The Female Quota and Performance Ambiguity:
The Effects on Women’s Career
FEB
A B S T R A C T
Two experiments were run to investigate whether the association with a gender quota and performance ambiguity, causes females to be rated less favorably. The participants in Experiment 2 had to evaluate the performance of the 14 winners of Experiment 1 on the basis of a noisy performance measure, by guessing their true performance and ranking them from best (rank 1) to worst performer (rank 14). It was found that the performance of female winners associated with the quota was more likely to be underestimated than for females without the quota association. However, the guessed ranking for women associated with the quota was not significantly lower than the ranking for those without quota association. Furthermore, women associated with the quota were not less likely to be perceived as qualified or less likely to be selected for cooperation. In spite of the limitations of this study, it could serve as an opening to further research.
Master Thesis Managerial Economics and Strategy
15ECTS Business Economics Dr. Silvia Dominguez Martinez
Statement of Originality
This document is written by Student Suzanne van Muijden who declares to take full responsibility for the contents of this document. I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it. The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.
Content
1 Introduction 4 2 Related literature 6 3 Methodology 12 3.1 Experimental Design 12 3.2Hypotheses 15 3.3 Empirical Data 17 4 Results 24 4.1 Main results 24 4.1.1 Differences in Means 24 4.1.2 Gender-neutral Task 254.1.3 Underestimated Guess of Performance 27
4.1.4 Underestimation of Rank 30
4.1.5 Competence 32
4.1.6 Cooperation 35
4.2 Limitations 37
5 Discussion and conclusion 37
6 Appendix 40
6.1Extra Tables 40
6.2 Instructions experiments 45
1 Introduction
Although the number of women in middle management has grown substantially in the last two decades, the number of female seats in top positions remains extremely low (Oakley, 2000; Ahern & Dittmar, 2011; Female Board Index 2015, Lückerath-Rovers). This awareness of the underrepresentation of females in top positions has induced many corporations to review their policies and practices (Oakley, 2000).
As a result, mandatory quotas are introduced in Europe, designed to enhance women’s representation in top positions. For example, in the Netherlands, a quota mandating at least 30% to be female employees in top positions was introduced in 2013. There are, however, no actual sanctions for not meeting the quota requirements. Related or not, fact is that the female representation is still only about 17% (Female Board Index 2015, Lückerath-Rovers). Moreover, Norway already introduced a female quota that requires a minimum of 40% female corporate board members in 2003, 10 years before The Netherlands introduced it in 2013 (Wang & Kelan, 2013). In 2003, women in Norway solely held nine percent of the board’s seats at that time (Ahern & Dittmar, 2011). In 2008, the reorganization was already largely implemented; the female representation was increased to 39.8% (Dale-Olsen et al., 2013).
Nonetheless, it is questioned whether the gender quota has no (severe) consequences (Ahern & Dittmar, 2011; Dale-Olsen et al., 2013). For example, Ahern & Dittmar (2011) studied if the required increase in women’s representation would improve or decline firm value. They found a negative relationship between the quota and firm value. Board compositions are chosen such that they are expected to maximize shareholder value and introducing the quota imposes a, possibly strong, constraint on the choice of the directors. This could explain the decline in firm value (Ahern & Dittmar, 2011). Dale-Olsen et al. (2013), however, did not find such relationship between the quota and the value of Norwegian firms. To complicate matters, Ahern & Dittmar (2011) also found that the gender quota led to substantial numbers of inexperienced women being appointed to boards. Experience may also be an explanatory factor for the found decrease in firm value.
Furthermore, because the quota mandates a certain percentage of women to be selected, it creates gender-based selection preference. When a woman is selected on the basis of her quality, her competence is affirmed and this will give her confidence a boost. However, when this selection commences with gender-based preference, information about her competence becomes of secondary importance and she may be left exposed to negative performance
expectancies (Heilman et al., 1997). If any doubt about competence is present, the woman (associated with gender-based selection preference) is likely to be viewed as incompetent or could even end up being socially rejected by co-workers (Heilman, 2001). Heilman et al.’s study (1997) on affirmative action pointed out that when women were associated with affirmative action they were not only rated less favorably than men, but also less favorably than women that were not associated with affirmative action.
Much of the literature focuses on the questions of why there are so little females in top positions, whether a quota increases the likelihood of women being appointed and how costly quotas (affirmative action) are (Ibarra et al., 2010; Wang & Kelan, 2013; Niederle et al., 2013; Kravitz, 2008; Ahern & Dittmar, 2011). However, besides Heilman et al. (1997) there is little research about the effect of the quota (affirmative action) on women’s career’s or representation in high-level positions in the long run. If the use of the quota raises the question of incompetence, this may cause women to receive bad evaluations or even prevent them from being eligible for promotion or a bonus, a possible barrier to climbing the ladder. As the main goal of the quota is to increase women’s representation, it is important to question what the consequences of the quota are and whether these quotas create the desired results. Therefore, it will be investigated here, by experiments, whether the association with the quota causes women to be rated less favorably than those without the quota association or not. I thus aim to contribute to the study of Heilman et al. (1997) by the use of actual performance information instead of hypothetical performance information. In these experiments, participants could be both performers and evaluators of the performance. This is done to create a more realistic work environment. Furthermore, as top positions do not allow for further promotions (you are as far up as you can go), this study will show the effects of quotas for women in middle-level positions. Whereas quotas were introduced to enhance female representation in top positions, quotas (affirmative action) may be unavoidable in middle-level positions (Insch et al., 2008). The reason being that, middle-level candidates eligible for a promotion to top levels are those that have had experience with so-called traditionally ‘masculine’ assignments, while women, are traditionally hired into ‘feminine’ jobs (Insch et al., 2008).
I conducted two experiments to study whether quotas causes women to be rated less favorably. In the first experiment participants performed a task, consisting of adding up 5 two-digit numbers. This task was found to be gender-neutral. The 30% best performers on the task were selected as winners and these 14 winners (14 in the treatment group and 14 in the control group) were evaluated in Experiment 2. The treatment group in Experiment 1 differed
from the control group by applying a quota that required at least 30% to be female in the winners group.
Subsequently, the participants in Experiment 2 were asked to evaluate the 14 selected winners of Experiment 1. The participants were given noisy performance measures, instead of the actual performance of the 14 winners of Experiment 1. For example, if the actual performance was 16.0, the noisy performance measure could take on any number between 15.0 and 17.0. With this unclear performance information, the participants in the second experiment were to evaluate the winners of Experiment 1 by guessing their actual performance. Next to this, the participants needed to rank the 14 selected winners from best performer (rank 1) to worst performer (rank 14), define whether the winners were qualified or not and if they were willing to cooperate with each winner or not. It was found that the performance of female winners associated with the quota was more likely to be underestimated compared to the performance of those without the quota association. However, the guessed ranking for women associated with the quota was not significantly lower than the ranking for those without quota association. Furthermore, women associated with the quota were not less likely to be perceived as qualified or less likely to be selected for cooperation.
