Convenient or Inconvenient Beliefs? A Review of Di Tella et al.’s (2015) “Conveniently Upset: Avoiding Altruism by Distorting Beliefs About Others’ Altruism” Corruption Game Experiment

Convenient or Inconvenient Beliefs?

A Review of Di Tella et al.’s (2015) “Conveniently

Upset: Avoiding Altruism by Distorting Beliefs About

Others’ Altruism” Corruption Game Experiment.

TABLE OF CONTENTS

1. Introduction 2

2. Theoretical Background 3

3. Experimental Design 5

4. Data Analysis Methodology 9

5. Results 14

6. Results Discussion 21

7. Conclusion and General Discussion 22

8. Appendix 26

9. Citations 28

Statement of Originality

This document is written by Klaus Fonseca Hoeltgebaum who declares to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document are original and that no sources other than those mentioned in the text and its references have been used in creating it.

The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

Introduction

Is it acceptable to take the last piece of pie without asking everyone else? And if it is not acceptable, do you find an excuse to do it anyways? These are the questions concerning the self-serving biases and motivated beliefs of agents. A self-serving bias, as defined by Kaplan and Ruffle (2007), is when individuals preferences affect their own beliefs to make their choices more palatable. A recent study by Di Tella et al. (2015) focuses on the direction of causality for self-serving biases, with the goal of finding the effect that a selfish choice has on this person’s belief about someone else. The study concerns with a variation of the dictator games. A novel experiment design is used whose goal is to only capture the cognitive dissonance arising from a person’s beliefs about others. The goal of the experiment is to test two hypotheses. The first one seeks an effect on beliefs of agents, predicting that they change beliefs about others to make an unfair action on the agent’s part be perceived as fair. The second hypothesis is concerned with selfish acts. Agents capable of self-deception will be more selfish. Di Tella et al. (2015) find evidence for both; however, these effects are not very robust. This paper will thus replicate and verify the robustness of their results for one version of their experiment and give an overview of how it stands in the current literature regarding motivated beliefs.

Theoretical Background

Psychological studies in the effect of self-esteem on a persons’ decision making (Hales, 1985) brought about new models for economic analysis. This intersection between psychology and economics led to the study of motivated beliefs which

investigates agents that were biased towards unethical actions yet still believed them be morally acceptable (Bersoff, 1999). Dana et al. (2007) found evidence in dictator games that agents acted generously due to wanting to be ‘fair’ and showed how agents can exploit context to act more selfishly. Follow up research on the subject has expanded on the extent of its effect, if it is generalizable to other games, and some of its limitations (van der Weele, Kulisa, Kosfeld, Friebel, 2014; Regner, 2017)

Di Tella et al. (2015) expand this area with a focus on how the opportunity for selfish choices can elicit negative self-serving biases. When trying to be moral these agents act against their morals because the perceive that the alternative action still feels fair (Gino, Norton, Weber, 2016). For example, suppose you are on a supermarket checkout and the cashier forgot to scan some small item such as a spice jar. This opportunity is the crux of the effect investigated by Di Tella et al. (2015), because now you can take the spice jar ‘for free’ and have the opportunity to convince yourself that the supermarket is a large greedy faceless corporation, even if you had no strong negative opinions over them up to that point. This is not to be confused with the opposite direction of the effect, where because you have a negative opinion of the supermarket chain you consider taking the spice jar ‘for free’.

This example brings us to Di Tella et al. (2015) more general investigation. The main research questions in their paper are concerned with the elicitation of the

previously described negative self-serving biases. To test these hypotheses, they create a modified dictator game called the ‘corruption’ game, where they split agents into two groups with differing opportunities to act selfishly. By making groups with different choice limitations, they argue that agents in these groups will also have differing negative self-serving belief elicitations. The agents, however, still have the same

incentive of maximizing monetary profits. The only difference between them then should be their beliefs about their partners.

Given this context, the first hypothesis is stated as:

“Hypothesis 1: Beliefs about others are affected by people’s own desire to be selfish”. (Di Tella et al., 2015).

It is also described in the paper of Ging-Jehli et al. (2019) as ‘strategic cynicism’ where agents hold negative beliefs to others if they want to justify a choice that is costly towards the other person. Agents maximizing personal gain may also concerned with their own self-image (Benabou and Tirole, 2002). Hypothesis 1 (Di Tella et al., 2015) therefore specifically tackles changes in individual beliefs of others in situations where a person has the opportunity to act selfishly.

