What Behaviors are Disapproved? Experimental Evidence from Five Dictator Games

Abstract

The literature on social norms has often stressed that social disapproval is crucial to foster compliance with norms and promote fair and cooperative behavior. With this in mind, we explore the disapproval of allocation decisions using experimental data from five dictator games with a feedback stage. Our data suggests that subjects are heterogeneous in their disapproval patterns, distinguishing two main groups: (1) Subjects who only disapprove choices that harm them, and (2) subjects who disapprove socially inefficient choices.

1. Introduction

Social scientists have often stressed that social disapproval can foster normative behavior and cooperation within groups—see [1] for numerous illustrations. Indeed, a growing experimental literature shows that the availability of approval/disapproval increases contributions in public goods games (see, for instance, [2,3,4]), cooperation in social dilemmas [5], or generosity in dictator games [6,7]. Given this evidence, it seems important to understand which kind of behavior tends to be disapproved, as this could clarify further how (and when) social disapproval affects cooperation and generous behavior.

Another, perhaps more speculative reason to study disapproval patterns has to do with social norms. In effect, much of the existing literature on social norms emphasizes that norms are partly sustained by the disapproval of others. [8], (p. 104), for instance, contends that “when people obey norms, they often want to avoid the disapproval—ranging from raised eyebrows to social ostracism— of other people.” Similarly, [9] (p. 366) notes that “norms are also enforced by expressions of disapproval, by ridicule, and in extreme cases, by ostracism. [...] If I dress bizarrely or spit in public, I may find myself shunned" (see also [10] and references therein, [11,12]. All this suggests that social norms might be elicited by studying what people disapprove [13], an important issue if we believe that behavior is often shaped by social norms.

This paper uses experimentation to study the disapproval of allocation decisions. For this, participants in our study play five dictator games with an additional stage in which the recipients can evaluate the dictator’s choice. We then apply the analysis of [14] in order to classify recipients according to their individual disapproval pattern. The results show that recipients are heterogeneous in this respect. In particular, we find that the best distributional model in terms of parsimony and empirical relevance distinguishes two types of recipients: those who disapprove of choices if they are harmed and those who disapprove of socially inefficient choices. Perhaps paradoxically, in our games equity considerations or payoff comparisons only play a relatively minor role. The data also provides insights into the content of social norms and more precisely on distributive norms, that is, norms regulating how to distribute income or other goods.

We proceed as follows. Section 2 describes and discusses our experimental design. In Section 3, we report some aggregate statistics and perform the classification analysis. We conclude in Section 4. The experimental instructions are relegated to the Appendix.

2. Experimental Design and Procedures

We consider five binary dictator games in which the recipient has the possibility to express her/his opinion about the dictator’s choice. The games are presented one at a time and each consists of two stages. In the first stage, the dictator chooses between two payoff allocations, which differ across games. In order to discriminate between different disapproval patterns (see next section), we employed the following ECU allocations (20 ECU = 1 Euro), where the first number in each parenthesis indicates the dictator’s payoff:

• $(180,140)$ and $(100,220)$
• $(180,120)$ and $(100,100)$
• $(240,120)$ and $(100,120)$
• $(180,120)$ and $(120,180)$
• $(200,120)$ and $(160,160)$

In the second stage, the recipient can send a closed–form, costly message (for a 10 ECU fee) to the dictator. The message depends on the dictator’s choice and must be chosen from the following: “Your choice was (1) good, (2) neither good nor bad, (3) bad.” In what follows, we interpret the first message as approval, the second as neutral, and the third as disapproval. One subtlety of our design is that subjects do not receive any feedback during the experiment about the choices they make of their match. For this reason, we elicited the recipient’s messages by means of the strategy method. In each allocation of each dictator game, more precisely, the recipient is first asked whether she/he wants to pay the message fee. Afterwards, and independently of her/his prior answer, the recipient is asked which of the three available messages she/he would like to send had she/he paid the fee. In this manner, we elicited 10 messages from each recipient (one for each allocation). Of course, the message is delivered only if the recipient decided to pay the fee at the allocation chosen by the dictator.

