1
Introduction
Every year, vast amounts of money are spent on projects that aim to influence peoples’ actions. For example, companies invest a large part of their budget in advertisements in order to entice people to buy their products or services. Ac-cording to Guttmann (2019), more than 560 billion U.S. dollars were spent on advertisements in 2019. Similarly, politicians spend large sums of money on their political campaigns in order to motivate people to vote for them. Finally, gov-ernments and private institutions spend significant amounts of money in order to regulate and monitor peoples’ behavior. Psychologists and behavioural economists have discovered that there are numerous subtle cues that systematically influence people’s behaviour, in a subconscious level (Kahneman, 2011; Reynolds, 2006).
The term used in the academic literature to describe research on this specific topic is “priming” (Bargh, Chen, & Burrows, 1996; Bargh & Chartrand, 2000; Cohn & Mar´echal, 2016; Thaler & Sunstein, 2009). As Cohn and Mar´echal (2016) describe, priming is about using subtle cues in order to activate mental concepts which can then influence a subject’s behaviour. There are various types of prim-ing and numerous methods are beprim-ing used by researchers to investigate different channels through which it can operate (Bargh & Chartrand, 2000). To illustrate, it has been reported that reminding people of the elderly stereotype can influence their walking speed Bargh et al. (1996).
The purpose of this paper is to contribute to the academic literature about priming dishonesty, mathematical ability and social identity. In particular, the type of priming that is investigated falls under the category of mindset prim-ing (Bargh & Chartrand, 2000). As Bargh and Chartrand (2000) explain, mind-set/conscious priming manipulations are when the participant actively engages in a type of thinking (e.g. reading a short paragraph about a specific topic) which subconsciously influences the participant’s behaviour in a later, seemingly unre-lated, context. This type of priming differs from conceptual/implicit/subconscious priming which is usually done in a way that the participants are completely un-aware that the priming effect is taking place (Banaji, Hardin, & Rothman, 1993; Bargh & Chartrand, 2000; Welsh & Ord´o˜nez, 2014). Another contribution of this thesis is on the investigation of potential gender differences that arise from
con-sciously priming social identity in terms of gender. To the best of the author’s knowledge, no other studies have researched the effect of simultaneously priming dishonesty or math-ability with social-identity in terms of gender in a conscious manner.
An experiment was conducted where participants played two short games: a real-effort game and a random draw game. The first one was a variation of the matrix task which was originally created by Mazar, Amir, and Ariely (2008). The second one, which was relatively similar to the experimental design of Fischbacher and F¨ollmi-Heusi (2013), asked participants to generate a random number and self-report their answer on a piece of paper. Both games allowed for the possi-bility to cheat and incentives were provided to motivate participants to act in a dishonest manner. The combination of these two games was chosen because it allowed the investigation of people’s behaviour on a task that was based on cog-nitive/quantitative ability and on a task that was based purely on luck. Subjects were split in three groups: the control treatment, the corruption treatment and the mathematical ability treatment.
In the control treatment, participants were asked to read the abstract from an academic paper about gender differences in personality traits prior to playing the two games (Costa Jr, Terracciano, & McCrae, 2001). Similarly, in the corruption treatment, participants were asked to read the abstract from an academic paper about differences in corruption (Swamy, Knack, Lee, & Azfar, 2001). Finally, in the mathematical ability treatment, the abstract was from an academic paper about gender differences in mathematical ability (Benbow & Stanley, 1980). The choice of the academic papers was done such that the control treatment had no obvious connection with cheating or math performance. Also, the paper for the corruption treatment was chosen in order to remind participants that they can cheat. As for the math treatment, its paper was used to highlight the competitive nature of the first game (the matrix task). The way the experiment was designed allowed for the investigation of priming dishonesty, mathematical ability and stereotypes about social identity (gender).
The main research question is whether different types of mindset priming can influence people’ s behavior in terms of (dis)honesty and even performance in quantitative, skill-based tasks. Evidence of such effects has been found in several
influential papers (Afridi, Li, & Ren, 2015; Gino & Ariely, 2012; Levy, 1996; Shih, Pittinsky, & Ambady, 1999; Steele & Aronson, 1995; Steele, 1997). However, all of these papers have used implicit/subconscious priming techniques. The techniques used vary from asking participants to answer questions about their social identity (e.g. gender, nationality and age-group), to asking participants to form sentences using scrambled words related to a concept that the experimenter is trying to prime. It can be argued that compared to implicit priming, mindset priming is equally as important, or perhaps, even more important in everyday life. For example, when a voter reads an article about a company’s CO2 emissions, he/she
could be primed, in the form of mindset priming, to vote for a candidate who will combat climate change (e.g. Bernie Sanders). Similarly, when a person reads or listens to something about corruption, this could prime him/her to misreport his/her taxes. Such mechanisms, that operate through mindset priming, need to be investigated further as they can have significant policy implications.
The remainder of the paper is structured in the following way. Section 2 pro-vides a literature review and states the hypotheses of the paper. Section 3 describes the experimental design and Section 4 is about the methodology used to test the hypotheses. Further, the results are analysed in Section 5. In Section 6 there is a discussion of the results, and the limitations of the experiment are in Section 7. Finally, the last section concludes.
2
Literature Review
2.1
Subconscious thinking
It is a common misconception that people use their intuitive thinking for all ac-tions in their everyday life. However, Reynolds (2006) coined a theory which he called the “neurocognitive model of ethical decision-making”. According to this theory, there are two different mechanisms that influence ethical decision-making processes. First, there is the higher-order system of conscious reasoning, which he names the “C-system”. Second, there is another system, which he calls “X-system”, that is based on implicit association and a reflexive pattern-matching. Reynolds (2006) argues that both these systems are used in ethical decision
mak-ing, despite the popular opinion that conscious reasoning (C-system) is in charge of such decisions.
A very similar theory was developed by Kahneman (2011). The main differ-ences between the two theories are the names that are given to each of those mechanisms. Instead of “C-system” and “X-system”, he calls them “System 1” and “System 2” respectively. Priming can affect either, or both, of these systems depending on the type of priming that is being used. Some experiments, such as the experiment described in this paper, use conscious priming techniques, while others try to use subconscious priming techniques (Banaji et al., 1993; Bargh & Chartrand, 2000; Welsh & Ord´o˜nez, 2014).
2.2
Subconscious priming techniques
The most popular method of subconscious priming is by asking subjects to form sentences out of scrambled words. The priming effect takes place by having many words that are directly related to a concept that the experimenter wants to prime. For example, Banaji et al. (1993) run two experiments where they primed subjects about aggression and dependence in order to see whether it would influence their perception of other people. They primed subjects by having many scrambled words describing behaviour that is stereotypical of dependence or aggression, which the subjects needed to use to form sentences. They found that when subjects were asked to rate how aggressive or dependent people of the opposite sex were, primed subjects gave significantly higher ratings than the control treatment. Furthermore, the authors showed in another experiment that the results hold even when subjects have no explicit memory of the primes.