The second hypothesis is stronger and is concerned over selfish actions of agents:

“Hypothesis 2: Selfish actions depend on people’s ability to manipulate their beliefs about others.” (Di Tella et al., 2015).

They ask if these actions depend upon belief manipulations about others. It is stronger as it asks if beliefs predicted by the first hypothesis affect actions of agents. The experiment replicated in this part of the paper does not tackle the latter hypothesis explicitly. Most of the focus will be in explaining Hypothesis 1.

Experimental Design

The ‘corruption’ game experiment is carried out in the z-Tree computer environment. It begins with a series of 5 short, around 1 minute, tasks which once completed endow each participant with 10 tokens. This part is not so relevant to the overall hypothesis tests. However, this makes the participants feel entitled to their tokens and aware that other participants worked for their tokens. The tokens are therefore not received for ‘free’ in the participants’ perspective. Later in the game one participant can redistribute these tokens, and the fact that they were ‘earned’

individually may affect his perceived self-image if he takes tokens from his partner. After this endowment game, participants go through instructions of how the base corruption game works. Players were matched randomly into pairs; one player was assigned the ‘allocator’ role and the other is assigned the ‘seller’ role.

The allocators’ task is to split the combined total token endowment of the pair. In addition, they are randomly split into two groups and receive additional instructions. They could drag yellow tokens from one player to another, however green tokens were blocked and could not be moved. In the group called Able=8 allocators were able to move up to 8 yellow tokens from each player, and in the other group called Able=2 they could only move up to 2 yellow tokens from each player.

Only the allocators are aware of these additional restrictions, none of the sellers are aware that the allocators received these additional instructions. Furthermore,

allocators are only aware of the limitations they individually received; they do not know if others had the same limitations in place as them. All allocators are aware that the

sellers are not aware of these limitations and they all received a copy of the seller's instruction sheet to further guarantee this knowledge. Given the setup, every allocator should form monetary profit maximizing beliefs independent of their Able group, unless they also take self-image into consideration. The self-serving bias effect can then be captured to test the first hypothesis.

The sellers meanwhile decide how much each endowment is worth and are given two options. Option A makes each token worth $2, while Option B makes each token worth $1, however if the latter option is chosen then they will also earn an additional $10, independent from how many tokens they are assigned by the allocator.

Both players make their choices and tasks simultaneously, and neither is aware of which choice the other player made until after they both committed to them. All participants had to complete a questionnaire with four questions on the instructions to see if they understood how the game works in practice and the rules of the game. After this, they could proceed into the actual game. In the final data processing, only

participants who answer more than 80% of the questions correctly were considered. The allocators received additional questions before selecting the number of tokens. Each allocator is additionally asked two questions over their beliefs. The first is if they believe their partnered seller would select Option B, the corrupt option. The second question, which also included an extra bonus monetary reward if they guessed correctly, is what percentage of all sellers would choose option B (2).

Once the game is completed, they fill in a survey about age, gender, and socioeconomic status. There are various steps to preserve anonymity throughout the experiment. This includes using anonymous number identities until the final survey, reminders that the experiment is anonymous, and individual work setups where

allocators did not know which seller they were assigned, and vice-versa. Anonymity was strictly enforced to avoid behavioral justifications from subjects to the experimenters, such as subjects who wish to appear fair to experimenters (Andreoni and Bernheim, 2009) and act more altruistically in-game, distorting the results.

The key to how the research question is tackled in this design is the difference in information between allocators and sellers. The allocator’s beliefs about how morality of the sellers is revealed when asked what their belief of the overall percentage of sellers that are corrupt (Di Tella et al, 2015). Since sellers have the choice of being corrupt, the allocator that can take more tokens can justify taking more by convincing themselves

that most sellers are corrupt. Note that here there is the implicit assumption that

allocators who believe that most sellers are corrupt also believe that their own seller is corrupt. This can be weakly verified by the individual belief statistic collected from each allocator. It may be less accurate, as it is not rewarded monetarily for a correct guess, but the intention here is to express beliefs without cost such that accuracy is less desirable then the subjects actual beliefs. Another reason for this is to avoid

conservative belief elicitation. Under the assumption that agents are risk-averse, the allocator may believe that his matched seller is not corrupt. Nonetheless they signal beliefs that the corrupt choice was chosen anyways as this is the rational monetary profit maximizing outcome without self-image considerations for the seller.