This paper is organized as follows. Section 2 summarizes the relevant literature used for this study. Section 3 describes the experimental design, hypothesis and empirical data. Section 4 presents the main results and limitations. Finally, Section 5 discusses, summarizes and concludes the paper.
2 Related literature
The awareness of the underrepresentation of females in top positions has induced many corporations to review their policies and practices (Oakley, 2000). To achieve gender equality, mandatory quotas were introduced in Europe, designed to enhance women’s representation in top positions (Ahern & Dittmar, 2011; Female Board Index 2015, Lückerath-Rovers). Next, several explanations for the underrepresentation of women and the development of women’s careers will be discussed.
Glass Ceiling or no Glass Ceiling
Some speak of a ‘glass ceiling’, an imaginary transparent barrier that limits women to advance to certain (higher) levels in the firm (Daily et al., 1999; Catalyst, 2000; Insch et al.,
2008). This glass ceiling is (supposedly) composed of corporate policies and practices in training and career development, compensation and promotion, while women are less likely to receive these components (Oakley, 2000). Women are traditionally hired into ‘feminine’ positions, such as administrative jobs or human resources (Insch et al., 2008). However, experience in areas such as finance, operations, manufacturing or sales, is often considered as an essential condition for senior positions (Oakley, 2000). Habitually, these so-called ‘pipe-line’ positions are not given to women (Insch et al., 2008).
Instead of a lack of sufficient experience, female executives mostly cite behavioral explanations such as stereotyping as primary barrier for women pursuing senior positions (Oakley, 2000). Managerial (leadership) positions could be perceived as being male sex-typed jobs, because they entail high levels of responsibility, authority and status (Oakley, 2000; Heilman, 2012). In other words, top positions have always been (almost) completely covered by men, as these qualities are considered to be male qualities. This ‘Old Boy Network’ functions well in sustaining rewards for males at the top (Oakley, 2000). As a consequence, male colleagues often perceive women in top positions as threats, because appointing a woman in a man’s position changes the status quo, which is considered undesirable (Oakley, 2000; Linehan & Scullion, 2008). Perceiving women as lacking person-job fit when it comes to managerial positions, solves that issue (Lyness & Heilman, 2006; Heilman, 2001). Also, the discomfort men feel when they are competing with women at the same level, could be caused by fear of ‘losing’ this competition with a woman (Oakley, 2000). Women are associated with having certain attributes, such as being kind and caring, and this stereotyping emerges expectations that women will fail in these positions (Lyness & Heilman, 2006). As a result, acting ‘feminine’ is associated with incompetence, while masculine traits are associated with competence. This yields the conclusion that women must be ‘un-feminine’ to be competent (Oakley, 2000). Women who ‘violate’ this prescribed set of gender norms (i.e. do not follow the stereotypically based attributes), are rated less favorably than their male counterparts (Heilman, 2012; Ibarra et al., 2013). In other words, women that act ‘un-feminine’, are considered cold, are disliked and seem to be less selected for bonuses or promotions (Heilman & Chen, 2005; Heilman, 2012).
In addition, managerial positions are full-time obligations and it is believed that women will turn down assignments due to their family obligations (Holmes, 2013). Moreover, firms become more and more international, and the possibility of assignments overseas, will cause women to be even less selected (Insch et al., 2008; Holmes, 2013). This is partly also because in some of these international locations prejudices against women are part of the local culture
(Insch et al, 2008). More generally, overseas assignments often act as a way to higher-level positions, and holding the beliefs that these assignments are ‘inappropriate for women’, could therefore serve as a crucial barrier for women to climb up the ladder (Insch et al., 2008).
Powell & Butterfield (1994) concluded, contrary to the foregoing, that the glass ceiling is no longer present for women. They studied whether the gender of the applicant influenced the selection decision. They found females to be favored in the hiring process. A possible reason could be that the women that applied were more qualified than the men that applied (Lyness & Heilman, 2006; Powell & Butterfield, 1994). Furthermore, most top ranked companies actively exhibit support for the promotion of women, where most of these firms target women to participate in executive education programs, and take steps to facilitate the movement of women into top positions (Oakley, 2000). The need for gender equality could have caused the findings of Powell & Butterfield (1994) that women were being favored in the selection process.
Mentoring versus Sponsorship
Ibarra et al. (2010) found an important difference to exist between mentoring and actual sponsorship and suggested that firms should make a stricter distinction between these two terms. Although they found men and women to be equally likely to receive mentoring, sponsorship is something that only men seem to get. A mentor offers ‘psychosocial’ support for both personal and professional development, plus career help including advice and coaching (Ibarra et al., 2010; Niemann & Dovidio, 2005). A sponsor offers the same as a mentor but, in addition, promotes actively for advancement (Ibarra et al., 2010). Women are less likely to have sponsors (Linehan & Scullion, 2008). In addition, they suffer from a lack of access to influential colleagues (Ibarra et al. 2010; 2013). Moreover, where men have CEOs or other senior executives as sponsors, women are more likely to be mentored by lower-level managers (Ibarra et al., 2010). Noteworthy however, is that both the lack of influential colleagues and the lack of senior executives as women’s sponsors could possibly be explained by the lack of women in senior positions. Another problem is that women need the sponsorship to climb up to senior positions, but male senior managers may possibly be unwilling to support young women as such relationship between a woman and her male superior could be misinterpret (Ibarra et al., 2010; Wang & Kelan, 2013).
Mentors can also serve as role models, which contribute to the construction of a worker’s identity (Sealy & Singh, 2009). Women, as well as men, need these role models to learn certain attributes, needed to rise to higher positions (Sealy & Singh, 2009). Both positive role
models, and negative role models help the learning process, the latter by showing how not to behave (Sealy & Singh, 2009). These role models could be mentors (sponsors), but individuals could also choose role models without involvement or without the role model’s permission (Sealy & Singh, 2009). Furthermore, women benefit from having female instead of male role models (Lockwood, 2006). As the number of women in senior positions is limited, scarcity of female role models becomes an additional barrier. The quota may be an effective way to enhance female representation in top positions, as these new ‘top’ females can eventually function as mentors or even role models for other women (Wang & Kelan, 2013; Sealy & Singh, 2009). Besides serving as a role model, the presence of females in senior positions signals the likelihood of a possible promotion. In addition, women in executive positions could possibly influence organizational policies and culture, making it more attractive for other women to stay within the organization (Sealy & Singh, 2009).
