Uppsala University
Department of Economics
Master thesis
Is the feeling mutual?
The effect of same-sex teachers: Disentangling teacher
bias from role model effects
Author:
Malin Backman
Supervisor:
Erica Lindahl
Contents
1 Introduction 4
2 Related literature 6
3 Theoretical model 10
3.1 Role model effect . . . 11
3.2 Stereotype threat . . . 12
3.3 Teacher bias/ Teaching styles . . . 13
3.4 Pygmalion effect . . . 15
3.5 A theoretical model for identifying mechanisms . . . 16
4 Data 17 4.1 Gender & Age . . . 18
4.2 Student characteristics . . . 19 4.3 Grade . . . 19 4.4 Long-term effects . . . 21 4.5 Teacher characteristics . . . 21 4.6 Afternoon sessions . . . 22 5 Empirical approach 24 5.1 Identifying assumptions . . . 26
5.2 Identifying mechanisms at play . . . 30
6 Results Part One – The effect of a female teacher 32 6.1 Separate by gender . . . 32
6.2 Pooled regressions . . . 38
6.3 Robustness and further analysis . . . 42
7 Results Part Two - Mechanisms 43 7.1 Teacher bias . . . 43
7.2 Role model effects . . . 44
9 Conclusion 51
10 References 52
Abstract
This paper studies the effect of same-sex teachers on students attending Economics A at Uppsala University from 2013 to 2018. Having a female teacher has no significant effect on female students’ exam performance but a significant and positive effect on the likelihood of attending the subsequent course in Economics. Certain types of female teachers have negative effects on male students’ exam performance in certain specifications and no effect on the likelihood of them attending the subsequent course. It was not possible to identify the underlying mechanism behind these results.
Special thanks
I would like to thank my supervisor Erica Lindahl for her help and wisdom. I
would also like to thank Per Engstr¨om, Charlotte Leyser, Tomas Guv˚a, Nina
Andersson & Johanna M¨ork for their invaluable help during the process of
1
Introduction
The role of female teachers, their prevalence in higher education and their potential effect on female students, is a heavily discussed topic. Results and conclusions regarding their influence on female students are however mixed as the research population and estimation methods have varied over time and between studies (Dee, 2007; Holmlund & Sund, 2008; Carrell et al. 2010 among others). Furthermore, although studies describing potential mechanisms are plentiful, disentangling one from the other has proved to be a difficult task where few have convincingly succeeded (Paredes, 2014; Hoffman & Oreopoulos, 2009).
The importance of female role models in academia stems from a number of arguments, especially regarding their influence on students. Although the positive/negative/nonexistent effect of female teachers on female and male students has a seat in this round table discussion, the equal recruitment of male and female teachers, especially in higher levels of schooling, has its own non-negligible role as well. As of 2017, according to Statistics Sweden, Upp-sala University housed 700 professors, where 199 of them, roughly 28%, were women. In light of Uppsala University’s ambitious equal recruitment policy, there seems to be some remaining work to be done. For example, should we see that a larger exposure to female teachers increase the likelihood that female students embark on similar academic paths, a more equal recruitment policy is self-sustaining.
The purpose of this paper is to quantify the effect of female teachers on female students, compared to male students, attending Economics A at Uppsala University. The effects of interest are in this case exam performance and likelihood of choosing to take additional courses in Economics later in their academic career. This study will also attempt to disentangle the effect of gender biased teacher behavior from the presence of a potential role model effect. In other words, the goal is to answer the following questions:
• Do female students perform better in the presence of same-sex teachers and what role does the behavior of the teacher play in this as opposed to the more distanced role model effects?
• Are female students more prone to pursue further studies in Economics if they are exposed to more female teachers?