The allocator’s belief that most sellers will pick option B - the ‘corrupt’ option - should on average be higher in the Able=8 compared to the Able=2 group. If the first hypothesis holds, the Able=8 allocators will predict this higher seller corruption to justify their choice of taking more tokens for themselves. On the other hand, the Able=2 group cannot move so many tokens and should therefore experience less cognitive

dissonance and express fewer selfish belief preferences. Since their token taking ability is highly restricted, they do not need to change their beliefs about sellers to preserve their self-image.

While the main goal of the experiment is to measure beliefs there is also an assumption in the second hypothesis that the allocators change these beliefs because they can justify taking more tokens for themselves. This implies that they would justify their selfish choice as morally acceptable in these cases. Di Tella et al. (2015) most likely did not explore this issue further in the base version of the game as they cannot differentiate which allocators justify their selfish choice as morally acceptable from the ones who are just selfish and do not have strong preferences to preserve their self-image but still believe that their partnered seller is corrupt. A novel way to verify the second hypothesis through statistical analysis is by selecting allocators who signal a below-average percentage corruption belief across all allocators yet still believe that their partnered seller is corrupt. These allocators are more likely to consider self-image

as valuable since they account for it when guessing the seller’s choice, however, still think their matched seller is corrupt. Chance may still affect the allocator’s choice, but it becomes more difficult to justify less cognitive dissonance as these allocators most likely are aware of their own belief divergence. The methodology will be further discussed in the end of the next section.

Data Analysis Methodology

Di Tella et al. (2015) present their data analysis in four main tables. Table 1 describes in words each of the variable names used in the analysis and their representations. The definitions given are restated here:

Is Corrupt: Dummy variable that takes the value of 1 if the allocator guesses that

the seller chose Option B.

% Corrupt: The allocator’s guess of the percentage of sellers in the game that

choose Option B.

Able=8: this Dummy value that takes a value of 1 if the allocator was in the group

that could take up to 8 tokens from the seller’s pile.

Tokens Taken: number of tokens taken from the seller by the allocator.

Age: Age of the individual

Female: A dummy variable that is 1 when the individual is a female, and 0 if they

are a male.

Socioeconomic class: “Lower class (1); Middle-lower class (2); Middle class (3);

Middle-higher class (4); Higher class (5).”

In the paper, Table 2 describes the summary of the statistics for the allocator subsample. Table 3 shows the average choices, beliefs, and characteristics by treatment group, and are the main tables used for the analysis. Table 4 shows the treatment effects in the form of an ordinary least squares regression on the variables. The effect of the Able=8 group is presented relative to the Able=2 group for each variable.

In the paper Di Tella et al. (2015) first show that the treatment group sample was effectively random with control variables. Participant characteristics, such as gender, age and socioeconomic status were not significant. They then verify that there is indeed a significant difference between the Able=8 and Able=2 treatment groups in the average number of tokens taken by allocators. To do this they measure the dependent variables, Tokens Taken, Is Corrupt and % Corrupt, with the independent variable being the

treatment group dummy Able=8 vs. Able=2. The dummy is assigned the value 1 for Able=8 (and Able=2 is assigned 0) and the variable coefficients measure the relative effect from one group to another.

The cognitive dissonance effect is quantified by comparing the coefficient measurements found in the regression described in the previous paragraph. These measure the average mean difference between the treatment groups for the tested variables. The null hypothesis for the test is that there are no self-serving biases, such that no belief differences are found between the two treatment groups. Else, if the average mean difference significantly different from zero they can show that agents do indeed change their beliefs given the opportunity to act selfishly.

To verify the significance of results, a Student’s T-test is employed. The T-test can use the estimator data in place of a population parameter. The population is not assumed to be normally distributed, however the sample means are assumed to be normal for the test to hold. The assumption of normality over the sample means is acceptable, a mean difference for a minimum of 80% power test between the treatment groups shows that the minimum sample size of 22 is necessary and the sample of allocators is 30. The test is valid for this case, as they are only testing for the average cognitive dissonance effect, which requires a single location test where the original population constants are unknown. Treatment groups should follow the same

distributions as sample selection was shown to be random, and allocators come from the same population. There should be an effect between treatment groups and both belief measurement variables, Is Corrupt and % Corruption.

The allocator’s sample is, however, subdivided into treatment groups of 15 each and mean comparison is between these groups the results may not have sufficient power. For this reason, a Wilcoxon Signed Rank test can be used as a non-parametric alternative to verify the robustness of the previous statistical results.