Competition Aversion
Niederle & Vesterlund (2007) provided another reason for the lack of women in high-level positions. They showed that while there are no gender differences in performance, men are more than twice as likely as females to enter a competitive environment. Instead of choosing to perform in a tournament, women were more likely to choose the opt-out option; women show a preference to perform under a non-competitive piece rate with a substantially lower return. They linked part of these findings to the explanation that men are substantially more overconfident about their relative performance than women. Besides the role of overconfidence, the gender difference in tournament entry could also explain part of the findings (Niederle & Vesterlund, 2007). Women are less likely to choose to compete than men. This does not only mean that less women apply for competitive jobs, but also that the number of women that eventually gets appointed (‘win’) is reduced. In other words, it decreases the probability of success for women when they are competing for promotions and more advantageous jobs (Niederle & Vesterlund, 2007). In addition, when the entry by women into competitive jobs is low, it reduces gender diversity, which - some stated - may harm the firm (Page, 2007; Weber & Zulehner, 2010). Weber & Zulehner (2010) studied the effect of female seniors on firm successfulness. Weber & Zulehner (2010) found that the presence of female seniors had a positive effect on a firm’s success. As this would be a factor in favor of having more females in leading positions, they suggested that glass ceilings are inefficient. Introducing a gender quota may increase female representation.
Niederle et al. (2013) looked at whether affirmative action could encourage more women to enter competitions. Affirmative action (a gender quota) requires the relative number of females that need to be selected to increase (while maintaining the total number of selected candidates constant). This may cause a lowering of the qualification standards, in order to attain the required number of females (Niederle et al., 2013). Niederle et al. (2013) found almost no overall performance difference between before and under affirmative action. Furthermore, the gender gaps in willingness to perform and confidence were reduced under affirmative action. Niederle et al. (2013) found a substantial increase of entry by women and a decrease of entry by men, once affirmative action was introduced. Increasing the chance of winning for women under affirmative action could affect entry by decreasing gender differences in beliefs about relative performance. In other words, due to the quota, women are more likely to win, which could change women’s beliefs about winning (and thus performing) and thereby will increase the willingness of women to compete (Niederle et al., 2013). Important to note is that the selection of winners was entirely based on the clear performance information. There was no discrimination present and the quota caused a need for lower qualifications for women to increase women’s representation. The quota therefore actually created reverse (positive) discrimination (Niederle et al., 2013). Luckily, the due to the quota resulting increase in entry of (high-performing) women, the performance was the same as before the quota was introduced, and therefore eliminated the problem of lowered qualifications.
When high-performing women are not applying, not even after the introduction of the quota, the quota could have (severe) negative consequences. In other words, suboptimal entry by female candidates can be costly for companies (Niederle et al., 2013). The unwillingness by high-performing women to apply for competitive jobs or high-level positions, keeps firms from hiring the best available candidates. As a result, firms have to lower the qualifications of those selected under preferential treatment (quota) (Niederle et al., 2013). This could explain why Ahern & Dittmar (2011) found that the introduction of the quota caused a large number of inexperienced women to be selected.
Stigma of Incompetence
Introducing the quota requires firms to select more women for high-level positions. In addition, Niederle et al. (2013) suggested that quotas result in an increase in the willingness to apply by female candidates. Normally, when a woman is selected on the basis of her quality, her competence is affirmed and this will give her confidence a boost. However, the
presence of the quota raises the question whether the woman was selected due to competence or due to preferential treatment. As stated before, if the women that apply would not have been selected without the quota, firms will have to lower the qualifications of those selected under preferential treatment (Niederle et al., 2013). In other words, it could have been that the woman was not selected without the quota, because she was considered as not qualified. Therefore, information about a woman’s competence is left ambiguous. As a result, this presence of competence uncertainty causes women hired under the quota (and thus possible preferential treatment) to be exposed to negative performance expectancies (Heilman et al., 1997). Furthermore, selected women (in the presence of the quota) could even be stigmatized as incompetent (Heilman et al., 1997).
Heilman et al. (1997) found that the knowledge of high ability of a woman, by others, soothes the negative effects of preferential selection on women. However, information that clearly predicts performance competence is not that conventional in work environments (Fiske et al., 1987). Work is much more done in teams nowadays, making it difficult to specify individual performance (Fiske et al., 1987). Furthermore, it is difficult to specify all aspects of a worker’s job in a contract and performance is therefore frequently subjectively evaluated (Fiske et al., 1987; Prendergast, 1999). Thus, if performance measurement is difficult and therefore information about performance is limited or poor, people can easily misinterpret the actual performance and fill in this gap by category-based stereotypes or ignore the performance information altogether (Fiske et al., 1987). People prefer to categorize others whenever possible; categorization empowers them to understand others efficiently (Fiske et al., 1987).
If a quota causes people to question a female’s competence, unclear performance information allows co-workers or superiors to dismiss or ignore potentially sufficient performance information to fit category-based expectations (Heilman et al., 1997). In addition, the quota may reduce the incentives for investment in acquiring the needed skills for climbing up the latter as women (the favored workers) can see themselves as more likely to succeed without these skills (Coate & Loury, 1993). The lower investment incentives under the quota will cause women to invest less in skills and experience. As a result, the selected women under the quota will actually prove the need for lower qualifications under the quota. This will only confirm someone’s potential beliefs about women being less qualified for the position. Therefore, the system of the negative stereotyping will remain in place (Coate & Loury, 1993).
People usually work alongside each other every day; people who can both support and antagonize their colleagues (Chiaburu & Harrison, 2008). The stigma of incompetence that arises due to the quota could therefore have a detrimental effect on the work environment of the female worker, e.g.: a woman associated with the quota could end up being socially rejected by co-workers (Chiaburu & Harrison, 2008; Heilman, 2001). Not only men, but also women can expose negativity towards other women, when they believe other women were preferentially selected due to the quota (Heilman et al., 1993). Moreover, if a woman is aware that another person considers them as being favored in the selection process, she is likely to assume that the other person holds the belief that she is incompetent (Heilman & Alcott, 2001). The assumptions about the other’s beliefs of her incompetence will be present, whether she was indeed favored or not.
Heilman et al.’s study (1997) on affirmative action pointed out that when women were associated with affirmative action they were not only rated less favorably than men, but also less favorably than women that were not associated with affirmative action.
3 Methodology
3.1 Experimental Design
To study whether the quota causes women to be rated less favorably than men, two different experiments were designed, using Qualtrics. They were distributed via email and Facebook, to reach as many different people as possible. To control for differences, background information was collected about age, sex, education and occupation.
