Moral Perceptions of Advised Actions∗ Lucas Coffman† Harvard University

Alexander Gotthard Real‡ Pontificia U Javeriana

March 28, 2017

Abstract Can an organization avoid some blame for an unpopular action by hiring an adviser to advise them to do it? We present experimental evidence suggesting this is the case advice to be selfish substantially decreases punishment of being selfish. Further, this result is true despite advisers conflicts of interest, known to all: Through a relational contract, advisers are motivated to tell the principals what they want to hear. In follow-up treatments, we show advice does not decrease punishment solely by affecting beliefs of how necessary the selfish action was. Finally, when advisers are available, selfish principals act more selfishly.

∗ We thank Katherine Coffman, Tobias Gesche, Paul Healy, John Kagel, Ian Krajbich, Muriel Niederle, Siqi Pan, and seminar participants at BABEEW, the UCSD Rady Morality conference, ESA Tucson 2016, and University of St. Gallen for helpful comments. This paper benefited greatly from the research assistance of Luke Fesko. † Department of Economics, Harvard University, 1805 Cambridge Street, Cambridge, MA 02138. [email protected]. ‡ Facultad de Ciencias Econ´ omicas y Administrativas, Pontificia U Javeriana, Carrera 7 #40-62, Bogot´ a, Colombia. [email protected].

1

1

Introduction

Suppose a board of directors proposes an exorbitant pay package for their CEO, a company announces extensive layoffs, or a government implements tightened fiscal policy measures. Such actions are often met with scorn, loss of goodwill, or protest. But what if such actions, however, were taken only after receiving advice – compensation consultants, managerial consultants, the International Monetary Fund, and so on? How does the presence of advisers change how we perceive the fairness of the subsequent actions? If the adviser is an independent third party offering unbiased advice to take an action, this could very reasonably change the moral calculus. However, advisers are not always unbiased. Many advisers have a meaningful conflict of interest (COI): They want more work with the same firm in many other capacities, so they are biased to provide advice that the firm wants to hear.1 For example, a compensation consultant with an eye towards a multiplicity of services with a firm might strongly advise that the CEO needs to be paid more than her peers to be competitive (e.g. Bebchuk and Fried (2003)). Could advice known to be slanted nevertheless alter how we perceive the morality of a firm’s actions? Recent work on moral judgments suggests the answer could be “yes”. Moral perceptions can be context-dependent, and one important contextual parameter is the number of actors involved. For example, the use of an agent or intermediary robustly reduces perceptions of responsibility. Unfair offers are more likely to be accepted when made by an agent (rather than oneself), even one contracted to make an unfair offer (Fershtman and Gneezy, 2001). Similarly, selfish actions are more likely to be taken (Hamman et al., 2010), and deemed less immoral when done through an intermediary (Paharia et al., 2009; Bartling and Fischbacher, 2012; Coffman, 2011; Oexl and Grossman, 2013), even one with zero agency. Further, partnerships also change our moral judgments. A pair of principles working together have been shown to behave substantially more selfishly than either one working on their own (Dana et al., 2007).2 Hence, thus far in the literature, when multiple actors are involved, seemingly regardless of their arrangement, responsibility can shift, total responsibility sometimes shrinks, and selfishness may increase. However, there has been no work studying moral perceptions of advised actions.3,4 It 1

This is similar to the COI discussion in Demski (2003) who note Arthur Andersen billed Enron $1 million per week, with less than half coming from auditing services, consistent with the idea they may have been lenient on auditing to maintain their other work with Enron. 2 Kagel and McGee (2016) show groups are also more uncooperative in a one-shot environment but look like individuals in repeated games. 3 There is a growing related literature on advice. In financial advising, Mullainathan et al. (2012) and Malmendier and Shanthikumar (2014) show financial advice is slanted by the advisers’ incentives rather than (purely) acting in their clients’ best interests. Foerster et al. (Forthcoming) show that advisers influence investment portfolios dramatically: portfolios are better predicted by advisers’ identities than fundamentals like stage of life cycle of the investor. Chen and Gesche (2016) show advisers who are incentivized to give bad advice do so, and continue to do so even after the incentive is removed. Charness and Schram (2015) demonstrate that unincentivized advisers encourage pro-social behavior and principles feel impelled to take this advice. In the Psychology literature, Kray and Gonzalez (1999) demonstrate that people feel less responsible for a bad outcome when they advised the action compared to when they took it themselves, and Harvey and Fischer (1997) interpret their results to suggest that an adviser and decision-maker feel responsibility is shared for an outcome. For a survey on the psychology literature on advice see Bonaccio and Dalal (2006). 4 There is a wider experimental economics literature on advice that ask questions unrelated to our pursuit.

2

could be that many of the mechanisms responsible for the changes in moral perceptions in the literature described above are absent in an adviser-advisee context. For example, it may be harder for a principal to distance herself from an outcome, or “keep her hands clean” by hiring an adviser (channels suggested by Paharia et al. (2009) and Coffman (2011)); she has to take the action herself. She cannot diffuse responsibility with another agent taking an action as in Dana et al. (2007). Relatedly, the theoretical model of responsibility in Bartling and Fischbacher (2012) would assign 100% of the responsibility to the principal since advice does not alter her choice set. Hence, existing results, while perhaps suggestive, do not necessarily predict an effect of advice on moral perceptions. To fill this gap in the literature, we turn to the laboratory. Though the laboratory is bereft of potentially important contextual parameters, it allows us to vary exogenously the existence of advisers, while cleanly controlling information, and measuring important ancillary data. This makes a laboratory investigation the right tool for a proof-of-concept. Follow-up work will be key in determining the robustness of our results, and how important they are in richer contexts, net of many other effects. The experimental design has five key features. First, a decision-maker has a material temptation to act selfishly at the expense of another. Second, we vary whether the decisionmaker has an adviser. In adviser treatments, a consultant, with no informational advantage, advises the decision-maker whether she should act selfishly or not. Third, to bias the advice towards what the decision-maker wants to hear, the adviser has a relational contract incentive. The decision-maker, after hearing the advice and after making her decision, can decide whether to retain the adviser’s services for future play (for which the adviser gets paid) or to fire him. Fourth, to measure perceived fairness, the player who is on the receiving end of the selfish decision can decide how much to punish the decision-maker. Fifth, we install some opaqueness as to how the decision-maker is treating the punisher. With a small probability, it is in everyone’s interests for the decision-maker to take the otherwise selfish action; however, the punisher does not know when this is true; only the consultant and decision-maker know the state of nature. In later treatments, we turn off this opaqueness by providing the punisher full information. This allows us to test whether advice reduces punishment simply by altering the beliefs of the punisher of how unfairly they were treated. In this simple set-up, we can begin to ask fundamental questions about how advised actions are perceived and what may be the root cause of these perceptions. The data from this paradigm support the main hypothesis: Selfish advice decreases punishment of selfish actions. As found in previous research, we find selfish actions are punished quite severely. However, average punishment of selfish actions decreases by 18% when the selfish action is advised (rank-sum p < 0.01). Despite the punishers knowing the advice is coming from a source with a substantial COI, it skews their perceptions of fairness in a meaningful way. Profit-maximizing decision-makers in our set-up have an optimal strategy of finding and keeping an adviser to tell them to act selfishly, and then acting selfishly; however, to maximize profits, the principal cannot hire any random adviser. She needs to hire an adviser who There is work on intergenerational advice and learning in coordination games (See e.g. Schotter (2003) and C ¸ elen et al. (2010)), on how advice is ignored when the adviser has misaligned incentives Kuang et al. (2007), and on the demand for advice in investment games Gehrig et al. (2010).

