Identity, Overconfidence, and Investment Decisions Francesco D’Acunto ∗ June 2015 Abstract Why are men more risk tolerant than women, and why do they invest more than women? I test whether identity stereotypes help explain this heterogeneity. I manipulate identity in a controlled environment by priming its salience to subjects. Men whose identity is primed take on more risk, and invest more often and more money than controls. The salience of male identity increases men's beliefs of experiencing good outcomes in a game of chance. Inducing overconfidence similarly makes men take on more risk and invest more. The effects are stronger for older cohorts of men, consistent with the notion that gender identity stereotypes have become less stark over the last decades. JEL: D81, G01, G11, J16 Keywords: Stereotypes, Experimental Economics, Heterogeneous Beliefs, Individual Investors. ∗ Haas School of Business, UC Berkeley (francesco [email protected]). UC Berkeley CPHS Approval Protocol number 2013-11-5805. For their invaluable help and guidance, I thank Stefano DellaVigna, Ulrike Malmendier, Gustavo Manso, Terrance Odean, and Ross Levine. For very helpful comments and discussions, I thank Nick Barberis, Dan Benjamin, James Choi, Xavier Gabaix, Luigi Guiso, George Korniotis (discussant), Samuli Kn¨ upfer, Rachel Kranton, Ross Levine, Don Moore, Adair Morse, Muriel Niederle, Christine Parlour, Matthew Rabin, Andrei Shleifer, Stephan Siegel (discussant), Michael Weber, and seminar participants at the 2014 Miami Behavioral Finance Conference, the 2013 NBER Behavioral Economics Fall meeting, the 2014 Whitebox Advisors Graduate Conference, and UC Berkeley. Financial support from the White Foundation is gratefully acknowledged. All errors are my own.. 1 Introduction Men are more risk tolerant than women (Croson and Gneezy, 2009). They are more competitive (Niederle and Vesterlund, 2007), and they invest more often and more aggressively than women in financial opportunities (Barber and Odean, 2001). Several explanations for the different risk attitudes across genders have been proposed. First, these differences have biological roots (Kuhnen and Chiao, 2009; Coates and Herbert, 2008; Mastripieri, Sapienza, and Zingales, 2009; Cesarini et al., 2010; Cronqvist and Siegel, 2014; Cronqvist et al., 2014). Second, social identity may additionally help to explain these differences, because identity stereotypes prescribe normative behaviors to men and women (Mischel, 1966; Akerlof and Kranton, 2000).1 Third, the differences may stem from other characteristics that correlate with gender. In this paper, I aim to provide a causal test for the effect of social identity on the risk attitudes and investment behavior across genders, and to study the economic channels that may mediate this effect. Running causal tests in the field would be hard (Camerer and Lovallo, 1999; Biais et al., 2005), thus I randomly manipulate the salience of identity stereotypes in a controlled environment (e.g., Steele, 1997; Shih et al., 1999), and I compare the risk tolerance and investment behavior of treated and control subjects. Similar priming techniques have been employed to address economic questions for which exogenous variation in the field is hard to detect (Benjamin, Choi, and Fisher, 2013; Cohn et al., 2015; Coffman, 2015). Results in social psychology show that priming and threatening male identity both induce risky behaviors in men due to salience and overcompensation (Maas et al., 2003; Willer et al., 2013). Salience means individuals tend to conform to the trait of their own identity that is primed. Overcompensation means individuals behave more in line with their own identity to reaffirm it when it is threatened.2 I find that men whose identity stereotypes are primed or threatened become more risk tolerant after the manipulations, and invest more often and more money in simple 1 See Table 1 for a list of identity stereotypes attributed to men and women by 200 online respondents. The identity channel is not necessarily uncorrelated with biology or genetics. In Willer et al. (2013), men with higher levels of basal testosterone are more responsive to identity threats. 2 Note that this notion of identity threat is unrelated to the stereotype threat of Carr and Steele, 2010. 1 investment decisions compared to control men and to women. I detect no effect of identity manipulation on women’s risk attitudes and investment behavior. Previous literature in Economics and in Psychology found conflicting results on the effect of gender identity on risk attitudes. In Meier-Pesti and Penz (2008), male gender salience increases the willingness to take risks in both men and women, but in a non-incentive-compatible setting, and in a subject pool in which men and women reported the same masculine attributes. To the contrary, in Benjamin, Choi, and Fisher (2010), there is no effect of gender salience on risk preferences in an incentive-compatible, between-subjects experimental design. Compared with previous literature, this paper proposes five innovations. First, I construct a within-subject experimental design, which controls for the heterogeneity of subjects’ preferences before the experimental manipulations. This design allows testing for the effect of identity on risk attitudes more precisely than between-subjects designs. If I only use the between-subjects variation in risk attitudes after the manipulations, I estimate effects of similar size but statistically indifferent from zero, which is consistent with the non-result of Benjamin, Choi, and Fisher (2010).3 The within-subject design raises concerns of demand effects. This is why I run artefactual field experiments with varied non-student subjects that are not aware of being part of an experiment throughout the paper. I also recruit additional subjects who perform the same experimental tasks as others, but guess the motivations of the tasks at the end of the experiment.4 Less than 10% of these subjects give answers that could be compatible with the true scope of the study, whereas I show that virtually all subjects exposed to the treatment conditions become more risk tolerant after the manipulations. Second, I test for the first time for the economic channels that transmit the effect of gender identity on risk attitudes. I find that beliefs are a channel of transmission, because men whose identity is manipulated become overconfident in a pure game of chance, 3 I report the between-subjects results for this paper in Table 5 of Appendix B. The priming procedure in this paper is also purposefully more aggressive than in Benjamin, Choi, and Fisher (2010): subjects read a text full of stereotypes related to male or female identity, and they recall and describe in detail a situation when they behaved in line with the stereotypes (see Appendix A). The more aggressive procedure possibly explains the larger estimated magnitudes of the effects compared to their test. 4 I thank Dan Benjamin for suggesting this test. 2 consistent with the illusion of controlling a random process. The effect of gender identity on men’s beliefs in a pure game of chance is a novel result that deepens our understanding of the sources of overconfidence, which is one of the most studied behavioral biases in Economics.5 Parallel to the lack of an effect of identity on women’s risk attitudes, the manipulations do not affect the beliefs of women. Third, I show that the effect of identity on men’s risk attitudes also obtains if subjects make investment decisions under risk, both on the extensive margin (the decision to invest or not), and the intensive margin (the decision of how much money to invest conditional on investing). The results are similar if the decisions are framed as delegated investments, that is, if subjects invest money on behalf of a principal. The delegated setting aims to mimic the decisions of asset managers and corporate executives in the field. These results could therefore be interpreted as a causal test for the large body of correlational evidence in Economics and Finance on overconfidence and investment decisions by individual investors and corporate executives.6 Fourth, I exploit the varied pool of subjects in terms of demographics to disentangle the effect of gender identity from the effects of biological and genetic characteristics on risk attitudes and investment decisions. I find that older men are the most affected by the identity manipulations, and the effect decreases monotonically the younger the subjects. This result is consistent with the notion that gender identity stereotypes have become less stark over the last four decades, and hence their manipulation should have lower effects on the behavior of younger cohorts of men. The biological and genetic characteristics of human beings, instead, cannot have changed significantly in such a short period of time. Fifth, I propose a coherent economic framework based on Bordalo, Gennaioli, and Shleifer (2014) to interpret the results in the paper. To the best of my knowledge, this is the first paper that proposes a theoretical foundation in Economics for the act of priming in a controlled environment. This framework could inspire the uprising economic research that uses priming techniques to discipline the predictions of their tests. In this paper, 5 Note that this test may have been run in the past with no results, and may have not been disclosed due to publication bias. 6 e.g., Barber and Odean, 2001; Malmendier and Tate, 2005; Malmendier and Tate, 2008; Gervais et al., 2011; Hirshleifer and Han, 2012; Ben-David et al., 2013. 3 the framework requires an effect of identity salience on men’s beliefs, which is exactly what I find. There are caveats to the interpretation of the evidence in the paper. By construction the analysis isolates reduced-form effects: a drawback is the causal tests do not inform on the expected size of the effects in the field. The experiments change the salience of identity stereotypes to gauge their causal effect on decision making, but the level of the perception of stereotypes should affect decisions in the field. Moreover, the subjects do not necessarily represent the average investor or the average CEO, and there is no uncertainty in the lottery choices and the investment decisions that the subjects perform. It is also important to stress what this paper does not show. The paper does not show that gender identity is more important than biology and genetics as a determinant of risk attitudes. Future research should be devoted to understanding how biology, genetics, and identity may interact to affect risk attitudes and investment decisions. Moreover, in this paper men and women react differently to the same identity stereotype manipulations. My results are therefore inconsistent with the claim that the differences between the sexes could be eliminated by manipulating men and women’s gender stereotypes, or that the only difference between the sexes is their gender identity. 2 Laboratory Environment The experiments are run on an online platform, Amazon Mechanical Turk (mTurk). Kuziemko et al. (2015) are among the first who use mTurk in Economics research. On mTurk Requesters post tasks, and a large pool of Workers can accept to perform them. Requesters are often private companies and Workers are registered users. Workers provide their fiscal address and social security number for tax purposes. Tasks are short, and the average pay is low ($1.39 per hour). Workers based in the United States, who are exclusively recruited in this paper, access mTurk mainly to spend their spare time constructively (Paolacci et al., 2010). The quality of answers is not lower than in human-subjects laboratories despite the lower pay (Casler et al., 2013).7 Recently, 7 Mason and Suri (2011) and Paolacci et al. (2010) describe mTurk and Workers. Berinsky et al. (2011) replicate laboratory results in political science on mTurk. 4 mTurk has gained interest as a means to recruit diversified subjects for artefactual field experiments, but a concern is Workers may not properly complete the tasks. To address the concern, (i) I restrict the subject pool to Workers with at least 95% positive rates on all the tasks accepted in the past; (ii) I track the time they take to complete each task; (iii) I read all essays to verify that subjects produce coherent statements; and (iv) I add implausible options to the lotteries and I verify that they are not picked.8 Advantages over laboratory. mTurk has a set of advantages compared to humansubject laboratories (Horton et al., 2011), especially for a study on identity priming and for within-subject designs: • Subjects come from the whole United States. Figure 1 plots their IP addresses. If subjects came from the same college town, they would live in a peculiar social environment that is likely dissimilar from the one average Americans face. Also, college and MBA students may sort into communities whose values they share. • As an online, double-blind platform, mTurk allows to run artefactual field experiments in which recruiting is simple (List, 2011). Subjects perform tasks in their environment and are not aware they are part of an experiment. This procedure helps to address the concerns of demand effects raised by within-subject designs. • Subjects’ demographics are well varied, whereas a sample of college students would only include individuals between 18 and 22 years of age with some college education. • Replicability of results is easy: any Requester accesses the same subject pool. Easy replicability allows for a transparent comparison of results across studies.9 • Workers’ anonymity makes the priming procedure most effective. Subjects may describe unconventional experiences with no fear of being identified. 3 Effect of Identity on Risk Tolerance In Experiment 1, I test whether gender identity stereotypes have a causal effect on risk attitudes in a within-subject experimental design, where subjects’ risk tolerance is elicited before and after being exposed to the manipulation of their gender identity. The manipulations include identity salience and threat, both of which have been shown to 8 For instance, I add a choice between $0 for sure and a lottery that pays $0 or a positive amount. On September 26, 2012, Daniel Kahnemann proposed a protocol for improving the credibility of research on priming, which is easy to implement on mTurk. 9 5 induce risky behaviors in men due to salience and overcompensation (Maas et al., 2003; Willer et al., 2013). Experimental Design. 