In the first experiment participants were asked to perform a real effort task. This task was the same for both the treatment and the control group, but was called task A and B, respectively. Participants were randomly assigned to either the treatment or the control group. The task consisted of adding up 5 two-digit numbers within 4 minutes. The numbers were randomly drawn and each sequence was presented in the following way:
28 92 33 91 42 +
The subjects’ performance was based on the number of correctly solved sequences. This specific task was chosen because it was found to be gender-neutral (Niederle & Vesterlund, 2007).
The participants did not learn the number of correctly solved problems. Before the actual round of 4 minutes started, participants were given the opportunity to practice their skills or to get used to doing the task in a practice round of 2 minutes. This round did not count for the experiment itself. 30% of the participants on the task were selected as winners in both the treatment group and control group. In addition, in the treatment group at least 30% of the selected winners had to be female. This reflects a quota for certain job positions. Afterwards, the participants were asked how they think they performed the task and whether they were familiar with the quota or not. To create incentives, the participants were told that two of the selected winners would randomly be chosen to receive a small reward. After the subtraction of incomplete responses, there were 47 participants in both groups, creating two selected groups of 14 winners in both the treatment and the control group.
Next, the participants who indicated that they were willing to participate in the follow-up experiment received an email with a personal link to the second experiment. In addition, the second experiment was again distributed via Facebook to reach more participants. The treatment group in the first experiment was assigned to the treatment group in the second experiment and vice versa. This was done to potentially match their answers. 43 of the participants in Experiment 1 also completed Experiment 2, of which 21 were in the treatment group and 22 were in the control group.
The second experiment was built upon the vignettes’ study of Heilman et al. (1997). However, instead of hypothetical performance information, here actual performance resulting from Experiment 1 was used. Heilman et al. (1997) used hypothetical performance information, such as: “this person was evaluated by his superior as being in the top 10% best performers”. Using actual objective performance information in Experiment 2 caused the participant’s choice to be entirely covered by own behavior.
The second experiment involved evaluating the selected group of the 30% winners from the previous experiment, which were 14 winners in total for each group. There were 5 female winners in the treatment group, and 9 male winners. This means that the quota requirement of at least 30% female winners was reached. In the control group, there were 6 female winners and thus 8 male winners.
Participants were informed about the selection procedure of the winners and what the task consisted of (the treatment group was also made aware of the requirement of at least 30% female participants among the winners). They were asked to rank the 14 winners from best performer (rank 1) to worst performer (rank 14), where, hypothetically, the top 5 best ranked were eligible for a bonus. Heilman et al. (1997) let the participants evaluate every subject separately. As people work alongside each other and have to compete for promotions or bonuses (Chiaburu & Harrison, 2008), it is more realistic to make use of relative performance evaluation. Therefore, I let the participants rank the selected winners.
The 14 winners were displayed, but to maintain further anonymity, they were given a fictive name. All females were referred to as ‘Sophie’ and all males as ‘Tom’. Numbers were included to the name to distinguish between different subjects. The participants were made aware of the fact that the numbering was random and did not contain any information about the actual performance. Besides the name, they received information about performance. However, there was a noisy performance measure available; the participants received no information about the true performance of the winners. This noisy performance follows from the study of Fiske et al. (1987), stating that performance information is often ambiguous. The participants were informed that this distortion of performance could both be positive or negative, and is added to the actual performance. It could take on any number between -1.0 and + 1.0 (with one decimal). An example was given to make it more understandable: “If an individual’s observed noisy performance measurement is 17.0, then the actual performance lies between 16.0 and 18.0”. They were made aware that the added distortion was randomly determined and independent of the actual performance. Besides the fictive names and noisy performances of the 14 winners, the 4 best ‘non-winning’ participants are displayed with their noisy performance to give some indication of what the cut-off is for the winners’ group. The individuals participating in this experiment have to first guess the actual performance of the 14 winners. They are told that one participant will receive a reward for accurately guessing the actual performance of the participants in Task A (B). After this, they are asked to rank the winners and answer two competence questions; for each winner whether they think they are qualified to be in the top 14 or not and if they would be willing to cooperate with the individual on the task in question. Each participant in this experiment has to do this exercise twice, with the same winners but with different performance distortions. At last, participants were asked to provide their main driver for choosing the ranking. There were 55 participants in total, 26 in the treatment group and 29 in the control group. Only 51 of the 55 participants finished the second time they had to evaluate the selected winners. Moreover,
76.5% of the participants filled in the guessed performance equal to the given noisy performance measure and 86.2% filled in the ranking based on the given order of the selected winners (ranked equal to the ranking based on the noisy performance). It was chosen not to include the second evaluation into the analysis. Table 1a and 1b are the frequency tables and are included in the Appendix.
3.2 Hypotheses
Hypothesis 1: There are no performance differences between men and women in Experiment 1; task is gender-neutral.
The task of adding up 5 two-digit numbers was found to be gender-neutral by Niederle & Vesterlund (2007), this task is chosen such that the ability of both men and women entering the task (‘tournament’) is equal. If this is the case, both men and women are equally likely of being part of the winners’ group and therefore initially equally qualified for a bonus.
Hypotheses 2-5 are designed to answer the research question: are women, associated with the quota rated less favorably than both those without quota association and men. Due to the female quota of 30% in the treatment group in Experiment 2, female winners in the treatment group are seen as women with a quota association.
Hypothesis 2: The performance of female winners in the treatment group is more likely to be underestimated than for females in the control group.
Hypothesis 3: The rank of female winners in the treatment group is more often underestimated than for females in the control group.
The presence of the quota creates possible reverse discrimination and can thereby lower the qualifications of those selected under the policy (Niederle et al., 2013). In other words, if the 30% quota is not met in the winner’s group of the treatment group, the least performing men selected for the winners’ group are substituted with female subjects next in line for the winners’ group, up until the 30% is reached. Moreover, in experiment 2, participants were given unclear information about the performance competence, i.e. there was only a noisy performance measure available. Therefore, due to the quota and the unclear performance
information, a stigma of incompetence could arise and - as a result - the participants in Experiment 2 could easily misinterpret the noisy performance information to fit category-based stereotypes (stigma of incompetence) or ignore the information completely (Chiaburu & Harrison, 2008; Heilman et al., 1997; Fiske et al., 1987). This would mean that participants in Experiment 2 will guess a lower performance than the noisy performance measure for women associated with the quota. Moreover, participants are more likely to give females associated with the quota a lower ranking than the ranking based on the noisy performance measure. Heilman et al. (1997) found some evidence for this; women associated with the quota were rated less favorably than not only men, but also women without the association with the quota.
Hypothesis 4: Female winners in the treatment group are more often seen as unqualified than female winners in the control group, due to the association with the quota.
