RECIPROCITY AND UNCERTAINTY: WHEN DO PEOPLE FORGIVE? Andrés Gago ∗† a
CEMFI, Madrid, Spain
15th December 2017
Abstract A sizeable proportion of individuals act reciprocally. They punish and reward depending on the (un)kindness of those with whom they interact. In this paper I study whether individuals still reciprocate looking at intentions when they are in an uncertain environment. By means of a dictator game with punishment opportunities, I show that when people intend to harm and fail, they are still punished. By contrast, if they intend to be nice and accidentally harm, they are forgiven. In order to isolate how uncertainty affects the assessment of intentions, I control for other possible departures from self-profit maximization such as inequity aversion, maximin preferences, spitefulness, envy, and efficiency maximization. Keywords: Reciprocity, Uncertainty, Blame, Intentions, Dictator, Punishment. JEL codes: C91, D63, C79.
∗
Postal address: c/ Casado del Alisal 5, 28014, Madrid (Spain) Email address:
[email protected] † I wish to thank G. Caruana, P. Rey-Biel, G. Llobet, M. Arellano, S. Bonhomme, A. Cabrales and seminar participants at CEMFI and the IMEBESS 4th Conference for their helpful discussions. I also want to thank Universidad Carlos III (Madrid) for letting me use the LEE UC3M lab, and especially O. Powell for his help there. Funding from CEMFI is gratefully acknowledged. CEMFI as an institution had no involvement in study design; in the collection, analysis and interpretation of data; in the writing of the report; or in the decision to submit the article for publication. An earlier version of this paper was published in CEMFI series [1105] with the title: “Reciprocity: Is it Outcomes, or Intentions? A Laboratory Experiment”. All consent required by Universidad Carlos III for experimentation with human subjects was obtained. Conflicts of interest: none.
1
1
Introduction
The literature on experimental economics has shown that around one third of people act reciprocally. They are willing to forgo wealth in order to reward those who have been nice to them and punish those who have been hostile (for a discussion on this see Fehr and Schmidt (2006) and Cooper and Kagel (2016)). Nevertheless, people are not always successful when they intend to be either nice or hostile, as the result of their actions is usually subject to uncertainty. In this paper I study how lack of full control over outcomes affects reciprocal responses. The classical view in psychology literature is that subjects are confounded when they judge actions with unintended results. In a seminal paper, Baron and Hersey (1988) find that people evaluate differently decisions that are ex-ante identical in probabilistic terms but lead to different outcomes. In this paper I challenge this view. I find that reciprocal individuals punish and reward according to the will behind an action, even if its result is the opposite of what was intended.1 When people intend to harm and fail they are punished. Similarly, if they intend to be nice and accidentally do harm they are forgiven. Adding uncertainty to human action does not change the fact that reciprocity is intention based. To reach this conclusion I run a game in which a dictator chooses between two lotteries that determine the allocation of e20 between herself and a recipient. After being informed about the choice of the dictator and the result of the lottery, the recipient chooses whether she wants to assign any punishment (or reward) to the dictator, at a cost. This enables me to compare the reaction to choices by the dictator with the same intentions but with different consequences. To isolate the effect on reciprocal responses of the lack of full control by dictators over consequences, I run an additional treatment. In this treatment (nature treatment), outcomes are decided by a random device from the very beginning. A reciprocal response is defined as one that judges the kindness/hostility of an action taken by another individual to reward or punish her (possibly against one’s own material interest). The fact that in the nature treatment the dictator takes no action prior to the recipient’s decision rules out reciprocity as an explanation. A comparison of the behavior of recipients in the reciprocal and nature treatments enables me to distinguish between a misjudgement of a dictator’s behavior and a response to outcomes that is independent of the dictator’s choice (motivated by welfare or distributional concerns). In this experiment more than one third of recipients are willing to forgo wealth to punish and reward hostile and kind actions. This confirms the relevance of reciprocity as a departure from traditional self-profit-maximizing behavior. Moreover, the finding that reciprocal individuals correctly assess intentions when dictators have only partial control over outcomes is consistent with the extension by Sebald (2010) of the theory of sequential reciprocity in Dufwenberg and Kirchsteiger (2004). Kindness perceptions depend on the ex-ante probability 1
I refer to individuals with reciprocal motivations as “reciprocal individuals”.