340 subjects were recruited on Amazon Mechanical Turk (mTurk) in April 2012 (first session) and September 2012 (second session). Subjects were proposed a creative writing task to earn $0.5, plus a chance to earn a bonus by picking lotteries. The nature or objectives of the study were not disclosed to ensure that (i) subjects were not primed with their identity before the experiment; (ii) subjects would not search for previous research on identity priming, which would raise concerns of demand effects; and (iii) the results could be replicated in subsequent sessions and by other researchers. The subject pool was restricted to users with a US tax identification, and with more than 95% of lifetime tasks approved in the past. Table 2 describes the experimental designs and samples for all the experiments in the paper. The design of Experiment 1 was a 2 (male, female) by 3 (control, male prime, female prime) factorial design. I excluded 20 subjects (5.8% of the full sample) because of inconsistencies in lottery choices, leading to a final sample of 320 subjects. Subjects were assigned in proportions of 2:2:1 to the control, male-prime, and female-prime conditions.10 The aim was to maximize the power for the male identity salience test, because my cheating-free threat had not been tested in previous research, and it was a weaker prime than the one used by Maas et al. (2003) and Willer et al. (2013).11 Procedure. The experimental procedure was as follows. In the first stage, subjects answered four background questions including the country of residence, gender, age group (18-22, 23-35, 36-45, 46-60, 60+), and education (high school or lower, some college, college degree, postgraduate degree). Then subjects faced two screens of lottery choices (Holt and Laury, 2002). Each choice included a degenerate lottery paying a positive outcome for sure (certainty equivalent), and a lottery paying a positive outcome with probability 1/2, and zero otherwise. In the second session, subjects faced three sets of lottery choices to make sure the results did not vary with effort, which indeed was not the case. Subjects had to complete the lottery task in full before proceeding, but they could 10 It is important to stress that on mTurk one cannot discriminate Workers at the recruitment stage by gender, that is, one cannot limit the sample to men only, or target the same number of male and female subjects. About 60% of U.S. mTurk workers are women (Mason and Suri, 2011). 11 They gave random feedback on a gender-identity survey suggesting that subjects are masculine or feminine, which implies an act of deception on the part of the experimenter. 6 leave the experiment at any time.12 The second stage consisted of the experimental treatments. Subjects read a short text (see Appendix A) and recalled and described an experience in line with the text. Subjects were asked to describe the situation and their feelings in detail in a short essay of 5 to 10 sentences. The texts were taken from online blogs, so as to be similar to the gender identity manipulations to which the subject pool, which is made of internet users, may be regularly exposed in their daily life. Note that this test does not aim to give subjects a scientific definition of identity stereotypes, but to activate their own interpretation of identity stereotypes once they face them explicitly. Control subjects read a text on ayurveda principles for a healthy lifestyle. Subjects in the male-identity-prime condition read a text full of stereotypes about how a masculine person behaves.13 Subjects in the female-identity-prime condition read a text full of stereotypes about how a feminine person behaves. The third stage consisted of two additional screens of lottery choices. The lottery choices changed to ensure subjects did not merely repeat their first-stage choices. All choices were incentive-compatible: subjects knew their bonus would be calculated by picking one screen and one line at random, running their choice (certainty equivalent or lottery), and dividing the final amount by 100. Manipulation Check. Across all experiments, the content of the essays produced by the subjects served for the manipulation checks. I verified that each essay reported words related to the identity stereotypes enlisted in Table 1 in the corresponding experimental condition (see Appendix A for essay samples). Ad hoc manipulation check tasks would have made subjects aware they were part of an experiment, and hence would have worsened the risk of demand effects. In Figure 7, I formalize this manipulation check for Experiment 1.14 In Panel (a) of Figure 7, I report the average number of times subjects wrote a stereotype associated with male individuals from Table 1 in their essays, 12 No one left the experiment before completing it in full. Consistent with Table 1, confidence entered the text. A concern is the treatment directly induces overconfidence instead of priming a generic set of manly characteristics. But the male-identity priming text explicitly states that ”the masculine side [...] includes [...] how to accurately weigh probabilities so that you know the most likely outcome to expect in situations you come across” (see Appendix A). Specifying the importance of a correct assessment of probabilities also speaks to the issue of demand effects in a within-subject design: if the subjects wanted to please the experimenter, they would pay attention to the objective probabilities in the lottery choices. 14 The (unreported) results are similar across all other experiments. 13 7 across experimental conditions. Subjects in the male identity prime condition reported on average 2.41 male identity stereotypes.15 Instead, subjects in the control or female identity prime condition do not write masculine identity stereotypes. In Panel (b) of Figure 7, I report the average number of times subjects wrote a stereotype associated with female individuals from Table 1 in their essays, across experimental conditions. Subjects in the female identity prime condition are likely to report stereotypes associated with female individuals, whereas others are not. Measuring Risk Tolerance and Non-parametric Analysis. I compute a non-parametric measure of risk tolerance at the subject level, described in Figure 2. I subtract the times the subject chose the certainty equivalent over the lottery from the times a risk-neutral agent would make this decision. I average the difference across the choices of the first stage to obtain a pre-treatment measure of risk tolerance: RiskT olerancepre,i = 1 X × (RN choicespre,l − choicespre,l,i ), L l where l ∈ (1, L) are the screens of lottery choices the subject faces, RN choicespre, l are the times a risk-neutral agent would choose the certainty equivalent over the lottery in screen l, and choicespre, l, i are the times agent i chose the certainty equivalent over the lottery in screen l. RiskT olerancepre,i is positive when the subject, on average, chose the lottery more often than a risk-neutral agent would do. I compute the analogous measure for choices made after the experimental treatments: RiskT olerancepost,i = 1 X (RN choicespost,l − choicespost,l,i ), × L l The within-subject change in risk tolerance after the treatment compared to before the treatment is ∆RiskT olerancei = RiskT olerancepost,i − RiskT olerancepre,i . Figure 3 plots the estimated densities and the cumulative distributions of RiskT olerancepre and of ∆RiskT olerance for the subsample of men, who are affected by the manipulations. Results for women are in Appendix B. Panel A of Figure 3 plots the estimated distributions of the level of risk tolerance of men across treatment conditions before the manipulations. The vast majority of subjects are risk averse: the mean of all 15 The figure only refers to the most common stereotypes, in Table 1. Subjects often reported other masculine stereotypes that do not enter the computation. 8 distributions and the largest part of their mass lie in the negative domain. Reassuringly, the distributions for men that will be exposed to the control and to the male prime conditions (green dot-dash and blue solid lines) are similar in terms of mean and standard deviation. The distribution of the risk tolerance of men in the female prime group spikes to a negative value and lies in the negative domain, but the standard deviation is lower than the one for the other experimental groups. Panel B of Figure 3 depicts the main result of Experiment 1. The plots report the density and cumulative distribution of ∆RiskT olerance for the male subsample across experimental conditions. The risk tolerance of control subjects (green, dot-dash line) does not change, on average, after exposure to the control condition. This evidence suggests that the control condition does not drive the results, and that the act of priming does not cause any change in behavior by itself. If the male prime and the female prime were increasing the risk tolerance of subjects, we would expect the distributions of the change in risk tolerance to shift to the right compared to the distribution for controls. Indeed, the change in the risk tolerance of the male prime group (blue, solid line) is on average positive (0.5).16 The distributions of the change in risk tolerance for the identity threat group (dashed, red line) also peak to a positive value (0.45). OLS Results. To assess the statistical significance of the effects, I estimate the following OLS equation on the subsample of men: ∆RiskT oleranceiae = α + γ1 × M aleP rimeiae + γ2 × F emaleP rimeiae + ηa + ηe + iae (1) where ∆RiskT oleranceiae is the within-subject change in risk tolerance for subject i, in age group a, and education group e; M aleP rime and F emaleP rime are dummies that equal 1 if the man is exposed to the male identity (salience) or female identity (threat) prime; ηa and ηe capture subjects’ age and education groups. The identification strategy is a difference-in-differences design: I look at subjects’ decisions before and after the manipulations, and across subjects in the treatment and control groups. In column (1) of Panel A of Table 3, men primed with male identity choose lotteries over certainty 16 The male group displays a fat right tail. In Table 7 of Appendix B, I show the results are not driven by the subjects in this right tail. 9 equivalents 0.51 times more often (s.e. 0.25) than controls after the prime.17 The result is robust to averaging out age- and education-group effects in column (2) of Panel A of Table 3. Men primed with female identity choose lotteries over certainty equivalents 0.43 times more often (s.e.0.21) after the prime compared to before it. In Panel B of Table 3, I add the subsample of women to the analysis, and hence I use a triple-differences estimator. I test if the results on men are robust to adding this additional layer of subjects by estimating the following: ∆RiskT oleranceiae = α + βM aniae + γ1 × M aleP rimeiae + γ2 × F emaleP rimeiae γ3 × (M aleP rime × M an)iae + γ4 × (F emaleP rime × M an)iae + ηa + ηe + iae (2) where ∆RiskT olerance is the same measure as in Equation 1, and M an equals 1 if the subject is a man. In Equation 2, β does not capture the difference in average risk tolerance between men and women ex ante, but the difference in the change in risk tolerance due to the exposure to the control condition. Hence, βˆ should be zero, unless being exposed to the control condition affects men and women’s risk tolerance differently. I cannot reject the null that βˆ is zero at any meaningful level of significance. The increase in risk-tolerant choices by men in the male-identity-prime condition is significantly more positive (0.84, s.e. 0.37) than that of women, and with respect to the control group. The result is robust to averaging out age- and education- group effects. Men also become more risk tolerant after the identity threat, but I can only reject the null of no effect at the 10% level of significance (0.59, s.e. 0.33). Demand Effects? The within-subject design raises concerns of demand effects: subjects may be willing to conform their behavior to what they believe is the aim of the experimenter. Demand concerns are the main reason why I run artefactual field experiments, in which subjects are not aware of the scope of the experimenter, or that they are part of an experiment, but they perform tasks on mTurk for which they have signed up for pay.18 I also test for the scope of demand effects more directly. In June and July 2014, I recruited 100 new subjects to perform the same tasks as in Experiment 1, but 17 Because individual-level choices are collapsed into their average before and after the treatment, robust standard errors are only corrected for White heteroskedasticity. 18 Note that mTurk’s primary function is for private companies to outsource simple tasks or to test the effectiveness of advertising campaigns. mTurk was not created as a tool for conducting scientific research. 10 also to guess the motivations of the tasks at the end of the experiment.19 Figure 8 plots the reported motivations for the lottery tasks and the writing task across six categories. Less than 10% of subjects gave answers that could be compatible with the true scope of the study, whereas in Figure 3 virtually all subjects exposed to the treatment conditions become more risk tolerant after the manipulations. Even if this test is run on a different pool of subjects that those in Experiment 1, the results provide direct evidence that demand effects are unlikely to be a relevant concern in this setting. Evidence against fabricated data. I run the test proposed by Simonsohn (2013) against fabricated data in experimental research. This test is important in light of the recent research misconduct scandals commented on by Kahneman (2012). In Table 8 of Appendix B, I verify for all the experiments in the paper that the standard deviations of the sample averages across experimental conditions do not differ from those one would obtain in random samples across conditions. In Experiment 1, the null that the observed distributions come from independently drawn random samples can only be rejected above the 99.99% significance level (99.98% in the male subsamples). 