On Women s Underperformance in Competitive Environments: When and Why

Transcription

1 On Women s Underperformance in Competitive Environments: When and Why Nagore Iriberri + and Pedro Rey-Biel First version: September 2011 This version: May 2014 Abstract We revisit gender differences in performance under competition by designing a laboratory experiment with two real-effort tasks, which differ in the existent stereotypes about which gender outperforms the other, and using several informational treatments. We find that women underperform when competing only when two conditions are met: 1) the task is perceived as favoring men and 2) the presence of a rival is strongly primed through the information provided. Subjects are highly sensitive to the information provided before the competitive task, which is consistent with the psychological evidence on stereotype-threat, suggesting reasons why women s underperformance has only been found under specific settings. Our results shed light on the sometimes contradictory evidence found on gender differences under competition. Keywords: gender differences, competition, gender perception, stereotype-threat. JEL classification: C72; C91; D81. We thank Carlos J. Gómez-Ariza for his valuable guidance in choosing the tasks, Pablo López Aguilar for programming the experimental software and Kurt Schmidheiny for valuable advice. We also thank Stefano Gagliarducci and seminar audiences for insightful comments. Nagore Iriberri acknowledges financial support from Ministerio de Economía y Competición, y Ciencia e Innovación (ECO ECO ) and from Departamento de Educación, Política Lingüística y Cultura del Gobierno Vasco (IT869-13). Pedro Rey-Biel acknowledges financial support from Ministerio de Educación (ECO ), Barcelona GSE Research Network and Generalitat de Catalunya. + University of the Basque Country UPV/EHU, IKERBASQUE, Basque Foundation for Science, Universitat Pompeu Fabra and Barcelona Graduate School of Economics. Av. Lehendakari Aguirre 83, Bilbao (Spain). Tel: (+34) nagore.iriberri@gmail.com. Universitat Autònoma de Barcelona and Barcelona Graduate School of Economics. Department d Economia i d Historia Econòmica, Bellaterra. Barcelona (Spain). Tel: (+34) pedro.rey@uab.es.

2 1. Introduction Gender differences in labor market outcomes persist, being a continuous object of study among economists. In addition to the classical explanations based on gender differences in human capital and preferences, or statistical discrimination, two seminal papers have proposed gender differences in competitiveness as a complementary explanation. Gneezy et al. (2003) showed that women underperform compared to men in competitive environments, referred to as gender differences in performance under competition, while Niederle and Vesterlund (2007) showed that women are more likely to avoid competitive environments than men, referred to as gender differences in tournament entry. Since labor markets are inherently competitive, these results would imply a gender gap in wages either because women may be less effective in performing in certain competitive environments, or because they may be less likely to seek promotions. The importance of these results has led to abundant follow-up studies which have been reviewed in Croson and Gneezy (2009) and in Niederle and Vesterlund (2011). Croson and Gneezy (2009) conclude that there are open questions on understanding women underperformance in competitive environment, both regarding the sensitivity to the task used to measure competitiveness, as well as the gender composition of the competing group. Indeed, follow-up studies to the seminal paper by Gneezy et al. (2003) have shown contradictory findings; women underperformance in competitive environment is not a robust finding but depends on the task (Günther et al., 2010, Dreber et al., 2011, Shurchkov, 2012, Cárdenas et al., 2012, and Wieland and Sarin, 2012), as well as on the gender composition of the competing group (Gneezy et al. 2003; Gneezy and Rustichini, 2004; Antonovics et al., 2009). This paper revisits gender differences in performance under competition. 1 Our purpose is to study the importance of the main factors which could potentially influence individuals effort when competing. Our starting premise is that individual performance is affected by knowledge about the competitive setting. Therefore, we ran an experiment in which we exhaustively varied what individuals know about themselves or their 1 Most previous gender studies about competition have focused on policies and institutions that can change gender differences in tournament entry and not on performance once individuals compete. For example, Cason et al. (2010), Wozniak et al. (2010), Ertac and Szentes (2010), and Berlin and Dargnies (2013) show that information on relative performance differences reduces the gender gap in tournament entry. Dargnies (2011) show that gender differences in tournament entry also depend on whether competition is at the individual level or at the team level. Other papers directly manipulate the gender composition of the competing group, making it visible to participating subjects, which results in women being more likely to enter competitions among groups with a higher proportion of women (Sutter and Rützler, 2010; Booth and Nolen, 2012; Gupta et al., 2011; and Grosse and Riener, 2010). 2