2
that the choices made by others attribute to either outcome. In a context in which participants are properly incentivized and competing explanations are controlled for, I find no evidence of the outcome bias documented by papers in psychology (see Baron and Hersey (1988) or Gino et al. (2010)). The result of this study is at odds with the no-harm-no-foul hypothesis proposed in Bartling and Fischbacher (2012). In an interesting study about attribution of responsibility they conclude, among other things, that when the outcome derived from an action is good people will not punish, no matter what the intention was. An important difference with Bartling and Fischbacher (2012) is that in my design the expected result of the hostile choice is markedly worse for the recipient than that of all alternative choices. Moreover, in my design there is no ambiguity about what the dictator expects when she makes the choice (the probabilities of the lottery are known to everyone). Both things together make intentions easy to identify. Arguably, when intentions are clear enough it is no longer possible to elude responsibility by hiding behind a good outcome.2 Closely related to my experiment is Charness and Levine (2007) . They find that people reward good intentions in a gift exchange game with uncertainty. Nevertheless, there are some important differences with my study. First and foremost, they look at positive reciprocity while in my design the focus is on negative reciprocity. There is evidence of the two being uncorrelated motivations that can be held by different individuals (Offerman, 2002; Dohmen et al., 2008). Focusing on negative reciprocity enables me to test the hypothesis that no harm implies no foul and observe whether individuals forgive after accidental damage. This has theoretical and practical applications and makes both studies interesting in their own right. Moreover, the inclusion of the non-reciprocal treatment enables me to distinguish an outcome bias in reciprocal responses from any other behavioral departure from self-profit maximization. Also related to this paper is Rubin and Sheremeta (2015). They study the effectiveness of reciprocity as an informal enforcement device in a double gift exchange game. Principals make an investment, state the returns that they expect, and assign a bonus/malus accordingly. They find that with random shocks to the effort level, no matter whether they are observable or not, individuals get further from the optimum. Principals pay lower wages, agents put in less effort, and bonuses are reduced. To disentangle strategic considerations from reciprocity, in my experiment I look at a simpler setting and avoid framing the situation as an investment opportunity (Stanca et al., 2009). The papers by Rand et al. (2015), Bereby-Meyer and Roth (2006), Xiao and Kunreuther 2
Bartling and Fischbacher (2012) derive the no-harm-no-foul hypothesis from the observed responses in two situations: In one of them individuals can choose a fair allocation or delegate the choice of the allocation to someone else. This means that the probability of each allocation after delegation is unknown, i.e. there is ambiguity. In the other they can choose a fair allocation or an unfair allocation, or they can delegate the choice to a random device with a known probability distribution. This means that delegation is the intermediate (and arguably neither kind nor unkind) choice. Altogether, these could make the intention behind the choice to delegate unclear in both situations, explaining why recipients choose not to punish if the outcome is good.
3
(2015), and Klempt (2012), which study how uncertainty affects the ability to cooperate, also stand at the intersection of social preferences and uncertainty. The same is true for those of Cappelen et al. (2013) (who study whether inequity-averse individuals care about people having equal chances or equal outcomes), Gurdal et al. (2013) (who let an agent decide whether to invest in a safe or a risky lottery on behalf of a principal), Krawczyk and Le Lec (2010), and Brock et al. (2013) (who examine the behavior of dictators after introducing noise into their decisions). However, none of them tackles the specific question that is central to this paper: Whether negative reciprocity is still intention-based in the presence of uncertainty. The conclusions of this study apply to any situation in which reciprocity has been found to be relevant and individuals do not have perfect control over outcomes. Consider for example a historically diligent entrepreneur who sees her sales plunging due to the global crisis. Compare this to an irresponsible board of directors earning astronomic bonuses while the company’s profits are suffering. According to the results of my experiment, if they imposed salary cuts the reactions of their workers would be rather different. In the latter case there would be much harsher opposition than in the former. A bad event (a salary cut), may or may not be forgiven depending on the intention behind it. Similarly, if roles are reversed entrepreneurs might be willing to maintain bonuses after not meeting targets as long as they consider that workers’ attitudes/efforts are satisfactory.3 The rest of the paper is organized as follows. Section 2 presents the design of the experiment, Section 3 gives the theoretical predictions, Section 4 analyzes the results, and Section 5 concludes. Further applications of the results are also presented in Section 5.