As the participants in Experiment 2 have also indicated whether they believed each winner was qualified for the task or not. In hypothesis 4 I will use this information to see whether female winners associated with the quota (hence, in the treatment group) will be significantly more often defined as unqualified. The presence of the quota causes uncertainty about whether a female was selected due to competence or because of preferential treatment, where preferential treatment is associated with lowered qualifications (Heilman, 2001) As a result, if performance information is unclear, people will be more likely to view females associated with the quota as less competent and therefore unqualified.
Hypothesis 5: Due to the association with the quota, participants are less willing to cooperate with female winners in the treatment group than with female winners in the control group.
Theory predicts that the favored disadvantaged group could end up being socially rejected (Chiaburu & Harrison, 2008; Heilman, 2001). In hypothesis 5 this will be tested by the ‘willingness to cooperate’. Following theoretical predictions, it is expected that participants are less willing to cooperate with female winners associated with the quota than those without quota association.
3.3 Empirical Data
To study whether the presence of the quota causes women to be rated less favorably, data was obtained by two experiments. For each hypothesis, a dependent variable is generated.
First, to test whether the task in Experiment 1 was gender-neutral (hypothesis 1), the dependent variable Correct is created. Correct defines the participants’ number of correctly solved sequences on the task in Experiment 1. Subsequently, several explanatory variables are made. Most importantly, a gender dummy, Female, that equals 1 if the participant was female, and 0 when male. As Niederle & Vesterlund (2007) found the task of adding up 5 two-digit numbers to be gender-neutral, I expect the same. Therefore, the effect of the gender dummy Female should be insignificant. Furthermore, the created Treatment dummy equals 1 if the participant was in the treatment group, and 0 if in the control group. The Treatment dummy is included to check whether there are no significant differences between the treatment and control group.
To control for possible individual characteristics, Education, Occupation and Field dummies were composed. Education accounts for the level of education the participant has completed, Occupation controls for the current occupation of the participant and Field covers the field of study of the participants. In addition, the participants in Experiment 1 were asked to fill in their Mathematics grade in High School. The variable Mathgrade is thereby created, which can take on any number between 1 and 10 (Dutch grading)1. Furthermore, the variable Age defines the participant’s age. Both of these variables are used as additional controls for individual characteristics. Table 3 presents the descriptive statistics of the variable Age, Mathgrade and Correct of the participants in experiment 1. There were 94 participants in Experiment 1, 47 in the treatment group and 47 in the control group.
1
It is possible that Mathgrade is highly correlated with the field of study (Field dummies). For example, someone who studied Mathematics or Economics could have scored higher for Mathematics in High School than someone who studied Arts. Therefore, a correlation table 2 is included in the Appendix. As is shown by the table, none of the correlations are above 0.5, such that multicollinearity is highly unlikely.
Table 3 – Descriptive Statistics Experiment 1
The table displays the number of observations, mean, standard deviation, minimum, median and maximum of Experiment 1 for the variables Age, Mathgrade and Correct.
Table 4 – Descriptive Statistics Experiment 2
The table displays the number of observations, mean, standard deviation, minimum, median and maximum of Experiment 2 for the Age of the participants.
The obtained data from Experiment 2 is used for hypotheses 2-5. Table 4 displays the number of observations, mean, standard deviation, minimum, median and maximum for the Age variable in Experiment 2.
Because every participant in Experiment 2 rated 14 selected winners, observations are not independent. Therefore, some differences-in-means t-tests are run. First, two variables are generated, that split up every dependent variable into a separate variable for the female
Treatment Obs Mean SD Min Median Max
Female Age 21 29.0 16.70 20 23 58 Mathgrade 21 7.4 1.33 5 7 10 Correct 21 9.0 4.20 2 8 18 Male Age 26 33.6 13.51 16 25 61 Mathgrade 26 7.2 1.20 5 7.5 9 Correct 26 9.2 3.58 3 9 18 Total Age 47 31.5 12.89 16 25 61 Mathgrade 47 7.3 1.25 5 7 10 Correct 47 9.1 3.83 2 8 18 Control Female Age 25 36.0 16.71 19 25 67 Mathgrade 25 6.8 1.29 5 7 9 Correct 25 8.7 4.01 3 8 18 Male Age 22 35.0 14.26 18 30.5 59 Mathgrade 22 7.3 1.24 6 7 10 Correct 22 10.1 4.98 4 11 21 Total Age 47 36.0 15.46 18 27 67 Mathgrade 47 7.0 1.28 5 7 10 Correct 47 9.4 4.49 3 8 21
Treatment Obs Mean SD Min Median Max
Age Female 14 28.0 10.59 22 24 58 Male 12 39.6 15.75 22 39.5 61 Total 26 33.4 14.22 22 24.5 61 Control Age Female 18 35.1 16.10 22 24 58 Male 11 38.6 14.20 23 25 59 Total 29 36.4 15.24 20 27 67
winners in the treatment group and a separate variable for the female winners in the control group. This is done to check whether there are differences in means between the female winners in the treatment group and female winners in the control group. In other words, if there is an effect of the quota present. Subsequently, the observations are collapsed by participant. This means that for every participant there is one observation left for both variables – which is the average of all observations of that participant. With the collapsed observations, a differences-in-means t-test is run. However, a lot of information gets lost by collapsing the variables. Moreover, the differences-in-means t-test does not take into account several control variables (e.g. individual characteristics). Therefore, some additional regressions are performed, with clustered standard errors to control for the non-independence of the observations.
First, it will be discussed how the dependent variables are generated for each hypothesis. After that, the explanatory variables are reviewed and it will be discussed what the expectations are.
The participants in Experiment 2 had to guess the actual performance, based on a noisy performance measure. For example, if the noisy performance measure is 15.0, the actual performance could take on any number between 14.0 and 16.0. Because the participants had no other performance information than this noisy performance measure, following the Expected Utility Theory, it is rational to guess the expected performance (Mongin, 1997). Therefore, it is expected that participants guess the actual performance equal to the noisy performance measure (15.0 in this example). Consequently, if the guessed performance is lower than the noisy performance measure, this will be defined as an underestimation of performance. In total, 39.2% of the guessed performances were underestimated, 41.2% of the guessed performance was overestimated and only 19.6% of the guessed performances were equal to the noisy performance measure.
Because only 19.6% of the guessed performances was equal to the noisy performance measure, there is chosen for a binomial dependent variable to test hypothesis 2. This binomial dependent variable, Negative Guess equals 1 if the performance was underestimated and 0 if the performance is either overestimated or equal to the noisy performance. To check for potential differences in effects of overestimation and no ‘misestimation’ (guessed performance is equal to noisy performance measure), a Multinomial Logit Regression on the variable Misguess is included in the Appendix. Misguess is equal to 0 if the guessed
performance was underestimated, 1 if the guessed performance was equal to the noisy performance measure and 2 if the guessed performance was overestimated.