3 competing rival, with the aim of understanding which pieces of information yield gender differences in competitive environments, that is, to understand when women underperform in competitive environments, as well as the reasons why this occurs. We present results from a laboratory experiment in which we combine the use of two tasks which differ in the existing perceptions regarding which gender outperforms the other with the provision of different pieces of information before individuals engage in a competition. Our experiment proposes two new and distinct cognitive tasks, one perceived to favor men and one perceived to favor women. We use incentivized elicitations of individuals perceptions in order to conclude and guarantee that the perception about the tasks indeed goes in opposite directions. In that respect, our paper contributes to the existing literature that finds that women s underperformance in competitive environments is task dependent (Günther et al., 2010, Dreber et al., 2011, Shurchkov, 2012, Cárdenas et al., 2012, and Wieland and Sarin, 2012). Although these papers hypothesize that the male and female nature of the task is behind the contradicting results found in the literature, none of them use incentivized elicitation of perceptions to classify the different tasks. Dreber et al. (2011) and Wieland and Sarin (2012) use a survey to elicit perceptions but not an incentivized one. We also obtained demographic variables, measures on individuals general attitudes toward competition, as well as incentivized measures of individuals confidence, which we use as additional controls in our analysis. Subjects perform each of the two tasks twice, first under a piece-rate incentive scheme and then under a competitive scheme. The provision of information right before individuals participate in the competitive stage is the treatment variable in our paper. In order to perform a comprehensive analysis of when performance is affected by the information provided, our experimental design includes eight different treatments. First, in the control treatment, we start with an environment which is abstract and aseptic, where competing individuals know nothing about their own performance under piecerate incentives nor anything regarding their rival. In the Ability Differences treatment, we provide subjects with information regarding the differences in the number of correct answers when performing the task under piece-rate, such that, subjects know whether they start with an advantage or disadvantage with respect to their rivals. In the Rival s Gender treatment, we tell participants the gender of their competing rival before they compete. In Own Ability Assessment, we simply inform agents whether they were 3

4 among the best or worst half of the session when performing the task under piece-rate. This relative measure gave the subjects an idea of how good they were performing the tasks. 2 In the Own Gender treatment, we use a subtle manipulation of gender, consisting in asking participants to fill in their gender before competing, which is used to prime gender. We also have a treatment, Rival Ready, in which we prime the presence of a rival, telling subjects that their rival is ready. Finally, we use additional treatments to combine the information regarding own gender and ability assessment, as well as ability differences and rival s gender. We find that women s relative underperformance in competitive environments depends not only on the task but also on the information provided (or omitted). More importantly, we show that the information or its omission is crucial, that it affects men and women differently and that it interacts with perceptions about the tasks. When competing in the task that is perceived to favor men, the omission of information, such as in the control treatment, results in no differences in performance under competition between men and women. We also find that it is in the informational treatments where the presence of a rival is primed, such as revealing the rival s gender or informing that rivals are ready to start competing, that differences in performance are most prominent, making women underperform compared to men. Finally, we find strongest effects when providing rival s gender in the male task, increasing men s performance by almost 75%, but reducing women s performance by 50%, compared to the absence of any information (control). In order to explain where our results come from, as well as to address the possibility of experimental demand effects, which is a concern in studies where the treatment variable is the provision of information, we use subjects individual characteristics and beliefs as additional controls in our analysis. While there are no significant differences between men and women in standard demographics in our sample, we do observe gender differences in perceptions about tasks, attitudes towards competition and overconfidence measures. When we add these variables as additional controls, gender differences in performance become weaker as they are partially explained by these variables, which is consistent with the presence of stereotype threat, (Steele, 1997), defined as the concern arising from a situation where a person confirms a negative stereotype about their social group. Consistent with our results, Steele and 2 We could alternatively have chosen to provide individuals with an absolute measure of their own ability. However, this information is not very informative as it is probably the first time subjects perform these cognitive ability tasks. That is why we chose a relative measure. 4

5 Aronson (1995) and Ryan and Ryan (2005) argue that very subtle manipulations can activate stereotype-threat and affect performance. In addition, the fact that these variables explain the strong gender effects also rules out the possibility that the gender differences observed were caused by pure demand effects. There are three closely related papers. Two of them, Freeman and Gelber (2010) and Gill and Prowse (2011) use relative performance information before the tournament to study whether women and men react to this information differently, which is a substantially different research question. While the former finds no significant gender effects, the latter does find that women and men react differently to positive and negative feedback. Geraldes et al. (2010) explicitly provide the gender of the opponent. The main difference from this paper is that they always provided the gender of the opponent and varied the information regarding the existing stereotype threat, while our treatment variable is precisely the information regarding gender. Actually, no previous paper has used gender information as a treatment variable. Our main contribution to the literature is that our treatment design allows us precisely to study how different degrees of manipulations regarding the information both on gender and ability, in combination with perceptions about which gender outperforms the other within a task, can explain gender differences in competitive environments. We show that the provision of information priming the presence of a competing rival explains women s underperformance in competitive environments and that this information provision interacts with the perception of the task, activating stereotype-threat. Our paper helps organize the existing and sometimes conflicting evidence of women s underperformance with respect to men in competitive situations. Our findings suggest that in order to observe women underperformance two conditions must be met: the task used is perceived to favor men and the presence of the rival must be primed. Indeed, it is precisely in environments in which the presence of rivals is most prominent (mixed gender head-to-head competitions) and in tasks that are perceived to be male (spatial ability and running), where women underperformance has proven to be strongest, such as in Gneezy et al. (2003) and Gneezy and Rustichini (2004). The paper is organized as follows. Section 2 describes the experimental design and procedures, giving detailed information regarding the tasks and individuals perceptions about them. Section 3 contains the results. Section 4 concludes with a discussion. The Appendix contains translations of the instructions and the postexperiment questionnaire. 5