4 Channels that Mediate the Effect of Identity on Risk Tolerance: Beliefs After having established that priming or threatening male identity increases men’s willingness to take risk in a within-subject, incentive-compatible research design, I move on to test for the economic channels that may explain this effect. The primes may act through two channels: (i) preferences: priming identity could be a positive shock to men’s risk tolerance. This shock would reduce the certainty equivalent men require to give up a chance to take part in a lottery; (ii) beliefs: priming identity could be a positive shock to men’s subjective probability of experiencing good outcomes if they take part in a lottery. They could believe they are more likely to obtain the good outcome conditional on participating than what the objective probabilities suggest. If individuals held subjective beliefs of experiencing good outcomes in risky settings, which might differ from the objective probabilities they are told, the identity manipulations could affect their subjective beliefs. Subjects would become overconfident, that is, believe they 19 I thank Dan Benjamin for suggesting this test. 11 are more likely than what the objective probabilities suggest to obtain the good outcome if they participate in the lottery. This form of overconfidence involves the illusion of controlling random processes, as opposed to overplacement or overprecision in interpreting signals (see Moore and Healy, 2008) In this section, I propose two novel tests to assess the possibility that a beliefs channel helps explain the effect of identity manipulations on men’s attitudes under risk.20 A Identity and Subjective Beliefs under Risk The first challenge to testing for an effect of identity manipulations on beliefs is to design a test that does not elicit at the same time the preferences for risk and the subjective beliefs of experiencing good outcomes under risk. I propose such a test in Experiment 2. Design and Procedure. I recruited 325 subjects on Amazon Mechanical Turk in May 2015, and invited them to work on a survey to earn $0.4.21 I excluded two subjects (1% of sample), because they did not write meaningful words or sentences in the priming task, leading to a final sample of 323 subjects. After answering four background questions, subjects were exposed to the experimental conditions of Experiment 1 (control, male identity prime, female identity prime). I assigned subjects in equal proportions to the three experimental conditions (see Table 2). I elicited subjective beliefs in two steps. First, subjects faced a lottery that won on average 5 out of 10 times. No additional details were disclosed, including the ”winning” or ”losing” payoffs. Subjects predicted the number of times they would win if the lottery was played 10 times. The number they reported was a direct measure of the subjective probability of experiencing a good outcome in a risky setting. To make the elicitation incentive-compatible, subjects knew that the lotteries would be played 10 times using a random number generator after the experiment, and they would receive a bonus based on the difference between the predicted number of wins and the actual number of wins.22 This design ensured that (i) subjects were told explicitly that the outcomes of the lotteries would be completely determined by 20 Assessing the relative contribution of the preferences and beliefs channels is beyond the scope of this paper. 21 A previous version of this experiment elicited beliefs in a non-incentive-compatible setting. I report the similar results for the previous experiment in the Online Appendix. 22 The bonus was 10 cents if the subject predicted the actual number of wins, 5 cents if the subject was one time long or short the actual number of wins, 2 cents for two times long or short, and 1 cent for three times long or short. 12 chance, and (ii) risk tolerance had no role in the predictions. In a second step, subjects imagined a person taken at random from their neighborhood participated 10 times in the same lottery. They predicted the number of times they thought this neighbor would win, and were paid a second bonus based on the accuracy of their predictions. The number was interpreted as a direct measure of the subjective probability that a peer would win the lottery. The comparison to peers is relevant, because if the primes were simply inducing a state of elation in subjects, they would reply that both they and the peers were likely to win more than 5 times. Instead, if subjects believed they could control a random process, and the primes made them confident they were better than average at the lottery, then primed subjects should have thought their peers were as likely to win the lottery as the objective probability suggested. Results. Panel A of Figure 4 plots the subjective beliefs of winning the lottery for men, across the control and primed conditions, where the primed condition pools together identity salience and threat.23 Primed men think they will succeed more than five times on average, whereas control men do not. A double-sided t-test for whether the average predicted number of wins by primed men (5.25) equals 5 rejects the null at the 1% level of significance, whereas a similar test for whether the average number of wins by control men (4.96) equals 5 cannot reject the null at any plausible level.24 Panel B of Figure 4 plots the average difference between men’s predicted wins by themselves and by peers across experimental conditions. The difference (0.04) is statistically and economically indistinguishable from zero for controls. It is significantly positive for primed men (0.25). Manipulating identity therefore does increase the subjective probability of experiencing a winning lottery outcome for men, even if they play a pure game of chance. This novel result can only be justified if subjects have the illusion of controlling a game of chance, and if priming makes subjects overconfident on their ability at this game. In Appendix B, I show that the manipulations have no effect on the beliefs of women: both control and treated women think on average that they will succeed five times if playing the lottery 10 times, in line with the objective probabilities they are told. The different results on the effect of the manipulations on beliefs across men 23 I merge the identity salience and threat groups to increase the power of the test, because the two primes have very similar effects on the willingness to take on risks in Experiment 1. 24 Note that a test for whether the averages across groups are equal only rejects the null at the 10% level of significance. 13 and women pair up with the asymmetric effect of identity manipulation on risk attitudes by gender in Experiment 1. Incidentally, the effect of identity on men’s beliefs may help to shed light on the sources of overconfidence, which is one of the most studied behavioral biases in Economics. Gender identity appears to causally affect beliefs in risky settings. B Inducing Overconfidence and Risk Tolerance The results of Experiment 2 suggest that indeed the beliefs channel could in part explain the effect of gender identity manipulations on the willingness to take risks by men. I therefore propose a further test for whether manipulating men’s beliefs about experiencing good outcomes in choices under risk may provide a similar effect on risk attitudes as the effect of the identity manipulation, and in an incentive-compatible setting. The methodological challenge for such test is to provide an experimental manipulation that plausibly affects men’s confidence in a pure game of chance, hence gives them the illusion of controlling the outcomes. To this aim, in Experiment 3 I build on the social psychology literature, and I induce overconfidence in the form of illusion of control asking subjects to recall a situation in which they had power over one or more individuals (e.g., Anderson and Galinsky, 2009; Fast et al., 2009; Fast et al., 2012). Table 1 shows that the stereotypes attached to powerful individuals by a pool of 200 respondents on mTurk are strikingly similar to those attached to men. Recalling a situation of power over others may also induce a state of elation in subjects, which would affect their risk attitudes for reasons unrelated to overconfidence or identity (Kuhnen and Knutson, 2011). I propose to prime the sense of success of subjects as a placebo manipulation. Recalling a situation when subjects were successful and describing it in detail plausibly induces a state of elation in subjects. But whereas powerful individuals are described with similar stereotypes as men, successful individuals are associated with a mix of male and female identity stereotypes (see Table 1). A concern with this test for elation states is that success itself may be expected to induce overconfidence in subjects, because confidence is a stereotype attached to successful individuals. But (i) there is no evidence in the social psychology literature that a 14 success prime induces overconfidence in subjects;25 and (ii) if the prime was inducing overconfidence, it would induce it in the form of overplacement as opposed to the illusion of control of outcomes. Contrary to the illusion of control of random outcomes, overplacement should not affect decisions under risk in a pure game of chance, because individual ability or effort cannot affect the outcomes. Hence, Experiment 3 could also be interpreted as a placebo test to show that overconfidence in the form of overplacement does not affect the risk attitudes of men in a pure game of chance. Design and Procedure. I recruited 340 subjects on Amazon Mechanical Turk (mTurk) in May 2012 (first session) and October 2012 (second session), and invited them to work on a survey on creative writing to earn $0.5, and a bonus by picking lotteries. I restricted the subject pool to mTurk Workers in the United States and with more than 95% of lifetime tasks approved by requesters. Table 2 describes the experimental design, which was a 2 (male, female) by 3 (control, overconfidence prime, success prime) factorial design. I excluded 17 subjects (5% of sample) because of inconsistencies in the lottery choices, leading to a final sample of 323 subjects. I assigned subjects in equal proportions to the three experimental conditions. The first and third stages of the experimental procedure were the same as in Experiment 1. In the second stage, subjects in the control condition recalled an event in which they felt relaxed. Subjects in the overconfidence-prime condition recalled an event in which they had power over an individual or individuals. Subjects in the success-prime condition recalled an event in which they felt successful. All subjects described the events in 5-10 sentences. Results. I measure subjects’ tolerance to risk before and after the treatment, as well as its change, as in Experiment 1. Panel A of Table 3 reports the estimated coefficients for the following OLS equation: ∆RiskT oleranceiae = α+γ1 ×Overconf idenceP rimeiae +γ2 ×SuccessP rimeiae +ηa +ηe +iae (3) Column (3) of Panel A of Table 3 shows that men induced with overconfidence choose lotteries over certainty equivalents 0.52 (s.e. 0.26) times more after the prime than before, compared to controls. The effect is robust to adding age- and education-group fixed 25 Clearly the missing evidence may be merely due to publication bias. 15 effects.26 The size of the effect of the overconfidence prime on men’s risk tolerance is similar to the size of the effect of the identity manipulations on risk tolerance in Experiment 1. This similarity is consistent with the similar characterizations of male and powerful individuals in Table 1. Once I add women in Panel B of Table 3, the effect of power on men is still positive (0.44), but I cannot reject the null that the coefficient is zero at plausible levels (s.e. 0.42). The estimated coefficient on the elation-state treatment, SuccessP rime, is always close to zero in magnitude and statistically indifferent from zero. This result suggests that inducing a state of elation in men, or inducing overconfidence in the form of overplacement, hardly drives the increase in risk tolerance after the identity or overconfidence primes. Contrary to Experiment 1, the effect of the control condition on the omitted categories is on average negative. A t-test for whether the mean is different from zero within the omitted categories cannot reject the null at any conventional level of significance. But this non-result may be driven by the low power of a test with few observations. This is why, in the next section that looks at risky decisions framed as investment opportunities, I ask all subjects to recall and describe a situation when they felt relaxed, which primes the same state as the control manipulation in Experiment 1 and Experiment 3.27 5 Identity, Overconfidence, and Investment Decisions In the previous sections, I found a positive effect of gender identity salience and threat on men’s willingness to take risk. I tested for whether a beliefs channel may help explain this effect, and I found evidence consistent with an effect of identity stereotypes on men’s subjective beliefs of experiencing good outcomes in a pure game of chance, and a positive effect of inducing overconfidence on their willingness to take risks. I move on to test whether the manipulations I have studied so far – identity and overconfidence – may also explain men’s choices in risky settings that are framed as investment decisions, instead of lotteries. Looking at investment decisions is relevant, because it can provide a causal test for the correlational evidence on the effects of 26 One subject did not report his education level. Moreover, Experiment 3 differs from all other experiments on the gender composition of the sample. Recruitment on mTurk cannot discriminate by gender, and about 60% of U.S. Workers are women (Mason and Suri, 2011). Contrary to all other experiments, in Experiment 3 there are more subjects who declare to be men than women (see Table 2) 27 16 gender and overconfidence on men’s investment decisions in the field, which have been documented in a large literature in Economics and Finance, both for individual investors (e.g., Barber and Odean, 2001), and for corporate executives (e.g., Malmendier and Tate, 2008; Malmendier and Tate, 2005; Gervais et al., 2011; Hirshleifer and Han, 2012; Ben-David et al., 2013). A Identity, Overconfidence, and Individual Investments In Experiment 4, I test if identity and overconfidence affect the behavior of men when facing simple investment decisions. Design and Procedure. I recruited 240 subjects on Amazon Mechanical Turk (mTurk) in September 2012, and invited them to work on a survey on creative writing for $0.5, plus a bonus by performing simple investment decisions. I restricted the subject pool to mTurk workers in the United States, and with more than 95% of lifetime tasks approved by requesters. The experimental design was a 2 (male, female) by 3 (control, male identity prime, overconfidence prime) factorial design. I excluded six subjects (3% of whole sample) who did not write meaningful words in the priming task, leading to a final sample of 234 subjects. I assigned subjects in equal proportions to the three experimental conditions. The experimental procedure was as follows. In the first stage, subjects answered four background questions. To address the concern that a change in the behavior of control subjects could drive the results, all subjects recalled a situation in which they felt relaxed, which is the condition that the control manipulation aims to induce. In the second stage, the control and male prime