As explained above, the variable Negative Guess is split up into a separate variable for female winners in the treatment group, and a separate variable for female winners in the control group. In other words, NGS-T (Negative Guess Sophie Treatment) equals 1 if the guessed performance of a female winner in the treatment group was lower than her noisy performance, and 0 otherwise. In addition, NGS-C (Negative Guess Sophie Control) equals 1 if the guessed performance of a female winner in the control group was lower than her noisy performance, and 0 otherwise. Subsequently, the observations are collapsed by participant. This means that for every participant, there is one observation left for both the NGS-T and NGS-C variable – which is the average of all observations of that participant. For example, the higher the average of the NGS-T variable is, the higher the possibility is that the performance of a female winner in the treatment group is underestimated. As it is previously showed that the presence of the quota causes women to be rated less favorably, it is expected that the mean of NGS-T is significantly higher than that of NGS-C.
After guessing the true performance, the participants in Experiment 2 had to rank the 14 selected winners, from best performer (rank 1) to worst performer (rank 14). As rational participants would guess actual performance equal to the noisy performance measure, the given rank should be equal to the ranking based on the noisy performance measure. Therefore, the rank is defined as underestimated when the given rank is lower than the ranking based on the noisy performance measure. In total, 15.1% of the given ranking was lower than the ranking based on the noisy performance measure, 13.4% of the ranking was overestimated, but the majority (71.6%) of the given ranking was equal to the ranking based on the noisy performance measure.
Again a binomial variable was generated to test for hypothesis 3, because the percentage that was overestimated was very small (13.4%). The dependent variable Negative Rank, equals 1 if the given rank is underestimated, and 0 if the rank was either overestimated or equal to the ranking based on the noisy performance measure. To check for potential differences in effects of overestimation and no ‘misestimation’, a Multinomial Logit Regression on the variable Misranking is included in the Appendix. Misranking is equal to 0 if the given rank was underestimated, 1 if the given rank was equal to the ranking based on the noisy performance measure and 2 if the given rank was overestimated.
Also the variable Negative Rank is split up into a separate variable for female winners in the treatment group, and a separate variable for female winners in the control group. In other
words, NRS-T (Negative Rank Sophie Treatment) equals 1 if the given rank of a female winner in the treatment group was underestimated, and 0 otherwise. In addition, NRS-C (Negative Rank Sophie Control) equals 1 if the given rank of a female winner in the control group was underestimated, and 0 otherwise. Subsequently, the observations are collapsed by participant. This means that for every participant, there is one observation left for both the NRS-T and NRS-C variable – which is the average of all observations of that participant. As it is previously showed that the presence of the quota causes women to be rated less favorably, it is expected that the mean of NRS-T is significantly higher than that of NRS-C.
The participants in Experiment 2 were asked to answer whether they thought the winner was qualified to perform the task or not. With this information, the Qualified dummy was created to test whether females associated with the quota are seen as less competent (hypothesis 4). Qualified equals 1 if the selected winner was rated as ‘qualified for the task’, and 0 otherwise.
To control for the non-independence of the observations, also the variable Qualification is split up into a separate variable for female winners in the treatment group, and a separate variable for female winners in the control group. In other words, QS-T (Qualified Sophie Treatment) equals 1 if the female winner in the treatment group was selected as qualified, and 0 otherwise. In addition, QS-C (Qualified Sophie Control) equals 1 if the female winner in the control group was selected as qualified, and 0 otherwise. Subsequently, the observations are collapsed by participant. This means that for every participant, there is one observation left for both the QS-T and QS-C variable – which is the average of all observations of that participant. As it is previously showed that the presence of the quota causes women to be rated less favorably (and thus less seen as qualified), it is expected that the mean of QS-T is significantly lower than that of QS-C.
At last, theory predicts that women associated with preferential treatment (quota) could be socially rejected by co-workers (Heilman, 2001). To test this, the participants in Experiment 2 were asked whether they were willing to cooperate with the selected winner or not. With this information, the Cooperation dummy was created, that equals 1 if the participant was willing to cooperate with the selected winner, and 0 otherwise.
Like all dependent variables of Experiment 2, the variable Cooperation is split up into a separate variable for female winners in the treatment group, and a separate variable for female winners in the control group. In other words, CS-T (Cooperation Sophie Treatment) equals 1 if the female winner in the treatment group was selected as qualified, and 0 otherwise. In addition, CS-C (Cooperation Sophie Control) equals 1 if the female winner in
the control group was selected as qualified, and 0 otherwise. Subsequently, the observations are collapsed by participant. This means that for every participant, there is one observation left for both the CS-T and CS-C variable – which is the average of all observations of that participant. As it is previously showed that the presence of the quota causes women to be rated less favorably (and thus less selected for cooperation), it is expected that the mean of CS-T is significantly lower than that of CS-C.
Next, the explanatory variables used to test the dependent variables of hypotheses 2-5. Because all dependent variables are dichotomous, Probit regressions are used.
First, a Sophie dummy is created, that equals 1 when the selected winner was female (‘Sophie’), and 0 if the selected winner was male (‘Tom’). This variable was included to test for gender-based differences. Secondly, Sophie x Treatment, is an interaction dummy, of the Sophie dummy and the Treatment dummy. This interaction variable is constructed to measure the effect of being a female winner in the treatment group. In other words, it is the effect of being a woman associated with the quota. As the quota requires at least 30% selected female winners, it could be that otherwise selected males are substituted with females. If the females were not selected without the quota, qualifications have to be lowered to reach the requirement. Together with the noisy performance information, it is unclear what the actual performance is of the selected winners and whether the female winners were selected due competence or preferential treatment. Therefore, it is expected that it is more likely that the performance of female winners associated with the quota (treatment group) will be underestimated than the performance of selected females without the quota association (control group). Moreover, it is also more likely that the given rank is underestimated for females associated with the quota. As a result, the Treatment x Sophie interaction variable will be significantly positive for both Negative Guess and Negative Rank. In addition, it is expected that females associated with the quota are less viewed as qualified and are less selected for cooperation. Because both the Qualification variable and the Cooperation variable are therefore lower for females with the quota association, it is expected that the Treatment x Sophie interaction variable will be significantly negative in both regressions.
The Education, Occupation and Field dummies and the Age variable are included to control for the participants’ individual characteristics. In addition, the Ranking variable is included, which is the rank the participants assigned to each winner. This variable is included into the model to control for the possible effect the given rank – and thus given rating – has on the probability of underestimation. In addition, it is expected that the lower the rank of a
selected winner is, the less likely the selected winner is viewed as qualified or is selected for cooperation.