6 2. Experimental Design and Procedures A total of 640 participants, 20 per session, were recruited using the ORSEE recruiting system (Greiner, 2004), ensuring that subjects had not participated in similar experiments in our laboratory in the past. 3 Experimental sessions were conducted in the Laboratori d Economia Experimental (LEEX) at Universitat Pompeu Fabra using z- Tree experimental software (Fischbacher, 2007). Our recruiting method ensured that half of the subjects were men and half were women, without subjects noticing there was a gender aspect involved in the experiment. Upon arrival, subjects were called into the lab in random order and were seated in individual cubicles separated by screens. Subjects could observe that individuals of both genders were participating in the experiment but there was no special emphasis on the gender composition of the subject pool. Each experimental session lasted one hour, including assignment of subjects to their seats and payment. Throughout the experiment we ensured anonymity and effective separation between subjects. They were paid individually and in private. All instructions appeared on screen and were read aloud to all subjects. Once the experiment had concluded, subjects filled in a voluntary questionnaire while they waited to be paid. The experiment consisted of two tasks, which subjects performed in a sequence of four-minute periods each, first under piece-rate incentives and then under a pair-wise tournament. For piece-rate incentives, one of the two tasks was randomly selected for payment and subjects were paid 15 euro cents for each correct solution they gave in such task. For the pair-wise tournaments, we followed a positive matching of participants based on their performance under piece-rate. In particular, participants were ranked according to their performance in each of the tasks under piece-rate and then the top performer was matched with the second highest performer, the third with the fourth and so on until the participant ranked nineteenth is matched with the one ranked twentieth. Even if, as we will see later, on average men perform better than women under piece-rate incentives in one of the tasks while in the other task performance is similar, the gender composition of the pairings using our matching protocol was as if it had occurred randomly and does not differ across treatments for both tasks (as it should 3 Some subjects suffered small computer glitches during the experiment, preventing them from having the full time to perform some of the tasks. We omit them from the analysis, which explains the small sample size variation across treatments. 6

7 be since all treatments are identical in the first two tasks of the experiment). 4 The matching protocol was public knowledge (see instructions in the Appendix). Based on the tournament literature, Lazear and Rosen (1981), in order to study the pure effect of competition on all participants it is important that the competition is similarly tight, which is ensured by our matching protocol. Also, ensuring it is public knowledge that individuals will be competing with individuals of similar ability makes the informational treatment effects harder to observe. For tournament payment, one of the two tasks was randomly chosen and the subject who performed best in each pair in such task earned 30 euro cents per correct answer, while the other subject earned nothing. 5 Additionally, once all tasks have concluded, subjects could earn 10 euro cents per each of 16 questions rewarding predictive accuracy. Finally, subjects also earned a 3 euro participation fee. Average total payments were euro with a large standard deviation, 6.34, due to the competitive environment. The two tasks were chosen based on the extensive literature in Psychology regarding inherent gender differences in cognitive abilities. Our objective was to find two distinct tasks in which under piece-rate incentives each gender would perform better than the other and, at the same time, where there existed common perceptions that one gender would outperform the other. Kimura (2004) argues that consistent gender differences in abilities are hard to find and that observed differences greatly depend on the specific details of the tasks. 6 Nevertheless, there exists a degree of consensus that men are better than women at tasks involving spatial skills, while women outperform men in tasks involving certain verbal and memory skills (Kimura, 1999). In particular, for a male favoring task we chose a mental rotation task, see Shepard and Metzler (1971) for a description of the task and Maccoby and Jacklin (1974) for a review of gender differences in this task. For a female favoring task we chose a symbol digit substitution task, see Wechsler (1958) for a description of the task and see Majeres (1983) for evidence on gender differences in this task. A non-incentivized pilot experiment conducted using paper and pencil at a different university with 184 subjects of the same age prior to our main experiment confirmed that not only men outperformed 4 For the Mental Rotation Task there were 87 male pairings, 87 female pairings and 146 mixed pairings. For Symbol Digit Substitution Task there were 80 male pairings, 81 all-female pairing and 160 mixed pairings. 5 Since subjects did not know until the end of the experiment whether, for each incentive-scheme, they would be paid according to their performance in the mental rotation task or symbol digit substitution task, independent of their attitude towards risk or their perceptions about how good they were in each of the two tasks, it was always in their best interest to always perform the best they could. 6 Hoffman et al. (2011) further show that nurture can be behind these gender differences. 7