conditions were the same as in Experiment 1. The overconfidence prime was the induction of a sense of power in Experiment 3. The third stage consisted of the investment decisions. I gave subjects a virtual endowment of $100 at the beginning of each of three periods. Each period, subjects faced an opportunity, and decided how much of their per-period endowment to invest, if any (Gneezy and Potters, 1997). The first opportunity succeeded with probability 1/2 and paid off three times the invested amount in case of success. The second and third opportunities succeeded with probability 1/6 and paid off seven times the invested amount. I presented the three opportunities to subjects in random order. They could not 17 invest less than zero, that is, pay to avoid a choice. A random-number generator calibrated to the objective probabilities of each opportunity determined good or bad outcomes. Subjects received feedback about the outcomes immediately after their choices. The aim was to verify the effects of identity and overconfidence were not confined to experiences of good outcomes, but they survived even after experiencing negative outcomes, which is what I find.28 Subjects were paid a scaled amount of what they were left with at the end of the experiment in the form of a bonus payment on top of the show-up fee. Results. Panel A of Figure 5 shows the average take-up rate across experimental conditions. Subjects in the male identity and power prime conditions were more likely to invest than controls. I estimate the following probit model to make statistical inference on the effects of primes on the extensive margin of investment: Pr(Invest = 1)iae = Φ(α + γ1 × M aleP rimeiae + γ2 × Overconf idenceP rimeiae + ηa + ηe ) (4) where Investiae equals 1 if subject i in age group a and education-level group e accepted to invest any positive amount, and zero otherwise. Φ(.) is the normal cdf, and covariates are defined as in Equation 1. Panel A of Table 4 reports the average marginal effects derived from the estimated coefficients. Standard errors are clustered at the subject level, because subjects made three investment decisions. Column (1) shows that men primed with male identity are 12 percentage points (s.e. 6.5 p.p.) more likely to invest than controls, who invest on average 65% of the times. The effect is robust to averaging out fixed age and education effects. Men primed with power invest 17 percentage points (s.e. 6.8 p.p.) more often than controls. Unreported results are similar if I estimate the marginal effects in a linear probability model. Panel B of Table 4 refers to the intensive margin of investment. The invested amount is censored to the right ($100) and to the left ($0), hence I estimate a tobit specification. Covariates are as in Equation 1. Standard errors are clustered at the subject level. Because estimated coefficients are within the censoring interval, I interpret them similarly to OLS coefficients. In column (1), subjects primed with male identity invest on average $16 28 The immediate feedback after the investments is also the reason why Experiment 4 and Experiment 5 have a between-subjects design: differentiating between treatment conditions across subjects that had experienced positive or negative outcomes in first stage investments would have increased the number of experimental cells, and eliminated the statistical power advantage of the within-subject design. 18 dollars more in each opportunity than controls, but the effect is not statistically significant. Averaging out age and education effects (column (2)) increases the size of the estimated coefficient to about $21 (s.e. $10), which is an economically and statistically significant effect. Subjects primed with power invest about $17 (s.e. $9) more than control subjects in the baseline specification of column (1). The size of the effect increases to about $23 (s.e. $10) once age and education fixed effects are averaged out. B Identity, Overconfidence, and Delegated Investments In Experiment 5, I test for the effects of identity on delegated investments in a setting where the conflict of interest between the principal and the agent is minimal. This setting is important to ensure that the overconfidence of agents predicts overinvestment, but there is only a minor scope for agency-based explanations of overinvestment. To reproduce this setting, I change the framing of the investment tasks to include a principal on whose behalf the subjects invest, which aims to mimic the investment choices of mutual fund managers or corporate executives in the field. If the subjects’ preferences are the same as those of the fictitious principal,29 the interests of the principal and agent are aligned in this setting, because the compensation of the agent is a fixed proportion of the principal’s revenues. Because of this feature, the experiment can be interpreted as a causal test of Malmendier and Tate (2005), where overconfident CEOs overinvest when they have access to cheap funds because they believe they act in the best interest of shareholders. I also allow for money-burning investment opportunities, because the field evidence on overconfidence and investment behavior by individual investors and executives has shown that the negative effects of overconfidence are mainly concentrated in money-burning investment projects, such as diversifying mergers (e.g., Malmendier and Tate, 2008) Design and Procedure. I recruited 220 subjects on Amazon Mechanical Turk (mTurk) in October 2012, and invited them to work on a survey on creative writing for $0.5, plus a chance to earn a bonus by performing simple investment decisions. I restricted the subject pool to Workers in the United States who had more than 95% of tasks approved by requesters. The experimental design was a 2 (male, female) by 2 (control, male prime) 29 The instructions for this experiment are in Appendix A. Importantly, the subjects were asked to imagine they were managing the money of Sally, a lady in their neighborhood who was unable to invest her own savings by herself, hence there was no deception on the nature of the principal in these investment games. 19 factorial design. Given the strikingly similar effects of overconfidence and identity primes on investment decisions in Experiment 4, and the similar categorization of stereotypes attached to a male individual and a powerful individual in Table 1, I increased the power of this test by only using one of the two treatment conditions, the male identity prime. I excluded seven subjects (3% of whole sample) who did not write meaningful words in the priming task, leading to a final sample of 213 subjects. I assigned subjects in equal proportions to the experimental conditions. The procedure was the same as in Experiment 4, except for the third stage. Investments were framed as delegated decisions: in each of two periods, subjects had to imagine they were given money by Sally, a lady in their neighborhood who wanted to invest her money instead of keeping it in her checking account, but was unable to do so. Each period, subjects faced an investment opportunity. They decided if and how much of Sally’s endowment they wanted to invest. Their compensation would be a performance-based fee, calculated as a fraction of the amount Sally ended up with after all the investment decisions. One opportunity succeeded with probability 1/2 and paid off 2.2 times the invested amount (sound investment). The other opportunity succeeded with probability 1/2 and paid off 1.8 times the invested amount (money-burning investment). A riskneutral agent would have invested her whole endowment in the first case, and nothing in the second case. The opportunities were presented in random order to subjects. Each period, subjects could not invest more than their endowment or less than zero. Results. Columns (3) and (4) of Panel A of Table 4 report the marginal effects implied by the estimated coefficients of the specification in Equation 4, but with the male-identityprime category only. Men whose identity is primed are 10.7 percentage points more likely than control men to invest (s.e. 5.0 p.p.). The effect is robust to controlling for age and education effects, and to computing the marginal effects with a linear probability model (untabulated). Panel B of Table 4 shows results for estimating a tobit model whose dependent variable is the invested amount in the two opportunities, censored at $0 at $100. Men whose identity is primed invest on average $27.4 more (s.e. $9.8) than controls. The effect on the intensive margin of investment is robust to controlling for age and education effects. In Panel B of Figure 4, I report the average invested amounts separately for the sound and the money-burning investments across experimental conditions. The 20 difference on the intensive margin is larger for the money-burning opportunity than for the sound opportunity, in line with the correlational evidence of Malmendier and Tate (2008). Note that this difference between opportunities is driven by the mistakes of primed men: whereas controls on average shy away from the money-burning opportunity, as they should, primed subjects invest similar amounts in the sound and the money-burning opportunities. 6 Identity vs. Biology: Effect Across Cohorts To better understand why identity affects risk attitudes and investment decisions, I move on to test if the effects change as the perception of gender stereotypes evolves. Alesina et al. (2013) show the perception of gender roles varies across countries. Fernandez and Fogli (2009) exploit the variation in cultural stereotypes across native and immigrant U.S. individuals. In my setting, all subjects are from the U.S., and I do not observe the immigration status of the subjects. I cannot exploit cross-regional variation in stereotypes within the U.S. because the IRB-approved protocol does not allow me to use the IP addresses of the subjects in the analysis. I therefore exploit the variation in gender identity stereotypes over time: with the gradual expansion of women’s rights, the roles of the two genders in society, and hence gender identity stereotypes, have become closer (Goldin et al., 2006: Doepke and Tertilt, 2009; and Glaeser and Ma, 2013). The genetic characteristics of individuals have not changed significantly over the last four decades. Instead, biological characteristics change as individuals age in the opposite direction of the perception of stereotypes. Since Williams (1966), evolutionary biology has argued that the risky behavior of male subjects of various species, including Homo Sapiens, is fueled by competition to access more mating opportunities. This fact is particularly true for young individuals, who have no established reputations within communities. A large body of neurological evidence (e.g., surveyed by Casey et al., 2008) finds that risk taking increases at puberty and decreases with age. Steinberg (2008) connects these results to the remodeling of the dopaminergic system at the time of puberty. Psychologists show the gender gap in risk taking is larger around puberty and decreases over time (e.g., see Byrnes et al., 1999). If the effect of identity salience is higher the higher the difference in male and female 21 stereotypes, the effects of male identity on subjects’ risk tolerance and investments should be larger for older cohorts of men. In Figure 6, I run the analyses of Experiment 1 and Experiment 4 separately for four cohorts of men. I look at subjects born before 1975, between 1976 and 1983, between 1984 and 1989, and between 1990 and 1994.30 Panel A of Figure 6 shows the effect of identity salience on the change in risk tolerance is highest for subjects born before 1975, and it fades away monotonically to become insignificantly negative for men born after 1990. A similar pattern exists for the effect of male identity on the probability to invest (Panel B of Figure 6). Panel C of Figure 6 shows the effect of male identity on the amount invested in Experiment 4 is also higher for those born earlier, although the decreasing pattern is not monotonic. These results also speak to the external validity of the effects, because older individuals are on average wealthier and are more likely to invest than the youngsters. The results across cohorts warrant two comments. First, I cannot disentangle the mediating effect of cohorts (change in the societal perception of identity stereotypes over time) from an effect of age (individual-level change in the perception of identity stereotypes from youth to adulthood). This caveat does not affect the interpretations of the results as long as individuals, if anything, become more conservative over time.31 But the two mechanisms have different implications for the identity effect over time. If the results are driven by cohort effects, the aggregate effect of identity on risk tolerance and investment decisions should fade away over time as more individuals with less stark views on gender roles replace the older cohorts (e.g., see Guiso et al. (2008) for a cross-country argument). If the results are partly driven by aging, the aggregate effects would fade away more slowly over time. Second, one may be concerned the results capture some form of regression to the mean. Young men are more likely to engage in risky behaviors than older men. Youngsters may be so risk tolerant that the manipulations have no effect on them. But the average risk tolerance of men before the manipulations, the average likelihood to invest, and the average invested amounts by controls do not decrease monotonically with age, which the regression-to-the-mean explanation requires. 