3

suggests she take the selfish action. Punishment of a selfish action is 16% higher than when the principal was advised to do the fair thing compared to no advice (rank-sum p = 0.01). Though the advice is coming from a slanted source, the punishers do not seem to discount it fully. Recall that in the game, when the decision-maker takes a selfish action, there is a small probability it is in everyone’s interests for her to do so. When punishers see the decision-maker was advised to act selfishly, they believe the probability of this state of nature quite high, much higher than its true likelihood. Hence, one potential mechanism for the main result is that punishers do not discount the information content of the advice as heavily as they should given the misaligned incentives of its source. We test this mechanism with a follow-up treatment. In this experimental arm, the punisher knows with certainty whether the decision-maker had to be selfish, or if she chose to be selfish. If advice affects punishment only through biasing beliefs regarding the state of nature, we should not see any effect of advice in this treatment as beliefs should not be manipulable. Nonetheless, advice continues to decrease punishment, by a similar magnitude as the previous treatment. The upshot of these results is that the presence of advisers allows selfish principals to act more selfishly, even in the presence of punishment. Among decision-makers who, absent punishment and no feedback to others, prefer the selfish outcome, the incidence rate of selfishness increases by 27% (17pp) when advisers are available, unconditional on the advice received (rank-sum p = 0.09). Despite the effectiveness of punishment in many contexts in encouraging pro-social behavior, we find its effectiveness is eroded by the availability of advisers. Previous work on advice considers the information and expertise provided (e.g. Green and Stokey (2007)’s setup). Our results provide a possible additional explanation of the value and ubiquity of advisers: They decrease the backlash faced by those they advise. If an organization wants to take an unpopular action, they may hire a consulting firm to advise that action, take that action, and face less loss of goodwill as a result. Absent an informational advantage, advisers may still provide value. The results also speak to the compensation consulting literature. Executive pay in the United States rose 937% in real dollars between 1978 and 2013.5 One often-cited culprit for this dramatic rise is compensation consultants. Though the empirical evidence is mixed (Bebchuk and Fried (2003), Murphy and Sandino (2010), Chu et al. (2015)), the theoretical argument is straightforward. Even when selected by the board of directors, compensation consultants have an agency problem as they want future work with the firm in other capacities. As a result, the consultant may be more inclined to tell the CEO what she wants to hear, that her pay should be quite high. The CEO and board present the compensation proposal to shareholders, who see the incentives were determined by a (biased) third-party, and they accept the proposal. Though compelling, one piece of the argument is missing: Why would the shareholders vote for the proposal, if there is a known agency problem? Does the simple fact that the decision of the CEO and board was advised change the moral perceptions? Our experimental evidence fills in this gap, suggesting the answer is yes. 5 Skapinker, Michael. 2015. Executive pay: The battle to align risks and rewards. Financial Times. April 30, 2015, accessed September 23, 2016, www.ft.com/content/9265406a-eaaf-11e4-96ec-00144feab7de.

4

2

Experimental Design

In the game, a principal will have a decision to be fair or selfish, and the affected player will have a chance to punish her. Additional treatments will vary the availability of an adviser and how much the punisher knows relative to the principal and adviser. This will leave us with four treatments: Control or Advice crossed with Uncertainty or Certainty. To aid in understanding, the motivating example for the game is the following: A company is going through hard times, and to save the company an executive must cut costs. She can cut employee pay a lot, executive pay a lot, or both a little bit. Most of the time any one of those cost-cutting measures will save the company; however, in some cases only certain cost-cutting measures will save the company. E.g. sometimes the only action that will save the company is cutting employee pay. A consultant is brought in to advise the executive on how to cut costs. In the Uncertain treatment in the next subsection, only the adviser and executive know which cost cuts can save the company. In the subsequent Certain treatment, everyone will know, including the employee.

2.1

No Advisors: The Uncertain-Control treatment

The baseline game involves two players: a principal and a punisher.6 Each player begins the game with an account endowed with three tokens, where each token is worth $6 at the end of the experiment. Both accounts will be completely wiped out unless they are “saved”. The accounts can be saved by spending two tokens. The principal has to decide which two tokens to spend. Some tokens can reduce the probability of both accounts being wiped out by 50 percentage points, which we will call “effective”; others by 0pp, “ineffective” tokens. If two effective tokens are spent, the accounts are saved for sure. If one effective token is spent, there is a 50% chance both accounts will be lost (the fates of the accounts are perfectly tied). Which tokens are effective varies. The computer picks one of four states of nature for each play of the game. With a 70% chance, both players’ tokens are effective. The other three states each have a 10% chance of being chosen: only the principal’s tokens are effective, only the punisher’s tokens are effective, and no tokens are effective. All four possibilities, and their likelihoods, are known by everyone regardless of treatment. In the control, only the principal will be told which tokens are actually effective, but the punisher will not. The different states of nature serve to provide a veil of the fairness of the principal’s actions. If the principal spends two tokens from the punisher’s account, it may be because she had to, even if that is unlikely. After the state of nature is revealed to the principal, she chooses which tokens to spend. If she chooses two (zero) effective tokens, the accounts are (not) saved. If she spends one effective token, the computer then determines if the accounts were saved. Subsequently, the punisher is told what tokens were spent and whether the accounts were saved, but not which tokens were effective. If the accounts are saved, the punisher then determines by how much he would like to reduce the principal’s payout. He can choose any whole cent amount between 6

Neutral language was used throughout, e.g. “Player A” and “Player B”. See Appendix F for an example of the actual experimental instructions used. All instructions were read aloud to everyone at once, and subjects had a paper copy in hand.

5

0 and 6.7 To understand how participants learn, participants played the game for 15 rounds. It could be that if a punisher sees both of his tokens were used in round one, they had to be, but if he sees this for eight straight periods, he might learn about the nature of the advice. They remained in the same role throughout each session, but were randomly rematched after each round. Subjects were paid for (the same) one randomly selected round, chosen after all rounds have been played (See Azrieli et al. (n.d.) for theoretical argument).