8 women in the mental rotation task we chose and that women outperformed men at the symbol digit substitution task, but also that subjects on average expected these results when asked which gender would on average do better at each task. 7 We adapted both the mental rotation and symbol digit substitution tasks to our computerized setting, which facilitated the provision of information and matching protocol. The mental rotation task (MRT) in our experiment consisted of showing pairs of three-dimensional figures to subjects who had to answer whether such figures were identical or mirror figures. Identical figures are those for which, after rotating one of them, you would get the other one. Mirror figures are those for which, no matter the number of rotations, one figure is never identical to the other one, and furthermore one is the reflection of the other figure. Figure 1 shows a pair of identical, shown in (a), and a pair of mirror images, shown in (b), from the experiment. (a) Identical Figures (b) Mirror Figures Figure 1. Mental Rotation Task (MRT) The symbol digit substitution task (SDST) in our experiment consisted of showing subjects codes, which associated nine numbers to nine letters and subjects had to de-codify sequences of three letters into numbers. Codes were changed every nine three letter sequences, so that the task would involve both memory and codification abilities. Our SDST adapts the original symbol digit substitution task to the computerized setting by changing two dimensions. 8 Figure 2 shows an example of one of the codes used in the experiment as well as one three letter sequence and the corresponding correct answer. 7 In our non-incentivized pilot, we gave subjects two minutes to perform each task. For the mental rotation task men on average solved figures correctly and women did For symbol digit substitution task men on average gave correct answers while women gave Both differences are significant at the 1% level. Regarding expectations and as shown in Figure 8, most subjects (43%) expected men to outperform women in the mental rotation task, while most expected women to outperform men in the codification task, where 42% of subjects expected women to outperform men. 8 Notice that when adapting this task to the computer we modified two elements. First, our codes associate numbers to letters, while in the original task codes associate numbers to symbols, and thus, subjects were asked to fill in numbers instead of symbols, since the z-tree software would only read numbers as variables. Second, sequences were presented in three letter strings instead of much longer strings commonly used. Shorter sequences allow us more precise performance measures. 8

9 Three letter sequence: KHR Correct answer: 925 Figure 2. Code used in symbol digit substitution task (SDST) and a three letter sequence with its solution Given our chosen tasks may differ in several dimensions, such as their level of difficulty, processing or available strategies, we always perform the analysis separately for each of the tasks. Our main measure of subjects performance is the number of correct answers subjects give for each of the tasks. We also use the number of submitted answers and the accuracy of submitted answers, which is calculated by the proportion of correct answers out of the submitted ones. Finally, to measure performance in competitive environments, we also look at the probability of winning each respective tournament. The experiment had eight treatments with 40 male and 40 female subjects in each treatment. 9 We used a between-subject design. In all treatments subjects perform both tasks, MRT and SDST, under a piece-rate scheme (called Tasks 1 and 2 in the experiment) and then they repeat both tasks under a tournament scheme (called Tasks 3 and 4 in the experiment). In the Control treatment subjects received no information before they participated in the tournament; they could only see the question Are you ready? before they moved to the tournament. In the remaining treatments, before the subjects saw the sentence Are you ready? they received information which either primed gender (own and rival s), information regarding their own performance, information regarding relative performance and the presence of a rival, prior to participating in the tournament. In particular, the information provided in each treatment was as follows (see the Appendix for specific instructions): 1. Ability Differences : subjects were informed of the sign and the difference in the number of correct answers provided by the two paired participants 9 In treatment Own Ability and Gender there were 30 male and 30 female subjects, due to one session having to be cancelled due to software problems. In treatment Own Ability we had 50 male and 50 female subjects, due to some of the subjects observations for Task 4 having to be deleted due to a software problem. 9

10 when the task was performed under piece-rate ( Your matched participant provided XX more/less correct answers than you did ). 2. Rival s Gender : subjects were told the gender of their rival ( Your matched participant is a female/male 10 ). 3. Ability Differences and Rival s Gender : subjects were told the information contained in treatments 1 and Own Ability Assessment : subjects were told whether they were among the best or worst half in the session when performing under piece-rate ( Of the 20 subjects in this session, you were among the best/worst 10 when performing in task 1/2"). 5. Own Gender : gender was primed by asking subjects about their gender before competing ( Please fill in your gender for administrative purposes ) Own Ability Assessment and Own Gender : subjects were told the information contained in treatments 4 and Rival Ready : the existence of a rival in the tournament was primed by indicating that the matched participant was ready to start the task ( Your matched participant is ready ). Figure 3 describes the treatment design. Info on performance differences Control T1 T2 T3 T4 T5 T6 T7 Ability Rival s Gender Ability Differences Own Ability Own Gender Own Ability Rival Ready Differences and Rival s Gender Assessment Assessment and Own Gender No Yes No Yes No No No No Info about rivals gender No No Yes Yes No No No No Info on relative performance No No No No Yes No Yes No Gender Priming No No No No No Yes Yes No Rival Presence Priming No No No No No No No Yes Figure 3: Treatment Design Once the four tasks concluded, subjects were given an incentivized questionnaire. This included questions regarding their number of correct answers, their 10 Experimental instructions were written in Spanish, using chico/chica as the actual wording indicating gender. 11 This question was asked both before task 3 and task 4 of the experiment, which could have been awkward for subjects, although no subject raised any concern about this during the experiment. However, it is important to ask it twice to maintain the strength of the priming condition. Obviously, no subject reported different genders between the two tasks. 10