30 The cohort-level analysis was designed after running the experiments, so that cohort boundaries are effectively exogenous for the purposes of the cross-cohort test. But I cannot perform robustness checks by varying the cohort boundaries, except from aggregating existing groups in alternative ways. 31 Biology may mediate the sensitivity to the priming: Samanez-Larkin et al. (2010) find nucleus accumbens activity explains the varying quality of financial choices by age. 22 7 Economic Interpretation of Priming and Threatening Identity Stereotypes In this section, I sketch an economic interpretation of the effects of priming identity on men based on Gennaioli, Bordalo, and Shleifer (2014), which helps to organize the results in the paper in a coherent setting. Table 1 reports the most common types a pool of 200 anonymous respondents associate with four identities: male, powerful, successful, and female individuals. The lists for male and female individuals are similar to those in the pioneering work of Williams and Bennett (1975). I interpret the enlisted characteristics as the most likely types associated with each identity. Because the most likely types for the male and female identities do not overlap, we can interpret them as the most representative types of each gender. In Bordalo, Gennaioli, and Shleifer (2014), a type is representative of a group if a Bayesian decision maker that observes the type thinks it was more likely drawn from the group than from its complement.32 I call M the set of representative types for men, and −M the set of representative types for women, where −M = Ω\M , and Ω, the union of male and female representative types, covers the universe of gender-related representative types.33 Note that the set of types in M needs not include the opposite of the types in −M . I follow Bordalo, Gennaioli, and Shleifer (2014) and define the stereotype for a group as a truncated probability distribution over types, where the most representative types are assigned a positive probability weight, and the decision maker neglects all other possible types. Thus, the stereotype is a set of beliefs decision makers hold about the probability distribution of types associated with the assessed group. The novelty of this paper is interpreting the act of priming as an exogenous shock to the probability distribution over the types the subjects associate with their identity. Before the act of priming, subjects assess their type based on all the social groups to which they belong, such as gender, social class, or religious affiliation. Because in general male subjects belong to different social groups, except for their gender group, we should observe that absent any intervention men have heterogeneous risk attitudes and beliefs. But after 32 33 This definition of representativeness builds on Kahneman and Tversky (1972.) There is no scope for gender-related types different from those associated with men or women. 23 the gender identity prime, subjects’ beliefs about their own types are determined by the most representative types of their gender identity, and the non-gender-related types for the other social groups to which the subject belongs are neglected. Importantly, I also interpret the act of threatening subjects’ gender identity as an exogenous shock to the probability distribution over the types the subjects associate with their own gender identity. This interpretation follows from the fact that the threats in the experiments consist of priming the other gender’s most representative types, hence for the case of men, −M . Making −M salient to men cannot have them relate to the types in −M , because those types are not part of men’s gender identity. Instead, priming −M makes it salient to subjects that they should assess their type based on the social group ”gender,” which, for the case of men, includes the representative types in M , and not in −M . This interpretation of gender identity threats is crucial to rationalize the same directional results of priming and threatening male identity on men’s willingness to engage in risky behaviors, which is documented in the social psychology literature (Maas et al., 2003; Willer et al., 2013), and is documented in this paper for the case of risk attitudes and investment decisions. Akerlof and Kranton (2000) argue that stereotypes prescribe normative behaviors to individuals. Hence, a link exists between gender-related types and the behaviors to which individuals conform. Suppose that any possible types associated with gender identities could be classified in a type associated with a ”winning behavior” or with a ”losing behavior.” Subjects will then form beliefs about the likelihood of winning as follows: P r(win) = p = f (πt ) where πt = πt(n) / P πt(n) for n ∈ {1, ..., N } and N are all possible types associated with the subject’s identity, across all the social groups to which he belongs. If a man faces a lottery with two possible outcomes, the definition above could be interpreted as his subjective probability of winning the higher payoff when entering the lottery. If the types associated with winning and losing have a similar weight in a man’s identity, this subjective probability would coincide with the objective probability of winning the higher payoff when entering the lottery. Based on the interpretation of priming described above, if a man is primed with male or female identity, he will assess his probability of winning 24 as follows: P r(win|M ) = pˆ = f (ˆ πt ) where π ˆt is now the male identity stereotype, that is, the probability distribution of P the most representative types that belong to the male identity: π ˆt = π ˆt(m) / π ˆt(m) for m ∈ {1, ..., M } ⊂ {1, ..., N }, and 0 otherwise. The set {1, ..., M } only includes the most representative types associated with the male identity group. The two subjective probabilities differ in general: P r(winner|M ) will be higher or lower than P r(winner), based on whether the most representative types for the male identity are associated with winning or losing. I propose the following hypothesis: Hypothesis 1. If the set of most representative male types {1, ..., M } includes more types associated with winning behaviors than with losing behaviors compared to the full set of possible types {1, ..., N }, then pˆ > p. If the subset of types {1, ..., M } is more likely to include types associated with winning than with losing, primed men become overconfident about their performance even in a pure game of chance, because they are more likely to believe they will experience high outcomes than what the objective probabilities for each state of the world suggest. This distortion of beliefs due to the priming is the crucial and novel hypothesis to interpret the results in this paper. I test directly for this hypothesis in Experiment 2, where I find that men whose identity is primed believe they are more likely than a group of peers to experience the winning outcome of a lottery, even if the objective probability of each future state of the world is known. Moreover, if men primed with their identity stereotypes become overconfident in a game of chance, they will choose lotteries over certainty equivalents more often after being primed than before. Intuitively, every time a subject decides whether to choose a lottery over a certainty equivalent, she will compare the expected utility from entering the lottery with the utility from the certain amount. Because pˆ > p, the subjective expected utility from entering the lottery after the prime is higher than the expected utility implied by the objective probabilities, as long as the subjects’ expected utility is strictly monotone: Hypothesis 2. If {1, ..., M } includes more types associated with winning than {1, ..., N }, 25 a primed man will require a higher certainty equivalent to give up the chance to take part in the same lottery compared to a non-primed man. I test Hypothesis 2 with a within-subject incentive-compatible design in Experiment 1, where I find that men primed either with salience or threat of their identity are more likely to choose lotteries over certainty equivalents after the manipulations compared to before them. Moreover, if priming male identity increases men’s subjective beliefs of experiencing higher outcomes, primed men should be more likely to invest in risky opportunities whose objective probabilities are known than non-primed men, because their expected utility from investing is higher than the one implied by the objective probabilities: Hypothesis 3. If {1, ..., M } includes more types associated with winning than {1, ..., N }, a primed man will be more likely to invest in a risky opportunity whose objective probability of success is known than a non-primed man. I test Hypothesis 3 in Experiment 4 for individual investment decisions, and in Experiment 5 for delegated investment decisions. In both cases, primed men invest more often and more money than controls. So far, the set of types associated with male identity, {1, ..., M }, was the same across men. Bordalo, Gennaioli, and Shleifer (2014) show how the change in the representativeness of types for each identity over time modifies the stereotypes associated with genders. I exploit the fact that identity stereotypes have become less stark over the last decades (Goldin et al., 2006; Doepke and Tertilt, 2009; Glaeser and Ma, 2013), to develop a hypothesis on the predicted magnitude of the effects of priming across cohorts with different perceptions of stereotypes: Hypothesis 4. If {1, ..., MOLD } includes more types associated with winning than {1, ..., MY OU N G }, the effects of priming identity on the behavior of men should be larger for older cohorts of men than younger cohorts of men. In Figure 6, I show evidence in line with Hypothesis 4 for the risk attitudes and investment decisions of men. 26 In the paper, I find no effect of priming or threatening gender identity on women. The non-result for women can be consistent with the setup above if the most representative types for female identity do not include more types associated with losing than with winning. The types enlisted in Table 1, as well as the more comprehensive lists of Williams and Bennett (1975), seem to include characteristics of both winners (smart, flirty, happy) and losers (passive, emotional, kind). But it is surely impossible to categorize the types, and there is no guidance on how to associate types to winning or losing behaviors in the theoretical framework. 8 Discussion and Conclusions I test whether social identity contributes to an explanation of the systematic heterogeneity in risk attitudes and investment behavior across genders, because gender identity prescribes opposite behaviors to men and women. Previous literature found conflictive answers to this question. In a set of artefactual field experiments, priming or threatening the identity of men increases their risk tolerance. Primed men invest more often and more money in risky opportunities, even when they act as agents of a principal, and especially in money-burning investments. The primes have no effect on the behavior of women. To the best of my knowledge, this is the first paper that tests for the economic channels that drive the effect of identity on risk attitudes. Manipulating identity salience affects men’s beliefs of experiencing positive outcomes in a game of chance. Moreover, the causal effects of identity on risk tolerance and investments decrease monotonically from older to younger cohorts of men, consistent with the attenuation of gender-related stereotypes over time. This result suggests different implications on the persistence of heterogeneous behaviors across genders than biology and genetics, which is effectively stable over time. I also analyze the interplay between identity and overconfidence, which I induce in subjects by priming their sense of power over others. Stereotypical behaviors attached to powerful individuals and to men are similar (Table 1). The investment behavior of men induced with overconfidence is qualitatively and quantitatively similar to that of men primed with male identity. 27 The results can be interpreted as a causal test of Barber and Odean (2001), who use male gender as a proxy for overconfidence. Results on delegated investments when the interests of principal and agent are perfectly aligned may be interpreted as a causal test of Malmendier and Tate (2005). At the same time, the results can be seen as an empirical test of Gervais, Heaton, and Odean (2011), who show that overconfident agents are less conservative than other agents when investing in sound opportunities. This finding is consistent with Hirshleifer, Low, and Teoh (2012). Consistent with the trade-off between the positive effects of self-serving beliefs and the risks of overconfidence of Benabou and Tirole (2002), the welfare implications of the results are driven by the quality of the opportunities individual investors and CEOs face.34 If the opportunities have a positive net present value, male identity and overconfidence benefit men by increasing their willingness to invest. But they cause men to overinvest in money-burning opportunities if not tied by liquidity constraints. These results can be the basis for research efforts that better characterize the sources of biased beliefs. Understanding the roots of biased beliefs has relevant implications for hiring firms who need to screen potential managers and employees, as well as to design policies that reduce the emergence of distorted beliefs if they hurt the decision makers. A fruitful avenue for future research is also to study the interactions between identity and biology to better understand financial decisions. mTurk does not allow to reliably collect information on the biological characteristics of subjects, but laboratories or field settings that the experimenter can physically reach may allow for such analysis. The priming techniques employed in the paper show that identity and overconfidence can be manipulated. Male identity and overconfidence cues may be used to design policies or compensation contracts that foster individual investors’ and executives’ willingness to take on financial risks. Figure 13 of Appendix B shows that financial firms seem to believe in the effects of priming the identity and overconfidence of potential clients to boost the take-up rates of their risky products. Future research should be devoted to testing for the causal effects of cues and priming on financial decisions in the field. 34 All claims of external validity are subject to the caveat that there is no uncertainty in the experiments in the paper 28 References G. Akerlof and R. Kranton. Economics and Identity. Quarterly Journal of Economics, 115(3):715–753, 2000. A. Alesina, P. Giuliano, and N. Nunn. On the Origins of Gender Roles: Women and the Plough. Quarterly Journal of Economics, 128(2):469–530, 2013. B. Barber and T. Odean. Boys will be Boys: Gender, Overconfidence and Common Stock Trading. Quarterly Journal of Economics, 116(1):261–292, 2001. I. Ben David, J. Graham, and C. Harvey. Managerial Miscalibration. Quarterly Journal of Economics, 128(4):1547–1584, 2013. R. Benabou and J. Tirole. Self-Confidence and Personal Motivation. Quarterly Journal of Economics, 117(3):871–915, 2002. D. Benjamin, J. Choi, and J. Strickland. Social Identity and Preferences. American Economic Review, 100(3):1913–1928, 2010. D. Benjamin, J. Choi, and G. Fisher. Religious Identity and Economic Behavior. Working Paper, 2013. A. Berinsky, G. Huber, and G. Lenz. Evaluating Online Labor Markets for Experimental Research: Amazon.com’s Mechanical Turk. Political Analysis, 20(3):351–368, 2012. B. Biais, D. Hilton, K. Mazurier, and S. Pouget. Judgemental Overconfidence, SelfMonitoring and Trading Performance in an Experimental Financial Market. Review of Economic Studies, 72(2):287–312, 2005. P. Bordalo, N. Gennaioli, and A. Shleifer. Stereotypes. NBER Working paper No. w20106, 2014. A. Boschini, A. Muren, and M. Persson. Constructing Gender Differences in the Economics Lab. Journal of Economic Behavior and Organization, pages 741–752, 2012. A. Boschini, A. Dreber, E. von Essen, A. Muren, and E. Ranehill. Gender and Economic Preferences in a Large Random and Representative Sample. Working Paper, 2014. T. Breda and S. Ly. Stereotypes, Discrimination, and the Gender Gap in Science. Working Paper, 2014. J.P. Byrnes, D.C. Miller, and W.D. Schafer. Gender Differences in Risk Taking: A Meta-Analysis. Psychological Bulletin, pages 367–383, 1999. C. Camerer and D. Lovallo. Overconfidence and Excess Entry: An Experimental Approach. American Economic Review, 89(1):306–318, 1999. P. Carr and C. Steele. Stereotype Threat Affects Financial Decision Making. Psychological Science, 21(10):1411–1416, 2010. 