This particular population of students have never before been studied within this field as the data is collected specifically for this purpose adding to the relatively small subgroup of studies within teacher effects that specif-ically study college students. This study focuses on students early on in their academic careers as to minimise certain selection issues that are more prevalent in later stages. Such selection issues include high performing fe-male students actively sorting into fields and courses where fefe-male teachers are more prevalent. Furthermore, the cohorts in this study are broken up into smaller groups and are randomly assigned teacher assistants for some parts of the course which holds several factors fixed whilst still allowing for within semester and cohort variation in the number of female teachers each student is exposed to. Relating to this, teacher assistants are seldom consid-ered in the related literature even though they compose a significant portion of the teachers that college students meet and interact with. Lastly, this study uses several novel approaches to attempt to separate between different mechanisms behind the effect of same-sex teachers.
The data used covers 11 semesters of the first course in Economics A, Principle of Micro- and Macroeconomics, with approximately 200 students attending each semester. For each student their results as well as their group and teacher assignments are collected. Two outcomes are considered; the points acquired on the first exam administered at the end of the course and the likelihood of attending the subsequent course in Economics. The outcomes are regressed on the gender of the student, the gender of the teacher and an interaction between the two as well as a set of covariates. Semester fixed effects are used to account for confounders that are semester specific.
The results show no significant effects of female teachers for female stu-dents when it comes to exam performance. For male stustu-dents the results are mixed; having a female exercise teacher has a seemingly negative effect on exam performance. However, this relationship is sensitive to changes in the specification, casting doubt over the validity of the result. Meanwhile, female students with female teachers attend Economics B at a higher rate than fe-male students with only fe-male teachers. The likelihood for fe-male students to attend Economics B is unaffected by teacher gender.
The remainder of this paper unfolds as follows: the next part describes the related literature and previous results of similar studies. The next part describes the relevant theoretical framework, mainly focusing on potential mechanisms behind eventual estimated effects. The next part describes in detail the data used followed by a presentation and discussion of the empirical strategy. The results are presented in two parts, one discussing the potential effects and the other part tackles the potential mechanisms behind these effects. This is followed by a discussion and conclusion and finishes with a reference list and the appendix.
2
Related literature
In this section I will describe the approaches and results of previous studies
within this topic. It is organized by type of student subpopulation and
outcome of interest. The last two decades have seen several studies with varying student populations, both with respect to age, type of schooling and subjects. Because of this, it is hard to argue for a general consensus within the field. Common for all studies however, is the importance of context, where seemingly small differences in setting and empirical approach renders highly different results.
Dee (2007) finds when using longitudinal data on a large number of American 8th graders that the assignment of a same-sex teacher has potential
to significantly improve the achievements of both female and male students in several subjects. Using subject fixed effects, Dee finds that the same-sex teacher assignment is shown to have positive effects for boys in mathematics and girls in history.
Paredes (2014) finds positive effects of female teachers on female stu-dents when studying a large quantity of 8th grade stustu-dents in Chile whilst simultaneously finding that being assigned a male teacher has no effect on male students. The results were found to be robust to controlling for teacher fixed effects.
Lindahl (2016) finds when studying if teacher assessments of Swedish 9th graders are gender biased, that female students perform better on a standardized mathematics test if their teacher is female. However, the main results show that, although female students are more likely to receive higher grades than the standardized tests should have granted them, they are less likely to do so if their teacher was female.
Contrary to the studies mentioned above, Holmlund & Sund (2008) find no strong support for the initial hypothesis that a same-sex teacher improves student outcomes when studying students attending the academic track of the Swedish upper-secondary school. An increased gender grade gap (where girls outperform the boys) is found when the share of female teachers are in-creased, a phenomenon shown in their OLS specification but the interaction between the gender of the student and the gender of the teacher loses power when controlling for subject and letting the main effect of the gender of the student vary between subjects. The relationship between girls outperforming boys when the share of female teachers is higher is instead attributed to a spurious relationship where the share of female teachers is higher in subjects where girls generally outperform boys due to reasons not applicable to the gender of the teachers. Holmlund & Sund (2008) also perform a specification where within student and subject gender variation is found (due to teacher turnover) but find no effect of same-sex teacher except for boys in math-ematics. The size of this estimate, a 0.5 grade point increase, is deemed
economically insignificant.