Two of the above mentioned features of our study merit some discussion. First, we used the strategy method because it maximizes the amount of statistical evidence (a crucial point for our classification analysis) and because we did not want subjects to receive any feedback. This was important to avoid variations in the players’ mood across games (e.g., a recipient might get angry at the dictator’s choice and that could affect her subsequent choices), and thus order effects which would complicate our classification analysis. While we believe that the use of multiple games and the strategy method is the most efficient manner to explore subjects’ disapproval patterns, some caution is warranted when interpreting our results, as the hypothetical character of this method might have an effect on its own. For instance, emotions could be attenuated and that could in turn affect the disapproval (see [15] for a survey of experiments using the strategy and the direct response method). Second, our design allowed us to know whether a recipient disapproved a choice even if the fee had not been paid. In other words, we elicited messages both when they are costly but also in the hypothetical case in which they were costless (as we asked the recipients even if they had not paid the fee). Although this is not typical in the experimental literature on non–monetary sanctions, which usually focus on costless messages, we find it an important design feature because a comparison between costly and costless messages allows us to see if the latter have been chosen carelessly or at random.

The experiment was computerized and programmed using z–Tree [16]. In total, 180 subjects from the Faculty of Economics and Business Administration at Maastricht University participated in the experiment. Since this experiment is rather short, we implemented it together with another experiment on a prisoner’s dilemma (PD) game, which consisted of three different treatments and whose results are detailed in [5].1 More precisely, after the corresponding PD treatment was finished, subjects received instruction sheets explaining the general structure of the dictator games (information about the payoffs was only provided for the first of the five games), were assigned a fixed role (dictator or recipient), and were randomly re-matched with another participant for the five games. All this was publicly announced. We decided not to re-match subjects after each dictator game because no repeated game effects were possible due to the absence of feedback; in fact, from a theoretical point of view, each game can be analyzed as one-shot. To ensure that the PD did not affect behavior in the dictator games, subjects were not informed about the results of the PD.2 In order to prevent income effects, only one game (among the PD and the five dictator games) was randomly selected for payment when all subjects had finished their choices (subjects knew this from the beginning). Finally, subjects were informed about their co–player’s actions in the payoff–relevant game. The average payoff of the 45 minutes session was about 9 Euros.

3. Results

This section is organized as follows. First, we briefly report on some aggregate data regarding costless messages (all messages, independently whether the fee was paid or not) and the dictators’ behavior. Then, we focus on our main objective, which is the elicitation and classification of the disapproval patterns of the recipients using again costless messages. Finally, we comment on the costly messages (the subset of messages the recipients decided to pay for).

3.1. Aggregate Statistics

Table 1. Approval [Disapproval] rates of the recipients in the five dictator games using costless messages.

Game					Costless Messages
	Left–Hand		Right–Hand		Left–Hand Allocation		Right–Hand Allocation
1	(180,140)	vs.	(100,220)		0.2778   [0.4000]		0.7333   [0.1667]
2	(180,120)	vs.	(100,100)		0.6889   [0.1333]		0.1556   [0.6444]
3	(240,120)	vs.	(100,120)		0.6222   [0.1667]		0.2000   [0.5444]
4	(180,120)	vs.	(120,180)		0.1778   [0.4556]		0.6222   [0.1667]
5	(200,120)	vs.	(160,160)		0.2000   [0.5444]		0.7333   [0.1222]

presents the costless messages of the recipients.3 For the left–hand allocation of game 1 (180/140 vs. 100/220), for instance, we see that 27.78 % of the recipients approve that choice, whereas 40 % disapprove it (the remaining recipients send a neutral message). We observe that in each game, one allocation receives mostly disapprovals and the alternative allocation mostly approvals. This suggests that recipients tend to follow some common rule.

Although our goal in this paper is to explore the recipients’ behavior, we can also make a brief reference to the actions of the dictators. In most games, they choose the allocation that maximizes their own payoff. More precisely, the frequency of that choice is 98.88 % in the first, 94.44 % in the second, 100 % in the third, 98.88 % in the fourth, and 81.11 % in the fifth game. One might be tempted to infer from this that dictators are selfish, but we note that in most of our games, many theories of other–regarding preferences make the same predictions as the standard model of selfish players.4 In any case, our games were not chosen to study the dictators’ behavior, but that of the recipients’. We do this in detail in what follows.