29 B.J. Casey, R.M. Jones, and T. Hare. The Adolescent Brain. Annals of the New York Academy of Sciences, pages 111–126, 2008. K. Casler, L. Bickel, and E. Hackett. Separate but Equal? A Comparison of Participants and Data Gathered via Amazon’s mTurk, Social Media, and Face-to-Face Behavioral Testing. Computers in Human Behavior, 29(6):2156–2160, 2013. D. Cesarini et al. Genetic Variation in Financial Decision Making. Journal of Finance, 65:1725–1754, 2010. J. Coates and J. Herbert. Endogenous Steroids and Financial Risk Taking on a London Trading Floor. Proceedings of the National Academy of Sciences, 105:623–628, 2008. K. Coffman. Evidence onl Self-Stereotyping and the Contribution of Ideas. Quarterly Journal of Economics, forthcoming. A. Cohn, M. Mar´echal, and T. Noll. Bad Boys: the Effect of Criminal Identity on Dishonesty. Working Paper, 2014. A. Cohn, J. Engelmann, E. Fehr, and M. Marechal. Evidence for Countercyclical Risk Aversion: An Experiment with Financial Professionals. American Economic Review, forthcoming. H. Cronqvist, A. Previtero, S. Siegel, and R. White. The Fetal Origin Hypothesis in Finance: Prenatal Environment and Financial Risk Taking. Working Paper, 2014. R. Croson and U. Gneezy. Gender Differences in Preferences. Journal of Economic Literature, 47(2):1–17, 2009. M. Doepke and M. Tertilt. Women’s Liberation: What’s in it for Men? Quarterly Journal of Economics, 124(4):1541–1591, 2009. T. Dohmen, A. Falk, D. Huffman, U. Sunde, J. Schupp, and G. Wagner. Individual Risk Attitudes: Measurement, Determinants, and Behavioral Consequences. Journal of the European Economic Association, 9(3), 2011. N. Fast, N. Sivanathan, N. Mayer, and A. Galinsky. Power and Overconfident Decision Making. Organizational Behavior and Human Decision Processes, 117(2):249–260, 2012. R. Fernandez and A. Fogli. Culture: An Empirical Investigation of Beliefs, Work, and Fertility. American Economic Journal: Macroeconomics, 1(1):146–177, 2009. A. Gandhi and R. Serrano-Padial. Does Belief Heterogeneity Explain Asset Prices: The Case of the Longshot Bias. Review of Economic Studies, forthcoming. N. Garleanu and S. Panageas. Young, Old, Conservative, and Bold: Implications of Heterogeneity and Finite Lives for Asset Pricing. Journal of Political Economy, forthcoming. 30 S. Gervais and T. Odean. Learning to be Overconfident. Review of Financial Studies, 14(1):1–27, 2001. S. Gervais, J.B. Heaton, and T. Odean. Overconfidence, Compensation Contracts, and Capital Budgeting. Journal of Finance, 66(5):1735–1777, 2011. E. Glaeser and Y. Ma. The Supply of Gender Stereotypes and Discriminatory Beliefs. NBER Working Paper No. 19109, 2013. U. Gneezy and Potters. An Experiment on Risk Taking and Evaluation Periods. Quarterly Journal of Economics, 112:631–645, 1997. C. Goldin, L. Katz, and I. Kuziemko. The Homecoming of American College Women. The Reversal of the College Gender Gap. Journal of Economic Persectives, 20(4):133–156, 2006. M. Grinblatt and M. Keloharju. Sensation Seeking, Overconfidence, and Trading Activity. Journal of Finance, 64(2):549–578, 2009. L. Guiso and M. Paiella. Risk Aversion, Wealth, and Background Risk. Journal of the European Economic Association, 6(6):1109–1150, 2008. L. Guiso, F. Monte, P. Sapienza, and L. Zingales. Culture, Gender, and Math. Science, 320:1164–1165, 2008. D. Hirshleifer, A. Low, and S. H. Teoh. Are Overconfident CEOs Better Innovators? Journal of Finance, 67(4):1457–1498, 2012. C. Holt and S. Laury. Risk Aversion and Incentive Effects in Lottery Choices. American Economic Review, 92:1644–1655, 2002. J. Horton, D. Rand, and R. Zeckhauser. The Online Laboratory: Conducting Experiments in a Real Labor Market. Experimental Economics, 14:399–425, 2011. D. Kahneman and A. Tversky. Subjective Probability: a Judgement of Representativeness. Cognitive Psychology, 3(3):430–454, 1972. C. Kuhnen and J. Chiao. Genetic Determinants of Financial Risk-Taking. PLosONE, 4(2), 2009. C. Kuhnen and B. Knutson. The Influence of Affect on Beliefs, Preferences and Financial Decisions. Journal of Financial and Quantitative Analysis, 46(3):605–626, 2011. I. Kuziemko, M. Norton, E. Saez, and S. Stantcheva. How Elastic are Preferences for Redistribution? Evidence from Randomized Survey Experiments. NBER Working Paper 18865, 2013. J. List. Why Economist Should Conduct Field Experiments and 14 Tips for Pulling One Off. Journal of Economic Perspectives, 25(3):3–16, 2011. 31 A. Maas, M. Cadinu, G. Guarnieri, and A. Grasselli. Sexual Harassment under Social Idetity Threat: the Computer Harassment Paradigm. Journal of Personality and Social Psychology, 85(5):853–870, 2003. U. Malmendier and G. Tate. CEO Overconfidence and Corporate Investments. Journal of Finance, 60(6):2661–2700, 2005. W. Mason and S. Suri. Conducting Behavioral Research on Mechanical Turk. Behavioral Research Methods, 44(1):1–23, 2012. D. Mastripieri, P. Sapienza, and L. Zingales. Gender Differences in Financial Risk Aversion and Career Choices are Affected by Testosterone. Proceedings of the National Academy of Sciences, pages 526–556, 2009. K. Meier-Pesti and E. Penz. Sex or Gender? Expanding the sex-based view by introducing masculinity and femininity as predictors of financial risk-taking. Journal of Economic Psychology, 29:180–196, 2008. W. Mischel. A Social Learning View of Sex Differences in Behavior. The Development of Sex Differences, E.E. Maccoby (Ed.) - Stanford University Press:56–81, 1966. M. Niederle and L. Vesterlund. Do Women Shy away from Competition? Do Men Compete too Much? Quarterly Journal of Economics, 122(3):1067–1101, 2007. G. Paolacci, J. Chandler, and P. Ipeirotis. Running Experiments on Amazon Mechanical Turk. Judgment and Decision Making, 5(5):411–419, 2010. E. Reuben, P. Sapienza, and L. Zingales. How Stereotypes Impair Women’s Careers in Science. Proceedings of the National Academy of Sciences, 111(12):4403–4408, 2014. G. Samanez-Larkin, C. Kuhnen, D. Yoo, and B. Knutson. Variability in Nucleus Accumbens Activity Mediates Age-Related Suboptimal Financial Risk Taking. Journal of Neuroscience, 30(4):1426–1434, 2010. M. Shih, T. Pittinsky, and N. Ambady. Stereotype Susceptibility: Identity Salience and Shifts in Quantitative Performance. Psychological Science, 10(1):80–83, 1999. U. Simonsohn. Just Post It: The Lesson From Two Cases of Fabricated Data Detected by Statistics Alone. Psychological Science, 24(10):1875–1888, 2013. C. Steele. A Threat in the Air: How stereotypes shape intellectual identity and performance. American Psychologist, 52:613–629, 1997. L. Steinberg. A Social Neuroscience Perspective on Adolescent Risk-Taking. Developmental Review, pages 78–106, 2008. R. Willer, C. Rogalin, B. Conlon, and M. Wojnowicz. Overdoing Gender: a Test of the Masculine Overcompensation Thesis. American Journal of Sociology, 118:980–1022, 2013. G.C. Williams. Adaptation and Natural Selection. Princeton University Press, 1966. 32 25 30 Latitude 35 40 45 50 Figure 1: Location of Subjects -120 -110 -100 Longitude Experiment 2 -90 -80 -70 Experiment 3 Figure 1 depicts the location of the subjects in Experiment 2 and Experiment 3 based on their IP address. 33 Figure 2: Risk Tolerance Elicitation and Measure Risk Neutral Agent choice: RN choices = 10 Subject’s choice: # choices = 13 Risk Tolerancei = 10-13 = -3 Figure 2 describes the construction of the Risk Tolerance measure based on the subject’s choices. Subjects face a set of screens, each including a variable number of choices between a decreasing certainty equivalent and a fixed lottery. The Risk Tolerance measure is the difference between the number of times a risk neutral agent would choose the certainty equivalent in the screen, and the number of times the subjects chooses the certainty equivalent in the screen. 34 Figure 3: Changes in Risk Tolerance induced by the Identity Manipulations Density Risk Tolerance before manipulations 0 .05 .1 .15 .2 Cumulative Distribution Risk Tolerance before manipulations 0 .2 .4 .6 .8 1 (a) Distribution of Risk Tolerance before the manipulations (male subsample) -10 -5 0 Control Female Prime 5 Male Prime -10 -5 0 5 risklevpre Control Female Prime Male Prime Cumulative Change in Risk Tolerance .2 .4 .6 .8 .4 .3 .2 0 .1 -4 0 Estimated Density Change in Risk Tolerance 1 (b) Distribution of ∆ Risk Tolerance induced by the manipulations (male subsample) -4 -2 0 2 4 6 -2 0 Control Female Prime 2 4 6 Male Prime Figure 3 plots the estimated densities and cumulative distributions for the risk tolerance before the manipulations (panel A) and the subject-level change in risk tolerance before and after exposure to the experimental conditions (panel B) in Experiment 1 for male subjects. The subject-level change in risk tolerance is measured as follows: ∆RiskT olerancei = RiskT olerancepost,i − RiskT olerancepre,i , where RiskT olerancepre,i and RiskT olerancepost,i are elicited via lottery choices a la Holt and Laury (2002). The smooth density plots use an Epanechnikov kernel with bandwidth of 0.7. The cumulative distribution plots interpolate the medians of 10 vertical bands with cubic splines. The plots of the distributions for the female subsample are in Appendix B. 35 Figure 4: Identity Manipulations and Beliefs under Risk 4.6 Predicted Wins if Lottery played 10 times (p=0.5) 4.8 5 5.2 5.4 5.6 (a) Subjective beliefs of experiencing the good outcome in a lottery that succeeds on average 5 out of 10 times Control Men Primed Men -.4 Own minus Neighbor Predicted Wins (p=0.5) -.2 0 .2 .4 .6 (b) Better-than-average beliefs across experimental conditions (men only) Control Men Primed Men Panel (a) of Figure 4 plots the subjective beliefs of experiencing the good outcome by gender for subjects who imagine to play 10 times a lottery whose good outcome obtains on average 5 out of 10 times. Bandwidths are 95% confidence intervals for estimated means. Panel (b) of Figure 4 plots the average subjective better-than-average beliefs for men in the control group and in the primed group, which collects the male identity (salience) and female identity (threat) primes. 36 Figure 5: Identity, Extensive and Intensive Margins of Investment 1 Number of Times Invested (out of 3) 2 3 (a) Times invested out of 3 opportunities (men only) Control Male Prime Power Prime 30 Amount Invested (out of $100) 40 50 60 70 (b) Dollars Invested in Delegated Opportunities (men only) Control Sound Male Prime Sound Control Bad Inv. Male Prime Bad Inv. Panel (a) of Figure 5 reports the average number of times that male subjects invest out of three potential opportunities in Experiment 3, where all the investment opportunities have a positive NPV, across experimental conditions. Panel (b) of Figure 5 reports the average amount of money (in experimental dollars) the subjects invest in Experiment 5 across the experimental conditions, and separately for sound and money-burning investment opportunities. 37 Figure 6: Identity vs. Biology: Effects of Identity across Cohorts -.2 Effect of Male prime on Change in Risk Tolerance 0 .2 .4 .6 .8 (a) Effect of Male Identity prime on Risk Tolerance by cohorts (men only) Born before 1975 Born 1976-1983 Born 1984-1989 Born 1990-1994 Marginal Effect of Male Prime on Probability to Invest -.1 0 .1 .2 .3 (b) Effect of Male Identity prime on Probability to Invest by cohorts (men only) Born before 1975 Born 1976-1983 Born 1984-1989 Born 1990-1994 -20 Effect of Male Prime on Amount Invested 0 20 40 60 (c) Effect of Male Identity prime on Amount Invested by cohorts (men only) Born before 1975 Born 1976-1983 Born 1984-1989 Born 1990-1994 Panel (a)-(c) of Figure 6 plot the effect of the male identity manipulations on the change in risk tolerance in Experiment 1, the likelihood that subjects invest in Experiment 4, and the average amount invested in Experiment 4, across four cohorts of subjects based on their year of birth. 38 Figure 7: Manipulation Check for Experiment 1 0 Number of Male Types from Table 1 in Essay 1 2 3 (a) Average Number of Male Types from Table 1 reported in Experiment 1 Control Male Prime Female Prime 0 Number of Female Types from Table 1 in Essay 1 2 3 (b) Average Number of Female Types from Table 1 reported in Experiment 1 Control Male Prime Female Prime Figure 7 reports the results for the manipulation check for Experiment 1. Panel (a) shows the average number of times each subject reported any of the types related to male individuals enlisted in Table 1 in their essays, by experimental conditions. Panel (b) shows the average number of times each subject reported any of the types related to female individuals enlisted in Table 1 in their essays, by experimental conditions. The bars represent 95% confidence intervals for the arithmetic mean within each group of subjects. 39 Figure 8: Testing for the Scope of Demand Effects (a) Reported Motives for the Reading and Essay Tasks (100 subjects) 35 28 21 14 7 0 No Idea Distraction Collect Opinions Check if Human Relate to Manipulation Lotteries (b) Reported Motives for the Lottery Tasks (100 subjects) 30 25 20 15 10 5 0 No Idea For Bonus Won't Use Measure Risk Relate to Change in text Risk Figure 8 reports results for 100 subjects recruited in June and July 2014 to perform the tasks in Experiment 1 following the same recruitment procedures as described in the paper. These subjects were also asked to guess the motivations of the tasks at the end of the study. Panel (a) reports the distribution of answers for why subjects had to read a text and write a short essay, split into six broad categories. The dark histogram to the right (8 subjects) shows the answers compatible with the aim of the study. Panel (b) shows the distribution of answers for why subjects had to make lottery choices before and after the writing task. The emphasis on before and after aimed to instill the idea that the order of tasks was related to their motives. The only category compatible with the true motives includes 9 subjects (dark, right), 8 of which are those depicted in dark in Panel (a). 