The Wilcoxon Signed Rank test ranks the absolute differences between the tested variable pairs and calculates a test statistic based on the sum of the differences, weighted according to their rank. Zero differences are excluded which further reduces the sample size. The test assumes that the paired variables have the same distribution, and it is symmetric around 0. Larger samples should converge towards a normal

distribution. Given that the distribution of the data may not be normal, a non-parametric test would allow an assessment of significance with a symmetry assumption over the distributions of the samples. Considering that the treatment group observations come from the same sample population this assumption holds in this case. Furthermore, measuring beliefs of agents with the normality assumption can make a cardinal interpretation difficult to assess for a relatively small sample size due to noise in the data. Thus, by not relying on normality assumptions of the overall sample and

population distributions, the non-parametric test results increase the overall robustness of the evidence found. The issue with this test is that a large amount of zero results could significantly decrease the already small sample size. While it has more power than a normal t-test under small samples, if too many zeros’ are present the results could be even less useful.

Initially, to quantify the effect they assume that the allocators in the Able=2 have no cognitive dissonance for selfishness such that once the assumption is dropped the value found serves as a lower bound. This provides an estimate for the variation in beliefs due to the cognitive bias. To add precision to this measurement they use an instrumental variable (IV) regression, with the dependent variable being the

independent guess about the seller’s choice, the endogenous variable being the tokens taken and the independent variable being the Able treatment groups.

The authors’ intention is to explain the allocators’ cognitive dissonance per token taken. The IV regression uniquely determines tokens taken with the treatment group and explain agent’s beliefs. From the assumption of hypothesis 1, the IV estimator should consistent. The experimental design has a goal of showing a difference in beliefs between treatment groups, so the instrumental variable is relevant for the dependent variable. The OLS regression results presented later in Table 4 indicate that the effect is indeed significant. The treatment groups and tokens taken are clearly related as the treatment groups are defined by restricted token taking abilities, so the endogenous and instrumental variable have a strong relation in the first stage. The relation between number of tokens taken and beliefs has a clear bias originated from the treatment group setup. From the assumptions, they should be relevant as beliefs change given the opportunity to take more tokens. As such, the 2nd _{stage relation between beliefs and}

instrumented tokens taken should therefore also be relevant, albeit weak. Statistical verification yields significant results as seen in Table 7 in the appendix,Is Corrupt has a significant relation with tokens taken and % Corrupt does as well however only at 10% significance. Even so, the IV regression’s small sample size does not guarantee a reliable asymptotic distribution. Therefore, the IV regression estimators may be biased, however the result derived from them are consistent (McFadden, 1999). Despite these issues, the results should contribute towards the robustness of the first hypothesis, however explanatory statistics derived from the estimations may be unreliable.

Note that the effects of these variables are not comparable with each other due to them representing different types of values, Is Corrupt is considered an indicator variable while % Corrupt is a probability belief (Di Tella et al., 2015). Furthermore, this statistical methodology should add robustness to Hypothesis 1; if there is a significant cognitive dissonance arising from self-serving biases.

Lastly, the methodology for the proposal to measure the second hypothesis mentioned before at the end of the design section. The dependent variable used is the individual seller belief, Is Corrupt, the independent variable is the proportion of total tokens taken over the number of tokens taken and conditional on the allocator having

below-average beliefs of the overall average sellers’ corruption. It may be the case that while allocators do have differences in beliefs about their partnered sellers, they still take the same proportion of tokens of those who do not. This would indicate that while beliefs change for different treatment groups, it does not change the allocator’s choice independently of which treatment group they are in. The largest issue with this test is that by selecting a sub-sample of the sample the number of observations will be insufficient for a reliable test under normality conditions. Therefore, a Wilcoxon Sign Rank test will be employed for more accurate statistics.

Results

Table 1

Treatments Mean Count

Able = 2 Tokens Taken 1.533 (0.834) 15 Is Corrupt 0.467 (0.516) 15 % Corrupt 0.493 (0.349) 15 Female 0.467 (0.516) 15 Age 21.067 (1.280) 15 Socioeconomic Class 3.467 (0.516) 15 Able = 8 Tokens Taken 6 (2.803) 15 Is Corrupt 0.867 (0.352) 15 % Corrupt 0.693 (0.240) 15 Female 0.6 (0.507) 15 Age 21.067 (2.890) 15 Socioeconomic Class 3.533 (0.743) 15 Total Observations 30