In a similar way, Welsh and Ord´o˜nez (2014) tested the difference between con-scious priming, subconcon-scious priming and no prime on (un)ethical behaviour. In the subconscious prime treatment, subjects were given five randomly positioned words and were asked to form four-word sentences. Eight out of the twenty sen-tences that had to be formed contained ethics-related words. This was in contrast to the control condition which contained no such words. Welsh and Ord´o˜nez (2014) found that participants from both subconscious and conscious priming manipula-tions did not demonstrate any significant differences in their (un)ethical behavior.
They also found that both priming treatments did exhibit statistically significant differences in behaviour from the participants who were not primed. Further de-tails about the conscious prime treatment will be given in the next subsection.
Several more striking findings have been discovered using the scrambled-words priming task. To illustrate, Bargh et al. (1996) find that implicitly priming people about rudeness made them act in a significantly more discourteous manner to the experimenter. This was measured by how quickly and frequently they interrupted the experimenter relative to the control treatment. Further, they also show that participants who were primed about the elderly stereotype walked at a signifi-cantly lower pace than those who were not primed. In addition, Srull and Wyer (1979) find that priming hostility and kindness using the scrambled-words task significantly impacted subsequent behaviour of participants. Also, it is important to note that the authors found that both the number of times the test concept had previously been triggered, and the time interval between the scrambled-words task and the task used to measure their actions, had a significant effect on the results. Despite the popularity of the scrambled-words priming task, academics have used other cunning methods to investigate the effect of subconscious priming. For instance, Bargh et al. (1996) primed subjects about aggression by flashing a sub-liminal picture of an African-American male face, which is said to be a stereotype of aggressive behaviour. They found that subjects who were exposed to pictures of an African-American male were significantly more aggressive than those who were exposed to pictures of a Caucasian male face.
Moreover, several studies have shown that priming someone’s social identity can influence his/her subsequent behaviour. For example, Afridi et al. (2015) primed participants about their social identity by asking them to fill-in a pre-experimental questionnaire which had a question about their hukou status. Hukou is China’s household registration system and it favours urban citizens, while simultaneously discriminating against rural residents in the allocation of resources. They find that priming the social identity of rural immigrants significantly reduces the per-formance relative to their local urban counterparts on an incentivized cognitive task. A similar method was used by Shih et al. (1999) in a very influential pa-per they wrote. In one of their treatment groups, they primed participants about their ethnic social-identity by asking them to indicate their ethnicity and to answer
questions related to their ethnic identity. In an analogous manner, they primed subjects about their gender social-identity by asking them gender-related ques-tions about themselves. They found that these implicit priming techniques had a significant effect on the performance of participants on cognitive tasks which were consistent with popular stereotypes about each social-identity. A more extensive discussion about the effects of priming social-identity will be given in an upcoming subsection.
2.3
Conscious priming techniques
It is clear that conscious priming techniques do not attract nearly as much in-terest in the academic literature as do unconscious priming techniques. Despite that, it is still a very important topic, with many real-life applications, that needs further investigation. Conscious or mindset priming manipulations are when the participant actively engages in a type of thinking (e.g. reading a short paragraph about a specific topic) which subconsciously influences the participant’s behaviour in a later, seemingly unrelated, context (Bargh & Chartrand, 2000). It is cru-cial to note that this does not mean that the participant is aware that he/she is being primed. The difference with subconscious priming is that in the latter, participants are completely unaware that they have been primed. However, in the former, participants can usually guess that the manipulation might have subcon-sciously influenced their subsequent behaviour.
There is a wide array of conscious priming techniques employed by researchers. First, Cohn, Engelmann, Fehr, and Mar´echal (2015) prime financial professionals by showing them scenarios of either financial booms or busts. They find that this manipulation had a significant impact on their willingness to take risk in a subse-quent task. Second, Fischer, Ferreira, Milfont, and Pilati (2014) primed Brazilian participants with corruption-related images and found that this increased their corruption intentions, compared with a control group. However, the method used in their paper to measure corruption intention is slightly problematic in terms of incentivizing people to give authentic responses. Participants were asked to indi-cate (on a scale of 0-10) to what extent they would behave like corrupt individuals in different scenarios. Evidently, this experimental design does not incentivize
subjects to give honest responses. Even though the experiment was anonymous, subjects could not earn anything by saying that they would be very likely to act like a corrupt individual in a certain scenario. On the other hand, by doing so, they would tarnish their perception of their self-image, something that is very impor-tant to people (Dana, Weber, & Kuang, 2007; Fischbacher & F¨ollmi-Heusi, 2013; Mazar et al., 2008; Pruckner & Sausgruber, 2013; Shalvi, Handgraaf, & De Dreu, 2011).
Moreover, Mazar et al. (2008) design a guileful method of conscious priming for (dis)honesty. They asked one group of students to recall the Ten Command-ments and then play the matrix task, where they had the opportunity to cheat. They compared their responses to those from another group of students who was asked to recall ten books they had recently read, instead of the Ten Command-ments. They found that those who were asked to recall the Ten Commandments exhibited a significantly more honest behaviour. Furthermore, Mazar and Zhong (2010) conducted a noteworthy experiment that show that participants who were exposed to environmentally friendly, “green”, products, demonstrated more pro-social behaviour, compared to those who were exposed to conventional products. Mazar and Zhong (2010) suggest that the exposure to “green” products triggered norms related to social responsibility and ethics that influenced future behaviour. Finally, as mentioned in the previous subsection, Welsh and Ord´o˜nez (2014) tested the difference between conscious priming, subconscious priming and no prime on (un)ethical behaviour. To test that, they added another element to the scrambled-words task, which was part of the unconscious priming treatment, in order to transform it to a conscious priming manipulation. Specifically, after finishing the scrambled-word task, subjects were asked to “think carefully about a time when you faced a decision in which you focused your attention on your moral standards (e.g., moral, religious, or personal values).”. The authors found that subjects from both priming treatments exhibited reduced unethical behaviour. However, they found no significant difference between the behaviour of subjects in the conscious and subconscious behaviour. This result could imply that the results of the experiment described in this thesis could externalise to subconscious primes as well.
partici-pants from the unconscious priming treatment were aware of the priming effect by giving them a specially designed questionnaire. Bargh and Chartrand (2000) orig-inally designed this questionnaire for exactly this purpose. It included questions in the form of “Did you notice any particular pattern or theme to the words that were included in the scrambled sentence task?” and “What do you think the true purpose of this experiment was?”. Welsh and Ord´o˜nez (2014) rightfully excluded participants who indicated signs of awareness. However, this method is not al-ways used by researchers which can raise doubts about the results of subconscious priming experiments.
2.4
Dishonesty
(Dis)honesty is one of the topics that has received a lot of attention in the literature about priming (Ariely, 2012; Gino & Ariely, 2012; Hill & Kochendorfer, 1969; Welsh & Ord´o˜nez, 2014). In this paper, one of the manipulations that were used in the experiment were aimed at priming participants about (dis)honesty. More details about the experimental design are available in the next section. Before that, it is important to define what is meant by (dis)honest behaviour in a context that is applicable to this paper.