2
Experiment design
The experiment is designed as a dictator game with punishment opportunities and three different treatments. In the first two treatments (reciprocal treatments 1 and 2) individual A (the dictator) chooses how to split e20 between herself and individual B (the recipient), with the particularity that the result of her choice is not deterministic (she chooses a lottery). After observing the choice and the outcome, individual B can choose whether to add or subtract points to individual A incurring in a cost. In the third treatment (nature treatment) a random device assigns the allocation from the very beginning. Again after observing the outcome, individual B can add or subtract points to individual A at a cost. Changing the lotteries that are available for the dictator in reciprocal treatments 1 and 2 allows me to test different hypothesis looking at the responses of B in different scenarios. Meanwhile the nature treatment allows me to distinguish distributional or efficiency concerns from the cognitive biases individual B might suffer when she judges individual A’s actions (see Section 3 for the details). 3 The relevance of reciprocity in labor market relationships is shown in Akerlof (1982); Dufwenberg and Kirchsteiger (2000); Krueger and Mas (2004); Dohmen et al. (2009); Brandes and Franck (2012); Kube et al. (2012); Cohn et al. (2014); Cobo-Reyes et al. (2017), and Gilchrist et al. (2016) among others.
4
Experimental sessions were conducted in the LEE UC3M lab at Carlos III University (Madrid, Spain). The subjects who participated in the experiment were undergraduate students of engineering, economics, business administration, law, and sociology. They were recruited using the mailing list of the laboratory at Carlos III (1100 students approx.). To run the experiment I used z-Tree experimental software (Fischbacher, 2007). Treatments were presented and explained one at a time to participants. In every treatment pairs were matched at random and the role of each participant was private information. Subjects were paid a show-up fee of e5 plus e0.20 for each experiment point that they earned during the game. The average payoff was e14.85. Sessions lasted on average 45 minutes, plus 30 minutes until everyone got paid.
2.1
Reciprocal treatment 1
In the first treatment individual A (the dictator) must decide between two options to split 100 points between herself and individual B (the recipient) with whom she has been paired. At a cost, individual B then chooses how many positive or negative points she wishes to award to individual A contingent on this decision. Figure 1 shows how this treatment works. If individual A chooses Option 1, 50 points go to herself and 50 to individual B. If she chooses Option 2 a die is rolled by the computer. There is a probability of 5/6 that the die will place them in Option 2 Left, where 80 points go to herself and 20 points to individual B. There is a probability of 1/6 that the die will place them in Option 2 Right, where 50 points go to herself and 50 points to individual B. This makes Option 1 the kind option and Option 2 the hostile option. Contingent on this decision, individual B chooses how many points she wants to award to individual A. Throughout the experiment, awarding 3x positive points has a cost for individual B of 1x point. Awarding 3x negative points has a cost for individual B of 1x point.4 This means that both adding and subtracting points to/from A is costly for B. Allowing B to allocate positive and negative points avoids any experimenter demand effect for punishments.
4
Awarding -9 points to A costs player B 3 points, awarding -21 points costs her 7 and so on so forth. Analogously, awarding +12 points to A costs individual B 4 points, awarding +33 costs her 11 points, etc.
5
Figure 1: Reciprocal treatment 1 Individual A
Option 1
Option 2 Random Device
Individual B
Prob. 5/6 (Left)
Individual B
Prob. 1/6 (Right)
Individual B
X Payoff A: 50+X Payoff B: 50-|X/3|
Y
Z
Payoff A: 80+Y
Payoff A: 50+Z
Payoff B: 20-|Y/3|
Payoff B: 50-|Z/3|
X, Y and Z represent points awarded by B to A in each situation.