11 relative ranking and whether women or men outperformed or not the other gender, for each task. For each correct answer subjects earned 10 euro cents. Additionally, in treatments where the information contained in such questions had not been provided in the past, subjects were asked questions regarding the gender of the rival and/or whether they believed they had outperformed or not their rival under piece-rate incentives. Finally, subjects filled in a final questionnaire regarding standard demographics (gender, age, nationality, mother language and studies), and questions regarding their attitude toward competition. All these variables will be used as controls when analyzing the results. See Figure 4 for the timeline of the experiment and see the Appendix for experimental instructions and for the questionnaire. Task 1: 4 minutes of MRT under piece-rate Task 2: 4 minutes of SDST under piece-rate Task 3: 4 minutes of MRT under competition Task 4: 4 minutes of SDST under competition Incentivized Questionnaire Final Questionnaire Feedback provided by treatment: Control: no information Ability Differences (in task 1) Rival s Gender (in Task 3) Ability Differences (in task 1) and Rival s Gender (in task 3) Own Ability (in task 1) Own Gender Own Ability (in task 1) and Own Gender (before task 3) Rival ready Feedback provided by treatment: Control: no information Ability Differences (in task 2) Rival s Gender (in Task 4) Ability Differences (in task 2) and Rival s Gender (in task 4) Own Ability (in task 2) Own Gender Own Ability (in task 2) and Own Gender Rival ready Figure 4. Timeline of the Experiment We can now check whether our choice of tasks satisfies the inherent gender differences we had aimed for with our design. Notice that under piece-rate incentives, performance in each task should not differ across all eight treatments. We tested for the validity of our randomization into treatments, confirming that we can aggregate data across treatments for both tasks (636 observations) 12. Figure 5 shows the cumulative distribution function (CDF) of the number of correct answers by gender in each of the tasks. For MRT, the performance by males statistically dominates the one by females (two-sample Kolmogorov-Smirnov test for equality of distribution functions yields a p- value of 0.00). However, for SDST, we cannot reject that the two cumulative distributions are equal (two-sample Kolmogorov-Smirnov test for equality of distribution functions yields a p-value of 0.27). This result differs from the results from our pilot, in which we observed that MRT was a male favoring task while SDST was a female favoring task. However, when we adapted SDST to our computerized setting this is no longer the case (see footnote 7). Canada and Brusca (1991) find that there is a 12 The average correct number of answers in MRT and SDST are and 36.94, respectively. The p- value for the null hypothesis that the performance is not statistically different across the eight treatments is 0.11 for the MRT and 0.10 for the SDST. 11

12 technological gender gap favoring men when tasks are computerized, which might explain the differences we find between the paper and pencil and computerized versions of this task. Mental Rotation Task under Piece-Rate (N=636) Symbol Digit Substitution Task under Piece-Rate (N=636) Figure 5. CDF of Number of Correct Answers in MRT and SDST by Gender under Piece-Rate More importantly, perceptions regarding which gender is favored by each task do not change when the tasks are adapted to a computerized setting. Figure 6 uses answers from the questionnaire administered after subjects concluded the experiment to graph the average frequency assigned by all subjects to each gender outperforming the other at each task under piece-rate incentives (see last question in Screen 11 of the instructions). Clearly, on average MRT is perceived to be a male favoring task while SDST is perceived to be a female favoring task, as they were perceived in the pilot. 13 Frequency Mental Rotation Task (N=637) Neutral Frequency Symbol Digit Substitution Task (N=637) Neutral Figure 6. Histograms of Perceptions in MRT and SDST under Piece-Rate The design of our experiment leaves us with two interesting cases. On the one hand, we have MRT, in which men not only outperform women but there exists a 13 Pearson s Chi-Squared tests, where the null hypothesis is that the both frequency distributions are consistent with a uniform distribution, conclude against the null with p-values smaller than