Table 2

Tokens Taken Is Corrupt % Corrupt Controls

Able=8 vs. Able=2 (1) 4.467*** 0.400** 0.200* No (0.755) (0.161) (0.110) Able=8 vs. Able=2 (2) 4.293*** 0.344** 0.182 Yes (0.603) (0.160) (0.111) Observations 30 30 30

Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1

Table 3

Is Corrupt % Corrupt Controls

Implied Treatment group effect of Tokens Taken (1) 0.090** 0.045* No (0.032) (0.025) Implied Treatment group effect of Tokens Taken (2) 0.080** 0.043 Yes (0.033) (0.026) Observations 30 30

Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1

The results of the original paper’s Table 3 and Table 4 are successfully replicated here with Table 1 and Table 2. The instrumental variable regression results are

replicated in Table 3. The regressors are found in the rows, while the regressands are represented in each column. To exemplify, for Table 2 the value 0.200 is the coefficient of the regression of % Corrupt on the dummy variable Able=8 vs. Able=2 which takes the value of 1 if the observation is in the Able=8 group, and 0 if in Able=2. The effect of the treatment group on the number of tokens taken is significant for all tests carried out here. For the % Corrupt coefficient, the significance is only at a 10% level and the regression with controls yields no significance. The effect of cognitive dissonance predicted by the authors in the base game is quite weak, if not insignificant. The ordinary least square’s regression seen in Table 2 coefficient for % Corrupt variable shows the cognitive dissonance effect on agents. The coefficient divided by the

average % Corrupt belief in Table 1 measures the relative cognitive bias that agents in the Able=8 treatment groups suffer. This indicates that 29% of the allocators’ beliefs about the sellers are negative self-serving biases.

As for the instrumental variable regression, using the IV coefficients as a measure for cognitive bias per token taken and the average tokens taken in each treatment group they conclude that the relative effect on cognitive dissonance for each token taken is on average 39% of the allocators’ beliefs for the Able=8 group, and 14% for the Able=2. This can be seen in Table 3.

The results for the non-parametric Wilcoxon signed-rank test can be seen in Table 4. The p-value for the treatment groups effect on % Corrupt found in Table 4c is not significant which reveals that this effect is not robust when normality is dropped. Significance at a 10% for the other belief measure, Table 4b, is found. It is also lower than when normality is assumed however sufficient to corroborate robustness of the effect.

Lastly the results for the proposal to check the IV regression results robustness were not significant as can be seen in the appendix Table 6. However, the

non-parametric test in Table 5 yielded significant results.

Table 4: a) Able=8 vs. Able=2 effect on Tokens Taken

Sign Obs. Sum Ranks Expected

Positive 1 5.5 229.500

Negative 26 453.500 229.500

Zero 3 6 6

All 30 465 465

unadjusted variance 2363.75 adjustment for ties -43.88 adjustment for zeros -3.50 adjusted variance

2316.38

Ho: treat = tokenstaken z = -4.654

Prob > z = 0.0000

* ‘treat’ refers to the treatment groups dummy variable where Able=8 is 1 and Able=2 is 0, referred to as Able=8 vs. Able=2 in text. ‘perc_corrupt’ is the variable that refers to %

b) Able=8 vs. Able=2 effect on Is Corrupt

Sign Obs. Sum Ranks Expected

Positive 2 52 117

Negative 7 182 117

Zero 21 231 231

All 30 465 465

unadjusted variance 2363.75 adjustment for ties -15.00 adjustment for zeros -827.75 adjusted variance

1521.00

Ho: treat = Is Corrupt z = -1.667

Prob > z = 0.0956

* ‘treat’ refers to the treatment groups dummy variable where Able=8 is 1 and Able=2 is 0, referred to as Able=8 vs. Able=2 in text. ‘perc_corrupt’ is the variable that refers to %

c) Able=8 vs. Able=2 effect on % Corrupt

Sign Obs. Sum Ranks Expected

Positive 15 201 229.500

Negative 12 258 229.500

Zero 3 6 6

All 30 465 465

unadjusted variance 2363.75 adjustment for ties -6.13 adjustment for zeros -3.50 adjusted variance

2354.13

Ho: treat = perc_corrupt z = -0.587

Prob > z = 0.5569

* ‘treat’ refers to the treatment groups dummy variable where Able=8 is 1 and Able=2 is 0, referred to as Able=8 vs. Able=2 in text. ‘perc_corrupt’ is the variable that refers to %