2.2

The effect of Advisers: The Uncertain-Advice treatment

To test the main research question, we insert an adviser into the above game. There is an equal number of advisers and principals in each session. In the first round, one adviser is randomly matched to each principal. Advisors and principals then observe which tokens are effective in that round.8 The adviser then submits his advice to the principal about what combination of tokens to use.9 After seeing the state of nature and the advice, the principal decides which combination of tokens to use. The punisher is told the principal’s choice, the advice she received, and whether the accounts survived (again, not the state of nature). To induce a COI, the advisers have a relational contract with their principal. Each period, the principal decides whether she wants to retain her adviser from the previous period or switch to a different, randomly assigned adviser. Advisors’ earnings are increasing in the number of principals that hire them. In particular, advisers start the game with a balance of $12. Each period, $1 is subtracted from their earnings but $1 is added for each principal to which they are assigned. Thus, advisers have an incentive to be retained, and thus to tell principals what they think principals want to hear. Their incentives are common knowledge to all players. In a sense, this creates a stress test of the hypothesis: If advised selfishness is punished less even when the advice is not an independent voice but rather slanted, it may also likely be the case in situations with unbiased advisers. In periods 1, 6 and 11, principals do not select their adviser. In those periods one adviser is randomly assigned to one principal. This exogenous assignment of advisers to principals in these periods helps to isolate the difference in punishment when principals have had the opportunity to select their adviser.

2.3 Removing Uncertainty for the Punisher: The Certain Treatments The effect of advice in the game above could be working through two channels. First, it may change a punisher’s belief of the realized state of nature. Observing that the principal 7

Punishment was costless to maximize nonzero punishment data. This punishment technology has been used before in the literature Coffman (2011), and as can be seen in the data, this design decision does not induce thoughtless decision-making. 8 We do not include a treatment where the adviser is informed of the state of nature and the principal is not. That the principal would be punished less when she is uninformed of the state of nature and is perhaps quite obvious. 9 The adviser simply types a number 0, 1, or 2 to indicate how many of the principal’s tokens should be used. Nothing else is transmitted.

6

was advised to use the punisher’s tokens may be a signal that only the punisher’s tokens are effective. Second, advice to be selfish may change the moral perception of the selfish act all else equal. In this treatment, we turn off the first channel, leaving only the second to drive any observed treatment effect. To do so, we repeat the control and treatment described above, but remove uncertainty for the punishers. All three players, including the punishers know which tokens are effective. The game now has zero information asymmetries. To identify which principals preferred the selfish outcome or the fair outcome absent all other concerns, in all sessions of the certain treatments, principals played a round “0”. In round 0, both tokens were effective, there was no punishment, and no feedback was given. The principal could dictate the outcome she preferred without any consequence.

2.4

End-of-Study Questionnaire

For all treatments, we elicited subjects’ beliefs after all participants had finished playing the game. They were asked their beliefs about the (i) likelihood advisers were retained based on their advice, (ii) average punishment for selfish behavior based on the advice received, and (iii) what was the state of nature based on the advice and use of tokens.10 To induce careful and truthful reporting, participants were rewarded for answers that were close to the empirical truth. Summary statistics for these beliefs are in Tables A-3 and A-3 . In addition, we collected (only) the following information at the end of each session: gender, nationality, participation in economics experiments, participation in psychology experiments, number of economics courses taken, SAT score, and perceived difficulty of the instructions (on a scale from 1 to 7). Table A-1 in the appendix shows the average of the exogenous variables for each treatment. Table A-2 shows these averages as well, but only considering punishers. 11, 12 Only one of the twenty four balance tests results in a statistical difference (proportion of punishers who were native speakers across Certain treatments); however, all F-tests for overall differences between treatments yield p > 0.6.

3

Results

A total of 716 students participated in our study, with roughly 100 punishers per treatment.13 The uncertain-control and uncertain-advice treatments were run between May 2014 and December 2014. The certain-control and certain-advice treatments were run between 10

Full text of elicitation and incentives can be found in online appendix. Some of our sessions lasted much longer than expected, and we had to either skip the elicitation procedure, the final questionnaire, or both. For the uncertain-advice treatment, session 7 had elicitation of beliefs but no questionnaire and session 11 had the questionnaire but no belief elicitation. For the certain-advice treatment, session 2 had no elicitation of beliefs and session 9 had no questionnaire and no elicitation of beliefs. 12 For SAT scores, unfortunately, our questionnaire did not specify whether the participant was supposed to report the reading and math score (up tp 1600 points) or the composite score including the writing section of the test (up to 2400 points). In addition, perhaps due to the combination of being sensitive and unincentivized many answers are meaningless: 25% are less than 600, the lowest possible composite score. Regressions including SAT as a control are included in the appendix, but not in the main text. 13 There were 102 punishers in the uncertain-advice treatment, 108 in the uncertain-control treatment, 100 in the certain-advice treatment, and 100 in the certain-control treatment. 11

7

6 5.5 5 4.5 4 3.5

3.5

4

4.5

5

5.5

6

Figure 1: Average Punishment of Selfish Behavior, by Round: Uncertain Treatments

0

5

Round

Selfish Advice

10

15

0

No Advice

5

Round

Fair Advice

(a) Selfish Advice vs. No Advice

10

15

No Advice

(b) Fair Advice vs. No Advice

Lines depict mean punishment, by round, following a principal using both of the punishers’ tokens.

February 2015 and May 2015. The software was developed with z-Tree (Fischbacher (2007)), and students were recruited using ORSEE (Greiner (2004)).

3.1

Uncertain Treatment Results

We will restrict the analysis to punishment decisions that occurred when principals were seemingly selfish – they used the punishers’ tokens to save the accounts – as well as only when the accounts were saved. Result 1: Selfish advice significantly decreases punishment of selfish actions and (in most models) decreases the likelihood that those actions are punished. Figure 1a displays per-round average punishment in the control (solid line) with punishment of principals that received advice to use the punisher’s tokens (dotted line). Average punishment of principals that received selfish advice is lower generally, and in thirteen of fifteen periods. Table 1 substantiates the visual using OLS regressions. The first two columns in panel A regress punishment on treatment and type of advice received, with and without demographics, clustering at the punisher level. Average punishment significantly decreases from $4.7 in the absence of advisers to $4.1 following selfish advice (p = 0.06), a decrease of 13.5%. This result holds controlling for demographic variables.14 We have many punishment observations per punisher that cannot be considered independent observations. In Panel A, we cluster the standard errors at punisher level. In Panel B we consider each punisher’s average punishment (per advice given) as an observation, also 14

The result holds when interacting advice with the round number. Though directionally positive (meaning the treatment effect is decreasing from round to round), there is no significant time trend of the treatment effect (See Table ?? in the Appendix).