Moving on to studies of higher levels of education, Carrell et al (2010) find when studying college students enrolled in the US Air Force Academy that female teachers have powerful effects on female students in mathemat-ics and science classes, both regarding their performance in class and more long-term outcomes such as applying for more STEM courses in the future. Effects on the long-term outcomes were only found among high-performing female students. They also found that having a female professor had lit-tle to no effect on male students and that female students performed worse than male students, even when controlling for pre-determined outcomes and merits.
Bettinger & Long (2005) study how the presence of female faculty at an Ohio 4-year college affects the likelihood of female students choosing subse-quent courses within a given field. The hypothesis is that a larger share of female faculty during an introductory course of a certain field will increase the probability of female students pursuing that field. The results are mixed, where the female presence increases the probability within fields such as ge-ology, psychge-ology, mathematics and statistics but decreases the probability within fields such as biology, physics and political science.
Rothstein (1995) finds, when using longitudinal data over American col-lege students from the 1970s, that the percentage of female faculty at a college had a positive and significant effect on the probability of female stu-dents attaining an advanced academic degree. The study fails to distinguish if this effect stems from female faculty acting as role models or if an in-creased percentage of female faculty is the result of an already more suitable environment for female students.
Hoffman & Oreopoulos (2009) find, when studying a large number of students from a Canadian college, no effect of having a same-sex teacher apart from a minor negative effect on male students from having an assigned female teacher. They also find that the probability of dropping a class decreases
with about 1% for both female and male students when assigned a same-sex teacher. For the standardized grades however, they find no effect for either male nor female students. They do however consider the possibility of gender biased drop off rates that could affect the estimates for standardized grades, as only students who finish the course do get a grade. If having a female teacher prevents female students from dropping a course, and these students are primarily from the left tail of the grade distribution, this will downward bias the OLS estimates for the standardized grades specifications. When the estimates are truncated to account for this drop out behavior, small positive effects from having a same-sex teacher arises, about 5 to 7 percent of a standard deviation.
Neumark & Gardecki (1998) find virtually no support for the hypothesis that more female faculty would improve the initial job placement for female Ph.D.-students when surveying all Ph.D.-granting economics departments in the US. The share of female faculty has no significant effect on the proba-bility of academic placement or placement at a Ph.D.-granting department. The share does seem to have significant, but not robust, negative effects for the tier of the hiring department. The study does however suffer from subpar response rates and crude measures of the level of female role models. De-spite this, they find that the share of female faculty significantly reduces the number of years in graduate school, indicating that female Ph.D.-students finish their schooling faster when the number of female role models are in-creased.
Canes & Rosen (1995) find similar results as Neumark & Gardecki (1998) when studying if the share of female faculty during a female student’s first and second year determines the share of female students within a particular major. They find no significant effects for any of the three universities that they include in their study. They do however find significant differences in the share of female students between majors, but that cannot be causally explained by the share of female faculty as the results are likely to suffer from omitted variable bias.
Lastly, a recent study by Lusher et al. (2018) uses race/ethnicity match-ing between college students and teacher assistants and look at the interaction effects of these. They find positive effects of having a same-race/ethnicity teacher assistant assigned to you, especially when teacher assistants where given advanced copies of the exam. Lusher et al. were able to audit sessions held by the teacher assistants as well as their office hours and found that students were more likely to seek assistance from their teacher assistant if they were of the same-race/ethnicity.