At the end of the experiments, the participants were asked what their opinion about the gender quota was, for which they could respond with 5 different answers: Good, Somewhat Good, Neutral, Somewhat Bad and Bad. These responses do not clearly demonstrate to what extent the participants have certain beliefs about the quota or to what extent they behave in line with their beliefs. However, to check whether at least some opinion (rather than the ‘neutral’ no opinion) about the quota explains the model better, two dummies were constructed to include in the regressions: a ‘positive’ Quota Opinion dummy and a ‘negative’ Quota Opinion dummy, where the ‘Neutral’ response is left out as baseline in the regression. The ‘positive’ dummy equals 1 if the participant responded with either ‘Good’ or ‘Somewhat Good’ and 0 otherwise and the ‘negative’ Quota Opinion dummy equals 1 if the participant responded with either ‘Bad’ or ‘Somewhat Bad’ and 0 otherwise. Experiment 2 consisted of 55 participants, of which 25% had some positive opinion about the quota, and 22% had some negative opinion about the quota. However, the majority had no opinion about the quota (55%). For example, someone with a positive opinion about the quota might think that women are wrongfully declined for competitive positions and the quota is therefore needed to enhance female representation. Consequently, it is expected that in general this participant is less likely to underestimate performance than someone with no opinion about the quota whatsoever, because this participant believes men and women are equally qualified to be selected as winners. On the contrary, someone that is negative about the quota might think that the quota causes unfair reverse discrimination. Women are, due to the quota, selected based on preferential selection and this participant could therefore be more skeptical about the noisy performance. It is expected that someone with a negative opinion about the quota generally is more likely to underestimate performance of women or to underestimate the rank assigned to the selected female winners. Moreover, they are less likely to view selected females as qualified and they are less likely to be willing to cooperate with the selected females. Therefore, in general, the effect of the Negative Quota Opinion on the underestimation of performance and rank is positive, and negative for the regressions on Qualified and Cooperation. However, it is questionable whether these effects will be significant, as it is measured what the effects are the underestimation of all selected winners in all groups, and the expectations are only clear for the females in the treatment group2.
2
There is no interaction effect included, to measure the effect of a participant having a certain opinion about the quota and the selected female to be associated with the quota. This is because of the relatively few observations available.
Table 5a and 5b present the descriptive statistics of the dependent variables of hypothesis 2-5. Table 5a displays observations that are collapsed by participant (average of dependent variable per participant) and table 5b displays the descriptive statistics for the total sample of observations.
4 Results
4.1 Main results
4.1.1 Differences in Means
Table 6 displays the Z-values for the Wilcoxon Ranksum Tests, for several important variables. This non-parametric test is used as by the use of histograms it was made clear that the variables are possibly not normally distributed, and most variables are ordinal. Table 6 shows that none of the displayed variables of Experiment 1 are significantly different between female and male participants, or between the control and treatment group. This could already indicate that there are no gender differences in performance on the task. Moreover, the treatment group and control group show no differences in performance and they seem therefore balanced.
Table 6 shows some significant Z-values for the variables in Experiment 2. Male participants are significantly older than female participants (at the 1% level), but the Age of the participants in Experiment 2 seems to have the same distribution both groups.
Next, I will look at the differences in means of the dependent variables. The performance is underestimated in this study when the guessed performance is lower than the available noisy performance measure in Experiment 2, i.e. the Negative Guess dummy variable equals 1. Column 1 shows that female participants seem to underestimate performance more often than male participants (significant at the 5% level). This indicates some gender differences in the choice for performance estimation. Moreover, with a significance level of 5%, the performance of male winners in Experiment 1 is more likely to be underestimated than that of female winners (column 3). On the contrary, the Z-value in column 1 for the Negative Ranking Dummy is significantly positive, indicating that male participants are more likely to underestimate a winner’s rank (lower given rank than rank based on the noisy performance measure) than female participants (at the 1% level). This is surprising, as logically the ranking will be lower as the guessed performance is lower.
Table 6 – Differences in means: Wilcoxon Ranksum Test
The table displays the Z-values for the non-parametric test for differences in means. *,**,*** Significantly different from zero at the 10%, 5% and 1% levels, respectively. Column 1 shows the differences in means for variables between female and male participants, column 2 the differences in means for variables between the treatment and control group and column 3 (in experiment 2) the differences in means for participants' actions between female and male winners.
Both Qualified and Cooperate are dummies that equal 1 for each winner if the participants think he or she is qualified for the task and if they would be willing to cooperate with the winner, respectively. Table 6 insinuates that male participants are significantly more likely to rate the selected winners as qualified than female participants are (at the 5% level), and female winners are significantly more likely to be rated as qualified than male winners (at the 1% level). Finally, participants seem to be significantly more likely to be willing to cooperate with female winners than male winners (at the 1% level). If indeed the performance of female winners is less likely to be underestimated, as follows from column 3, it seems plausible that female winners are more likely to be rated as qualified than male winners. However, individual differences of the participants might explain the potentially found differences. Therefore, subsequently, I have run some regressions to account for individual characteristics.
4.1.2 Gender-neutral Task
Table 7 presents the regression on dependent variable Correct. The dependent variable Correct was transformed to a natural logarithm to make it more normally distributed. The most appropriate transformation was chosen by using the ‘ladder’ and ‘gladder’ commands in
Male-Female=0 Control-Treatment=0 Tom-Sophie=0
Variables (1) (2) (3) Experiment 1 Age 0.95 1.04 Mathgrade 0.82 -1.21 Correct 1.01 0.02 Experiment 2 Age 2.37*** 0.47 Negative Guess -2.44** 0.85 2.15** Negative Ranking 3.12*** -0.76 -0.20 Qualified 2.04** -0.73 -3.02*** Cooperation -1.29 -0.68 -2.85***
Stata. Moreover, the Education, Occupation and Field dummies were included in column 2-4 to control for the participant’s individual characteristics.
The coefficient of the Female dummy is insignificant in both columns. This provides evidence for the task to be gender-neutral and therefore supports Hypothesis 1. Moreover, the Mathgrade coefficient seems significantly positive, even after controlling for individual characteristics in column 2. The higher the mathematics grade in High School of the participant was, the more sequences the participants solved correctly. It seems plausible to say that someone who is good with numbers, scores higher on the task of adding up 5 two-digit numbers. Likewise, the coefficient of Age is significantly positive. This suggests the older the participant is, the more sequences he or she solves correctly.
I also used a subjective measure for the participants’ beliefs about gender differences. All participants in Experiment 1 were asked both to compare their performance to the performance of their male co-workers who performed on the same task and next to compare their performance to that of their female co-workers. I have compared these answers and found that 100% of the participants answered the question the same for their male co-workers, as for their female co-workers. Hence, the participants did not carry the beliefs that there were gender differences in performance on the task.