The maximum number of positive points that individual B can assign to A is +48 and the
maximum number of negative points is -48.
Imposing these upper- and lower-bounds ensures
that nobody ends up having negative payoffs. These limits hold for the three treatments.
2.2
Reciprocal treatment 2
This treatment (Fig. 2) is similar to reciprocal treatment 1, but the options available to individual A are now different. Randomness is introduced in the kind option rather than in the unkind one. As shown in Section 3 below, this enables additional predictions to be tested. If Individual A chooses Option 1, 80 points go to herself and 20 to individual B. If she chooses Option 2 a die is rolled. There is a probability of 5/6 that the die will place them in Option 2 Left, where 50 points go to herself and 50 to individual B. There is a probability of 1/6 that the die will place them in Option 2 Right, where 80 points go to herself and 20 to individual B. Then individual B chooses how many points she wants to award to individual A.
6
Figure 2: Reciprocal treatment 2 Individual A
Option 1
Option 2 Random Device
Individual B
Prob. 5/6 (Left)
Individual B
Prob. 1/6 (Right)
Individual B
X Payoff A: 80+X Payoff B: 20-|X/3| Y
Z
Payoff A: 50+Y
Payoff A: 80+Z
Payoff B: 50-|Y/3|
Payoff B: 20-|Z/3|
X, Y and Z represent points awarded by B to A in each situation.
2.3
Nature treatment
Finally, in the nature treatment, 100 points are again divided between A and B, but in this case a die determines how points are split without player A’s participation (see Fig. 3). The die is rolled by the computer from the very beginning. There is a probability of 2/3 that individuals will be placed in Option 1, where 80 points go to individual A and 20 to individual B. There is a probability of 1/3 of their being placed in Option 2, where 50 points go to A and 50 points to B. Following Falk et al. (2008), these probabilities were chosen so as to roughly mimic the decisions taken by dictators in some initial pilots that I ran, so the random device is perceived as neutral.5 After the die is rolled, individual B has the possibility of adding or subtracting points to/from individual A at a cost. 5
Bolton et al. (2005) describes how people react to procedural fairness. The authors show that people judge random devices as “fair” or “unfair” and that this has an impact on decisions of players B (for more information on procedural fairness see also Mertins et al. (2013); Mertins (2008)). Following Falk et al. (2008) I assume that people would judge as neutral a random device that imitates reality. Another possible solution would have been to choose a 50-50 random device, a plausible focal point for a neutral device.
7
Figure 3: Nature treatment Random Device
Option 1
Prob. 2/3
Option 2
Individual B
Prob. 1/3
Individual B
X
Y
Payoffs A: 80+X
Payoffs A: 50+Y
Payoffs B: 20-|X/3|
Payoffs B: 50-|Y/3|
X and Y represent points awarded by B to A in each situation.
Notice that in this treatment there are no friendly or unfriendly options, given that player
A is taking no decision.
Hence, points assigned by B should not be thought of as punishments
orrewards. As shown in Section 3 below, they capture any of the competing behavioral explanations different from reciprocal behavior that could explain departures from self-profit maximization.
2.4
Experiment procedure
66 subjects participated in the experiment, 33 as individual A and 33 as individual B. Before the game started, each subject was randomly assigned to the role of player A or B. Participants maintained the same roles throughout the experiment. They played once in each treatment. This is a key point in the design, as the impact of reciprocity is measured as deviations relative to behavior in the nature treatment (see Section 3 for details). Using a within subject design allows me to control for individual fixed effects. Even though the diagrams may give the impression that this was a dynamic game, in practice subjects played simultaneously. Individuals B had to give a full contingent plan of action before knowing what A had chosen (this is normally referred to as the strategy method). Contingent choices for each treatment were given one at a time. In a survey, Brandts and Charness (2011) find preponderant evidence that the strategy method has no impact on results. Moreover, as it was used in all three treatments, any impact that it had would have been constant across them all. In order to prevent learning effects or reputation building, participants were not told the 8
result of any treatment or any decision taken by any player until the experiment was over. To control for possible order effects I used a counterbalanced design, so there were 34 individuals playing first reciprocal treatments and afterwards the nature treatment, and 32 individuals playing in the reverse order. However, the posterior statistical analysis suggests that order effects are not an issue in this setting.6 After playing all three treatments participants were paid for one of them chosen at random. Every treatment had the same probability of being the payment treatment, and that was public knowledge. Individuals were seated in front of their terminals and given the instructions, which were read aloud so that there was no doubt that they were common to everyone. Instructions also included illustrative examples. Participants were not permitted to get up or talk, and any doubts were solved in private. Once the instructions had been read and any doubts clarified, but before the experiment actually started, they had to answer correctly a number of control questions to make sure that everybody had understood the experiment rules. The experiment instructions and the control questions can be found in the appendix.