In an attempt to summarize, the role of same-sex teachers seems to be more prominent for students at lower levels of schooling whereas studies of college and university students have been more mixed in their results and conclusions. This could in part be due to differences between colleges and college students being potentially larger than differences between high schools and average high school students. This could also be attributed to younger students being more prone to change or more sensitive to teacher influence. In regards to this study, the previous literature predicts either limited and/or mixed results that are hard to generalize over different universities and age cohorts. Nevertheless, the mixed results provide motivation to expand the literature to include more students in different settings and different types and levels of exposure to teachers of different genders.
3
Theoretical model
As several studies have shown the effects and non-effects of same-sex teach-ers on student outcomes, attention has now turned to study in what ways teachers matter. In other words, what mechanisms explain the effect of being assigned a same-sex teacher.
Dee (2005) divides these mechanisms into two groups, passive and active mechanisms. The passive mechanisms pertain to those not directly related to the student-teacher interaction, but rather those that arise from simply
being assigned a teacher of a certain sex and being exposed to them at a distance. Such mechanisms include the role model effect and the stereotype threat. The active mechanisms on the other hand, pertain to those found through the active student-teacher interactions such as teacher bias, teaching styles and the Pygmalion effect. Even though these effects are referred to as active, this is not to say that they are at all times conscious, neither by teachers nor students.
3.1
Role model effect
Paredes (2014) presents a formal theoretical model of role-model effects on students’ academic outcomes by assuming that students’ scores are produced through the following learning function:
Sit= I(fi(hit, hjt), βiriths, abilityi) + (1)
That states that the outcome, S, of student i, with teacher t, is a function of the effect of teachers’ time allocation to each gender, i and j, fi(hit, hjt),
the role-model effect, riths, and the student’s ability.
fi(hit, hjt) = hit Ni + γi hjt Ni+ Nj (2)
The time allocation function, f , is assumed to be a function of the time
allocated to gender i, hit divided by the number of students with gender i,
Ni, as well as the time allocated to students of the other gender, j. γi < 1
captures the idea that students of one gender can learn from hours allocated by the teacher to the other gender, but not as much as had the teacher allocated the hours to students of their own gender.
the student’s household, h, and the society, s:
riths = g(rit, rih, ris) (3)
Where g is an increasing function in all its components. The first com-ponent is equal to 1 if the teacher is of the same gender as the student and the second two captures the idea that the role model effect is stronger for those lacking role models in the household and stronger within fields that tend
to have negative stereotypes regarding gender i. If βi > 0 in the learning
function, then students who are exposed to role models have more
favor-able outcomes, such as higher test scores. If βi = 0, then there is no role
model effect present. With the data available to me, I will attempt a similar approach to this by using the within semester variation in previous courses attended by the students that will proxy the level of female role models that each student has been exposed to.
Erkut & Mokros (1984) study to what extent male and female students choose same-sex teachers as their role models by using survey data where students were asked about a professor who has made a significant impact on them. They find that female students mention female professors when asked about this type of role model in proportion to availability (i.e. in proportion to the share of female faculty) whilst men disproportionately disregard female professors as role models.
3.2
Stereotype threat
Steele (1997) defines the stereotype threat as the phenomenon where groups, a certain gender or race, who have been negatively stereotyped within a certain domain, such as a field of study, are hindered as they fear that their actions may further stereotype the group to which they belong or are assumed to belong to. In other words, the stereotype threat predicts that women, aware of the negative stereotypes about female performance within a field,
such as mathematics or economics, will perform worse than had there been no or positive stereotypes regarding their performance.
Hoffman & Oreopoulos (2009) argue that the stereotype threat mech-anism does not apply to their study as the gender composition of the class is close to equal. I however, dispute this claim as even though your fellow students may be of the same sex as you as a female student, everyone else, such as faculty, might not be. The stereotype threat may work through the roles held by the people around you and having only female peers, and not role models, may have negative effects on female students through the stereotype threat. However, should Hoffman & Oreopoulos be correct in this assumption, this would hold for my subjects as well due to the equal division between male and female students during the time period I am studying. Should they not be correct, the stereotype effect may be very present among the female students in my study as the head teachers and professors have all been male, the majority of authors in the mandatory literature are male and most successful economists are or have been men.