8

clustering at the punisher level.15 In these empirical models, average punishment decreases by 18% (p < 0.01) and by 21% (p < 0.01) with controls. In every model, selfish advice decreases punishment substantially and significantly. Is the decrease in punishment coming from some punishers deciding not to punish altogether? Columns 3 and 4 of Table 1 report results for a linear probability model, regressing a dummy equal to 1 if punishment is greater than zero. Again, Panel A clusters at the punisher level while Panel B considers punisher-level average decisions of whether to punish. In Panel A, selfish advice reduces the likelihood of punishing by 8 pp without controls (p = 0.16) and by 11 pp when controls are included (p = 0.05). Similarly the regressions in Panel B estimate likelihood of punishing decreases by 13 pp (p = 0.01) and 16 pp (p < 0.01). Some of the decrease in punishment seems to be coming through some punishers deciding not to punish when the selfish action was advised. Result 2: Fair advice significantly increases punishment of selfish actions and significantly increases the likelihood that those actions are punished. A related result is also illustrated in Figure 1b: On average, punishment of selfish principals is greater when they are advised to behave fairly than punishment in the control. A look at Table 1 tells us that there is a strong and significant effect of fair advice on punishment and the decision to punish: Average punishment is estimated to increase by between $0.66 and $0.78 across the four regressions in Panels A and B with and without controls when selfish principals contradict the advice they received (p < 0.01 for all four). Further, the likelihood of observing positive punishment increases by between 9 and 12 pp across the specifications (p < 0.01, p = 0.01, p = 0.02, and p = 0.04 for panel A, columns 3 and 4, and panel B, columns 3 and 4 respectively). Result 1 suggests that consultants can be hired to decrease blame: Punishment of selfish principals is lessened by the presence of validating advice even if advisers have a conflict of interest. However, Result 2 shows that it is not an adviser per se, but the advice matters as well. If the adviser suggests the principal act fairly and she does not, backlash actually increases. The moral perception of the act hinges on the content of the advice. How does advice change moral perceptions? One possible channel is that the punisher does not know the state of nature – whether the action taken by the principal is indeed selfish or if it is necessary – and she updates her beliefs based on the advice given. For example, if the adviser suggests the principal use the punisher’s tokens, she may reasonably believe it is more likely that the principal has to use the punisher’s tokens to save the accounts. Hence, believing it less likely the principal was selfish, she may reduce her punishment.

Though not dispositive, the beliefs data are consistent with this hypothesis. Figure 2 shows the likelihood that a principal had to use the punisher’s tokens to save the account, conditional 15

On average, there are 5.6 punishments per punisher in the control treatment, 5.2 in the advice treatment with selfish advice, and 1.6 in the advice treatment with fair advice.

9

Table 1: The Effect of Selfish and Fair Advice on Punishment: Uncertain Treatments (OLS) A - Round-level Punishment (clustered s.e.) Dep. Var.

Selfish Advice Fair Advice Constant

Controls Observations Clusters R-squared

(1) Punishment

(2) Punishment

(3) Any Punishment

(4) Any Punishment

-0.63* (0.34) 0.78*** (0.27) 4.69*** (0.23)

-0.81** (0.36) 0.79*** (0.29) 4.44*** (0.57)

-0.08 (0.05) 0.12*** (0.04) 0.82*** (0.04)

-0.11* (0.06) 0.12** (0.04) 0.80*** (0.08)

NO 1,298 209 0.04

YES 1,210 197 0.07

NO 1,298 209 0.02

YES 1,210 197 0.06

(1) Punishment Avg

(2) Punishment Avg

(3) Any Punishment Avg

(4) Any Punishment Avg

-0.84*** (0.31) 0.74*** (0.26) 4.70*** (0.20)

-1.01*** (0.32) 0.66** (0.27) 4.74*** (0.52)

-0.13** (0.05) 0.10** (0.04) 0.83*** (0.03)

-0.16*** (0.05) 0.09** (0.04) 0.84*** (0.08)

NO 273 209 0.08

YES 257 197 0.11

NO 273 209 0.07

YES 257 197 0.11

B - Punisher-level averages (clustered s.e.) Dep. Var.

Selfish Advice Fair Advice Constant

Controls Observations Clusters R-squared

Note: ∗ ∗ ∗p < 0.01, ∗ ∗ p < 0.05, ∗p < 0.1. Robust standard errors clustered at Punisher level in parentheses. “Punishment” is the amount of money deducted in a given round; “Punishment Avg” is the average level of punishment for for a punisher; “Any Punishment” is a dummy for whether “Punishment” is greater than 0; “Any Punishment Avg” is a punisher’s mean likelihood of punishing. Only scenarios where principal took selfish action considered throughout. Controls include: Gender, participation in economics experiments, participation in psychology experiments, number of economics courses taken, perceived difficulty of the instructions, a native-speaker dummy, and round number.

10

Figure 2: Likelihood Principal Had to Take Selfish Action when She Did, by Content of Advice Actual versus beliefs Actual

.4

*

*** n.s. n.s.

0

.2

Likelihood

.6

Elicited

No Advice Selfish Advice Fair Advice

No Advice Selfish Advice Fair Advice

Bars represent average likelihood that spending the punishers’ tokens was necessary, not selfish. Three bars on left show empirical averages. Three bars on right are elicited beliefs of subjects at the end of the session. ∗ ∗ ∗p < 0.01, ∗ ∗ p < 0.05, ∗p < 0.1. Significance obtained by comparing to the control (no advice), using rank-sum. Standard errors shown.

on the principal using the punishers tokens. The three bars on the left are empirical averages from the experiments. The three bars on the right are beliefs of the punishers elicited at the end of the session after playing all fifteen rounds of the game. First, note subjects are overly optimistic of the principals’ intentions: Comparing the bars on the right to those on the left, subjects believe principals had to take the selfish action roughly twice as often as the data bear out. Second, and more germane to the main hypothesis, we find evidence that advice does change the punishers’ beliefs of the intentions of the principals. As shown in the bars on the right, without advice, punishers believe, on average, that there is a 46% chance that a “selfish” principal is not really being selfish. This belief goes up to 53% when the seemingly selfish action follows selfish advice (rank-sum p = 0.08). Fair advice has a stronger effect on beliefs: The average belief that a “selfish” principal had to act selfishly goes down to 31% (rank-sum p < 0.01). That advice changes the punishers’ beliefs of the principals’ intentions may explain the differences in punishment with and without advice. In the next section, we will describe our results for the Certain treatments, where we eliminate punishers’ uncertainty about how the principal treated her. If this treatment arm turns off the effect of advice on punishment, the effect was operating solely through the beliefs channel. If the effect remains, beliefs cannot be the sole thrust of the main result.

3.2

Certain Treatment Results

In the Certain treatments, punishers could observe which players’ tokens were effective and thus knew whether or not the principal was indeed acting selfishly. As before, the data we are interested in occur when the principal used the punishers’ tokens. Recall the [Certain]

11

6 3.5

3.5

4

4

4.5

4.5

5

5

5.5

5.5

6

Figure 3: Average Punishment of Selfish Behavior, by Round: Certain Treatments

0

5

Round

Selfish Advice

10

15

0

No Advice

5

Round

Fair Advice

(a) Selfish Advice vs. No Advice

10

15

No Advice

(b) Fair Advice vs. No Advice

Lines depict mean punishment, by round, following a principal using both of the punishers’ tokens when all tokens ware effective.

treatment sessions were run later than the Uncertain sessions, and are not meant to be directly compared. Rather this treatment arm asks a question revealed in the results of the first. Result 3: Even without any informational role for advice, selfish advice significantly reduces punishment of selfish actions and significantly decreases the likelihood that those actions are punished. Figure 3a is analogous to Figure 1a: we compare the per-round average punishment of selfish principals in the control with punishment of selfish principals that received advice, by the content of that advice. Average punishment for selfish principals that are advised to behave selfishly is lower in every single round than selfish principals in the no-advice control treatment. Table 2 formalizes this result. The first two columns in panel A show OLS regressions where punishment in a given period is the dependent variable and standard errors are clustered by subject. Average punishment significantly decreases from $5.35 in the control to $4.43 in the treatment with advice (p < 0.01), a decrease of about 17.4%. This result holds if we control for demographic variables.16 Selfish advice also significantly reduces the probability of being punished. While in the control 93% of all selfish actions are punished, selfish actions validated by selfish advice have a lower, 79% chance of being punished (p < 0.01). If we use averages per punisher as independent variables, all of these results hold. 17 16 The result holds when interacting advice with the round number. Additionally, in this linear specification, the treatment effect is decreasing by an estimated 4.6% per round, which is marginally significant (See Table ?? in the Appendix). 17 On average, there are 5.5 observations per punisher in the control treatment, 3.8 in the advice treatment with selfish advice, and 1.6 in the advice treatment with fair advice.