13 consensus that this is the case and, on the other, we have SDST, where perceptions regarding a female advantage are not confirmed by performance data. This will allow us in section 3.3 to further explore the role of perceptions in explaining gender differences in performance under competition Results 3.1. When Do Women Underperform in Competition? We start by exploring under which circumstances women perform worse than men in competitive environments. As the two tasks are inherently different and the goal of the paper is testing for treatment effects, we will always analyze the two tasks separately. First, we look at cumulative distribution functions (CDFs), in order to evaluate whether there exist gender differences in performance across tasks, separately in the control and seven treatments, as shown in Figure 7. We draw the CDFs of the individual improvement from performing under piece-rate to performing under a tournament, that is, the difference between the number of correct answers in the competitive environment and the number of correct answers under the piece-rate incentive scheme. The average improvement in the number of correct answers in MRT and SDST are 5.23 and 3.03, respectively. Visually, women s underperformance would imply that the CDF of men s improvement stochastically dominates the CDF of women s improvement. Two results are noteworthy in Figure 7. First, we find no evidence for women underperformance in SDST, shown in 7(b), as there is no treatment in which men s CDF dominates women s CDF. Second, we do find some evidence of women underperformance in some of the treatments in MRT, shown in 7(a). In particular men s CDF for MRT stochastically dominates women s CDF in three treatments ( Rival s Gender, Ability Differences and Rival s Gender and Rival Ready). For all other CDFs, men and women s CDFs intersect in several points and, in fact, there are no statistically significant differences 14 Our measures of gender perceptions about the tasks are obtained once subjects have performed both tasks under piece-rate and tournament schemes, such that they could be interpreted as some type of expost justification of their individual experiences. First, notice that subjects had monetary incentives to express their true perceptions. Second, perceptions were elicited before subjects were provided with their performance results and thus only in treatments Ability Differences, Ability Differences and Rival s Gender, Own Ability and Own Ability and Gender treatments subjects could have a partial indication of whether they had an ex-ante advantage with respect to their rival in each tournament. We tested whether the mean perception is different across the treatments and this is ruled out for MRT (pvalue of ), but cannot be rejected for SDST (p-value of ). With respect to the latter, the distribution that is really different is in treatment 8 where most subjects (almost 55%) believe it is a neutral task. Finally, the correlation between the number of submitted responses to each task and perceiving the task as favoring the opposite gender, although negative, it is always below

14 between them. 15 These results can be interpreted as initial evidence that women may only underperform under competition in tasks which are thought to favour men and only when the presence of a rival, through mentioning its gender or its readiness to compete, is strongly primed. Control Control Improvement in Mental Rotation Task Improvement in Symbol Digit Substitution Task Ability Differences Ability Differences Improvement in Mental Rotation Task Improvement in Symbol Digit Substitution Task Rival's Gender Rival's Gender Improvement in Mental Rotation Task Improvement in Symbol Digit Substitution Task 15 One-way ANOVA test for equality of means across gender can be rejected for Rival s Gender (pvalue of ), Ability Differences and Rival s Gender (p-value of ) and Rival Ready (pvalue of ) but not for the control and the rest of the treatments. The Mann-Whitney test for the null hypothesis that the median of the distributions is the same across gender rejects that the male and female improvement is the same for the treatments Rival s Gender (p-value of ), Ability Differences and Rival s Gender (p-value of ) and Rival Ready (p-value of ) but cannot reject it for the Control and the rest of the treatments. Finally, two-sample Kolmogorov-Smirnov test for equality of distribution functions rejects that male and female improvement is the same for Rival s Gender (pvalue of 0.064), but not for the control and the rest of the treatments. 14

15 Ability Differences and Rival's Gender Ability Differences and Rival's Gender Improvement in Mental Rotation Task Improvement in Symbol Digit Substitution Task Own Ability Own Ability Improvement in Mental Rotation Task Improvement in Symbol Digit Substitution Task Own Gender Own Gender Improvement in Mental Rotation Task Improvement in Symbol Digit Substitution Task Own Ability and Gender Own Ability and Gender Improvement in Mental Rotaion Task Improvement in Symbol Digit Substitution Task 15

16 Rival Ready Rival Ready Improvement in Mental Rotation Task Improvement in Symbol Digit Substitution Task (a) Mental Rotation Task (b) Symbol Digit Substitution Task Figure 7. CDF of Improvement in the Number of Correct Answers from Piece-Rate to Competition by Tasks and Treatments We now turn to regression analysis using improvement from piece-rate to tournament, that is, the difference in performance between tournament and piece-rate, as the dependent variable, and a constant and a dummy variable for gender as the only independent variables, as shown in Table 1. The upper part of Table 1 corresponds to MRT while the bottom part corresponds to SDST. 16