Table 5: Is Corrupt effect on proportion of tokens taken

Sign Obs. Sum Ranks Expected

Positive 0 0 15

Negative 4 30 15

Zero 5 15 15

All 9 45 45 unadjusted variance 71.25

adjustment for ties -0.50 adjustment for zeros -13.75 adjusted variance

57.00

Ho: Is Corrupt = prop z = -1.987

Prob > z = 0.0469

Results Discussion

While the statistical analysis appears valid, the results do not seem so robust for the base version of the game as % Corrupt is only significant at a 10% level. The least squares regression and instrumental variable is used to quantify the effect, however the number of observations in the base game is quite low and as such there is a wide margin of error for this measure. Furthermore, the non-parametric tests found no

significant difference between the two groups for % Corrupt beliefs again increasing the evidence that the effect is too small. The Is Corrupt belief, expressed at no cost and potentially less accurate, does yield significant differences between treatment groups for all tests. As such, there is evidence for belief differences, but not sufficient to affect behavior. The effects derived by Di Tella et al. (2015) are therefore not so robust. As mentioned earlier they may also have an interpretation mistake by assuming no cognitive bias from the Able=2 group, consequently a relative bias can be found and measured but is not necessarily in the direction predicted by them, nor is the size of the effect accurate for the base version. The IV regression results are consistent but suffer the similar issues as the OLS ones in addition to low reliability and other issues

mentioned previously. Nevertheless, both of their hypothesis holds for a significance of 10% albeit, with some additional robustness found in the additional non-parametric tests seen in Table 4 and, and the additional proposed test found in Table 5.

In the modified game, statistical results for these variables are significant and similar to the results of the base game, which increases the robustness of their results, and makes the statistical analysis of the effect more meaningful. Nonetheless, too many changes in the design occur to make a meaningful comparison between these two versions.

Conclusion and General Discussion

Di Tella et al. (2015) paper contributes to existing literature on motivated beliefs. They find evidence in favor of strategic cynicism, that is agents who act selfishly but are conscious about their self-image will negatively change their beliefs over others to preserve their self-image. Their experimental design for the base game, replicated in this paper, successfully captured the belief change predicted in their first hypothesis, but the experimental results were weak in explaining the direction and size of the effect. Other experiments are carried out later in their paper to contribute to the robustness of their results, however these see various design changes and as such are not

considered in this paper.

Past literature (Hales, 1985) and recent literature (Gino, Francesca, Norton, Weber, 2016) pointed towards agents reconciling selfish decision making with moral

correctness and justifying their actions with their beliefs about their actions not being morally wrong. Measurements of fairness in experiments (Konow 2000) have yielded evidence of belief changes from negative self-serving biases of agents. Another issue coming from the literature is the direction of the bias, if it self-serving belief that justifies profit maximization or a ‘self-destructive’ belief that lends itself towards the agent sacrificing his own profit for fairness or morality (Kaplan and Ruffle 2007). This links back to Di Tella et al.’s (2015) interpretation over the variation in beliefs due to the cognitive dissonance. The measure of the bias holds; however, the direction of the bias is in favor of fairness which would indicate that the behavior of subjects is

self-destructing, not self-serving (Kaplan and Ruffle, 2007). Agents are therefore more likely to sacrifice profits and deviate from the strategic choice, to appear fairer, rather than justify increased profits from the strategic choice by adopting negative beliefs over others. Nevertheless, both papers find that there is relative cognitive bias arising between treatment groups and their results for the relative bias are consistent.

Kaplan & Ruffle (2007) find that the bias tends in the direction of fairness rather than selfishness. In particular, they use a guessing game experiment and find that agents are

more likely to have beliefs that their opponents are more fair, such that they act less selfishly, which again provides evidence for the opposite direction of the cognitive bias. Moreover, their experimental results show that players attempt to predict each other’s biases. If they believe that a selfish bias exists, then they will act accordingly. Therefore, while in general they find that the bias tends to the fairness direction, they do support Di Tella et al.’s (2015) conclusion that the belief of a selfish partner would lead to selfish acts from the agent. However, they reinforce the fact that it is hard to differentiate selfish belief and action originated from self-image manipulation from strategic behavior.

Nonetheless, due to most games not resulting in self-serving bias by agents (Kaplan & Ruffle, 2007), their literature tends to corroborate with the weak effects found in the base game of Di Tella et al. (2015) paper and Ging-Jehli, Nadja, Schenider and Weber’s (2019) replication and alternative interpretation of Di Tella’s et al. (2015) results.