12

Table 2: The Effect of Selfish And Fair Advice on Punishment: Certain Treatments (OLS) A - Round-level Punishment (clustered s.e.) Dep. Var. Selfish Advice Fair Advice Constant

Controls Observations Clusters R-squared

(1) Punishment

(2) Punishment

(3) Any Punishment

(4) Any Punishment

-0.93*** (0.26) -0.44 (0.29) 5.35*** (0.13)

-0.90*** (0.25) -0.39 (0.28) 5.27*** (0.31)

-0.13*** (0.04) -0.09* (0.05) 0.93*** (0.02)

-0.13*** (0.04) -0.08* (0.04) 0.95*** (0.05)

NO 1,094 199 0.04

YES 1,094 199 0.07

NO 1,094 199 0.03

YES 1,094 199 0.07

(1) Punishment Avg

(2) Punishment Avg

(3) Any Punishment Avg

(4) Any Punishment Avg

-0.85*** (0.25) -0.33 (0.28) 5.27*** (0.14)

-0.77*** (0.26) -0.27 (0.27) 5.45*** (0.36)

-0.14*** (0.04) -0.08* (0.05) 0.92*** (0.02)

-0.12*** (0.04) -0.07 (0.04) 0.99*** (0.06)

NO 273 199 0.04

YES 273 199 0.08

NO 273 199 0.04

YES 273 199 0.10

B - Punisher-level averages (clustered s.e.) Dep. Var. Selfish Advice Fair Advice Constant

Controls Observations Clusters R-squared

Note: ∗ ∗ ∗p < 0.01, ∗ ∗ p < 0.05, ∗p < 0.1. Robust standard errors clustered at Punisher level in parentheses. “Punishment” is the amount of money deducted in a given round; “Punishment Avg” is the average level of punishment for for a punisher; “Any Punishment” is a dummy for whether “Punishment” is greater than 0; “Any Punishment Avg” is a punisher’s mean likelihood of punishing. Only scenarios where principal took selfish action considered throughout.Controls include: Gender, participation in economics experiments, participation in psychology experiments, number of economics courses taken, perceived difficulty of the instructions, a native-speaker dummy, and round number.

While selfish advice reduces punishment, fair advice has no clear effect. First note, however, that punishment of selfish behavior with no advice is already quite high, with a mean of $5.35 per punisher. Since the maximum possible punishment is $6, this leaves very little room for an increase. Hence the lack of results for the fair advice could be because of the new treatment, or it could be due to a “ceiling effect”. That said, Figure 3b shows that average punishment of selfish actions with fair advice is actually lower than punishment in the control in about half of periods (9 out of 15); however, this effect is not significant (see Table 2). On the other hand, in 3 out of our 4 specifications, fair advice marginally significantly reduced the probability of a principal being punished (p < 0.1 in all four). This could due to the combination of a ceiling effect and noise, or it could be that this treatment flips the effect for the likelihood of punishment (but not the size) for fair advice. To sum up, selfish advice decreases punishment even when the punisher knows the principal

13

1 .4

.4

.6

.6

.8

.8

1

Figure 4: Per-Round Prevalence of Selfishness

0

5

Round Advice

10

15

0

No Advice

5

Round Advice

(a) All Principals

10

15

No Advice

(b) Selfish Principals

Lines depict mean rate of taking the selfish action (using only the punisher’s tokens when all tokens are effective), split by availability of advisers, unconditional on advice given.

acted selfishly. Hence Result 1 cannot be driven by beliefs alone. Advice to act in a selfish way does not just change beliefs that it was necessary; it actually changes how we perceive the morality of that act knowing that it was not necessary.

3.3

Selfishness

Result 4: The availability of advisers increases selfishness (marginally significantly) among those principals who prefer the selfish outcome, but not overall. Does the availability of advisers increase unfair actions? There are two channels that could produce this result. First, having an adviser, or making decisions in a group can nurture selfishness regardless of the preferences of the individuals ex ante: An adviser-advisee pairing breeds selfishness. We will look at this possibility by considering a fully unconditional look at how the presence of advisers affects the incidence of selfishness for all principals. The second channel is that advisers decrease sanctioning, thus relaxing constraints on behavior, allowing those who prefer to be selfish to be selfish. To look at this possibility, we will analyze how the presence of advisers affects the incidence of selfishness among those advisers who prefer the selfish outcome, the “selfish types”.18 Overall, only 40% of our principals preferred the selfish outcome absent punishment and feedback to others. We see evidence for the second channel but not the first. Figures 4a and 4b compare the per-round prevalence of selfishness by the availability of advisers. Figure 4a shows selfishness 18

Recall in the Certain treatment, we measure preferences for fair versus selfish outcome by including a round 0 where the principal chooses whose tokens to use when all are effective, there is no punishment, and no one else learns what she did.

14

for all principals, and Figure 4b only shows the choices for the principals who prefer the selfish outcome. Visually, Figure 4a does not reveal a strong difference in selfishness with advisers present. Indeed, including all principals, the rate of selfishness increases directionally from 52% to 55% when advisers are available (p = 0.7). However, Figure 4b shows a different story. With the exception of round 3, selfishness was higher when advisers were available. Among the selfish types, selfishness increased by 27% (17pp) when advisers were available, unconditional on the advice received, and this increase is marginally significant (rank-sum p = 0.09). This may be a result of selfish type principals hiring selfish advisers, but the results are inconclusive. When both tokens are effective, selfish principals re-hire advisers who tell them to do the fair thing 59% of the time but re-hire those who tell them to take the selfish action 82% of the time (Sign-rank p = 0.21, N = 12), compared to 76% and 77% respectively for the fair type principals. Hence we see marginal significance for the availability of advisers increasing selfishness for those who prefer the selfish action absent sanctioning. However, we find no evidence that an adviser-advisee relationship changes preferences towards the selfish action.