40 Table 1: Characterization of Stereotypes MALE POWERFUL SUCCESSFUL FEMALE 41 Strong 37 Strong 50 Confident 46 Caring 40 Loud 27 Confident 45 Smart 36 Emotional 34 Aggressive 26 Arrogant 19 Happy 26 Flirty 16 Confident 18 Aggressive 18 Proud 19 Passive 15 Arrogant 15 Assertive 15 Determined 15 Smart 13 Dumb 12 Dominant 13 Focused 12 Loving 13 Stubborn 11 Greedy 9 Motivated 11 Happy 12 Assertive 11 Cocky 8 Assertive 10 Gossipy 12 Proud 10 Proud 8 Ambitious 10 Helpful 10 Dominant 10 Smart 7 Strong 10 Kind 10 1 reports the words mentioned most frequently by 200 survey respondents recruited on mTurk to describe how Table they think a typical male individual, a powerful individual, a successful individual, or a female individual behave. Respondents wrote 3 to 5 words for each type of individual. Types were presented in a random order. The numbers next to each stereotypical semantic area are the number of times respondents reported a word in that area. Table 2: Experimental Designs and Sample Descriptions Treatment Design Sample Experiment 1 Identity Salience, 2 (male-M, female-F) Risk Attitudes Threat X 3 (control-C, male identity-MI, female identity-FI) MC:55, FC:74, MMI-61, FMI:64, MFI:32, FFI:34. Total:320 Experiment 2 Identity Salience, 2 (male-M, female-F) Threat X 3 (control-C, male identity-MI, female identity-FI) Experiment 3 Induction of 2 (male-M, female-F) Risk Attitudes Overconfidence X 3 (control-C, overconfidence-O, success-S) Experiment 4 Identity Salience, 2 (male-M, female-F) Investments Induction of Overconfidence X 2 (control-C, male identity-MI, overconfidence-O) Experiment 5 Identity Salience 2 (male-M, female-F) Beliefs Investments 42 X 3 (control-C, male identity-MI, overconfidence-O) MC:50, FC:58, MMI:49, FMI:53, MFI:53, FFI:60. Total:323 MC:56, FC:51, MMI-54, FMI:53, MO:59, FO:50. Total:323 MC:35, FC:43, MMI:38, FMI:41, MO:32, FO:45. Total:234 MC:48, FC:60, MMI:49, FMI:55. Total:212 Table 2 describes the treatments, experimental designs, and sample compositions for all the experiments in the paper. On mTurk, one cannot discriminate by gender at the recruitment stage, and about 60% of U.S. mTurk Workers are women (Mason and Suri (2011)). Table 3: Identity, Overconfidence, and Risk Taking Experiment 1: Identity Experiment 3: Overconfidence (1) (2) Male Prime 0.507 (0.246)** 0.577 (0.264)** Overconfidence Prime Female Prime 0.431 (0.205)** 0.564 (0.258)** Mean omitted Median omitted Session f.e. Age group f.e. Education f.e. Observations R2 -0.212 0 148 0.03 -0.212 0 X X X 148 0.09 (1) (2) Male Prime -0.337 (0.280) -0.286 (0.268) Female Prime -0.163 (0.262) -0.089 (0.271) Man* Male Prime 0.844 (0.373)** Man* Female Prime Man Panel A. Men Only Panel B. Full Sample Mean omitted Median omitted Session f.e. Age group f.e. Education f.e. Observations R2 (3) (4) 0.519 (0.259)** 0.595 (0.270)** Success Prime 0.014 (0.260) -0.005 (0.262) Mean omitted Median omitted Session f.e. Age group f.e. Education f.e. Observations R2 -0.332 -0.167 169 0.03 -0.332 -0.167 X X X 168 0.12 (3) (4) Overconfidence Prime 0.077 (0.331) 0.122 (0.348) Success Prime 0.263 (0.321) 0.274 (0.328) 0.836 (0.375)** Man* Overconfidence Prime 0.442 (0.420) 0.394 (0.444) 0.594 (0.333)* 0.659 (0.344)* Man* Success Prime -0.249 (0.414) -0.292 (0.417) -0.257 (0.212) -0.329 (0.223) Man 0.048 (0.320) 0.103 (0.331) 0.045 0 0.045 0 X X X 320 0.10 Mean omitted Median omitted Session f.e. Age group f.e. Education f.e. Observations R2 -0.536 -0.333 -0.536 -0.333 X X X 322 0.08 320 0.02 323 0.02 Columns (1) and (2) of Table 3 report the estimated coefficients from: ∆RiskT oleranceiae = α + γ1 × M aleP rimeiae + γ2 × F emaleP rimeiae + ηa + ηe + iae Columns (3) and (4) report the estimated coefficient from: ∆RiskT oleranceiae = α + γ1 × Overconf idenceP rimeiae + γ2 × SuccessP rimeiae + ηa + ηe + iae where ∆RiskT oleranceiae is the within-subject change in risk tolerance for subject i, in age group a and education group e, before and after being exposed to the treatments in Experiment 1 and Experiment 2. ηa and ηe are fixed effects for subjects’ age and education-level groups. Hubert-White s.e. are reported in brackets. Statistical significance is reported as follows: ** 5%, * 10%. 43 Table 4: Identity, Overconfidence, and Investment Decisions Experiment 4: Individual Investments Experiment 5: Delegated Investments (1) (2) (3) (4) Male Prime 0.123 (0.065)* 0.136 (0.063)** 0.107 (0.050)** 0.105 (0.048)** Overconfidence Prime 0.170 (0.068)** 0.173 (0.069)** 0.65 1 0.84 1 315 0.03 0.65 1 X X X 315 0.09 194 0.05 0.84 1 X X X 194 0.06 (1) (2) (3) (4) Male Prime 15.66 (9.92) 20.59 (9.72)** Overconfidence Prime 16.97 (9.32)* 22.77 (9.77)** 41.7 29.0 41.7 29.0 X X X 315 0.03 Panel A. Probability to Invest Mean omitted Median omitted Success prob. f.e. Age group f.e. Education f.e. Observations pseudo-R2 Panel B. Amount Invested Mean omitted Median omitted Success prob. f.e. Age group f.e. Education f.e. Observations pseudo-R2 315 0.01 Male Prime Mean omitted Median omitted Success prob. f.e. Age group f.e. Education f.e. Observations pseudo-R2 Male Prime Mean omitted Median omitted Session f.e. Age group f.e. Education f.e. Observations pseudo/R2 27.35 (9.82)*** 24.99 (9.57)** 53.3 50.0 53.3 50.0 X X X 194 0.02 194 0.01 Panel A of Table 4 reports the estimated coefficients from: Pr(Invest = 1)iae = α + γ1 × M aleP rimeiae + γ2 × Overconf idenceP rimeiae + ηa + ηe ) where ηa and ηe are fixed effects for subjects’ age and education-level groups. Panel B of Table 4 reports the estimated coefficients from a tobit regression of the amount invested in each opportunity (censored at $0 and $100) on the Male identity prime and Overconfidence prime indicators from Experiment 3 and Experiment 4. S.e. are clustered at the subject level. Statistical significance is marked as follows: ** 5%, * 10%. 44 Appendix A - Experimental Materials This Appendix provides part of the materials used in Experiments 1 through 4. Figure 8 shows the screenshot subjects faced any time they had to perform a risk aversion elicitation lottery task a` la (Holt and Laury (2002)). For each line, subjects were asked to choose between a lottery paying a strictly positive amount with probability 50%, zero otherwise, and a certain amount. The certain amount decreases from one line to the other, while the lottery is fixed. The number of times a subjects chooses the certainty equivalent is subtracted from the number of times a risk neutral agent would pick the certain amount over the lottery to obtain a measure for the subject’s risk tolerance (see Section 5.B). In Experiments 1 and 2, subjects perform lottery tasks before and after being exposed to the experimental condition, so as to obtain a within-subject measure of the change in risk tolerance due to the experimental treatment. Figure 9 shows the screenshot subjects faced if assigned to the male identity priming condition. It is an excerpt from an internet blog, which makes the treatment as close as possible to daily-life experiences by subjects, who are regular internet users. The excerpt describes a series of characteristics attributed to the “masculine side” of life, which are similar to the traits survey responders in Table 1 think best describe how a typical male individual behaves. Figure 10 shows an analogous screenshot subjects faced if assigned to the female identity priming condition. The excerpt describes a series of characteristics attributed to the “feminine side” of life. Finally, Figure 11 shows the screenshot subjects faced if assigned to the control condition. The excerpt, also from a blog, describes lifestyle according to the “ayurveda principles”. In all three cases, after reading the excerpt, subjects were asked to recall a situation when the behaved in line with the principles presented in the text, and describe the situation, their thoughts and feelings in detail in a short essay (5 to 10 sentences). Computing a within-subject measure of change in risk tolerance helps to ensure results are not driven by individuals exposed to the control condition. To further address this concern, all subjects were asked to recall a situation when they felt relaxed, and describe it in detail before being exposed to any experimental condition in Experiment 3 and Experiment 4. 44 Survey | Qualtrics Survey Software Ignore Validation Click Here to Start Over Do Not Show Hidden Questions Previewing Survey Figure 8: Set of lottery pairs choice This picture reports the screenshot of one of the sets of lottery pair choices subjects were faced with in Experiment 1 and Experiment 2. Lottery pairs are a degenerate version of those introduced by Holt and Laury (2002). For each line, please pick one choice by filling the corresponding circle: $6 for sure 50% -- $7 $5.7 for sure 50% -- $7 $5.4 for sure 50% -- $7 $5.1 for sure 50% -- $7 $4.8 for sure 50% -- $7 $4.5 for sure 50% -- $7 $4.2 for sure 50% -- $7 $3.9 for sure 50% -- $7 $3.6 for sure 50% -- $7 $3.3 for sure 50% -- $7 $3 for sure 50% -- $7 $2.7 for sure 50% -- $7 $2.4 for sure 50% -- $7 $2.1 for sure 50% -- $7 $1.8 for sure 50% -- $7 $1.5 for sure 50% -- $7 $1.2 for sure 50% -- $7 $0.9 for sure 50% -- $7 $0.6 for sure 50% -- $7 $0.3 for sure 50% -- $7 $0 for sure 50% -- $7 Survey Powered By Qualtrics https://dc-viawest.qualtrics.com/SE/?SID=SV_9H92b78shkIfoeV&Preview=Survey&BrandID=ucbpsych[10/7/2012 2:48:31 PM] 45 Previewing Survey Ignore Validation Do Not Show Hidden Questions Click Here to Start Over Figure 9: Male Identity Priming text This picture reports the screenshot of the text subjects in the male identity prime treatment were asked to read. It is an excerpt from a blog entry on the internet. After reading the excerpt, subjects were asked to recall a situation when the behaved in line with the “masculine side” as presented in the text, and describe the situation, their thoughts and feelings Please read the following short textincarefully. In (5 the screens you will be asked three in detail a short essay to next 10 sentences). easy questions based on the text. It should take you about 10 minutes to complete this section. The Masculine Side The Masculine Side deals with the strength of the self. It is what causes you to act either timidly or self-confidently. The thing that is most important in determining the strength of the masculine side, is the value that you, at a deep level, place on yourself. This is a value you know within yourself that you have really and truly earned. It could be thought of as a sort of self esteem. Placing a high value on yourself affects your whole being and helps you feel strong and confident in operating your life. And, in the reverse direction, when you are able to operate your life confidently, things can really turn around for you because you get more out of life, and this automatically makes you place a higher value on yourself. You can build the masculine side through progress and small wins, through positive reinforcement, by practicing, and by doing things and generally taking an active part in operating your life. If you have a strong masculine side, you are in charge of your own life because you are internally controlled. You tend to look people in the eye. You stand straight, and you usually command attention when you walk into a room, whether you say anything or not. This happens because of the strength within. […] if you have a strong masculine side you are self-confident, and don't feel it is necessary to show off. The masculine side is full of things that you have to be strong and self-confident in order to do. These include being able to claim your basic rights, such as the right to feel free to operate independently of others, and the right to belong or fit into society in any way you please. Claiming your rights also includes being able to stand up to people who try to take away your rights, either by force or intimidation, or by manipulation, or by trying to hinder you in choosing your own direction in life. The masculine side also includes the ability to take risks when appropriate, to be decisive when necessary, and to focus or concentrate in order to get something done. In addition, it includes being able to figure out how to accomplish things so you can get more of what you want out of life. Part of this is being able to figure out how to operate your life in a responsible manner, how to reason without distorting reality and without fooling yourself, and how to accurately weigh probabilities so that you know the most likely outcome to expect in situations you come across. Extract from http://www.lovesedona.com/02.htm. https://new.qualtrics.com/SE/?SID=SV_9LlRDmIFHq6VSqU&Preview=Survey&BrandID=ucbpsych[5/4/2012 8:02:25 PM] 46 Previewing Survey Do Not Show Hidden Questions Click Here to Start Over Figure 10: Female Identity Priming text This picture reports the screenshot of the text subjects in the female identity prime treatment were asked to read. It is an excerpt from a blog entry on the internet. After reading the excerpt, subjects were asked to recall a situation when the Please thethe following text carefully. In the screens you will asked three behaved in read line with “feminine short side” as presented in the text, andnext describe the situation, theirbe thoughts and feelings in easy questions based on the text. detail in a short essay (5 to 10 sentences). It should take you about 10 minutes to complete this section. The Feminine Side […] the Feminine Side is based, […] on a value that you place on others. It could be thought of as a sort of other esteem. The value you place on others affects your whole being. If you have a strong feminine side and place a high value on others, you are often giving and unselfish. You usually know what is good for people, and you tend to operate in ways that help others get what they want out of life. You happily let people operate their own life without interference from you, but when asked, you are also willing to help by supporting, cooperating, and giving advice. People feel comfortable with you because you give them who you are without pushing yourself on others. If you have a strong feminine side, people also feel comfortable being around you because there is no selfishness for them to detect. [… ] If you have a strong feminine side, you often behave in ways that are considered feminine in nature. You do things you have to be giving and unselfish in order to do. These include recognizing the basic right of all people to use their own will to operate their own life, for example, by allowing them freedom to operate independently, freedom to fit in where and how they want, and freedom to choose what things to confront or face up to in life. Allowing people their basic rights also includes allowing them to control their own life without interference from you, to choose their own obligations in life without being manipulated by you, and to choose their own path or direction