3.2. Classification Analysis

We apply the classification procedure from [14] to analyze the motives behind the recipients’ disapproval. More precisely, we posit that the recipients follow deterministic decision rules when disapproving (the rules may differ from subject to subject), but also that they deviate with probability $\epsilon >0$. This classification procedure has several favorable attributes. By selecting the decision rule that best fits each subject’s behavior, we can classify subjects by types. It also helps us finding the best single decision rule or the combination of two, three, etc. decision rules that best account for the behavior in our dictator games. Given this, we can then apply the Akaike Information Criterion (AIC) to infer the number of decision rules necessary to provide an accurate but parsimonious explanation of the observed disapproval pattern. Importantly, this procedure circumvents the multicollinearity problems that would appear in a classical regression analysis if the decision rules were entered as independent variables and allows for appropriate inferences even when testing all possible decision rules —no matter how similar their predictions are—at the same time.

3.2.1. Individual Decision Rules

For simplicity, we restrict our analysis to binary decision rules; that is, rules that indicate for each allocation whether or not the subject expresses disapproval (without specifying the message sent if the subject does not disapprove). More precisely, any rule can be formally described as a vector of ten numbers, one for each allocation of each game. The rule takes value one if it predicts disapproval at the corresponding allocation, and value zero otherwise. We consider seven rules that seem especially appealing. Letting $(x_1^l,x_2^l)$ refer to the left–hand and $(x_1^r,x_2^r)$ to the right–hand allocation at any of our dictator games (1 denotes the dictator and 2 the recipient), the rules are defined as follows:

• The No–Disapproval rule never predicts disapproval;
• the Envy rule predicts disapproval at allocation $a\in \{l,r\}$ if, and only if, $x_1^a>x_2^a$ –that is, if the dictator gets a larger payoff than the recipient;
• the Harm rule predicts disapproval at allocation $a\in \{l,r\}$ if, and only if, the alternative allocation $al\ne a$ is such that $x_2^a<x_2^{al}$ –that is, if the recipient was harmed by the choice of the dictator;
• the Spite rule predicts disapproval at all allocations;
• the Efficiency rule predicts disapproval at allocation $a\in \{l,r\}$ if, and only if, $x_1^a+x_2^a<x_1^{al}+x_2^{al}$, where $al\ne a$ –that is, subjects disapprove socially inefficient choices;
• the Equity rule predicts disapproval at allocation $a\in \{l,r\}$ if, and only if, $|x_1^a-x_2^a|>|x_1^{al}-x_2^{al}|$, where $al\ne a$ –that is, subjects disapprove inequitable choices; and,
• the Maximin rule predicts disapproval at $a\in \{l,r\}$ if, and only if, $min\{x_1^a,x_2^a\}<min\{x_1^{al},x_2^{al}\}$, where $al\ne a$ –that is, subjects disapprove deviations from a maximin or Rawlsian norm.

Table 2. Predictions of disapproval in the five dictator games for the Non–Trivial Decision rules. The following notation is used: EN=Envy, HA=Harm, EQ=Equity, EF=Efficiency, MA=Maximin.

Game					Predictions of Disapproval
	Left–Hand		Right–Hand		Left–Hand Allocation		Right–Hand Allocation
1	(180,140)	vs.	(100,220)		EN, HA		EQ, MA
2	(180,120)	vs.	(100,100)		EN, EQ		HA, EF, MA
3	(240,120)	vs.	(100,120)		EN, EQ		EF, MA
4	(180,120)	vs.	(120,180)		EN, HA
5	(200,120)	vs.	(160,160)		EN, HA, EQ, MA

presents the predictions of the Non–Trivial rules in the five dictator games.

For instance, the Efficiency rule only predicts disapproval at the right–hand allocation of games 2 ($180/120\mathit{vs}.100/100$) and 3 ($240/120\mathit{vs}.100/120$). In these allocations, the social surplus is lower than in the corresponding alternative allocation (in contrast, games 1, 4, and 5 are surplus–preserving). As another example, the Maximin rule does not predict any disapproval in game 4 ($180/120\mathit{vs}.120/180$) because the minimum payoff is the same at both allocations. Note also that Table 2 motivates our selection of games: As the reader can confirm, the games allow us to discriminate between the above mentioned rules; that is, no two decision rules make the same predictions for all five games. For instance, game 5 ($200/120\mathit{vs}.160/160$) is useful to distinguish between Harm and Efficiency; the first rule predicts disapproval at the left–hand allocation, whereas Efficiency predicts no disapproval. As another example, game 2 ($180/120\mathit{vs}.100/100$) can be used to discriminate between rules like Efficiency and No-Disapproval.