4

Discussion

This paper provides first evidence that the morality of actions can be perceived very differently when those actions are advised, even when the adviser has a substantial COI. When unfair actions are preceded by advice to act unfairly, they are punished substantially less. However, when unfair actions are preceded by advice to act fairly, they are punished substantially more. In a follow-up treatment, we find these results are not wholly due to the advice suggesting the principals had to act in a certain way; holding fixed the intentions and outcomes of their actions, advice changes how the act is morally perceived. Like any paper, whether experimental or empirical, we are only able to study one context. While we only proffer the results herein as a proof of concept, understanding the robustness obviously requires varying the contextual parameters. One potentially important parameter is the identity of the advisers. Often, advisers are authority figures, experts, like the IMF or various consulting firms. Perhaps advice would change moral perceptions even more when coming from such sources. On the other hand, perhaps advice was able to change norms in our context precisely because the advisers were similar to the punishers, both randomly assigned from the undergraduate study pool. With the increasing ubiquity of consulting and advising, understanding how these institutions affect moral perceptions is increasingly important. The results suggest that organizations may be more likely to get away with acting unfairly if they have an adviser, and second, that they have an incentive to hire an adviser to help them get away with it. Finally, and more specifically, compensation consultants, even with a known conflict of interest, could have a role in inflating CEO pay.

15

References Azrieli, Yaron, Christopher P Chambers, and Paul J Healy, “Incentives in experiments: A theoretical analysis.” Bartling, Bjorn and Urs Fischbacher, “Shifting the Blame: On Delegation and Responsibility,” Review of Economic Studies, 2012, 79, 67–87. Bebchuk, Lucian Arye and Jesse M Fried, “Executive Compensation as an Agency Problem,” Journal of Economic Perspectives, 2003, 17 (3), 71–92. Bonaccio, Silvia and Reeshad S. Dalal, “Advice taking and decision-making: An integrative literature review, and implications for the organizational sciences,” Organizational Behavior and Human Decision Processes, 2006, 101, 127–151. C ¸ elen, Boga¸ chan, Shachar Kariv, and Andrew Schotter, “An experimental test of advice and social learning,” Management Science, 2010, 56 (10), 1687–1701. Charness, Gary and Arthur Schram, “Inducing Social Norms in Laboratory Allocation Choices,” Management Science, 2015, 61 (7), 1531–1546. Chen, Zhuoqiong and Tobias Gesche, “Persistent Bias in Advice-Giving,” University of Zurich Working Paper No. 228, 2016. Chu, Jenny, Jonathan Faasse, and P. Raghavendra Rau, “Do compensation consultants enable higher CEO pay? New evidence from recent disclosure rule changes,” Working Paper, University of Cambridge, 2015. Coffman, Lucas C., “Intermediation Reduces Punishment (and Reward),” American Economic Journal: Microeconomics, 2011, 3 (4), 77–106. Dana, Jason, Roberto A. Weber, and Jason Xi Kuang, “Exploiting moral wiggle room: experiments demonstrating an illusory preference for fairness,” Economic Theory, 2007, 33, 67–80. Demski, Joel S, “Corporate conflicts of interest,” The Journal of Economic Perspectives, 2003, 17 (2), 51–72. Fershtman, Chaim and Uri Gneezy, “Strategic delegation: An experiment,” RAND Journal of Economics, 2001, pp. 352–368. Fischbacher, Urs, “z-Tree: Zurich Toolbox for Ready-made Economic Experiments,” Experimental Economics, 2007, 10 (2), 171–178. Foerster, Stephen R, Juhani T Linnainmaa, Brian Melzer, and Alessandro Previtero, “Retail Financial Advice: Does One Size Fit All?,” Journal of Finance, Forthcoming. Gehrig, Thomas, Werner G¨ uth, Ren´ e Lev´ınsk` y, and Vera Popova, “On the evolution of professional consulting,” Journal of Economic Behavior & Organization, 2010, 76 (1), 113–126. Green, Jerry R and Nancy L Stokey, “A two-person game of information transmission,” Journal of Economic Theory, 2007, 135 (1), 90–104. Greiner, Ben, “An Online Recruitment System for Economic Experiments,” 2004.

16

Hamman, John R., George Loewenstein, and Roberto A. Weber, “Self-Interest through Delegation: An Additional Rationale for the Principal-Agent Relationship,” American Economic Review, 2010, 100 (4), 1826–1846. Harvey, Nigel and Ilan Fischer, “Taking advice: Accepting help, improving judgment, and sharing responsibility,” Organizational Behavior and Human Decision Processes, 1997, 70 (2), 117–133. Kagel, John H and Peter McGee, “Team versus individual play in finitely repeated prisoner dilemma games,” American economic Journal: microeconomics, 2016, 8 (2), 253– 276. Kray, Laura and Richard Gonzalez, “Differential weighting in choice versus advice: Ill do this, you do that,” Journal of Behavioral Decision Making, 1999, 12 (3), 207–218. Kuang, Xi Jason, Roberto A Weber, and Jason Dana, “How effective is advice from interested parties?: An experimental test using a pure coordination game,” Journal of Economic Behavior & Organization, 2007, 62 (4), 591–604. Malmendier, Ulrike and Devin Shanthikumar, “Do security analysts speak in two tongues?,” Review of Financial Studies, 2014, 27 (5), 12871322. Mullainathan, Sendhil, Markus Noeth, and Antoinette Schoar, “The market for financial advice: An audit study,” 2012. Murphy, Kevin J and Tatiana Sandino, “Executive pay and independent compensation consultants,” Journal of Accounting and Economics, 2010, 49 (3), 247–262. Oexl, Regine and Zachary J Grossman, “Shifting the blame to a powerless intermediary,” Experimental Economics, 2013, 16 (3), 306–312. Paharia, Neeru, Karim S Kassam, Joshua D Greene, and Max H Bazerman, “Dirty work, clean hands: The moral psychology of indirect agency,” Organizational Behavior and Human Decision Processes, 2009, 109 (2), 134–141. Schotter, Andrew, “Decision making with naive advice,” American Economic Review, 2003, pp. 196–201.

17

FOR ONLINE PUBLICATION

A

Additional Statistics and Regression Results

Table A-1: Questionnaire Variables - Averages per Treatment (All Participants) UncertainControl

UncertainAdvice

p-value H0 : equality

CertainControl

CertainAdvice

p-value H0 : equality

Male

0.62 (0.04)

0.61 (0.04)

0.82

0.63 (0.04)

0.59 (0.04)

0.41

Econ. Exp.

4.43 (0.42)

4.15 (0.40)

0.61

2.87 (0.27)

2.95 (0.20)

0.31

Psych. Exp.

2.63 (0.33)

2.81 (0.41)

0.62

2.76 (0.33)

2.76 (0.36)

0.46

Native Speaker

0.80 (0.03)

0.73 (0.03)

0.11

0.8 (0.03)

0.83 (0.03)

0.47

Econ. Courses

2.82 (0.20)

3.08 (0.33)

0.89

2.81 (0.22)

2.56 (0.19)

0.63

SAT

1228 (78.10)

1119 (73.97)

0.20

1571 (67.14)

1490 (61.05)

0.47

F-test

0.64

F-test

0.65

“Male” and “Native Speaker” are dummy variables. “Econ. Exp.”, “Psych. Exp.”, and “Econ. Courses” reflect the number of times a subject has participated in an economics experiment or psychology experiment, or taken an economics course. p-values are from tests of proportions for dummies and rank-sum tests otherwise. Standard errors in parentheses.