3.3
Teacher bias/ Teaching styles
The effect of teacher bias is one that stems from teachers having certain bi-ases towards students of the opposite sex. Teacher bibi-ases could also be found relating to other traits, such as race. Dee (2005) finds that teachers are more likely to perceive students of the other sex and/or other race as disruptive, inattentive or found to rarely to complete their homework. Teacher biases could also be present when grading students, where equally able female and male students are given different grades due to the biases held by the teacher. As mentioned in the section about previous literature, Lindahl (2016) does not find support of these same-sex favoring tendencies but rather that female teachers tend to restrain themselves when assessing female students’ accom-plishments. On the other hand, other studies have found biased tendencies regarding how teachers grade same-sex students (Lavy, 2008; Hinnerich et
al., 2011).
The presence of teacher bias could also affect the teaching style of the teacher. The teaching style may well be an inherent trait correlated to gender but nevertheless affect the outcomes of the students. A female teacher might be advantageous for female students simply because their teaching style is more appropriate for their learning or the teaching style may be consciously altered by the teacher to help female students more.
Paredes (2014) presents a formal model of the teacher’s decision where the teacher is assumed to maximize their utility of teaching subject to the maximum number of hours that they can spend teaching. Depending on the presence of bias, a “taste” for teaching same-sex students, the teacher will adjust time spent with each gender as to maximize utility, V .
max h1t,h2t Vt = 2 X i=1 αit Ni N1+ N2 U (hit) s.t. ¯h = h1t+ h2t (4)
Where the utility is the sum of the utility gained by teaching each gender, U (hit), weighted by the share of each gender within the student group, N1N+Ni 2,
and the level of bias towards each gender, αit. What this model tells us is
that the teacher will choose how much time to spend on either female or male students depending on the gender composition of the class and the level of
bias towards their own gender, αit. From this follows:
αit = 1 if i=t αt ≤ 1 otherwise.
If the teacher has a preference towards their own gender, then αt < 1,
and following the maximization the problem, they will allocate more time to students of their own gender. If there is no bias, there should be no difference in the time allocated towards female or male students.
The magnitude of the teacher bias and the effect of gendered teaching styles could also work through decisions made by the students. Lusher et al. (2018) find in their study of same-race/ethnicity student-teacher assistant interactions that students were more willing to seek the help from a TA more similar to them. A teacher assistant for example, unable to actively seek out students, can only “live out” their bias if they are first approached by a student. In other words, the level of teacher bias could potentially be regulated by the level of student bias and vice versa. As only teacher assistants vary in gender in this study, and their level of contact with students is more limited, we should expect to see limited, if any, effect pointing to the presence of teacher bias even though it may be present.
3.4
Pygmalion effect
The Pygmalion effect, named after George Bernard Shaw’s play Pygmalion (1912) describes the self-fulfilling prophecy where individuals act in accor-dance to what is expected of them (A Dictionary of Psychology, 2015). In this context, the Pygmalion effect could be realised as teachers of different genders voicing different expectations for female and male students - the Pyg-malion effect describes how the students adapt to these expectations. One example could be that female teachers, more so than male teachers, are vocal about the positive expectations for female students and thus giving them the encouragement to perform better. Should this effect be in place, this exam-ple would create differences between female students having male or female teachers, all else equal.
The Pygmalion effect and stereotype threat are on the surface level very similar to each other and cannot be distinguished from one another in this setting. Furthermore, they could very well co-exist as negative stereotypes of one gender may influence the teachers’ behavior and induce a Pygmalion effect by being vocal about these negative stereotypes. However, they are theoretically not the same as the stereotype threat works through societal
norms and the Pygmalion effect works through the student-teacher interac-tion.