According to DePaulo, Kashy, Kirkendol, Wyer, and Epstein (1996), a lie hap-pens when you purposefully attempt to mislead someone. It is necessary that both the intention to deceive, and the deception itself must occur. For example, in the experimental setting of the matrix task, a participant has committed a lie if he has purposefully misreported the number of matrices he/she solved (Mazar et al., 2008). The results from Mazar et al. (2008) show, that when participants are given the ability to cheat in the matrix task, by checking the test themselves and then shredding their answers, they overreport the number of matrices solved by 10% -in the absence of any prim-ing effects. Also, DePaulo et al. (1996), f-ind that college students lie in approximately 33% of their social interactions. This finding further proves the importance of this topic in real life.
Hypothesis 1: Priming individuals about dishonesty in a conscious way will make them act in a more dishonest manner.
conscious priming manipulation about dishonesty (by reading an abstract from a paper about gender differences in terms of corruption), would report higher answers in both games that were subsequently played. The first game requires basic quantitative skills, while the second game relies only on luck. This hypothesis is in line with the findings of Fischer et al. (2014). However, in the current experimental setting, participants are incentivised to act in a dishonest manner which is a more realistic scenario.
It should be noted that it is not clear whether reminding people of dishonest behaviour will lead to an increase or a decrease of cheating. Welsh and Ord´o˜nez (2014) find evidence that seems contradictory to Hypothesis 1. The authors hy-pothesised that consciously priming individuals about moral standards, by asking them to think about a situation when they faced a decision that required concen-tration on moral standards, would entice them to act in a more ethical manner. The intuition behind their hypothesis is that when people are more attentive to their moral standards, they are both more likely to categorize morally ambiguous situations as unethical, and less likely to overlook the ethical aspect of a decision. There are some reasons that can explain the difference in these two hypotheses. First, Welsh and Ord´o˜nez (2014) measure ethical behaviour in terms of responses given by participants to abstract questions. In their experimental setting, there is no room for cheating, which is a very important kind of unethical behaviour. Secondly, Welsh and Ord´o˜nez (2014) do not incentivize their participants to cheat or act in a dishonest manner as their focus and experimental setting are aimed at a different kind of unethical behaviour. It can be argued that cheating and dishon-esty are the most important kinds of unethical behaviour because they can have severe real-life implications. In the experimental design of this paper, participants are incentivized to act in a dishonest way which reflects a more real-life setting.
2.5
Social identity and Stereotypes
Another goal of this paper is to investigate the effect of priming social identity on behaviour. Social identity is a person’s perception of who they are, based on their membership on a group. For instance, a person who is 70 years old might self-identify as “old” just because he belongs in the age-category of elderly,
despite not feeling old. Several researchers have investigated the impact of making someone’s social identity salient, to subsequent performance or behaviour (Afridi et al., 2015; Benjamin, Choi, & Strickland, 2010; Cohn, Mar´echal, & Noll, 2015; Hoff & Pandey, 2014).
For example, Cohn, Mar´echal, and Noll (2015) find that when prisoners were primed with their criminal identity, they were much more likely to cheat on a subsequent task than prisoners who were in the control group. Also, Hoff and Pandey (2014) show that Indian students who were primed about their low so-cial status performed significantly worse in a latter incentivized cognitive task. Furthermore, Benjamin et al. (2010) find that priming someone’s ethnic identity influences his/her patience and risk aversion. Finally, as it was mentioned previ-ously, Afridi et al. (2015) demonstrate that priming the institutionally imposed social identity of an individual can have a significant impact on his/her cognitive performance.
Despite the importance of all these findings, other researchers have focused on the effect of priming stereotypes about someone’s social identity on subsequent behaviour (Levy, 1996; Shih et al., 1999; Steele & Aronson, 1995; Steele, 1997). As defined by Hilton and Von Hippel (1996), stereotypes are “beliefs about the characteristics, attributes, and behaviours of members of certain groups. More than just beliefs about groups, they are also theories about how and why certain attributes go together.”. The authors explain that there are two main routes that lead to the formation of stereotypes. First, stereotypes can be created because of real differences between groups (e.g. food preferences between people of different cultures). Second, another route is when stereotypes are formed independent of any real group differences. This usually happens when spurious and inconsistent differences between groups are observed, which for some reason gain rapid popular-ity. These inconsistent stereotypes can then be entrenched and gain some validity through the mechanism of self-fulfilling prophecies (Hilton & Von Hippel, 1996; Levy, 1996; Steele & Aronson, 1995; Steele, 1997).
The second mechanism responsible for the formation of stereotypes is very rele-vant for this paper because it is posited that priming individuals about a stereotype of their social identity (gender), will influence their subsequent behaviour. In par-ticular, it is hypothesised that males who are asked to read an abstract from a
popular paper, saying that males are more corrupt than females, will cheat more in both games. This result would be in-line with the effect of priming social identity (gender) and the theory of self-fulfilling prophecies.
Hypothesis 2: Priming individuals about a stereotype that is applicable to their social identity (gender) would influence their subsequent behaviour, even if it is not necessarily believed by them.
Steele and Aronson (1995) assert that the threat of possibly being judged stereotypically can affect anyone, even if the stereotype is not necessarily believed by the person. The authors explain that for the self-fulfilling prophecy to occur, it is a sufficient that the person is simply aware of the existence of the stereotype. A striking result that they find is that black participants, who were primed about their race, exhibited impaired cognitive performance, even when it was made clear that the task they completed was not indicative of intellectual ability. The authors argue that this result arises from the stereotype that African-Americans are poor students, compared to white students (Steele & Aronson, 1995). The intuition is that participants primed about their social identity (race), would subsequently be vulnerable to stereotype-threat which can influence their performance.
This is exactly what is expected in the experimental design of this paper. The main difference is that participants need not necessarily be aware of the stereotype before attending the experiment. Reading the stereotype that males are more corrupt than females would make participants subject to stereotype-threat and can in turn influence their behaviour. Similarly, reading the stereotype that males have better mathematical ability than females could also lead to males performing better in a task requiring basic quantitative skills than females. Shih et al. (1999) wrote a prominent paper, which was mentioned earlier in this thesis, and find evidence in favour of a similar hypothesis. They found that, in-line with the common cultural stereotype, Asian-American women performed better on a mathematics test after being primed about their ethnic identity than Asian-American women who were not primed about their ethnicity. However, they also found that when they were primed about their gender identity, instead of their ethnicity, they performed worse when compared with a control group (Shih et al., 1999).
fake-news, about a certain group (gender, ethnicity, age, etc.) could influence the group’s behaviour. This in turn could intensify problems that are prevalent throughout the world, such as gender inequality.
2.6
Why do people cheat?