17 Table 1 Improvement from Piece-Rate to Competition by Treatment Mental Rotation Task (MRT) Control Ability Differences Rival's Gender Ability Differences and Rival's Gender Own Ability Assessment Own Gender Own Ability Asses. And Own Gender Rival Ready *** ** * (1.638) (1.751) (1.637) (1.593) (1.409) (1.577) (2.120) (1.843) Constant 4.769*** 4.282*** 8.400*** 8.077*** 3.980*** 4.850*** 6.200*** 7.350*** (1.158) (1.238) (1.157) (1.134) (0.991) (1.115) (1.499) (1.303) Observations R-squared Control Ability Differences Rival's Gender Symbol Digit Substitution Task (SDST) Ability Differences and Rival's Gender Own Ability Assessment Own Gender Own Ability Asses. And Own Gender Rival Ready (0.823) (0.840) (0.821) (0.690) (0.925) (0.814) (1.166) (0.857) Constant 1.974*** 3.564*** 3.725*** 1.897*** 3.590*** 3.300*** 2.400*** 3.475*** (0.589) (0.598) (0.581) (0.491) (0.658) (0.576) (0.825) (0.606) Observations R-squared Notes: the dependent variable is the Improvement from piece-rate to tournament, defined as the difference between the number of correct answers in the tournament and the number of correct answers in the piece-rate. is a dummy that takes value of 1 if the subject is female. Robust standard errors in parenthesis. * p<0.1; **; ** p<0.05; *** p<0.01. Table 1 shows a negative and statistically significant coefficient for the gender regressor ( ) in the three experimental treatments in which we observe men s higher improvement in MRT ( Rival s Gender, Ability Differences and Rival s Gender and Rival Ready ), but never for SDST. This result confirms that women s underperformance is highly dependent on the task and the information primed with the treatments. 16 In those cases, on average, women show a lower improvement from piecerate to competitive environment performance compared to men. In particular, a 37% 16 We performed similar regressions with alternative dependent variables such as the improvement in the number of submitted answers and the improvement in the number of mistakes. We find qualitatively the same results, although the alternative dependent variables show higher variance and the results are slightly weaker. We have also performed an alternative specification such as having the number of correct answers in the competitive stage as the dependent variable and including the number of correct answers under piece-rate as the independent variable. With this alternative specification we find both quantitatively and qualitatively identical results. 17

18 ( Rival s Gender ), a 53% ( Rival Ready ), and a 58% ( Ability Differences and Rival s Gender ), of the improvement in performance shown by men. Three observations are noteworthy. First, the observed treatment effect when providing rival s gender cannot be reconciled with any learning effect because subjects in the control group face exactly the same tasks in the same sequence such that if learning effects were present they should also be present in the control. Second, we have already seen that a crucial difference between MRT and SDST is the perception about which gender is favoured by the task, such that, MRT is perceived as a task favouring men, while SDST is perceived as a task favouring women. Our results show that one necessary condition for women to underperform in competitive environments is to have a male favouring task, as we observe that in SDST no women underperformance is observed. However, this is not sufficient as even when performing the MRT, there are several treatments in which women do not underperform compared to men, including the control treatment, where no information is provided. Third, the information primed before the competitive environments is also crucial. The only cases in which we observe women underperforming compared to men is when rival s presence is primed, either by priming rival s gender or the existence of them, such as in the Rival ready treatment. Note that, in the case of Ability Differences and Rival s Gender two pieces of information are provided, one that by itself has no effect (ability differences) and one that has the expected negative effect on women s improvement in performance (rival s gender). We now compare each informational treatment with the control. Table 2 reports the results, the MRT in the upper part and the SDST in the lower part. The dependent variable is, as in Table 1, the improvement in the number of correct answers from the piece-rate to the tournament. The omitted group is the control group while the treatment includes the one variable indicated by the column s title. The main variable of interest is the interaction term between female and the treatment. 18

19 Table 2 Treatment versus Control: Differential Treatment Effect for Women Mental Rotation Task (MRT) Ability Differences Rival's Gender Ability Differences and Rival's Gender Own Ability Assessment Own Gender Own Ability Asses. and Own Gender Rival Ready (1.638) (1.638) (1.638) (1.636) (1.638) (1.641) (1.638) Treatment ** 3.308** (1.661) (1.734) (1.674) (1.605) (1.663) (1.778) (1.779) *Treatment *** ** * (2.398) (2.316) (2.287) (2.158) (2.274) (2.677) (2.466) Constant 4.769*** 4.769*** 4.769*** 4.769*** 4.769*** 4.769*** 4.769*** (1.181) (1.181) (1.181) (1.179) (1.181) (1.183) (1.181) Observations R-squared Ability Differences Rival's Gender Symbol Digit Substitution Task (SDST) Ability Differences and Rival's Gender Own Ability Assessment Own Gender Own Ability Asses. and Own Gender Rival Ready (0.824) (0.824) (0.824) (0.824) (0.824) (0.825) (0.824) Treatment 1.590* 1.751** * (0.849) (0.835) (0.804) (0.863) (0.831) (1.006) (0.914) *Treatment (1.177) (1.163) (1.076) (1.237) (1.158) (1.427) (1.189) Constant 1.974*** 1.974*** 1.974*** 1.974*** 1.974*** 1.974*** 1.974*** (0.610) (0.610) (0.610) (0.610) (0.610) (0.611) (0.610) Observations R-squared Notes: the dependent variable is the Improvement from piece-rate to tournament, defined as the difference between the number of correct answers in the tournament and the number of correct answers in the piece-rate. The upper part of the table refers to MRT while the lower part of the table refers to SDST. is a dummy that takes value of 1 if the subject is female. Treatment takes value of 1 if the treatment is the one given by the column. The omitted group is the control group. Robust standard errors in parenthesis. * p<0.1; **; ** p<0.05; *** p<0.01. In accordance with what we saw in Table 1, the treatment differential effects that depend on gender are only present in MRT but not in SDST. Furthermore, when competing in MRT, we clearly see that women and men are very differently affected when the rival s presence is primed, either explicitly mentioning rival s gender (treatments Rival s Gender and Ability Differences and Rival s Gender shown in columns 2 and 3) or when the subjects are told that the rival is ready (treatment Rival 19

