B}\cdot [10-0]\Leftrightarrow \theta _{B}\ge \frac{1}{5}$ which is satisfied if $\theta _{B}\ge \frac{2}{5}$.

Claim 2

The strategy profile ((In, Convey $X=0$, Not Convey $X=6$), (Right $\vert C^{0}$, Left $\vert C^{N}$, Left $\vert C^{6}$)) can be supported as a SE of $\Gamma _{\mathrm{PPT}}$ for $\theta _{B}\in [\frac{2}{7} ,1]$.

Again, to verify that this strategy profile is an equilibrium, we need to check that neither player has a profitable deviation. For Player A, playing the equilibrium strategy yields an expected payoff of 7; therefore, Player A cannot profitably deviate to Out. Player A would not deviate and Not Convey $X=0$, which would result in a payoff of 0 compared to a payoff of 10 from following the equilibrium strategy to Convey $X=0$. Player A is indifferent between Convey $X=6$ and Not Convey $X=6.$ Therefore, Player A has no profitable deviation from the prescribed equilibrium strategy. For Player B, we need to consider deviations at each of the possible conveyance states. Given consistent beliefs in equilibrium, we have that ${\varvec{\alpha }}_{A}={\varvec{\beta }}_{B}=(1,0,0), E_{A}=E_{B}|h=7\ \forall h\in \{C^{0},C^{N},C^{6}\},$ and ${\widehat{m}}_{A}=5.$ Player B will not deviate to Right $\vert C^{6}$ so long as: $6-\theta _{B}\cdot [7-6]\ge 4\Leftrightarrow \theta _{B}\le 2$. Player B will not deviate to Right $\vert C^{N}$ so long as: $6-\theta _{B}\cdot [7-5]\ge 4\Leftrightarrow \theta _{B}\le 1.$ Similarly, Player B will not deviate to Left $\vert C^{0}$ so long as: $4\ge 6-\theta _{B}\cdot [7-0]\Leftrightarrow \theta _{B}\ge \frac{2}{7}$.

1.4 Incomplete information regarding sensitivity to guilt

As I stated in the previous section, equilibrium Claims 1 and 2 assume common knowledge of Player B’s sensitivity to guilt, $\theta _{B}.$ In a recent paper, Attanasi et al. (2015) have extended the B&D model to allow for incomplete information regarding guilt sensitivities in a 2-player trust game.³³ A formal application of their model to $\Gamma _{\mathrm{PPT}}$ would require extensive theoretical preliminaries regarding the formulation of hierarchies of beliefs for each player via “type” structures, which is beyond the scope of this paper and would not add a proportionate increase in the main conclusions drawn. In addition, because I do not elicit any information about beliefs or types in the experiment (as the sessions were run prior the Attanasi et al. 2015 paper), I would be unable to derive precise predictions of the incomplete information model. As a result, I forgo a formal equilibrium analysis of $\Gamma _{\mathrm{PPT}}$ under incomplete information; rather, I appeal to insights from the Attanasi et al. (2015) paper to provide some general discussion regarding how incomplete information about types could impact behavior in $\Gamma _{\mathrm{PPT}}.$ ³⁴

The main difference between the complete information and incomplete information extension of the guilt aversion model (as is most relevant to the discussion that follows) is that in an incomplete information setting where Player As and Player Bs are matched at random (as in the current experimental procedure), it is likely that participants hold heterogeneous and dispersed beliefs (Attanasi et al. 2013, p. 18). The important behavioral implication is that under complete information, there is likely to be much more polarization of behavior because the common knowledge of $\theta _{B}$ enables coordination. Specifically, if $\theta _{B}$ is small enough then Player B will not be susceptible to guilt induction and will choose Left, and Player A will choose Out (the standard payoff maximizing equilibrium outcome); on the other hand, if $\theta _{B}$ is sufficiently large then Player B will be susceptible to guilt induction by Player A, and Player A could increase his/her payoff by choosing In and then inducing guilt (consistent with Claims 1 and 2 above).