Having briefly described what is meant by dishonesty and social identity, the nat-ural next step would be to provide a theory that can explain people’s (dis)honest behaviour. Becker (1968) developed an economic model where self-interested, per-fectly rational individuals choose their actions based on an external cost-benefit analysis that considers the rewards of dishonest behaviour, the probability of being caught and the consequences of punishment. A similar model was developed by Allingham and Sandmo (1972). Their economic model is tailored to explain tax-evasion, which is another form of dishonesty. Economic models similar to the two described above are typically given the name “Simple Model of Rational Crime” or “SMORC” (Ariely, 2012). Despite the assumption of self-interested and perfectly rational individuals, this model has been supported by some empirical evidence (Hill & Kochendorfer, 1969). The authors conducted an experiment with 60 sixth grade boys and their results were in line with the predictions made by a simple cost-benefit model, similar to the one developed by Becker (1968).
However, the consensus amongst more recent literature lies in two alternative theories. First, Ariely (2012) coined the “Fudge-factor” theory which basically argues that people have a desire to acquire the rewards of cheating, while simul-taneously want to view themselves as honest individuals. People manage to fulfil both desires simultaneously by acting in a dishonest way to a small extent, which they can justify to themselves. This way, they are able to accrue some benefits from their dishonest behaviour, and retain a positive image about themselves.
Second, Mazar et al. (2008) coined the “self-concept maintenance theory” which can provide very similar predictions to the “Fudge-factor” theory. This theory ar-gues that people balance many objectives when making ethical judgments. Specif-ically, people desire to maximize the benefits that can be obtained in a certain scenario, while maintaining a positive self-concept.
example. According to the “Fudge-factor” theory, a person who cheated in the matrix task would try to convince herself that she honestly didn’t realise she mis-reported the number of matrices solved. However, according to the “self-concept maintenance theory”, that same person would instead try to convince herself that misreporting was a fair action because she solved less matrices than she could have solved because of a certain excuse (e.g. someone was making noise and she couldn’t concentrate).
Furthermore, Mazar et al. (2008) run six large experiments which confirmed their theory that individuals lie to some degree to increase their profit, while still being able to maintain a positive self-concept as honest individuals. Across their six experiments, only 0.6% of the 791 participants cheated by the maximal amount. The majority of the participants cheated to a limited degree, which presumably allowed them to maintain their positive self-concept.
To make matters more complicated, the game-theoretical literature that fo-cuses on dishonesty has developed a mechanism that is similar to the self-concept maintenance mechanism. “Perceived cheating aversion” is the name Dufwenberg and Dufwenberg (2018) have given to this mechanism. In simple terms, perceived cheating aversion is when an individual gets a disutility because others believe that the individual is acting in a dishonest manner. For example, participants might prefer to act in an honest manner and avoid cheating, which would give them the possibility to earn higher monetary rewards, in order to ensure that others will not infer that they are dishonest. Similarly, Geanakoplos, Pearce, and Stacchetti (1989) and Battigalli and Dufwenberg (2009) have developed game-theoretical models that allow for the influence of such belief-dependent emotions. In addition, Fischbacher and F¨ollmi-Heusi (2013) find support for the self-concept maintenance theory, although their results are slightly different from Mazar et al. (2008). Fischbacher and F¨ollmi-Heusi (2013) asked 746 participants to play an anonymous die roll experiment and found that 20% of inexperienced subjects (who had never played the die roll game before) lie to the fullest extent. Given that the probability of getting caught was zero, this behaviour is compatible with the rationality assumption made by economists. On the contrary, 39% of the participants were fully honest, and a very large share of participants were partial liars. Homogeneous findings were discovered in a field experiment run by Pruckner
and Sausgruber (2013) in Australia. The authors placed boxes of newspapers in public places where people could take a newspaper and pay for it by themselves. Even though there was purposefully no one checking whether people paid the cor-rect amount for the newspaper, the cost of a newspaper was made explicitly clear to anyone taking one. The authors found that two-thirds of customers did not pay for the newspaper. Harmoniously with the self-concept maintenance theory and the fudge-factor theory, from those who did pay something for the newspaper, the majority deposited much less than the indicated price.
Furthermore, Shalvi, Dana, Handgraaf, and De Dreu (2011) shows that individ-uals derive value from self-justification which allows them to lie for money, while at the same time retaining an honest self-concept. The authors conducted a lab experiment where participants were asked to report the outcome of a die role done privately and obtain monetary awards according to the number reported. When participants were allowed to roll a die three times, but were explicitly instructed to only report the first roll, they found that the reported number is significantly higher than when participants were allowed to privately roll only one die. Eliminat-ing the ability to observe more than one roll reduces lyEliminat-ing, which can be explained because people find it harder to self-justify dishonesty. One notable result was that people avoided major lies, even when the chance of getting caught was essen-tially zero (Shalvi, Dana, et al., 2011). In another paper, Shalvi, Handgraaf, and De Dreu (2011) find support for the self-concept maintenance theory by running an experiment where participants had to anonymously report a die roll and were awarded according to the number reported. The authors find that even though people avoid lying to the maximum possible degree, they do lie to intermediate levels that imply a substantial increase in payoff. Their results indicate that lying is psychologically costly. To describe this behaviour, the authors use the term “ethical manoeuvring”. In short, this term is about the inclination of people to compromise between honesty and morality, on the one side, and self-interest on the other (Shalvi, Dana, et al., 2011).
In contrast, in a more recent paper, Gneezy, Kajackaite, and Sobel (2018) find that the largest fraction of lies came from people who lied to the maximum degree, rather than partial liars. Even though they find a significant number of partial-liars, their results indicate that partial lying arises primarily because of social
identity concerns, rather than the self-concept maintenance mechanism. In partic-ular, partial lying was significantly higher in a treatment where the experimenter could not observe the actual outcomes of participants (this treatment allowed participants to lie without influencing the way their social identity towards the experimenter), compared to a treatment where the outcomes were observable. If the self-concept maintenance mechanism played a dominant role, lying behaviour should have been the same in both treatments.
To test whether individuals avoid major lies because of intrinsic motivation or instead shy away from bigger (unjustified) rewards due to the possibility of drawing attention and suspicion, Hilbig and Hessler (2013) conducted an experiment with 765 participants. By creating a clever variation of the die-under-the-cup game, they find that dishonesty is a decreasing function of the distance between the lie needed to obtain higher rewards, and the actual truth. Therefore, their results show that people do indeed avoid major lies not because of fears of raising suspicion but instead because of intrinsic motivation.
Also, Gneezy (2005) provides evidence against the Simple Model of Rational Crime (SMORC) by showing that the consequences of lying matter to the person who decides whether or not to commit a lie. When imposing an externality on another subject, participants who could lie, did so at a smaller degree than in the scenario without the externality. This finding contradicts the mainstream economic assumption of self-interested individuals.