20 Ready, shown in column 7). While the effect of information regarding rivals gender is positive for men, it is highly negative for women, as the interaction term between female and the treatment is negative and significant. The strongest effect is observed for the treatment Rival s Gender. When no information is provided, men on average improve in about 5 answers while men improve in about 6 answers, this difference not being significant. However, when information about rival s gender is provided, men on average improve on 8 answers while women only improve on 3 answers, this difference being highly significant. In other words, in the treatment including information about rival s gender men s performance improves by 75% in the tournament, while women s performance decreases by almost 50%, compared to the control when no information is provided. We find very similar effects for Gender Differences and Rival s Gender and Rival Ready treatments. Although, the effects seem weaker, we cannot reject that the effects are the same. Given it has been observed that women react very differently depending on the gender of the rival, that is, whether it is a competition among the same gender or a mixed competition, we will further analyze these effects in the case of the two treatments in which the gender of the opponent is explicitly revealed ( Rival s Gender and Ability Differences and Rival s Gender ). Estimation results are shown in Table 3. The variables of interest are Treatment, which captures the treatment effect, *Treatment, which captures the differential treatment effect for females, Treatment* Rival, which captures the differential treatment effect when the rival is male, and finally, *Treatment* Rival, which captures the differential treatment effect when females compete against males. For Rival s Gender treatment, even though men are encouraged by the rival s gender information and women are discouraged by this information, as we have already seen, we find no significant differential treatment effect that depend on whether the competitor is male or female, as we find that Treatment* Rival and *Treatment* Rival are insignificant. For Ability Differences and Rival s Gender treatment, we again see that men react positively to the informational treatment while women react negatively. However, we now see that the variable Treatment* is negative and highly significant while *Treatment* Rival is not. This means that while women 20

21 react negatively independently of the gender of the rival, men in fact react positively only when learning they are competing against women. 17 These differential effects are further confirmed in Figure 8, where we draw the CDFs of improvement in MRT in the two main treatments we are considering. In the Rival s Gender figure, placed on the left, we observe that the two CDFs for male subjects stochastically dominate the ones for female subjects, irrespective of the gender of the competitor. On the other hand, in the Ability Differences and Rival s Gender figure, placed on the right, we see that the only CDF that clearly stochastically dominates the rest, is the one where male subjects compete against female subjects. We can draw two conclusions from these results. First, the actual content of the information, i.e. knowing whether the rival is a man or a woman, only matters when information on ability differences is also provided. Second, the differential gender effect that depends on the gender of the opponent is mostly coming from men who face women, who are encouraged by this type of information, which is consistent with the results found by Antonovics et al. (2009). 17 We have also tested for differential gender and treatment effects that depend on the positive and negative nature of the information (advantageous/disadvantageous difference over the opponent) but found no significant effects. 21

22 Table 3 Treatment versus Control: Differential Treatment Effect that Depend on the Gender of the Competitors Rival s Gender Ability Differences and Rival s Gender 5.706*** 5.706*** (1.493) (1.493) Rival 3** 3** (1.397) (1.397) * Rival (2.641) (2.641) Treatment 9.222*** 10.26*** (2.118) (1.558) *Treatment *** *** (2.925) (2.631) Treatment* Rival *** (2.970) (2.680) *Treatment* Rival (4.277) (4.099) Observations R-squared Notes: the dependent variable is the Improvement from piece-rate to tournament, defined as the difference between the number of correct answers in the tournament and the number of correct answers in the piece-rate, in MRT in the "Rival's Gender" and "Ability Differences and Rival's Gender" treatments. The omitted group is the control group. is a dummy that takes value of 1 if the subject is female. Treatment takes value of 1 if the treatment is the one given by the column. Robust standard errors in parenthesis. * p<0.1; **; ** p<0.05; *** p<0.01. Rival's Gender Ability Differences and Rival's Gender Improvement in MRT Improvement in MRT vs. vs. vs. vs. vs. vs. vs. vs. Figure 8. CDF of Improvement in the Number of Correct Answers from Piece-Rate to Competition in MRT in Mixed and Same Gender Competitions To summarize, women underperformance compared to men when competing is only observed when two conditions are met: when performing in MRT, a task that is perceived to favor men, and when the presence of the rival is primed, either through explicitly mentioning rival s gender or when subjects are told that the rival is ready. 22