3
Predictions
When other-regarding preferences are studied the same observed behavior can often be explained by different motivations. Hence, to make sure that the role of intentions in reciprocal behavior under uncertainty is pinned down it is highly important to control for competing explanations. To that end in this section I list the predictions of intention-based negative reciprocity together with what competing theories in the literature would predict in all the possible scenarios. Then, comparing the behavior of individuals in different situations, I will be able to identify whether recipients suffer an outcome-bias, as has been argued by psychology literature, when they judge dictators’ choices. Table 1A presents the full game. In each situation, different motivations lead to different optimal strategies for the recipient. Even though a subject’s preferences might actually be a mixture of some of these motivations, for the sake of clarity in the exposition I group them under three categories and describe them individually. Below, when I present the tests, I account for the possibility that some of them might jointly affect the decisions of the recipient. Table 1B summarizes the predictions of each type of preference. Considering that assigning positive or negative points reduces one’s own profits, a traditional self-profit-maximizer recipient would always assign zero points to the dictator. 6
After running the experiment, I performed a Mann-Whitney U-test to check whether there were actually any order effects. I found that in 7 out of the 8 possible final situations for individuals B, the answers were not statistically different for the two groups. Only in reciprocal treatment 2 when Option 1 is chosen and the outcome is (80,20) punishments differ.
9
Table 1 Panel A: Full Game Reciprocal Treatment 1 Kind
Unkind
Nature Treatment Option 1
Option 2
Pr. 5/6 Left
Pr. 1/6 Right
Pr. 2/3
Pr. 1/3
Reciprocal Treatment 2 Unkind
Kind Pr. 5/6 Left
Pr. 1/6 Right
Individual A
50
80
50
80
50
80
50
80
Individual B
50
20
50
20
50
20
50
20
PR1 K50
PR1 U20
PR1 U50
PN20
PN50
PR2 U20
PR2 K50
PR2 K20
Points given by B to A*
*In the lower line, P stands for ‘Points’; the super-index R1, N, R2 denotes “reciprocal treatment 1”, “nature treatment” and “reciprocal treatment 2”; the sub-index U, K stands for “unkind”, “kind” – therefore in the nature treatment there is neither of them-; and the sub-index 20, 50 denotes an (80,20) or a (50,50) outcome.
Table 1 Panel B: Predictions* PR1 K50
PR1 U20
PR1 U50
PN20
PN50
PR2 U20
PR2 K50
PR2 K20
Self-Profit Maximization
0
0
0
0
0
0
0
0
Spitefulness
(-)
(-)
(-)
(-)
(-)
(-)
(-)
(-)
Efficiency Maximization
(+)
(+)
(+)
(+)
(+)
(+)
(+)
(+)
Distributional Concerns**
0
(-)
0
(-)
0
(-)
0
(-)
Intention-based Negative Reciprocity
0
(-)
(-)
0
0
(-)
0
0
Outcome-biased Negative Reciprocity
0
(-)
0
0
0
(-)
0
(-)
*What different behavioral explanations prescribe in each possible final situation: 0 stands for giving zero points, (-) for giving maximum negative points, (+) for giving maximum positive points. ** Inequity aversion, maximin preferences and envy.