Finally, the impact of money on behaviour has been investigated by several re-searchers and the results seem to support the self-concept maintenance theory, the fudge-factor theory and the ethical manoeuvring theory (Ariely, 2012; Hsee, Yu, Zhang, & Zhang, 2003; Mazar et al., 2008). In his book, Ariely (2012) describes a variation of the matrix experiment where in one of the treatments, participants were rewarded in the form of tokens, which they could later swap for real dollars. The results showed that people cheated significantly more in the token condition. This implies that people are more dishonest with the non-monetary rewards. The intuition could be that people find it more difficult to justify dishonesty to them-selves and maintain a positive self-concept when it is tied to real money.
When it comes to gender differences in terms of dishonesty, Dreber and Johan-nesson (2008) find significant variation between men and women. In particular,
in their experiment, where 156 people participated, males lied more than females for monetary benefits. In contrast, some psychological studies find that women lie more than men (DePaulo et al., 1996; Tyler, Feldman, & Reichert, 2006). Dreber and Johannesson (2008) explain that the reason why their results differ from the psychological literature is because their experimental setting allowed participants to remain fully anonymous, whereas in the aforementioned studies there was non-anonymous interaction amongst participants. The lack of anonymity introduces the potential that reputation creation influences people’s behaviour. Given that the experimental design for this thesis ensured anonymity amongst participants, the results of Dreber and Johannesson (2008) can explain any potential gender differences in dishonesty.
In summary, simple cost-benefit models appear to be ineffective when it comes to predicting cheating behaviour and dishonesty. The “fudge-factor” theory, the “self-concept maintenance” theory and the “ethical manoeuvring” theory are amongst the most popular theories that help predict dishonest behaviour by taking into ac-count findings from behavioural economics and psychology. Some of the most important and striking empirical findings that can be explained by these theo-ries is that people avoid cheating to the maximum and that they regularly act in an honest manner towards others, even in scenarios that guarantee complete anonymity.
2.7
Priming cognitive ability
As mentioned in the previous subsection, priming social identity has been shown to have an impact on cognitive ability (Afridi et al., 2015; Hoff & Pandey, 2014; Jamieson & Harkins, 2012; Levy, 1996; Steele & Aronson, 1995). To the best of the author’s knowledge, there has been no research on the effect of priming cognitive ability on cognitive ability. For instance, there have been no experiments asking participants about their IQ, which would be a prime of cognitive ability, and then asking them to perform a task which requires intellectual strain.
Another contention of this paper is that priming individuals about mathemat-ical ability would improve their subsequent performance on a task that requires basic quantitative skills. This is tested by asking participants to read an abstract
from a popular scientific paper about gender difference in mathematical ability and then measuring their performance in the matrix task (Mazar et al., 2008). Hypothesis 3: Priming individuals about mathematical ability would lead to im-proved performance on a subsequent task that requires basic quantitative skills.
The intuition behind the aforementioned hypothesis is that participants would take the task more seriously and perhaps become more competitive against them-selves. Also, another explanation could be that the improved performance arises because of dishonesty. Perhaps participants feel pressured to exhibit good quanti-tative skills and resort to over reporting the number of matrices solved. However, it is also possible that the effect runs in the opposite direction in case the majority of participants believe that their quantitative skills are lacking. Instead of boosting their performance, the prime could lead to a self-fulfilling prophecy where partic-ipants assume that because they have poor quantitative skills, they will perform poorly in the matrix task. The mere existence of such a thought could lead to the actualization of such a scenario. (Hilton & Von Hippel, 1996; Levy, 1996; Steele & Aronson, 1995; Steele, 1997). Jamieson and Harkins (2012) find that in the case of females, the negative effect on cognitive performance caused by the unfavourable math stereotype prime arose because of withdrawn effort.
Given that the matrix task requires very basic quantitative skills, it is expected that the prime will actually improve performance. A negative effect would most likely be observed in case the task required mathematical knowledge that the subjects are not very familiar with.
In a very influential paper, Gneezy, Niederle, and Rustichini (2003) conduct a lab experiment where they show that competition increases the performance of males, while females’ performance remained the same. In an extension of their previous paper, Gneezy and Rustichini (2004) show that their results are robust even in a field study setting, instead of solely in the laboratory. Further, they show that the gender differences in the effect of competition on performance comes from a very young age as the average age of participants in their later paper was less than 10 - as opposed to 23 in the former paper. In the experimental design of this thesis, there is no direct form of competition which means that the results of the two aforementioned papers cannot generalise to this thesis. In addition,
scholars have found evidence in favour of the hypothesis that males tend to be more competitive than females (Andersen, Ertac, Gneezy, List, & Maximiano, 2013; Buser, Niederle, & Oosterbeek, 2014). This difference in competitiveness can lead to a corresponding variation in effort because the more competitive a person is, the more effort he/she will exert. Thus, these findings can help explain any potential gender differences in the outcomes of Game 1.
3
Experimental Design
The experiment consisted of three treatments: control treatment, corruption treat-ment and math treattreat-ment. In all three treattreat-ments, subjects were given a leaflet consisting of four single-sided pages (The leaflets can be found in the Appendix). Subjects were told that they were going to take part in an experiment about be-havioural economics which needed to be run as part of the experimenter’s thesis. They were not given any details about the goal of the experiment or about what was being investigated in order to avoid bias. In the first page (see page 45), subjects had to specify their age, gender and nationality.
The only difference between the three treatments was in the second page of the leaflet (see page 46,49 and 50). Subjects from the control treatment were asked to read an abstract from an influential paper about gender differences in terms of personality traits (Costa Jr et al., 2001). This paper was chosen because it does not say or imply anything about gender differences in terms of dishonesty or math-ability. In contrast, subjects from the corruption treatment were asked to read an abstract from a famous paper about gender differences in terms of corruption (Swamy et al., 2001). This way, they were consciously primed about dishonesty, which could affect their subsequent behaviour. Also, they were consciously primed about stereotypes related to their social-identity (gender) which could lead to self-fulfilling prophecies. By reading the abstract, male participants saw that there is evidence that males are generally more corrupt than females. The realisation that there is evidence for such a stereotype could then incite them to act in a more dishonest manner, when given the opportunity. Finally, people in the math-ability treatment were asked to read an abstract from a popular paper about gender differences in terms of mathematical ability (Benbow & Stanley, 1980). As it can
be seen in the Appendix, all three of the abstracts showed the number of times that the paper had been cited. This was done to ensure that participants found the papers legitimate and authentic. All three of the papers had a large amount of citations and differences between them are not likely to play a role.
The next part of the experiment begun at the end of a statistics class which was approximately 45 minutes after participants read the abstract from the leaflet they were given. The procedure was the same across all three treatments. Participants were asked to turn to the third page of their leaflet, where they were given instruc-tions about the matrix task (Mazar et al., 2008). In the matrix task, participants are given a set of four-by-four matrices that contain a number between zero and 10 in each cell (the sheet of paper with the matrices is available in the Appendix). In each matrix, there are exactly two numbers that when added together, sum-up to 10. An example of one such solved matrix can be seen in Figure 1. The numbers 2.22 and 7.78 have been made in bold to show that they are the answer to the matrix.