in life without hindrance from you. The feminine side also includes having enthusiasm and zest for life, and recognizing what things are worth getting enthusiastic about. And it includes having the persistence and tenacity to stay with things to the end, while still knowing when to give up on something if your energy is better used elsewhere. In addition, the feminine side also includes being kind, compassionate, patient, responsive to the needs of others, and it includes knowing how much energy you can put into each of these without hurting yourself by draining your own energy. Extract from http://www.lovesedona.com/02.htm. https://new.qualtrics.com/...icsSurveyEngine/?SID=SV_9LlRDmIFHq6VSqU&SVID=&Preview=Block&ID=BL_0oG4krywq4d8CQQ[5/4/2012 8:19:51 PM] 47 Ignore Validation Previewing Survey Do Not Show Hidden Questions Click Here to Start Over Figure 11: Control Condition text This picture reports the screenshot of the text subjects in the control condition were asked to read. It is an excerpt from a blog entry on the internet. After reading the excerpt, subjects were asked to recall a situation when the behaved in line with the “ayurveda principles” as presented in the text, and describe the situation, their thoughts andasked four feelings in detail in Please read the following short text carefully. In the next screens you will be a short essay (5 to 10 sentences). easy questions based on the text. It should take you about 10 minutes to complete this section. Dincharya [Daily Routine] In Sanskrit, the word 'dincharya' means daily routine. According to Ayurveda, one should follow the dincharya in order to lead a healthy and disease-free life. Everyday, two cycles of change pass through the human body, each bringing a Vata, Pitta, or Kapha predominance. Based on the cycles of vata, pitta and kapha, our daily routine should be divided into morning, noon, evening/twilight, dinner and bedtime. In the Ayurvedic texts, it is written that a person should wake up two hours prior to the sunrise, if he/she is not suffering from any diseases such as fever or diarrhea. Very young, very old and sick people are some of the exceptions. According to dincharya, the day should be kick-started by eliminating the colon and the bladder, followed by a through cleaning of the senses - ears, eyes, mouth etc. This should be followed by an oil self massage. Exercise in the morning, just after the massage, helps rejuvenate the body and soul. After bathing, one should head towards the dining table for breakfast. The day follows by activities like studying, working or traveling. During the lunch, one should consume nutritious meal […]. Dinner should consist of a light meal. Before going to bed, one should sit back and relax. By following the dincharya of Ayurveda, one can ensure a healthy life. Though it is difficult to follow a stringent dinacharya in this fast moving life, it is highly recommended by Ayurvedic physicians, because a number of health benefits are associated with it. The dinacharya makes one to lead a healthy and disciplined life. According to the latest studies in the field of medical science, people who stick to the daily routine are more fit than those, who do not have a particular time to perform their everyday activities. It is said that dinacharya reduces the stress level to a great extent. In addition to this, the person's body is purified and detoxified. Therefore, barring a few exceptions like sickness, very old and young age, Ayurvedic dinacharya is recommended for everyone. Extract from http://ayurveda.iloveindia.com/dincharya/index.html I am now ready to answer four questions based on the text above. https://new.qualtrics.com/...ricsSurveyEngine/?SID=SV_9LlRDmIFHq6VSqU&SVID=&Preview=Block&ID=BL_b9HHSp9g5S0ZnFi[5/4/2012 8:14:48 PM] 48 This page reports sample essays written by subjects in the four conditions across the experiments presented in the paper: priming relax, male identity, female identity or power over other individuals (any typos as in the subjects’ entries. Control condition: I was at our lake house. The kids were reading and taking care of themselves. The lake was calm and the bugs were scarce. All you could hear was the occasional boat, which to me is relaxing. The sound of the tiny waves lapping at the shore in the twilight was music to my ears. I had a book, but spent most time looking out at the mountains and clear water. Priming male identity: It was a typical weekend after work. [...] I was with a group of friends in a bar and in walked the hottest group of girls that night. Every guy immediately turned their heads to look, including me, and the girls knew it and loved it. However, everyone quickly averted their glances thinking they were out of league. However, I remained calm and kept my steady gaze until the one I was eyeing saw me in her sweep of the room. She looked back, cocked her head slightly, and I saw a faint smile forming at the edge of her lips. After about half an hour, I confidently walked up to the group of girls, straight to the girl I picked out, and asked her to join me at the bar for a drink. The rest is history. Priming female identity: My best friend was recently laid off from his job and his lease to his apartment was set to expire. Although I live with 3 other roomates, I allowed my friend to stay in my apartment for 4 months until he got back on his feet with a full-time job. I felt a sense of responsibility to be compassionate and help him in this situation. Priming Power over other individuals: When I became the executive director of my actual company, my attributions were many. I was almost the most powerful person. I could do whatever I wanted, but I did only what was fair for every employee. Everyone got what they deserved and they seemed satisfied 49 with me. I felt very capable and skillful. I did exactly what was right. 50 Appendix B - Additional Results 51 Table 5: Experiment 1 - Using Post-manipulation Risk Tolerance only (No Within-subject Design) The left panel shows results for the level of risk tolerance after the experimental manipulations in the male subsample in Experiment 1. The right panel shows results for the level of risk tolerance after the experimental manipulations for the all subjects in Experiment 1. Statistical significance is marked as follows: *** 1%, ** 5%, * 10%. (1) Male Prime Female Prime (2) Men only (3) (4) (6) 0.803 0.821 0.918 -0.134 -0.138 -0.008 (0.488) (0.501) (0.526)* (0.474) (0.474) (0.462) 0.937 0.907 0.934 -0.337 -0.407 -0.371 (0.431)** (0.491)* (0.500)* (0.399) (0.396) (0.400) Man*M. Prime Man*F. Prime Man 0.894 0.866 0.733 (0.681) (0.689) (0.689) 1.034 1.039 1.040 (0.596)* (0.587)* (0.594)* 0.087 (0.483) Mean omitted Median omitted Session f.e. Age group f.e. Education f.e. Observations R2 (5) Full Sample X 148 0.08 -2.01 -1.67 X X 148 0.08 X X X 148 0.09 52 X 320 0.06 0.080 (0.482) -2.25 -2.5 X X 320 0.07 0.074 (0.484) X X X 320 0.11 Experiment 2 20 40 60 80 100 120 Experiment 1 61+ 18-22 23-28 29-35 36-45 46-60 61+ Experiment 4 0 20 40 60 80 100 120 Experiment 3 0 20 40 60 80 100 120 18-22 23-28 29-35 36-45 46-60 0 0 20 40 60 80 100 120 Figure 12: Age Distribution of Subjects across the Experiments 18-22 23-28 29-35 36-45 46-60 61+ 18-22 23-28 29-35 36-45 46-60 53 61+ Figure 13: Identity and Overconfidence cues in financial products ads The set of pictures below are examples of financial products ads across countries and for different types of products with male identity and overconfidence cues. 54 Figure 14: Risk Tolerance Change and Beliefs in the Female Subsamples Panel A depicts estimated densities for the change in risk tolerance measured at the subject level across experimental conditions in Experiment 1 (Identity prime) for female subjects. The change in risk tolerance is measured as follows: ∆RiskT olerancei = RiskT olerancepost,i − RiskT olerancepre,i , where RiskT olerancepre,i and RiskT olerancepost,i are elicited via lottery choices a la Holt and Laury (2002) before and after exposure to the experimental conditions, respectively. Panel B depicts the average measure of “better-than-average” beliefs constructed in Experiment 4 for the female subsample across experimental conditions. Estimated Density Change in Risk Tolerance - Women 0 .1 .2 .3 .4 A. Distribution of ∆Risk Tolerance (female subsample) -6 -4 -2 0 2 4 x Control Female Prime Male Prime -.6 -.4 -.2 0 .2 .4 .6 B. Better-than-Average Beliefs (female subsample) Control 55 Male and Power Prime Table 6: Investment Decisions by the Female Subsamples Panel A shows results for the investment decisions made by women in Experiment 3. The left panel reports marginal effects computed from the following probit model: P r(Invest = 1)iae = Φ(α + β × M aleP rime + γ × Overconf idenceP rime + ηa + ηe ) where Invest is 1 if subject i in age group a and education level group e invests any positive amount, zero otherwise. Φ(.) is the normal cdf. M aleP rime and Overconf idenceP rime equal one for subjects exposed to the male identity or power prime; ηa and ηe are fixed effects for subjects’ age and education level groups for female subjects. The right panel reports results for a tobit model whose dependent variable is the amount of money a subject invests in each opportunity, which is censored at $0 and $100. Standard errors are clustered at the subject level, and computed with the delta method. Statistical significance is marked as follows: *** 1%, ** 5%, * 10%. A. Experiment 3: Investment Decisions (female subsample) Female subsample (1) (2) (3) (4) (5) Probability of Investing Male Prime 0.113 Overconfidence Prime 0.137 0.135 (7) (8) Amount Invested 0.136 3.554 5.638 5.961 5.763 (0.076) (0.074)* (0.074)* (0.073)* (5.667) (5.321) (5.384) (5.359) -0.026 -0.017 -0.004 -0.003 -7.751 -6.813 -6.616 -6.778 11.94 14.87 -13.54 2.803 (4.48)*** (5.64)*** (15.82) (15.32) X X X X X X 387 0.01 384 0.01 384 0.03 (0.073) (0.073) (0.074) (0.073) Constant Age group f.e. Education f.e. Success prob. f.e Observations (pseudo-)R2 (6) 387 0.01 X X X X X X 387 0.02 384 0.02 384 0.09 (6.060) 387 0.01 (5.851) (5.864) (5.848) B. Experiment 4: Investment Decisions (female subsample) Female Subsample (1) (2) (3) (4) Probability of Investing Male Prime 0.005 (0.055) 0.001 (0.053) (6) Amount Invested 0.004 (0.052) Constant 0.54 (5.37) 0.50 (5.22) 0.84 (2.18) X X X 230 0.01 228 0.01 38.26 (3.92)*** Age group f.e. Education f.e. Observations (pseudo-)R2 (5) 230 0.00 X X X 230 0.03 228 0.06 56 230 0.00 Table 7: Experiment 1 - Contribution of right tail and average to results Panel A shows results for the change in risk tolerance in the male subsample in Experiment 1. Panel B shows results for the change in risk tolerance for the full sample of subjects in Experiment 1. The outcome variable (∆RiskT olerance) is winsorized at the 5-95 percentiles in the left Panels, at the 10-90 percentiles in the right panels. Hubert-White s.e. are reported in brackets. Statistical significance is marked as follows: ** 5%, * 10%. A. Experiment 1: Winsorizing outliers (male subsample) Male Only (1) M. Prime F. Prime (4) (5) (6) Winsorize 10-90 perc. 0.319 0.336 0.335 0.293 0.306 0.304 (0.197) (0.195)* (0.195)* (0.165)* (0.161)* (0.163)* 0.481 0.447 0.450 0.419 0.377 0.373 (0.208)** (0.215)** (0.221)** (0.189)** (0.194)* (0.199)* Mean omitted Median omitted Session f.e. Age group f.e. Education f.e. Observations R2 (2) (3) Winsorize 5-95 perc. -0.170 0 X X X 148 0.03 148 0.09 X X X X 148 0.09 148 0.04 -0.133 0 X X 148 0.09 X X X 148 0.09 B. Experiment 4: Winsorizing outliers (full sample) Full Sample M. Prime F. Prime Male*M. Prime Male*F. Prime Male (1) (2) (3) Winsorize 5-95 perc. Observations R2 (5) (6) Winsorize 10-90 perc -0.265 -0.269 -0.237 -0.189 -0.188 -0.169 (0.194) (0.196) (0.196) (0.154) (0.155) (0.156) -0.117 -0.162 -0.153 -0.088 -0.127 -0.122 (0.202) (0.209) (0.209) (0.162) (0.167) (0.167) 0.590 0.588 0.566 0.489 0.484 0.469 (0.276)** (0.277)** (0.278)** (0.225)** (0.225)** (0.227)** 0.628 0.650 0.643 0.543 0.557 0.553 (0.276)** (0.280)** (0.281)** (0.235)** (0.239)** (0.241)** -0.253 -0.262 -0.271 -0.244 -0.248 -0.252 (0.179) Mean omitted Median omitted Session f.e. Age group f.e. Education f.e. (4) X 320 0.03 (0.179) 0.050 0 X X 320 0.04 (0.182) (0.151) X X X X 320 0.06 320 0.03 57 (0.151) 0.072 0 X X 320 0.05 (0.154) X X X 320 0.10 Table 8: Just Post It - Evidence vs. Fabricated Data Panel A refers to the full sample, Panel B to the male subsample. For each cell the sample mean is reported, the standard deviation is reported in round parentheses, and the number of subjects in curly parentheses. Panel C shows results for running the test of Simonsohn (2013) for each experiment and sample. The test consist of: (i) estimating 100,000 samples randomly drawn from a normal distribution with mean equal to the one in the condition, and standard deviation equal to the average across conditions in each experiment (samples are drawn from a binomial distribution with probability equal to the mean of the extensive margins in Experiment 3 and Experiment 4); (ii) Computing the standard deviation of the standard deviations in each drawn sample; (iii) reporting the ratio of times that the simulated standard deviation of standard deviations are lower than the actual one in the data. A. Summary Statistics (full samples) Experiment 1 Experiment 2 Experiment 3 (extensive) Experiment 3 (intensive) Experiment 4 (extensive) Experiment 4 (intensive) Experiment 5 Male Identity Prime Female Identity Prime -0.01 (1.74) {125} 0.05 (1.07) {66} Overconfidence Prime -0.21 (1.43) {107} 0.680 (0.468) {231} 23.34 (28.55) {231} 0.747 (0.436) {237} 27.95 (29.81) {237} 0.904 (0.296) {208} 49.93 (32.17) {208} 0.079 (1.711) {214} Success Prime -0.38 (1.37) {109} Control -0.06 (1.22) {129} -0.51 (1.62) {107} 0.624 (0.485) {234} 23.41 (29.40) {234} 0.852 (0.356) {216} 42.04 (30.07) {216} -0.090 (1.743) {110} B. Summary Statistics (male subsamples) Experiment 1 Male Identity Prime Female Identity Prime 0.30 (1.54) {61} 0.22 (0.81) {32} Experiment 2 Experiment 3 (extensive) Experiment 3 (intensive) Experiment 4 (extensive) Experiment 4 (intensive) Experiment 5 0.781 (0.416) {114} 34.92 (35.42) {114} 0.949 (0.221) {98} 60.88 (35.03) {98} 0.405 (1.407) {84} Overconfidence Prime 0.03 (1.32) {54} 0.823 (0.384) {96} 35.47 (33.97) {96} Success Prime -0.47 (1.39) {59} Control -0.21 (1.09) {55} -0.49 (1.40) {56} 0.648 (0.480) {105} 26.99 (33.41) {105} 0.844 (0.365) {96} 44.96 (32.75) {96} -0.024 (1.968) {41} C. P-values for random draw of samples across conditions Full Sample Male Subsample Experiment 1 0.99991 0.99982 Experiment 2 0.82831 0.10523 Experiment 3 (ext. margin) 0.45736 0.41635 Experiment 3 (int margin) 0.20449 0.17365 0.50212 0.47808 Experiment 4 (int. margin) 0.67175 0.49040 Experiment 5 0.17617 0.98543 Experiment 4 (ext. margin) 58
© Copyright 2024