The first category covers mutually exclusive preferences that always prescribe the same
behavior. I refer to them as unconditional preferences. In both cases the reason that explains the behavior is the same: Positive and negative points have a cost of 1 to 3. This means that recipients can harm others more than they harm themselves by always assigning negative points, so if they are guided only by spitefulness (Klempt, 2012) they should always give negative points to individual A. Similarly, total welfare could be maximized by giving positive points, so individuals B that only care about efficiency maximization (Kamas and Preston, 2012; Engelmann and Strobel, 2004, 2006; Charness and Rabin, 2002; Dufwenberg and Gneezy,
10
2000) should always give maximum positive points.7 The second category covers distributional preferences: Envy (Kirchsteiger, 1994; CoboReyes and Jiménez, 2012), inequity aversion (Cox, 2004; Bolton and Ockenfels, 2006; Fehr et al., 2006; Falk et al., 2003), and maximin preferences (Engelmann and Strobel, 2004, 2006). Under such motivations, recipients maximize their utility by taking points away from A whenever they are behind -[80, 20] situations-, and assigning zero points when they are even -[50,50] situations-. Distributional preferences also rely on the 1 to 3 cost for negative points, but in this context it becomes a mechanism for reducing payoff inequality. They are conditional on outcomes but as in the case of unconditional preferences, they do not depend on the dictator’s participation. Finally, the last category covers reciprocity (Brandts and Sola, 2001; Bolton et al., 1998; Falk et al., 2003; Nelson, 2002; Cox, 2004; Blount, 1995; Falk et al., 2008). Reciprocity does depend on the dictator’s participation, as it prescribes different behaviors depending on her actions. It relies on the 1 to 3 cost as a punishment device. If negative reciprocity is intentionbased recipients will maximize their utility by punishing dictators after ill-intended choices, and giving zero points otherwise. However, if it is true that people do not punish ill-intended choices after good outcomes (no-harm-no-foul), and/or it is true that they punish others after accidental damage, then once uncertainty is introduced reciprocal individuals become not intention-based but outcome-biased.8 In the next sub-sections I present the four tests that answer these questions. Table 1A details the acronyms used in the text to denote the points given by recipients in each situation.
3.1
Does reciprocity matter?
First I want to test whether the results in this experiment are in line with previous evidence. PUR1 20
and PUR220 in Table 1A represent the situation that previous papers studying negative re-
ciprocity have addressed, i.e. how people react to bad outcomes that result from ill-intended choices. Following McCabe et al. (2003), Falk et al. (2008), Blount (1995), and Gächter and Thoni (2010), I disentangle reciprocity from distributional preferences by testing whether punishments in reciprocal treatments are harsher than punishments in the nature treatment when the roll of the die leads to [80,20]. R2 N H0 : PUR1 20 , PU 20 = P20 7
For the sake of clarity I am describing the utility-maximizing behavior of individuals who are guided only by spitefulness/welfare concerns. However in the literature these are never the sole motivations of agents. Later in the tests I account for the fact that non-mutually exclusive motivations could appear jointly in the utility functions of participants. 8 Whether reciprocal individuals responding to outcomes are really outcome-biased or consequence-based is a rather philosophical question. In the latter case they would consciously disregard intentions and reward and punish the other’s actions depending on outcomes. In the former they would do so by mistake. Psychology papers such as Baron and Hersey (1988) or Gino et al. (2010) portray this behavior as a cognitive bias. I stick to this interpretation.
11
N If PUR120 and PUR220 turn out to be more negative than P20 , this would be evidence that negative
reciprocity is making B punish A more aggressively, over and above any possible distributional concerns or spiteful behavior.
3.2
Is intention-based reciprocity robust to uncertainty?
Once I confirm that negative reciprocity is relevant, I test whether it is still intention-based in a context of uncertainty. This is the main contribution of the paper. I test whether subjects are forgiven after accidental damage, and whether they are punished if they intend to harm and fail. N First, I compare PUR150 and P50 .