Figure 1: Solved example from the matrix task
Participants were first given verbal instructions of the matrix task and were then given time to read the instructions by themselves. They were told to raise their hand if they had a question. All questions were answered privately to ensure that a single participant could not bias the entire group with his/her question. For example, if a participant asked how the experimenter would be able to detect cheating, this would bias the entire group of participants because they would infer that cheating is socially acceptable.
the participant had to fill-in how many matrices he/she solved. They were told that at the end of the task, they would be shown the solutions of each matrix and they would then have to self-report how many matrices they managed to solve. Participants were also told that they were not allowed to write anything on the sheet of paper that had the matrices. These three measures were taken to enable them to infer that they would be able to act in a dishonest manner and over-report the number of matrices they solved.
Dishonest behaviour in the form of over-reporting the number of matrices solved was incentivized with monetary rewards. The instructions stated that one person from the group would be randomly selected and would receive money according to how many matrices he/she solved. In particular, they were told that the selected person would receive payment “equal to 0.50e multiplied by the answer you gave in the field “How many matrices did you solve?””. The purpose was for the par-ticipants to understand that they were able to cheat, without making it explicitly clear.
After everyone had read the instructions and there were no more questions, subjects received a new one-sided piece of paper with 12 matrices (see page 51). They were told not to turn the sheet of paper the other way around to ensure that everyone had the same amount of time in the matrix task. They were then given three minutes to solve as many matrices as they could. A stopwatch with a 3-minute countdown was displayed on a large screen which allowed participants to manage their own time however they wanted. When the time was up, everyone was asked to stop working on the task.
As promised, the solutions of each matrix were displayed on a large screen and participants were given a minute to check their answers and report how many matrices they solved by filling in an empty field that was on the third page of the leaflet.
After everyone was done with the matrix task, they were asked to turn to the fourth page of the leaflet to play the second game. This game asked participants to use their phone and go to a website where they could generate a random number between zero and 100. They were requested to generate a number once, and to then write it down in the space provided.
dishonest behaviour. The participants could easily infer that the experimenter would not be able to detect if the number they generated was different from the number they reported. Dishonest behaviour, in the form of reporting a higher number than was generated, was incentivised using non-monetary rewards. In particular, participants were told that they would be given chocolates according to the number they reported that they generated, divided by 10. Also, non-integer numbers would be rounded down. For example, a person who generated the number 38 would receive three chocolates (because 3.8 rounds down to three). This measure was taken in order to encourage dishonest behaviour further.
After everyone was done with the second task, two people were selected ran-domly and were rewarded as promised: the first person received money for his/her performance on the first task, while the second person received chocolates for his/her performance on the second task. In the end, participants were given a brief description about what was being investigated and the goal of the experi-ment.
4
Methodology
In order to test this paper’s hypotheses, first-year students from the faculty of Economics and Business Economics of the University of Amsterdam were asked to participate in the experiment. It was conducted in four classes where students signed up to follow a free Statistics training. They were organized by a company offering lessons to prepare Bachelor’s students for their midterm and endterm exams. Three of the four classes were each assigned randomly into one of the three treatments (control, corruption and math ability). Classes differed only in terms of the time when they took place, ranging from 10:00-11:30 until 16:00-17:30. To make each treatment have approximately the same number of participants, a fourth class was assigned the math-ability treatment in a non-random manner.
In total 53 students participated: 16 in the Control treatment, 20 in the Cor-ruption treatment and 17 in the Math ability treatment. Participants had to report their gender, age and nationality in the first page of the leaflet they re-ceived. Information about participants’ gender identity was necessary in order to test the two hypotheses that had been made ex-ante. However, information about
age was asked in order to check whether there were any significant differences in age between groups. Although it is not very likely, differences in age could per-haps explain differences in performance in the matrix task, which requires basic quantitative skills. Amongst the three treatments, the minimum average age was 18.8 (Control treatment), while the maximum average age was 19.9 (Corruption treatment). The difference between them is not significant at a 5% significant level (p-value=0.053).
Overall, there were 29 female participants and 24 male participants. There was an equal representation of males and females across treatments. The proportion of females in each of the 3 treatments ranged from 53% to 56% and the difference is not significant (p-value=0.57).
The methodology used to test Hypothesis 1 was by means of a difference in the number of reported matrices solved. Given that all participants were first-year students, with relatively similar demographic characteristics, it would be expected that in the absence of the priming effect, the average reported number of matrices solved would be the same across all three treatments. If the average reported number of matrices solved in the corruption treatment (XCorruption), is higher than
the control treatment (XControl), it would be evidence in favour of Hypothesis 1.
Intuitively, it would imply that priming dishonesty enticed participants to act in a more dishonest manner.
Similarly, for Game 2, it would be expected that in the absence of priming, the average reported number would be the same across all three treatments. That’s because participants were simply asked to generate a random number between zero and 100, and report it. The expected value, in the absence of dishonest behaviour would be 50. Because it was incentivized to over-report, it is very likely that participants, even in the control treatment, would cheat and report a higher number than what they actually generated. This would mean that the average reported number is probably going to be much higher than 50. Comparison between the average number reported in the control and the corruption treatment is needed to test Hypothesis 1. If the average number reported in the corruption treatment (YCorruption) is different than in the control treatment (YControl), it would
HYPOTHESIS 1: The average reported number of matrices solved in Game 1 and the average reported number in Game 2 is different in the Corruption treatment than in the Control treatment.
Therefore, the statistical setting for Hypothesis 1 is: For Game 1: H0 : XCorruption− XControl = 0 H1 : XCorruption− XControl 6= 0 For Game 2: H0 : YCorruption− YControl = 0 H1 : YCorruption− YControl 6= 0
With reference to Hypothesis 3, it was contented that participants that had been primed about mathematical ability would perform better in the matrix task than those who had not been primed - or at least report that they performed better. If the average reported number of matrices solved in the math treatment
(XM ath), is higher than the control treatment (XControl), it would be evidence in
favour of Hypothesis 3. However, it could also be the case that people from the math treatment were more dishonest than those from the control. To ensure that this did not happen, it should hold that from Game 2, the average number reported in the math treatment (YM ath is the same as in the control treatment (YControl).
Of course, even if this second condition holds, the possibility that the improved performance in Game 1 arose because of dishonesty cannot be disregarded com-pletely. Participants might feel increased pressure to lie about their quantitative skills because of the math prime.
HYPOTHESIS 3: The average reported number of matrices solved in Game 1 will be higher in the Math-Ability treatment than in the Control treatment, while the average reported number in Game 2 will be the same between the two treatments. Therefore, the statistical setting for Hypothesis 3 is:
For Game 1:
H0 : XM ath− XControl = 0
For Game 2:
H0 : YM ath− YControl = 0
H1 : YM ath− YControl > 0
Hypothesis 2 stated that consciously priming individuals about a stereotype that is applicable to their social identity (gender) would influence their subsequent behaviour, even if it is not necessarily believed by them. There are many ways to test this hypothesis.