To finish, we note that our games do not only permit disapproval behavior to be studied and whether subjects are heterogeneous in this respect. In effect, we can also infer something about social norms if one assumes that deviations from social norms are disapproved. More precisely, we are thinking here about distributive norms; that is, norms regulating how to distribute income or other goods. We can think of many examples: (i) Efficiency, or “choose the allocation that maximizes the social surplus", (ii) Equity, or “choose the most equitable allocation", (iii) Maximin, or “choose the allocation that maximizes the minimum payoff”, (iv) No–Harm, or “choose an allocation so that the co–player’s payoff is not reduced below a reference point" (this is inspired by [23] and the literature on reciprocity), and (v) No–Envy, or “choose any allocation that does not leave your co–player disadvantaged.”5 Most of our rules described before have a clear relation with one of these norms. For instance, a person who disapproves according to the Envy rule might do so because she finds the No–Envy norm binding, while following the Harm rule might be a signal that the corresponding player places importance on the No–Harm norm. In turn, a person who does not care about norms should act according to the No–Disapproval rule.6

3.2.2. Methodology

Let D be the set of all recipients (of cardinality d) and let $r_k$ be any set of k rules ($k=1,\dots ,7$) from the seven decision rules introduced above. Recipients are indexed by i, decision rules by j. Further, $x_i\left(j\right)$ denotes the number of allocations in the five dictator games where the rule j correctly predicts the choice (disapproving/not disapproving) of recipient i. Obviously, $x_i\left(j\right)$ cannot be larger than 10. Similarly, $x_i\left(r_k\right)$ is the number of allocations for which recipient i follows the rule(s) in $r_k$ that best replicates her behavior; that is, $x_i\left(r_k\right)={max}_{j\in r_k}{x}_i\left(j\right)$. For example, if $r_2=\{Harm,Envy\}$ and recipient i only disapproves the dictator at the left–hand allocation of game 1, then it follows from Table 2 that $x_i\left(\{Harm,Envy\}\right)=7$ because the Harm rule correctly predicts her behavior for 7 allocations whereas the Envy rule predicts only for 6 allocations. Given any set $r_k$, let $f_{r_k}:D\to r_k$ be a mapping that assigns each recipient i to the decision rule j in $r_k$ that replicates her/his behavior best—i.e., $x_i\left(j\right)=x_i\left(r_k\right)$.7 A model$(r_k,f_{r_k})$ consists of a set $r_k$ and a mapping $f_{r_k}$.

No model $(r_k,f_{r_k})$ is likely to perfectly replicate the actual disapproval pattern of all recipients, so we must allow for some error. More precisely, we posit that every recipient deviates from her corresponding rule with probability $\epsilon >0$ at each allocation. Therefore, the probability that recipient i follows her assigned rule exactly $x_i\left(r_k\right)$ times out of the 10 choices is ${(1-\epsilon )}^{x_i\left(r_k\right)}·{\epsilon }^{10-x_i\left(r_k\right)}$.8 Since the decision of the recipients are independent, the joint probability that the data is generated by the model $(r_k,f_{r_k})$ is then equal to

$$\prod _{i\in D}{(1-\epsilon )}^{x_i\left(r_k\right)}·{\epsilon }^{10-x_i\left(r_k\right)}$$

Given the joint probability distribution, one can show by standard optimization techniques that the maximum likelihood estimator of the error rate for the model $(r_k,f_{r_k})$ coincides with the proportion of overall deviations:

$$\widehat{\epsilon }(r_k,f_{r_k})=\frac{10·d-{\sum }_{i=1}^d{x}_i\left(r_k\right)}{10·d}$$

To finish, observe that for any two models $(r_k,f_{r_k})$ and $(r_s,f_{r_s})$ such that $s<k$, it follows $\widehat{\epsilon }(r_s,f_{r_s})\le \widehat{\epsilon }(r_k,f_{r_k})$, as $x_i\left(r_s\right)$ can never be larger than $x_i\left(r_k\right)$ for any i. However, the model $(r_k,f_{r_k})$ contains more rules and it is hence more complex than $(r_s,f_{r_s})$. For this reason, it makes sense to introduce a penalty that depends on the number of decisions rules in a model. For this, we use the Akaike Information Criterion (AIC), according to which the best model ($r_k,f_{r_k})$ has to maximize the value of the log–likelihood function minus the number of parameters $(10+d)·k$.