3.5
A theoretical model for identifying mechanisms
Paredes (2014) ventures to identify which mechanism could explain poten-tial differences in outcomes between male and female by modeling scenarios where we could distinguish role model effects from teacher bias. Recalling the learning function defined above:
Sit= I(fi(hit, hjt), βiriths, abilityi) + (5)
Differences in outcomes between male and female students with a male or female teacher, expressed as follows:
E[Sii− Sij] > 0 (6)
Where the expected outcomes of students with a same-sex teacher, Sii,
are higher than the expected outcomes of students with a teacher of the oppo-site sex, Sij. This could be prevalent either because fi(hii, hji) > fi(hij, hjj)
or because riihs > rijhs. In other words, we cannot tell if this difference
in expected outcomes stems from teacher bias, where biased teachers spend more time on same-sex students, or differences in role model effects, where same-sex teachers provide more of a role model compared to those of the other sex.
Paredes proposes exploiting variation in students’ parents’ educational background to quantify the level of role model effects, following the definition and discussion of role model effects above. If the role model effect is stronger for students with parents with lower levels of education, and keeping teacher bias constant, the expected differences in outcomes for student-teacher gender
matches should be larger for those with weaker educational backgrounds. If there is no role model effect at work, the difference should be equal.
Paredes also proposes exploiting variation in the gender compositions of different classrooms to identify teacher bias effects, as the effects of teacher bias should be larger if the preferred gender belongs to the minority, thus allowing more hours per student that is favored. This is very much in line with what I intend to do with the data collected.
4
Data
The data used in this study is collected by me from Uppsala University. The study focuses on students attending Economics A during the time period 2013-2018. This amounts to 11 semesters with approximately 200 students per cohort. The data is the result of combining several different data sources from enrolment lists, the results database and the schedule for each cohort. After removing observations with missing social security numbers, the final observation count lands at roughly 2490. A table of attrition as well as a compressed variable summary is included in the appendix, tables A1 & A2. The dependent variables of interest are the points from the exam and if the student attended the subsequent course in Economics within one year of completing Economics A. The main explanatory variables of interest are the students’ and teachers’ genders.
Economics A is divided into three distinct components where they are exposed to or interacting with a teacher.
1. Lectures with head teachers 2. Optional exercise sessions 3. Mandatory seminars
The lectures with head teachers are held several times a week and have been taught by male lectors or professors during the whole time period that this study is based on. This means that variation in teacher gender is sourced solely from teacher assistants. The optional exercise sessions are held, for the most part, by master students attending the Economics master’s program at
the university. Exercise sessions are held 7 times during the course and
are 2 hours long each time. The mandatory seminars are held 4 times by Swedish speaking graduate students. If there aren’t enough Swedish speaking graduate students to cover all seminar groups, teachers are recruited from the master’s program as well.
The lectures are held for the entire cohort, but the seminars and exercise sessions are held in smaller groups. These groups are not randomly assigned as students get to choose what group they want to be in. However, they do not know prior to choosing their group which teacher is assigned to which group. The seminars are held separately for each group and the exercise sessions are held for two groups at a time (Group 1 & 2, Group 3 & 4 and so forth)
4.1
Gender & Age
The gender of the students is determined by their social security number,
where the second to last number indicates legal sex. The gender of the
teacher is determined by their social security number or name when the social security number was missing. The student gender variable is defined
as a dummy equal to one if the student is female, zero otherwise. The
teacher gender variable is defined as two separate dummy variables, one for the exercise teacher and one for the seminar teacher, both equal to one if that particular teacher type is female and zero otherwise. An alternative to this definition is the share of teachers (that do vary in gender) that are female and this variable can take on the values 0, 0.5 or 1.
The age of the students at the time of attending the course is also de-termined by their social security number. Almost half, 49%, of students are female during the time period studied and the average age is roughly 23 years.
For each semester, group and combined groups (for the exercise ses-sions) gender compositions are calculated and included in several specifica-tions.