First, assuming that both genders would perform equally well in the absence of priming, if the average number of matrices solved for the males in the cor-ruption treatment (XCorruption,M ale) is higher than the females in the corruption
treatment (XCorruption,F emale), this could be evidence in favour of Hypothesis 2.
That’s because male participants in the corruption treatment read an abstract from a paper that basically argued that males are more corrupt than females. They were effectively primed about a stereotype of their social identity. Similarly, if the average number reported in the corruption treatment by male participants
(YCorruption,M ale) is higher than in the female participants in the corruption
treat-ment (YCorruption,F emale), it could be evidence in favour of Hypothesis 2.
Unfortu-nately, the experimental design does not allow for a clear conclusion to be made on what drives the effect of priming. The effect could be driven either by the dishon-esty prime, or by the social-identity prime. However, if YCorruption was higher than
YControl for both genders, it would be some indication of dishonesty prime, because
the higher outcomes of females should not have been a result of the social-identity prime.
Second, if the average number of matrices solved by males in the math treat-ment (XM ath,M ale) is higher than females in the math treatment (XM ath,F emale),
this could be evidence in favour of Hypothesis 2. That’s because male participants in the math treatment read an abstract from a paper that argued that males have superior mathematical ability than females. They were thus consciously primed about a stereotype of their social identity which could theoretically improve their performance on the matrix task. To say that it was the improved performance that caused the increase in the average number of matrices solved, instead of more dishonest behaviour, it should also hold that the average number reported in the
math treatment by male participants (YM ath,M ale) is the same as that of female
par-ticipants from the math treatment (YM ath,F emale). However, even if such a trend is
observed, it is also possible that the results did not arise because of improved per-formance, as predicted by Hypothesis 2. It could be the case that males felt more pressure to lie about their performance in Game 1 and therefore over-reported the number of matrices solved more than females from the same treatment group. The experimental design cannot specify which of the two mechanisms actually occur. HYPOTHESIS 2A: The average reported number of matrices solved in Game 1 and the average reported number in the subgroup of male participants from the Corruption treatment in higher than in the subgroup of female participants from the Corruption treatment. At the same time, no such differences are observed in the Control treatment between the two subgroups.
HYPOTHESIS 2B: The average reported number of matrices solved in Game 1 in the subgroup of male participants from the Math-Ability treatment is higher than in the subgroup of female participants from the Math-Ability treatment, while the average reported number in Game 2 is the same between the two subgroups. At the same time, no such differences are observed in the Control treatment between the two subgroups.
In more technical terms, the regression model used to compare the control to the corruption and the math treatment is:
Y = β0+ β1C + β2M + β3G + β4(C · G) + β5(M · G) + β6A + (1)
From the regression model, Y is the outcome variable and shows the number of matrices solved and the random number generated for Game 1 and 2 respectively. Second, β0 is the constant and shows the number of matrices solved and the
random number generated by males from the control treatment in Game 1 and 2 respectively. The term C shows whether an observation was from the Corruption treatment and therefore, β1shows the effect of the corruption prime on the outcome
variable. Similarly, the term M shows whether an observation was from the Math ability treatment and thus, β2 shows the effect of the math prime on the outcome
came from a female participant and zero when it came from a male participant. The coefficient β4 and β5 are the coefficient of two interaction terms. These two
terms can help check whether the priming effect was different for females than males in the corruption and in the math treatment respectively. In particular, if β4 and β5 are significant, this would be evidence in favour of Hypothesis 2. The
term A controls for the age of the participants because it might influence the outcome variable. Older participants might have superior quantitative skills and lie to a different extent than younger participants. Finally, is the error term.
5
Results
To proceed with the analysis of the gathered data, it is useful to first have a visual inspection of the average responses across all three treatments for each game. As seen in Figure 2, participants from the control group solved fewer matrices than in the Corruption and Math treatment. Due to the fact that students were randomly assigned into a treatment, the difference can be attributed to the priming effect. More specifically, participants from the corruption and math treatment solved on average 0.61 and 0.52 more matrices, respectively, than the control treatment. However, the difference between the corruption and the control treatment is not significant (p-value=0.155). This means that the results from Game 1 do not sup-port Hypothesis 1. Potential reasons for the lack of significance are discussed in the next section. Also, the results do not support Hypothesis 3 because the difference between the control and the math treatment is not significant (p-value=0.273).
Figure 3 shows the average reported number generated per treatment in Game 2. In this game, the results are in line with Hypothesis 1. Participants from the corruption treatment reported an average number that was larger than the con-trol treatment by 17.375. The difference is significant at a 5% significance level (p-value=0.041). This result is in line with Hypothesis 1. The effect remains signif-icant at a 10% significance level, even when controlling for the age of participants (p-value=0.079). Further, the math treatment reported an average number that was larger than the control treatment by 5.05. As anticipated, this difference is not significant (p-value=0.603).
Figure 2: Average reported number of matrices solved per treatment in Game 1. The lines represent the standard error bars for a 95% confidence interval.
treatment appeared to act in a perfectly honest way. The expected average number generated in the absence of dishonesty was 50 because numbers were generated from a uniform distribution [0,100]. As it can be seen from Figure 3, the average reported number from the control treatment was 52.125. This shows that despite the incentive that was provided, participants preferred to maintain an honest self-concept rather than be rewarded a few extra chocolates.
5.1
Gender differences
When looking at the effects of priming for males and females separately, some interesting and unexpected realisations can be made. First, as seen in Figure 4 and 5, the difference between males and females is not consistently significant. It appears to be marginally significant only in Game 1 in the Math-ability treatment, and if a one-tailed test was used, it would have been significant at a 5% significance
Figure 3: Average reported number generated per treatment in Game 2. The lines represent the standard error bars for a 95% confidence interval.
level. Alternatively, when looking at Table 1 and 2, it is clear that the coefficient of Gender is insignificant in every game except from Game 1 in the Math-ability treatment. This implies that in the absence of priming, it appears that there are no consistent innate differences between male and female participants in terms of responses given in the two games. Second, as Figure 4 shows, males from the Math-ability treatment solved on average 1.39 more matrices than males from the control treatment (p-value=0.014). Also, in Game 2, no such differences are observed which could mean that the difference in Game 1 did not arise because increased dishonesty. Potential explanations are discussed in the following section. Further, as it can be seen from Figure 5, males from the corruption treatment reported an average number that was higher than males from the control treatment by 25.19. The difference is significant at a 5% significance level (p-value=0.038). It can also be observed that there appear to be no such differences amongst the
Figure 4: Average reported matrices solved in Game 1 per gender group in each treatment. The lines represent the standard error bars for a 95% confidence inter-val.
subgroup of female participants. Therefore, the effect of consciously priming dis-honesty and math-ability appears to have a stronger effect on male participants than female participants.