Additional evidence that the interpretation of the results points to a different direction comes from Ging-Jehli, Nadja, Schenider and Weber’s (2019) paper, which replicates Di Tella et al.’s (2015) experiment with an additional ‘neutral observer group’. They find that while the relative bias in Di Tella et al.’s result is replicable; the absolute effect of the bias actually runs in the opposite direction that is predicted. Agents accurately

predict selfishness of their opponents when they can act selfishly, however become less accurate when this ability is restricted. Therefore, agents display a positive bias towards their partners when their choices are restricted, rather than a selfish one when they have fewer choice restrictions. They reconcile these results by arguing that Di Tella et al. (2015) measured a relative bias between groups, such that indeed strategic cynicism is found from one group to another.

A possible design change that can corroborate Ging-Jehli, Schenider and Weber’s (2019) replication results that the absolute bias effect runs in the opposite direction is to test a separate case when allocators are fully unrestricted, such that, according to Di Tella et al. (2015) they can fully express their potential selfishness. As such, a separate independent sample with an unrestricted version of the ‘corruption’

game which can compare with the sample from the corruption game can serve as robust evidence for both of their hypotheses. The unrestricted version can be defined as the same game as Di Tella et al. (2015) base corruption game, where all players assigned to the allocator roles can move all tokens. There is no information difference between sellers and allocators, and as such no cognitive bias should arise from

allocators in this scenario. All tokens would be colored ‘yellow’. The results would serve as an absolute baseline for selfish belief expression in the two treatment groups.

If beliefs show significant differences between the unrestricted treatment and the original treatment groups, then the results would add robustness to Di Tella et al. (2015) hypothesis that opportunity to be selfish affects beliefs about others. If beliefs in the unrestricted treatment and the Able=8 group do not strongly differ, or the former is less cynic than the latter then there is also evidence in favor of Ging-Jehli’s interpretation that the bias runs in the other direction. Overall, there should be almost no difference in allocator beliefs over the percentage of corrupt sellers in the unrestricted group

compared with the Able=8 one, however a large difference with the Able=2 group. The main issue is that the samples are independent, it is difficult to know if the observations in the samples have the same or similar preferences and views over self-image. If the samples are recruited from the same general population, then this should be less of an issue and the results will be consistent.

Further literature was also carried out in suggestions mentioned in Di Tella et al.’s (2015) discussion, which expands the scope of self-serving biases effects. Issues regarding memory and learning behavior for this game have been explored by Saucet & Villeval (2018). They use a dictator game setting to verify if motivated memory is used to maintain self-image and implications that it may have in choices. They find evidence that dictators manipulate their memory ex-post the games, however find no clear evidence of biased memory. Consequently, agents selectively recall and make recollection mistakes over amounts when they are selfish, however they do not manipulate their memory to make themselves seem altruistic when they were selfish.

Furthermore, the study focuses on short-term memory recollection, while in the long run the biases may disappear or reinforce themselves.

Another experiment that tests memory and learning effects on self-serving bias was carried out on Zimmerman’s paper (2020), where he tests for feedback effects on self-serving beliefs. He additionally contributes to the current literature by testing for selective recall effects in a longer period of one-month post experiment. The evidence found in the paper points towards stronger selective recall if subjects received negative feedback regarding their actions in the longer run, and further sometimes either

suppress or erase feedback that negatively impacts their self-image. Thus, the results of his paper validate the view that most likely cognitive dissonance for negative

self-service biases increases in longer periods of time, agents prefer to suppress undesirable views of their self-image.

To finalize, this paper contributes to the current literature by re-examining the ‘corruption game’ experiments carried out in Di Tella et al.’s (2015) paper. In this paper, I show that the relative bias of Di Tella et al. (2015) basic game experiment are

replicable. Additional non-parametric tests are carried out to provide robustness towards a cognitive bias. The findings give mixed evidence towards the bias being a selfish self-serving bias. While an effect on the change in beliefs of agents in different treatment groups can be found, this effects only supports the relative direction of the bias for the experimental treatment group and is not generalizable. Overall, I find evidence with the same data that agents display a relative self-serving bias, and their behavior does affect agent’s beliefs. Nonetheless, more recent literature continues to find mixed evidence for the direction of the effect for these beliefs. Further research is required to show that the ‘strategic cynicism’ hypothesized by Di Tella et al. (2015) does indeed hold in the direction they predict and is not, in-fact, a positive self-serving bias.