Table 7 – Regression on (Log-transformed) Correctly Solved Sequences
Total sample Total sample
Variables (1) (2)
Coefficients on dependent variable lncorrect
Female -0.083 -0.047 (0.103) (0.100) Treatment -0.027 -0.007 (0.105) (0.115) Mathgrade 0.101** 0.071* (0.044) (0.041) Age 0.008* 0.011** (0.004) (0.005) Intercept 1.192*** 0.511 (0.403) (0.508)
Education Dummy No Yes
Occupation Dummy No Yes
Field Dummy No Yes
Observations 94 94
*,**,*** Significantly different from zero at the 10%, 5% and 1% levels, respectively. The robust standard errors are presented between brackets under each coefficient.
4.1.3 Underestimated Guess of Performance
Because every participant rated 14 selected winners, the observations are not independent. Therefore, a differences-in-means test was run, to test whether the presence of the quota causes the performance of selected females to be more likely to be underestimated. NGS-T (Negative Guess Sophie Treatment) equals 1 if the guessed performance of a female winner in the treatment group was lower than her noisy performance, and 0 otherwise. In addition, NGS-C (Negative Guess Sophie Control) equals 1 if the guessed performance of a female winner in the control group was lower than her noisy performance, and 0 otherwise.
The observations are collapsed by participant, which means that for every participant there is one observation left for both the NGS-T and NGS-C variable – which is the average of all observations of that participant. For example, the higher the average of the NGS-T variable is, the higher the possibility is that the performance of a female winner in the treatment group is underestimated. The differences-in-means t-test is shown in table 8.
Table 8 – T-test for differences in means
The table displays the number of observations, mean and standard error of the variables
NGS-T and NGS-C. Moreover, a t-test is performed on whether the mean of NGS-T is
significantly higher than that of NGS-C.
The means are not significantly different (p-value is 19%), which means that the performance of females associated with the quota (treatment group) seems not more likely to be underestimated than the performance of selected females not associated with the quota (control group). This result suggests that the quota association does not cause the performance of selected women to be more likely to be underestimated. Next, some regressions are performed, to control for potential individual characteristics.
Variable Obs Mean SE
NGS-T 26 0.34 0.060
NGS-C 29 0.27 0.049
Difference 0.07 0.076
Degrees of freedom 53
Table 9 presents the Probit regression on the ordinal variable Negative Guess. The standard errors are clustered by participant, to control for the non-independence of the observations. The coefficient of Treatment x Sophie is significantly positive at the 1% level in the first column, and this result remains after controlling for individual characteristics. This means that that the performance of a female winner in the treatment group is more likely to be underestimated (based on the noisy performance measure) than the performance of a female winner in the control group. Furthermore, column 4 and 5 show the regressions done with only female participants and only male participants, respectively. The coefficient of Treatment x Sophie remains significantly positive, which means that there are no gender differences.
The consistently significant coefficient of Treatment x Sophie supports hypothesis 2, the performance of women associated with the quota is more likely to be underestimated than those without the quota association. Moreover, this result is in line with theoretical predictions; when women are associated with the quota and performance information is ambiguous, people can easily misinterpret the noisy performance information altogether (Chiaburu & Harrison, 2008; Fiske et al., 1987). Moreover, it supports the study of Heilman et al. (1997), who found that women associated with affirmative action were rated less favorably.
In addition, the coefficient of Sophie is significantly negative in all 5 columns. It seems that, in general, the performance of female winners is less likely to be underestimated than that of male winners. Because this effect holds for both female participants and male participants (column 4 and 5), this does not follow from discrimination in favor of own sex (Matsa & Miller, 2011). In other words, participants – in general – seem to favor female winners in the evaluation. Please note, when this female winner is associated with the quota (in the treatment group), the opposite occurs3.
Furthermore, selected winners in the treatment group seem to be more likely to get an underestimated performance evaluation (based on noisy performance). This would indicate that there are preference differences between the groups. However, looking at column 4 and 5, it seems that this effect comes from female participants in the treatment group. This suggests that the female participants in the treatment group differ from those in the control group.
3 The coefficients of Sophie and Treatment x Sophie together are found to be significantly different from zero.
This means that, taken together, the performance of females associated with the quota is not only significantly more likely to be underestimated than the performance of females without the quota associated, but also than the performance of males.
Table 9a – Probit Regression on ordinal variable Negative Guess
*,**,*** Significantly different from zero at the 10%, 5% and 1% levels, respectively. The clustered robust standard errors are presented between brackets under each coefficient.
Surprisingly, looking at column 3, the coefficient of the Negative Quota Opinion is significantly negative at the 10% level. This indicates that the participants with a negative opinion about the quota are less likely to underestimate the performance of the selected winners, which is against expectations. However, as shown by column 4 and 5, this significance comes from female participants with at least some negative opinion about the quota. Women could think the quota unnecessarily harms a woman’s reputation, due to preferential treatment. If, additionally, these women believe that females do not need the quota to get selected, female participants could choose less for underestimation. In other words, they treat the female and male winners equally, and therefore base their evaluation on the noisy performance measure. However, this is solely a possible explanation, as the Quota Opinion dummies do not cover to what extent someone holds certain beliefs.
Negative Guess equals 0 when both the guessed performance is equal to the noisy performance measure and when the guessed performance is overestimated. The ordinal
Total sample Total sample Total sample Only women Only men
Variables (1) (2) (3) (4) (5)
Coefficients on dependent variable Negative Guess
Sophie -0.646*** -0.694*** -0.708*** -0.756*** -0.866*** (0.140) (0.149) (0.150) (0.198) (0.279) Treatment x Sophie 0.961*** 1.042*** 1.086*** 1.267*** 1.170*** (0.177) (0.191) (0.194) (0.283) (0.319) Treatment -0.400* -0.385* -0.436** -0.888*** 0.188 (0.233) (0.226) (0.213) (0.280) (0.418) Female 0.307 0.288 0.207 (0.192) (0.244) (0.229)
Positive Quota Opinion 0.280 0.446 1.203
(0.251) (0.304) (0.742)
Negative Quota Opion -0.575* -0.983** 0.400
(0.302) (0.380) (0.441) Ranking 0.011 0.014 0.015 0.026* 0.005 (0.011) (0.012) (0.012) (0.015) (0.023) Age 0.009 0.004 -0.002 -0.005 0.009 (0.007) (0.014) (0.013) (0.015) (0.019) Intercept -0.593* 0.456 1.253 2.158*** -3.073 (0.347) (0.914) (0.950) (0.736) (2.358)
Education Dummy No Yes Yes Yes Yes
Occupation Dummy No Yes Yes Yes Yes
Field Dummy No Yes Yes Yes Yes
Observations 770 770 770 448 322