18

Table A-2: Questionnaire Variables - Averages per Treatment (Punishers) UncertainControl

UncertainAdvice

p-value H0 : equality

CertainControl

CertainAdvice

p-value H0 : equality

Male

0.65 (0.05)

0.62 (0.05)

0.71

0.61 (0.05)

0.61 (0.05)

1.00

Econ. Exp.

3.79 (0.32)

4.12 (0.36)

0.34

3.03 (0.35)

3.07 (0.29)

0.30

Psych. Exp.

2.76 (0.41)

2.51 (0.42)

0.51

2.68 (0.41)

2.55 (0.49)

0.33

Native Speaker

0.82 (0.04)

0.74 (0.05)

0.17

0.76 (0.04)

0.86 (0.04)

0.07

Econ. Courses

2.69 (0.20)

2.94 (0.28)

0.94

2.7 (0.27)

2.45 (0.22)

0.94

SAT

1159 (94.73)

1063 (105.33)

0.50

1502 (85.05)

1444 (88.98)

0.73

F-test

0.73

F-test

0.85

“Male” and “Native Speaker” are dummy variables. “Econ. Exp.”, “Psych. Exp.”, and “Econ. Courses” reflect the number of times a subject has participated in an economics experiment or psychology experiment, or taken an economics course. p-values are from tests of proportions for dummies and rank-sum tests otherwise. Standard errors in parentheses.

Table A-3: Elicited Beliefs - Averages (All Participants)

Uncertain

Certain

Punishment Selfish Actions

Likelihood Using Punisher’s Tokens Necessary

Morally Unacceptable

Retained Adviser

Control

4.49 (0.14)

45.5 (2.2)

-

-

Selfish Advice

3.39 (0.16)

47.5 (2.3)

-

70.9 (1.9)

Fair Advice

4.75 (0.12)

31.9 (1.9)

-

62.8 (1.9)

Control

5.07 (0.12)

-

2.96 (0.13)

-

Selfish Advice

4.61 (0.12)

-

2.42 (0.11)

72.4 (1.7)

Fair Advice

4.83 (0.13)

-

3.16 (0.12)

60.9 (2.0)

“Punishment” is subjects’ beliefs of the level of punishment following a selfish action. “Likelihood Using Punisher’s Tokens Necessary” is subjects’ beliefs that using the punishers’ two tokens was necessary to save the accounts. “Morally unacceptable” measures how morally unacceptable subjects believe a selfish action to be (1=not unacceptable, 5=very unacceptable). “Retained adviser” measures beliefs about the likelihood the adviser was retained following a round where all tokens were effective. Standard errors in parentheses.

19

Table A-4: Elicited Beliefs - Averages (Punishers)

Uncertain

Certain

Punishment Selfish Actions

% Chance Using Punisher’s Tokens Necessary

Morally Unacceptable

% Chance Retained Adviser

Control

4.67 (0.16)

45.9 (2.7)

-

-

Selfish Advice

3.21 (0.22)

53.2 (3.2)

-

71.4 (2.9)

Fair Advice

4.80 (0.16)

31.0 (2.7)

-

67.7 (2.7)

Control

5.12 (0.14)

-

3.28 (0.16)

-

Selfish Advice

4.69 (0.16)

-

2.68 (0.16)

72.6 (2.5)

Fair Advice

5.02 (0.18)

-

3.65 (0.15)

65.0 (2.8)

“Punishment” is subjects’ beliefs of the level of punishment following a selfish action. “Likelihood Using Punisher’s Tokens Necessary” is subjects’ beliefs that using the punishers’ two tokens was necessary to save the accounts. “Morally unacceptable” measures how morally unacceptable subjects believe a selfish action to be (1=not unacceptable, 5=very unacceptable). “Retained adviser” measures beliefs about the likelihood the adviser was retained following a round where all tokens were effective. Standard errors in parentheses.

Does Confusion Drive the Results? The data herein show that selfish advice reduces punishment despite the adviser’s COI. However, one may be worried that some punishers were not aware of the adviser’s incentives. This does not seem to be the case. First, punishers were reminded in two different waiting screens that had no other information other than the fact that principals were deciding whether to retain their adviser from the previous period or switch to a new one. These screens were shown in the 12 rounds where principals made hiring decisions. The screens are shown in appendix C. Second, we can look at subsets of our subjects for whom the results might not hold if the confusion hypothesis is correct. Though these subsamples may be underpowered, such tests could be indicative of the underlying channel for the results overall. We find our results qualitatively hold if we only consider subsets of punishers that might have been more likely to understand the adviser’s incentives – Native speakers, and those who self-reported the instructions were not difficult. For the uncertain treatments, table A-7 shows the effect of selfish advice on punishment for these subsamples and compares with the effects shown in table 1 (“Original Coeff.”). Likewise, table A-8 shows the effect of selfish advice on punishment for the same subsamples and compares with the effects shown in table 2 (“Original Coeff.”). All the subsample effects go in the same direction and are of similar magnitude than the effects estimated with the complete sample. These subsamples are underpowered for statistical significance; however, since the magnitudes are roughly the same it is clear the overall results are not being driven by confused, uninterested subjects.

20

21 0.11

YES 257

4.74*** (0.52) YES 191 148 0.13

5.05*** (0.74)

-0.79** (0.38) 0.75** (0.33)

Punishment Avg

Punishment Avg -1.01*** (0.32) 0.66** (0.27)

(2) Incl. SAT

YES 928 148 0.08

4.92*** (0.74)

-0.64 (0.43) 0.98*** (0.37)

Punishment

(2) Incl. SAT

(1) Main

YES 1,210 197 0.07

4.44*** (0.57)

-0.81** (0.36) 0.79*** (0.29)

Punishment

(1) Main

YES 250 194 0.11

4.91*** (0.53)

-0.96*** (0.33) 0.63** (0.28)

Punishment Avg

(3) Excl. t=1,5

YES 1,060 194 0.07

4.60*** (0.57)

-0.75** (0.37) 0.80*** (0.30)

Punishment

(3) Excl. t=1,5

YES 1,210 197 0.07

4.59*** (0.59)

-1.17*** (0.45) 0.73* (0.41) 0.04 (0.03) 0.00 (0.04)

Punishment

(4) t interactions

0.11

YES 257

0.84*** (0.08)

-0.16*** (0.05) 0.09** (0.04)

Any Punishment Avg

(1) Main

YES 1,210 197 0.06

0.80*** (0.08)

-0.11* (0.06) 0.12** (0.04)

Any Punishment

(1) Main

YES 191 148 0.13

0.89*** (0.11)

-0.12** (0.06) 0.11** (0.05)

Any Punishment Avg

(2) Incl. SAT

YES 928 148 0.07

0.88*** (0.11)

-0.08 (0.07) 0.15*** (0.06)

Any Punishment

(2) Incl. SAT

YES 250 194 0.1

0.85*** (0.08)

-0.15*** (0.05) 0.09* (0.04)

Any Punishment Avg

(3) Excl. t=1,5

YES 1,060 194 0.05

0.82*** (0.08)

-0.10* (0.06) 0.12** (0.05)

Any Punishment

(3) Excl. t=1,5

Note: ∗ ∗ ∗p < 0.01, ∗ ∗ p < 0.05, ∗p < 0.1. Robust standard errors clustered at Punisher level in parentheses. “Punishment” is the amount of money deducted in a given round; “Punishment Avg” is the average level of punishment for for a punisher; “Any Punishment” is a dummy for whether “Punishment” is greater than 0; “Any Punishment Avg” is a punisher’s mean likelihood of punishing. Only scenarios where principal took selfish action considered throughout. Controls include: Gender, participation in economics experiments, participation in psychology experiments, number of economics courses taken, perceived difficulty of the instructions, a native-speaker dummy, and round number. (1) = Model in the Results section, (2) = Controls for reported SAT score, (3) = Removes rounds 5 and 10, the only rounds without an adviser COI (round 15 had a bonus paid after giving advice), (4) = Interacts round with selfish and fair advice.