However, with incomplete information about $\theta _{B}$ (a reasonable assumption with regard to Player As in the experiment) more dispersion in beliefs and mis-coordination can arise, which can possibly lead to interesting patterns of behavior in $\Gamma _{\mathrm{PPT}}$. Namely, many Player As could have held beliefs that Player Bs were not sensitive to guilt (i.e., low $\theta _{B}$), which may have motivated them to choose Out, which could explain why there was only a marginal 14 % increase in In rates between $\Gamma _{\mathrm{PPT}}$ and $\Gamma _{\mathrm{UPT}}.$ Additionally, for those Player As that did choose In (for other possible reasons besides attempting to exploit the guilt aversion of Player B), they would not be motivated to convey $X=0$ and/or not convey $X=6$ as a means of inducing guilt, as it would not be effective given their beliefs about $\theta _{B}$ (which is consistent with the analysis of the questionnaire data in Table 4); these Player As may opt, for example, to Convey $X=6$ to display honesty to Player B, consistent with the discussion in Sect. 4.3. On the other extreme, some Player As could have believed that Player Bs were very sensitive to feeling guilt (i.e., high $\theta _{B}$), such that Player B would be motivated to choose Right regardless of the conveyance state; in this case, these Player As would, again, not necessarily be motivated to convey $X=0$ and/or not convey $X=6$ as a means of inducing guilt, as it would not be necessary to motivate Player B to choose Right.

If we assume that Player As have incomplete information about Player B’s guilt sensitivity, $\theta _{B},$ then multiple equilibria can arise in $\Gamma _{\mathrm{PPT}}$, as in the case with perfect information. In particular, the “selfish” equilibrium exists that results in Player A always choosing Out since Player B would always choose Left, if Player Bs have sufficiently low $\theta _{B}$s.³⁵ In addition, by placing some restrictions on the distribution of $\theta _{B}$ (e.g., assuming $\theta _{B}\in [\frac{2}{7},1]$), equilibria that correspond to effective guilt induction (those specified in Claims 1 and 2 above) can also arise. However, with incomplete information regarding $\theta _{B}$, players do not have the ability to coordinate based on $\theta _{B}$; hence, it is much more likely to observe heterogeneity in behavior because of the dispersion in beliefs (as shown by Attanasi et al. 2013). As a result, the fact that participants playing the role of Player A posses incomplete information regarding Player B’s sensitivity to guilt could explain some of the observed patterns in the data; namely, why only marginally more Player As chose In in $\Gamma _{\mathrm{PPT}}$, as well as the conveyance behavior inconsistent with attempted guilt induction, where Player As choose to Not Convey $X=0$ and Convey $X=6$.

Appendix B.1: Copy of experimental instructions

1.1 Sample instructions: PPT treatment

Welcome and thank you for participating. Your participation is VOLUNTARY, and you may leave at any time. Feel free to raise your hand and ask questions at any time, and you may refer back to these instructions at any time during the session. Please remain seated and quiet for the remainder of the session. All decisions are to be completed individually and interaction with other participants is strictly PROHIBITED. Thank you for your cooperation.

Each person will receive a $5 show-up payment for participating. In addition, you can receive additional compensation based on the decision(s) that are made in the decision task described below. After the task is complete, you will be privately paid the amount of money you have earned. Upon completions of the decision task, please remain quietly seated in your carrel until you have been paid.

1.1.1 The decision task

You will be participating in a 2-person decision task. Each person will be randomly and anonymously paired with another person in the lab. In each of the 2-person decision making pairs, one person will be randomly assigned the role of PLAYER A and the other person will be randomly assigned the role of PLAYER B. You will remain in your assigned role for the entire session. The earnings of each Player will depend on the decision(s) he/she makes, and/or the decision(s) of the Player with whom they are paired. A brief outline of steps of the decision task will first be provided, followed by a detailed description of each step and the corresponding earnings for Player.

Step 1: PLAYER A begins by first choosing IN or OUT

• If PLAYER A chooses OUT, the task ends

• If PLAYER A chooses IN, then the task proceeds to Step 2

Step 2: PLAYER A might privately learn some payoff information that was initially unknown to both players. If PLAYER A does learn the information, the PLAYER A will then have an opportunity to convey the information to PLAYER B. Then the task will proceed to Step 3.