With reference to Hypothesis 2, it was contended that gender differences would be observed, mainly because of self-fulfilling prophecies. The results obtained do not provide conclusive evidence in favour of this hypothesis. Using the regression model described in the previous section, it is found that β4, which is the coefficient
of the interaction term Gender · Corruption, is not significant in none of the games (Table1). This means that consciously priming the gender stereotype about corruption did not influence peoples’ behavior in terms of dishonesty. However, as seen in Table 2, β5, which is the coefficient of the interaction term Gender·M ath, is
Figure 5: Average reported number generated in Game 2 per gender group in each treatment. The lines represent the standard error bars for a 95% confidence interval.
is equal to -1.597 which means that female participants from the math-ability treatment reported that they solved 1.597 fewer matrices, on average, than males from the same treatment. The coefficient remains significant even when controlling for Age (p-value=0.084). At the same time, as predicted by Hypothesis 2, β5 is
Table 1: Regression results for the Corruption treatment Game 1 Game 2 1 2 3 1 2 3 Corruption 0.613 0.698 0.537 17.375 25.190∗∗ 22.669∗ (0.422) (0.626) (0.695) (8.158) (11.640) (12.906) Gender 0.587 0.546 7.079 6.425 (0.556) (0.556) (12.565) (12.923) Gender · Corruption −0.143 −0.044 −14.049 −12.500 (0.849) (0.868) (16.489) (17.416) Age 0.097 1.527 (0.089) (1.897) Constant 3.188∗∗∗ 2.857∗∗∗ 1.047 52.125∗∗∗ 48.143∗∗∗ 19.778 (0.291) (0.334) (1.707) (6.302) (8.308) (35.121) N 36 36 36 36 36 36
Notes: The outcome variable for Game 1 is the average reported number of matrices solved and for Game 2 it is the average reported number. The robust standard errors are in parentheses. ∗p < 0.1;∗∗p < 0.05;∗∗∗p < 0.01.
Table 2: Regression results for the Math treatment
Game 1 Game 2 1 2 3 1 2 3 M ath 0.518 1.393∗∗ 0.978 5.051 9.107 11.200 (0.465) (0.531) (0.688) (9.620) (14.194) (15.553) Gender 0.587 0.396 7.079 8.045 (0.559) (0.573) (12.637) (13.299) Gender · M ath −1.615∗ −1.597∗ −7.218 −7.307 (0.886) (0.873) (19.698) (19.947) Age 0.447∗ −2.254 (0.250) (5.102) Constant 3.188∗∗∗ 2.857∗∗∗ −5.450 52.125∗∗∗ 48.143∗∗∗ 89.996 (0.291) (0.336) (4.649) (6.319) (8.356) (93.708) N 33 33 33 33 33 33
Notes: The outcome variable for Game 1 is the average reported number of matrices solved and for Game 2 it is the average reported number. The robust standard errors are in parentheses. ∗p < 0.1;∗∗p < 0.05;∗∗∗p < 0.01.
Despite these results, when looking at the math ability treatment, there seems to be evidence in favour of Hypothesis 2. First, for Game 1, it is observed that males who were primed about math ability solved on average 1.39 more matrices than males from the control treatment. The difference is significant at the 5% significance level (p-value=0.014). The data collected from Game 2 show that males from the math treatment did not report higher numbers on average than males from the control (p-value=0.526). Two potential explanations could be that either the math prime improved the performance of the male participants on the specific task, or that male participants felt increased pressure to lie and
over-report the number of matrices solved because of the priming effect. Even though Hypothesis 2 is supported from the evidence found for males in the math treatment, it is not clear whether the social-identity prime or the math prime is responsible. Similarly, for the females, it could be that neither the social-identity prime nor the math prime had a significant effect, or that both of them had a significant effect which cancelled each other out.
6
Discussion
As it was seen in Figure 2 and 3, the experimental results seemed to support Hypothesis 1 and 3. However, some of the differences found were not significant at a 5% significance level. A big reason for the lack of significance could be the small sample size. Each of the three treatments had no more than 20 participants which complicated the statistical analysis of the results.
It is important to note that the experimental data support the self-concept maintenance theory, the fudge-factor theory and the ethical manoeuvring theory, which were discussed in the Literature Review. None of the participants cheated to the maximum possible degree in either games. To illustrate, participants could report that they solved all 12 matrices in Game 1 and that they generated the number 100 in Game 2 in order to maximize their payoff. This finding is supported by the 3 theories mentioned above. It seems to be the case that participants valued the maintenance of a positive and honest self-concept more than the rewards they could gain by acting in a dishonest way.
However, there are some alternative explanations that can perhaps justify this result. First, it could be the case that participants were afraid of getting caught. Even though it was made implicitly clear that they could cheat without getting caught, participants might have attributed a non-zero value to the probability of getting caught. This reason is not a limitation of the experiment because it is reflective of the real world. In the real world, people are never explicitly told that they can cheat without getting caught.
Second, it could be the case that participants exhibited dishonesty to a small degree because they cared about their image to outsiders. The experiment was done in a classroom setting and participants were not complete strangers to each
other. Maybe the reason why they did not cheat to the maximum degree was not because they wanted to maintain a positive self-concept, but because they wanted to be seen as fair and honest by their peers. B´enabou and Tirole (2011) develop a theory and a general mathematical model which shows that agents place an intrinsic value on “social identity”. Similarly, Akerlof and Kranton (2000) posit that the way others, even strangers, perceive an agent, determines his/her social identity. Social identity may in turn influence the agent’s actions, even if the agent expects to have no further interactions with these people. In the experimental setting of this thesis, each participant was seated close to each other, which meant that a participant’s responses in the leaflet could be seen by other participants who were seated next to him/her. Therefore, some participants might have preferred to be seen as honest by people seated near them, instead of obtaining a larger reward.
With reference to Hypothesis 1, the experimental results suggest that con-sciously priming dishonesty seems to exasperate subsequent behaviour. As ex-pected, the increased attentiveness to moral standards caused by the prime was not significant, relative to the negative effect of the prime. Even though the ex-perimental setting is very different, this finding contradicts, to a certain extent, the conclusion made by Welsh and Ord´o˜nez (2014). Despite the lack of significant evidence in favour of the dishonesty priming effect in Game 1 (p-value=0.155), participants cheated significantly more in Game 2 (p-value=0.041).
One reason for the lack of significance in Game 1 could be that participants could not maintain an honest self-concept by over-reporting the number of ma-trices solved. According to the fudge-factor theory, participants could convince themselves that they made an honest mistake and reported a larger number of matrices that what they actually solved. Evidently, such a mistake is very hard to justify, especially because participants were given sufficient time to check their own answers and make sure that they accurately report the number of matrices solved. In contrast, in Game 2, participants could justify dishonest behaviour by “accidentally” generating a random number a second time, and convincing them-selves that it was unintentional. The point made is that in Game 2 it was easier to act in a dishonest manner and maintain a positive/honest self-concept, than in Game 1. Kajackaite (2018) conducts an experiment and finds that