N H0 : PUR1 50 = P50
If the distribution of the former is below that of the latter, then when player A tries but fails to be greedy she is still punished and the no-harm-no-foul hypothesis is rejected. Notice that N P50
N then hostile intentions captures welfare maximizing concerns (if any). If PUR150 is below P50
are making those players B who are not efficiency maximizers punish, and making those who are give lower or negative points. N R2 is not below P20 then when people try to be nice and unwittingly do harm Likewise, if PK20
they are forgiven. Accidental damage is not punished beyond distributional concerns. N H0 : PUR2 20 = P20
Both things together would indicate that adding observable uncertainty does not change the fact that unkind intentions are a necessary and sufficient condition for explaining the decision to reciprocate. The outcome bias documented by psychologists would not affect reciprocal punishments.
3.3
Do consequences affect the intensity of punishments?
After establishing whether intentions are a sufficient and/or necessary condition for reciprocal punishments, I study the effect of outcomes on punishment intensity. Recipients could evaluate ill-intended decisions that lead to a bad outcome as worse, making their punishments N harsher. To test this hypothesis I compare the difference between P20 and PUR120 with the differN R1 ence between P50 and PU 50 . N R1 N H0 : PUR1 20 − P20 = PU 50 − P50
Notice that the expected outcome of the dictator’s choice in PUR120 and in PUR150 is the same. This means that when I make the comparison I keep intentions constant. Moreover, by subN N tracting P20 and P50 from PUR120 and PUR150 , I control for inequity reduction and welfare maxim-
ization opportunities. If different motivations enter additively into the utility function, sub-
12
tracting points given in the nature treatment is a way of isolating reciprocal behavior from inequity aversion, spitefulness, and efficiency maximization. N N If the difference between P20 and PUR120 turns out to be the same as the difference between P50
and PUR150 then the hypothesis that an outcome bias in reciprocal behavior affects punishment intensity can be rejected.
4
Results
There are 264 observations of individuals B. Using non-parametric tests and mean tests Section 4 assesses the robustness of intention-based reciprocal behavior and how it works under uncertainty. Table 2 shows the average points given by B to A in every possible situation.9 Table 2 - Average Points Given by B to A* Reciprocal Treatment 1 Unkind Kind 50 80 50 50 20 50
Nature Treatment Option 1 Option 2 80 50 20 50
Reciprocal Treatment 2 Kind Unkind 80 50 80 20 50 20
6 4 2 0 ‐2 ‐4 ‐6 ‐8 ‐10 ‐12 ‐14 ‐16 *Black lines show 90% confidence intervals
Table 3 shows the results of the tests introduced in Section 3. Paired t-tests compare the averages shown in the table, while Wilcoxon tests make a non-parametric comparison of the whole distributions. They both take account of dependence within individual. Altogether, they provide a clear picture. Intention-based negative reciprocity matters and is robust to uncertainty. If observable noise is introduced, reciprocal individuals still look at intentions when they judge a choice, and do not show any cognitive biases. In tests 1, I find evidence that negative reciprocity matters. The points assigned to the dictator when the outcome is bad are more negative when intentions are unkind. Even though the Wilcoxon p-value for test 1B is unexpectedly high, both the Wilcoxon and the t-test in 1A, 9
Information on aggregate choices of individuals A can be found in Tables A2 and A3 in the Appendix.
13
and the t-test in 1B give support to this result.10 Having someone behind an action makes recipients’ reaction stronger. This confirms previous evidence in McCabe et al. (2003); Falk et al. (2008); Blount (1995) and Gächter and Thoni (2010). Test 2 shows that actions with unkind intentions trigger punishments independently of the outcome. The points assigned by B in reciprocal treatment 1 when the unkind option is chosen and [50,50] is the outcome are significantly lower than those in the nature treatment when [50,50] is the outcome. The no-harm-no-foul hypothesis can be rejected. In the same fashion, given the results of test 3, the notion that people are forgiven when they try to be nice but accidentally do harm cannot be rejected. Punishments after a bad outcome when the intention was kind are not bellow points assigned to correct for payoff inequality in the nature treatment. Table 3 - Tests Wilcoxon**
t-test
0.02
0.01
= −1.73
0.14
0.05
= 1.09
0.01
0.05
0.77
0.97
0.655
0.98
Tests 1: Baseline Negative Reciprocity 1A:
= −10.64∗