Step 3: PLAYER B chooses between RIGHT or LEFT, and the task ends

A full description of each step and the corresponding earnings for Player follows:

• 1. PLAYER A first chooses between IN or OUT.

  • If PLAYER A chooses OUT, then the decision task ends. PLAYER A will receive $6 and PLAYER B will receive $2.

  • If PLAYER A chooses IN, the task proceeds to step (2) where PLAYER A might privately learn the unknown information, and then have an opportunity to convey that information to PLAYER B. After step (2), the task will proceed to step (3) where PLAYER B will then be asked to decide between RIGHT or LEFT.

• 2. I postpone the details about the information that PLAYER A can possible learn, and convey to PLAYER B until after step (3) is described. Describing step (3) first will help clarify step (2).

• 3. If PLAYER A chooses IN, at step (1), then PLAYER B must choose between RIGHT or LEFT.

  • If PLAYER B chooses RIGHT, then the decision task ends. PLAYER A will receive $10 and PLAYER B will receive $4.

  • If PLAYER B chooses LEFT, then the decision task ends and PLAYER A will receive $X and PLAYER B will receive $6. There is a 50 % chance that X $=$ $0 and a 50 % chance that X $=$ $6. That is, X $=$ $0 and X $=$ $6 are equally likely.

NOTE: When the decision task begins, neither PLAYER A nor PLAYER B knows the value of X. Therefore, PLAYER A does not know the value of X when he/she decides between IN or OUT in step (1).

Back to the description of step (2):

• 2. If PLAYER A chooses IN, there is an 80 % chance that PLAYER A will privately learn the value of X, and a 20 % chance that PLAYER A will not privately learn the value of X.

  • If PLAYER A does learn the value of X (80 % chance), then PLAYER A must then decide whether or not to convey the value of X to PLAYER B before PLAYER B makes his/her decision in step (3).

    • If PLAYER A does convey the value of X, then PLAYER B will know the value of X before he/she decides between RIGHT or LEFT in step (3).

    • If PLAYER A does not convey the value of X, then PLAYER B will not know the value of X before he/she decides between RIGHT or LEFT in step (3).

  • If PLAYER A does not learn the value of X (20 % chance), then PLAYER A will not have an opportunity to convey the value of X to PLAYER B. The task will proceed to step (3) where PLAYER B will then choose between RIGHT or LEFT without knowing the value of X.

If PLAYER A chose to convey the value of X, then when PLAYER B makes his/her decision in step (3), the following message will appear on PLAYER B’s screen:

“PLAYER A has chosen to convey the value of X to you”

“The value of X is: [actual value of X]”

If either (a) PLAYER A did not learn the value of X after choosing IN, or (b) PLAYER A did learn the value of X but chose not to convey the value of X, then when it is time for PLAYER B to make his/her decision in step (3), the following message will appear on PLAYER B’s screen:

“The value of X remains unknown”

Payoff Table:

The table below summarizes the earnings of each Player for each of the possible outcomes in the decision task:

DECISION OUTCOME	Earnings of PLAYER A	Earnings of PLAYER B
PLAYER A chooses OUT	$6	$2
PLAYER A first chooses IN and then:
PLAYER B chooses RIGHT	$10	$4
PLAYER B chooses LEFT	$X	$6
There is a 50 % chance X = 0 and a 50 % chance X = 6

Each person will participate in this decision making task ONE time. After the task has ended, the decision(s) of each Player and the corresponding earnings of each Player will be revealed to both Players. Additionally, the value of X will be revealed to both PLAYER A and PLAYER B regardless of the decisions made in the task. You will then be asked to fill out a short questionnaire that will take about 3 min to complete. Your answers to the questionnaire are confidential and will not be shared with any other participants. After completion of the questionnaire, an Experimenter will then come by and privately pay you your total experimental earning which equals your earnings from the decision task PLUS the $5 show-up payment. After you have been paid, you may quietly exit the lab.

1.2 Appendix B.2: Sample screen shots

Player A’s Initial In/Out Decision

Player A’s Conditional Convey/Not Convey Decision

Player B’s Conditional Left/Right Decision (X Known)

Player B’s Conditional Left/Right Decision (X Unknown)