3.2.3. Results

Table 3. Error rates, AIC, and the optimal assignment of subjects in the best models of 1, 2, and 3 decision rules (models with more decision rules have a very negative AIC). The following abbreviations are used: ND = No–Disapproval, EN = Envy, HA = Harm, EF = Efficiency, EQ = Equity, and MA = Maximin.

Model			Error Rate	AIC	Assignment of Subjects
Efficiency			0.2967	–647
Harm			0.3256	–668
No–Disapproval			0.3344	–674
Maximin			0.3544	–685
Envy			0.4944	–724
Equity			0.5322	–722
Spite			0.6656	-764
Efficiency	Harm		0.1900	–638	49 EF	41 HA
No–Disapproval	Harm		0.2122	–665	46 ND	44 HA
Efficiency	Envy		0.2177	–672	63 EF	27 EN
Harm	Maximin		0.2388	–695	49 HA	41 MA
Efficiency	No–Disapproval		0.2455	–702	53 EF	37 ND
Harm	Efficiency	No–Disapproval	0.1566	–690	35 HA	33 EF	22 ND
Efficiency	Harm	Envy	0.1566	–690	46 EF	30 HA	14 EN
Efficiency	Harm	Equity	0.1655	–703	46 EF	34 HA	10 EQ
No–Disapproval	Harm	Maximin	0.1744	–717	40 ND	40 HA	10 MA
Harm	Maximin	Equity	0.1811	–725	41 HA	36 MA	13 EQ

presents the results of the classification analysis. We first consider all models where $r_k$ consists of just one rule—i.e., all models assuming homogeneous players—and report for each one the corresponding error rate and the value of the AIC. Note that the models are ordered according to the value of the AIC. We also report the error rate and the AIC for the five best models with two and three rules, together with the assignment of the 90 recipients to their respective rules.9 For instance, the model with three rules and the maximum AIC includes the Efficiency, the Harm, and the No–Disapproval rule: 33 subjects predominantly followed the first rule, 35 the second, and 22 the third rule. Note also that this model has an error rate of 0.1566, which means that the model can explain almost 85 % of the choices of the average recipient.

If we focus on models with just one rule, the reader can observe that the Efficiency rule describes behavior best, followed by the Harm and the No–Disapproval rule. If we consider models with two rules, the best model has an estimated error rate of 0.1900 and includes the Efficiency and the Harm rule (49 subjects are assigned to the Efficiency and 41 subjects to the Harm rule). Overall, this is the best model in terms of parsimony and accuracy, as it maximizes the Akaike information criterion (–637.60). Models with three rules, in contrast, have a lower AIC. Taking into account both empirical relevance and parsimony, therefore, our analysis suggests that there are two main types of recipients in our dictator games: those who disapprove of inefficient choices and those who disapprove harming ones.

3.3. Costly Messages

The previous analysis has focused on all messages, irrespective of whether the message fee was effectively paid—i.e., we studied what would happen in case messages were costless. Now, we concentrate on the costly messages. There were two main reasons to include them into the design. First, some people might be more likely to pay the fee than others, and this can be studied with our design. Second, an exclusive focus on costless messages as it is usual in the existing literature is potentially problematic because selfish types might choose messages randomly, which could contaminate our analysis of which behaviors are disapproved. Our elicitation of costly messages can clarify this issue. In effect, extremely different patterns for the costless and the costly messages could be interpreted as evidence that a lot of noise has been introduced by the selfish types choosing randomly.10

For several reasons, we highly doubt that (many) recipients chose their hypothetical, costless messages in a random or careless manner. Note first that a model in which all recipients choose randomly in an uniform manner should have an error rate of 0.5000. However, even models with only one rule have much lower error rates. For example, the Efficiency rule alone can explain more than 70 % of the data. This would not be the case if many players acted randomly. To provide another example: the best model in terms of accuracy and parsimony, the one consisting of the Efficiency and the Harm rules, can explain more than 80 % of the recipients’ choices. Again, this appears to be totally at odds with the idea that many recipients chose carelessly. To make a final point along those lines, note that the Spite rule has the least predictive power among all of our rules (in fact, it performs even worse than a totally random rule). Since this rule predicts an arguably incoherent behavior (disapproval always), its failure is another apparent signal that people took some care when making their choices.