Controls Observations Clusters R-squared

Constant

Fair Advice

Selfish Advice

Dep. Var.

B - Punisher-level averages (clustered s.e.)

Controls Observations Clusters R-squared

Constant

Fair Advice * t

Selfish Advice * t

Fair Advice

Selfish Advice

Dep. Var.

A - Round-level Punishment (clustered s.e.)

Table A-5: The Effect of Selfish and Fair Advice on Punishment: Uncertain Treatments, additional regressions (OLS)

YES 1,210 197 0.06

0.82*** (0.09)

-0.14** (0.07) 0.09 (0.06) 0.00 (0.01) 0.00 (0.01)

Any Punishment

(4) t interactions

22 0.08

YES 273

5.45*** (0.36)

-0.77*** (0.26) -0.27 (0.27)

Punishment Avg

(1) Main

YES 1,094 199 0.04

5.27*** (0.31)

-0.90*** (0.25) -0.39 (0.28)

Punishment

(1) Main

YES 160 118 0.06

5.24*** (0.65)

-0.22 (0.31) 0.12 (0.32)

Punishment Avg

(2) Incl. SAT

YES 657 118 0.06

4.90*** (0.63)

-0.53 (0.35) 0.21 (0.28)

Punishment

(2) Incl. SAT

YES 270 199 0.08

5.53*** (0.37)

-0.78*** (0.27) -0.30 (0.28)

Punishment Avg

(3) Excl. t=1,5

YES 960 199 0.08

5.42*** (0.32)

-0.91*** (0.26) -0.47 (0.30)

Punishment

(3) Excl. t=1,5

YES 1,094 199 0.05

5.59*** (0.31)

-1.51*** (0.42) -1.03** (0.51) 0.07* (0.04) 0.07 (0.05)

Punishment

(4) t interactions

0.1

YES 273

0.99*** (0.06)

-0.12*** (0.04) -0.07 (0.04)

Any Punishment Avg

(1) Main

0.07

YES 1,094

0.95*** (0.05)

-0.13*** (0.04) -0.08* (0.04)

Any Punishment

(1) Main

YES 160 118 0.08

0.96*** (0.10)

-0.06 (0.05) -0.01 (0.05)

Any Punishment Avg

(2) Incl. SAT

YES 657 118 0.07

0.89*** (0.10)

-0.10* (0.06) 0.02 (0.04)

Any Punishment

(2) Incl. SAT

YES 270 199 0.11

1.01*** (0.06)

-0.12*** (0.04) -0.08* (0.04)

Any Punishment Avg

(3) Excl. t=1,5

YES 960 199 0.08

0.98*** (0.05)

-0.13*** (0.04) -0.09* (0.05)

Any Punishment

(3) Excl. t=1,5

Note: ∗ ∗ ∗p < 0.01, ∗ ∗ p < 0.05, ∗p < 0.1. Robust standard errors clustered at Punisher level in parentheses. “Punishment” is the amount of money deducted in a given round; “Punishment Avg” is the average level of punishment for for a punisher; “Any Punishment” is a dummy for whether “Punishment” is greater than 0; “Any Punishment Avg” is a punisher’s mean likelihood of punishing. Only scenarios where principal took selfish action considered throughout. Controls include: Gender, participation in economics experiments, participation in psychology experiments, number of economics courses taken, perceived difficulty of the instructions, a native-speaker dummy, and round number. (1) = Model in the Results section, (2) = Controls for reported SAT score, (3) = Removes rounds 5 and 10, the only rounds without an adviser COI (round 15 had a bonus paid after giving advice), (4) = Interacts round with selfish and fair advice.

Controls Observations Clusters R-squared

Constant

Fair Advice

Selfish Advice

Dep. Var.

B - Punisher-level averages (clustered s.e.)

Controls Observations Clusters R-squared

Constant

Fair Advice * t

Selfish Advice * t

Fair Advice

Selfish Advice

Dep. Var.

A - Round-level Punishment (clustered s.e.)

Table A-6: The Effect of Selfish and Fair Advice on Punishment: Certain Treatments, additional regressions (OLS)

YES 1,094 199 0.08

1.02*** (0.05)

-0.28*** (0.07) -0.20** (0.08) 0.02** (0.01) 0.01* (0.01)

Any Punishment

(4) t interactions

Table A-7: Comparison of Treatment Effects: Overall and Different Subsamples, Uncertain treatments [Original Coeff.]

Punishment [-0.63*]

Punishment [-0.81**]

Any Punishment [-0.08]

Any Punishment [-0.11*]

Native speaker (960 obs, 155 clusters)

-0.60 (0.40)

-0.52 (0.40)

-0.08 (0.07)

-0.07 (0.07)

Perceived difficulty ≤ 3 (837 obs, 134 clusters)

-0.79* (0.45)

-0.71 (0.46)

-0.14* (0.07)

-0.12 (0.08)

Controls

NO

YES

NO

YES

Note: ∗ ∗ ∗p < 0.01, ∗ ∗ p < 0.05, ∗p < 0.1. Robust standard errors clustered at Punisher level in parentheses. “Punishment” is the amount of money deducted in a given round. “Any Punishment” is a dummy for whether “Punishment” is greater than 0. Only scenarios where principal took selfish action considered throughout.

Table A-8: Comparison of Treatment Effects: Overall and Different Subsamples, Certain treatments [Original Coeff.]

Punishment [-0.93***]

Punishment [-0.90***]

Any Punishment [-0.13***]

Any Punishment [-0.13***]

Native speaker (896 obs, 161 clusters)

-0.99*** (0.29)

-0.99*** (0.28)

-0.14*** (0.05)

-0.15*** (0.04)

Perceived difficulty ≤ 3 (754 obs, 138 clusters)

-0.82** (0.33)

-0.62** (0.31)

-0.12** (0.05)

-0.10** (0.05)

Controls

NO

YES

NO

YES

Note: ∗ ∗ ∗p < 0.01, ∗ ∗ p < 0.05, ∗p < 0.1. Robust standard errors clustered at Punisher level in parentheses. “Punishment” is the amount of money deducted in a given round. “Any Punishment” is a dummy for whether “Punishment” is greater than 0. Only scenarios where principal took selfish action considered throughout.

23

B

Screenshots: Waiting Screens

Figure A-1: Screen shown at the start of a round.

Figure A-2: Screen shown when principals were making hiring decisions.

24