Appendix

Table 6:

OLS regression for the effect of Is Corrupt on the Proportion of Tokens Taken

(1) (2)

VARIABLES Is Corrupt Is Corrupt

Proportion of Tokens Taken 0.417 -0.235 (0.294) (0.248) Age -0.256* (0.102) Class -0.265 (0.178) Female 1.115** (0.277) Observations 9 9

Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1

Table 7:

(1) (2) (3) (4)

VARIABLES Is Corrupt Is Corrupt % Corrupt % Corrupt

Tokens Taken 0.0916*** 0.0932*** 0.0351* 0.0369* (0.0242) (0.0256) (0.0182) (0.0215) Age 0.0515 0.0116 (0.0338) (0.0284) Socioeconomic Class -0.00351 -0.0699 (0.114) (0.0953) Female 0.293** 0.123 (0.141) (0.118) Observations 30 30 30 30

Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1

Citations

Andreoni, J. & Bernheim, B. (2009). Social Image and the 50-50 Norm: A

Theoretical and Experimental Analysis of Audience Effects. Econometrica, 77(5), 1607-1636. doi: 10.3982/ecta7384

Benabou, R., & Tirole, J. (2002). Self-Confidence and Personal Motivation. The Quarterly Journal Of Economics, 117(3), 871-915. doi: 10.1162/003355302760193913

Bénabou, R., & Tirole, J. (2016). Mindful Economics: The Production,

Consumption, and Value of Beliefs. Journal Of Economic Perspectives, 30(3), 141-164. doi: 10.1257/jep.30.3.141

Berman, J., & Small, D. (2012). Self-Interest Without Selfishness. Psychological Science, 23(10), 1193-1199. doi: 10.1177/0956797612441222

Bersoff, D. (1999). Why Good People Sometimes Do Bad Things: Motivated Reasoning and Unethical Behavior. Personality And Social Psychology Bulletin, 25(1), 28-39. doi: 10.1177/0146167299025001003

Dana, J., Weber, R., & Kuang, J. (2006). Exploiting moral wiggle room:

experiments demonstrating an illusory preference for fairness. Economic Theory, 33(1), 67-80. doi: 10.1007/s00199-006-0153-z

Di Tella, Rafael, Ricardo Perez-Truglia, Andres Babino, and Mariano

Sigman. (2015). "Conveniently Upset: Avoiding Altruism by Distorting Beliefs about Others' Altruism." American Economic Review, 105 (11): 3416-42

Dušek, L., Ortmann, A., & Lízal, L. (2005). Understanding Corruption and Corruptibility Through Experiments. Prague Economic Papers, 14(2), 147-162. doi: 10.18267/j.pep.259

Ging-Ging-Jehli, N., Schneider, F., & Weber, R. (2020). On self-serving strategic beliefs. Games And Economic Behavior, 122, 341-353. doi: 10.1016/j.geb.2020.04.016

Gino, Francesca, Michael I. Norton, and Roberto A. Weber. (2016). "Motivated Bayesians: Feeling Moral While Acting Egoistically." Journal of Economic

Perspectives, 30 (3): 189-212.

Hales, S. (1985). The Inadvertent Rediscovery of Self in Social

Psychology. Journal For The Theory Of Social Behaviour, 15(3), 237-282. doi: 10.1111/j.1468-5914.1985.tb00056.x

Kaplan, T., & Ruffle, B. (2004). The Self-serving Bias and Beliefs about Rationality. Economic Inquiry, 42(2), 237-246. doi: 10.1093/ei/cbh057

Konow, James. (2000). "Fair Shares: Accountability and Cognitive Dissonance in Allocation Decisions." American Economic Review, 90 (4): 1072-1091.

McFadden, D. (1999). Econometrics 240B Second Half Reader. University of California Berkeley, 65-80.

Regner, T. (2017). Reciprocity under moral wiggle room: Is it a preference or a constraint?. Experimental Economics, 21(4), 779-792. doi: 10.1007/s10683-017-9551-2

Saucet, C., & Villeval, M. (2018). Motivated Memory in Dictator Games. SSRN Electronic Journal. doi: 10.2139/ssrn.3170866

van der Weele, J., Kulisa, J., Kosfeld, M., & Friebel, G. (2014). Resisting Moral Wiggle Room: How Robust Is Reciprocal Behavior?. American Economic Journal: Microeconomics, 6(3), 256-264. doi: 10.1257/mic.6.3.256

Zimmermann, Florian. (2020). "The Dynamics of Motivated Beliefs." American Economic Review, 110 (2): 337-61.