For further evidence, consider now costly messages.

Table 4. Approval [Disapproval] rates of the recipient in the five dictator games using costly messages. The total number of messages being paid for is presented to the right of the disapproval rates.

Game					Costly Messages
	Left–Hand		Right–Hand		Left–Hand Allocation		Right–Hand Allocation
1	(180,140)	vs.	(100,220)		0.3636   [0.5455]   11		1.0000   [0.0000]   13
2	(180,120)	vs.	(100,100)		0.4444   [0.5556]      9		0.2000   [0.7000]   10
3	(240,120)	vs.	(100,120)		0.3333   [0.6666]      6		0.5000   [0.5000]      4
4	(180,120)	vs.	(120,180)		0.2857   [0.2857]      7		1.0000   [0.0000]      9
5	(200,120)	vs.	(160,160)		0.2500   [0.7500]      4		1.0000   [0.0000]      5

shows the approval and disapproval rates for costly messages, together with the total number of costly messages sent at each allocation. We observe that the fee was paid in about 8.67 percent of all possible cases, and that 44 (29) messages approved (disapproved) the corresponding dictator’s choice. Although the relative scarcity of observations warrants some caution in analyzing these results, some preliminary evidence suggests that motives of efficiency and harm remain important factors. For instance, costly disapprovals are totally absent for some allocations in which there is no loss of social efficiency and/or the recipient is not harmed, as at the right–hand allocation of the first, fourth, and fifth games. In contrast, costly disapprovals are relatively frequent for some allocations in which the recipient is harmed, like the right–hand allocation of game 2 and the left–hand allocation of game 5. Yet, there is some evidence that other factors also play a role. For example, the choice of the socially efficient (even Pareto efficient) allocation 240/120 is disapproved by 66 % of the costly messages, and a similar situation is observed in game 2. This suggests that payoff comparisons play a non–negligible role for the subjects who are most willing to pay in order to express disapproval (note however that the strictly egalitarian allocation 100/100 in game 2 is also disapproved by 70 % of the choices), a point that future research should address.

In order to gain more confidence that the costless messages were not chosen randomly, we have also performed the classification analysis (using again costless messages) on (i) the 65 recipients who never used costly messages and (ii) the remaining recipients who used messages at least once (20 out of these 25 subjects used a costly message more than once).

Table 5. Error rates, AIC, and the optimal assignment of subjects in the best models of 1 and 2 decision rules (models with more decision rules have a very negative AIC) for the two subgroups using hypothetical messages. The following abbreviations are used: ND = No–Disapproval, EN = Envy, HA = Harm, EF = Efficiency, EQ = Equity, and MA = Maximin.

Model		Error Rate	AIC	Assignment of Subjects
Subjects who sent at least one costly message
Efficiency		0.2920	–186
Harm		0.3000	–188
No–Disapproval		0.3000	–188
Maximin		0.3640	–198
Envy		0.4480	–207
Equity		0.5000	–208
Spite		0.7000	–188
Efficiency	Harm	0.2040	–196		14 EF	11 HA
Efficiency	Envy	0.2040	–196		17 EF	8 EN
No–Disapproval	Harm	0.2120	–199		13.5 ND	11.5 HA
Efficiency	No–Disapproval	0.2280	–204		13 EF	12 ND
No–Disapproval	Envy	0.2440	–209		17 ND	8 EN
Subjects who never sent a costly message
Efficiency		0.3090	–471
Harm		0.3454	–486
No–Disapproval		0.3575	–495
Maximin		0.3606	–496
Envy		0.5196	–525
Equity		0.5515	–527
Spite		0.6575	–494
Efficiency	Harm	0.1969	–461		35 EF	30 HA
No–Disapproval	Harm	0.2242	–486		33 ND	32 HA
Efficiency	Envy	0.2348	–495		46 EF	19 EN
Harm	Maximin	0.2454	–503		34 HA	31 MA
Efficiency	No–Disapproval	0.2636	–517		40.5 EF	24.5 ND

presents the error rates, the AIC, and the optimal assignment of subjects within the two subgroups in question.

Most importantly, the results of the subgroup analysis are arguably very similar to the overall findings and across subgroups. In particular, we find that the ordering of the seven decision rules in terms of their explanatory power is the same within both subgroups (Efficiency > Harm > No–Disapproval > Maximin > Envy > Equity > Spite). It also turns out that the corresponding error rates are not significantly different across subgroups at the 5 percent level: the two–sided p–values of the Mann Whitney U tests are equal to 0.9160 for the Efficiency rule, 0.4432 for the Harm rule, 0.0512 for the No–Disapproval rule, 0.4069 for the Maximin rule, 0.3237 for the Envy rule, 0.3338 for the Equity rule, and 0.0512 for the Spite rule. Overall, the evidence seems to be at odds with the idea that the costless messages were chosen in a careless manner, so that we conclude that all our prior analysis based on this type of messages is informative.

4. Conclusion

Our analysis indicates that subjects are heterogenous in regards to which behavior they disapprove and we identify two main behavioral types in this respect: one disapproves when she has been harmed, while the other disapproves when a socially inefficient outcome has been chosen. Further, we observe that other factors, such as envy and equity only play only minor role in our games (at least when messages are costless). This is somehow surprising. Since abundant evidence suggests that inequity–aversion or some kind of egalitarian motives are important to explain monetary punishment (see [17,18,24,25]), we expected a similar result with non–monetary punishment (i.e., disapproval). Perhaps the causes of monetary and non–monetary punishments are different.11 Another possibility has simply to do with our set of games. In this respect, we hope that our study motivates additional research on disapproval patterns so as to see whether our results generalize to other games or other socioeconomic environments.

While our study attempts to elicit norms by looking at disapproval patterns, related literature has studied other–regarding preferences by looking at how agents allocate payoffs. For instance, [19,27], and [20] report data from several simple games like our dictator games, and classify dictators according to their choices in the games. Since allocation decisions might be driven by fairness norms, the relation to our study is clear. More speculatively, one could guess that allocation decisions are somehow the mirror image of disapproval decisions: a person choosing allocation a as a dictator should disapprove a choice different from a as a recipient. While this idea has obviously some flaws (e.g., self–serving agents are not likely to care about disapproval when being recipients; see also our comment later on self-serving biases), some preliminary evidence seems to be in line with it. For instance, [20] emphasize the importance of social welfare maximization, which seems to coincide with our result that many people disapprove socially inefficient choices.

Even though our focus was not on the effect of the disapproval on the dictators’ behavior, we also note that our findings qualify previous results in this respect in the experimental literature. For instance, [28] reports that about one–third of the dictators are willing to pay one dollar to “exit” the game and hence ensure that the recipient never knows that a dictator game was played. Exiting might reflect a desire to not violate the other’s expectations, or possibly a desire to not feel disapproved of. In [6,7] ex post messages significantly foster generous behavior in the dictator game. Also, [29] observe that cursing a cheater is a widely–used and effective method to deter cheating in collusive agreements. In contrast, generosity in our game 5 remains at an arguably low level even when approval/disapproval is available. One possible reason for these contrasting results is the particular kind of messages that were available, as the latter three cited studies used open–form messages. Maybe it is not only disapproval per se that people care about, but more particularly the way in which it is framed.

There are other issues that merit future attention. A particularly appealing one concerns self–serving biases [30]: do people have a biased view of norms (maybe inadvertently) depending on their position in the game? In our experiment, for instance, we conjecture that dictators would probably exhibit rather different disapproval patterns than the recipients if they were asked what choices merit approval or disapproval (we doubt, for instance, that harm would play such a significant role). However, since dictators might have an interest to manipulate the answers to such questions in order to provide a justification of their choices, the challenge of future research is to obtain sincere answers in this respect.