4.2
Student characteristics
Several students attend Economics A as a part of their bachelor’s program and what program they attend is included in the controls for some specifica-tions as their previous exposure to female teachers vary depending on what program they are enrolled to. The type of bachelor’s program could also affect their performance during the course as previous courses will vary as well. Although their program should not affect the assignment of a teacher of a certain gender, it is included to either increase precision or to iden-tify mechanisms at play (see the next chapter ”Empirical Approach”). The largest share of students, 53%, come from the business bachelor. There is also a significant gender difference in the type of bachelor attended, where significantly more women attend the political or social science program.
4.3
Grade
The main outcome of interest is the points acquired from the first ordinary exam, held shortly after the students’ interaction with both seminar and exercise teachers. The data containing points and the date of the exam is matched through the students’ social security number. If the scores are from exams administered during any other semester than the one where student was first enrolled they are re-coded as missing values. Multiple scores from a
single individual are removed and only the attempt from the student’s first semester is kept. Students without any recorded scores or only recorded scores from later attempts are assumed to not have attempted the exam and this also considered as an outcome for alternative specifications.
The maximum number of points has always been 40 meaning the scores are comparable over time. However, to ensure that the results from this study are quantifiably comparable to other studies, this outcome is re-coded as a percentile ranking for each semester. The exam could be unequally diffi-cult from semester to semester, which is why the main specification includes semester fixed effects. The grading of exams is done in part by teachers and in part by teacher assistants where the identity, and thereby gender, of the student is hidden from the person grading. Only when final scores are recorded is the identity revealed. This should eliminate possible gender biases to occur during grading. This is reassuring as previous studies have shown that differences in grades between male and female students may be due in part to teachers giving more favourable grades to same-sex students (Lavy, 2008, Hinnerich et al., 2011).
A potential threat to the unbiasedness of the grading is the possibility of detecting gender through other signals, such as handwriting, grammar or phrasing. Should this be the case one might see gender bias in the grading. It is however unlikely that the magnitude of gender bias should depend on the gender of the student’s teachers but rather be constant for students of a certain gender and/or gender signals. For this reason, the potential presence of gender bias in grading should not disrupt the identification in this study. However, as the gender composition of teachers, and thus likely the gender composition of the graders, vary between semesters there is reason to believe that the magnitude of gender bias (should there be one) could vary as well. This presents a problem in cases where more female teachers introduce more favourable gradings of female students during semesters where female stu-dents are more likely to have a female teacher. This motivates further the use of semester fixed effects.
As can be seen in the descriptive statistics, female students outperform male students on average and do so almost every semester. Male students also have more spread in their outcomes, with their standard deviation being larger, both in total and for most semesters.
4.4
Long-term effects
Another outcome of interest, following the example of Bettinger & Long (2005) is the more long-term effects of having a same-sex teacher, in this context defined as attending the subsequent course in Economics, Economics B. This is a dichotomous variable, equal to one if the student is found to be registered to the course during any subsequent semester to the one when they first registered to Economics A. To limit the influence of other factors, only students that attend the course within one year of finishing Economics A are considered. Students not attending Economics B at any point in the future are assumed to have chosen a different major or ended their studies entirely. 16% of all students attend Economics B within one year of attending Economics A, the majority of them being men.
4.5
Teacher characteristics
As the effect of teachers will also depend on the quality of said teacher, some sort of quality control is preferable. This can be done in two ways from the data available. Firstly, we can use teacher evaluations written by the students at the end of the course. This is not ideal in a number of ways, evaluations may be gender biased and not accurately reflect the quality of teacher. Furthermore, because they are written at the end of the course, they may be the result of the grade that the student received. Students with lower grades may be more “disgruntled” and report lower evaluations because they are not pleased with the help they received, regardless of the quality of the teacher. A second approach to quality control is to use the grades
(or grade point average) of the teachers as they are all students at the time of teaching. This is preferred over the evaluations as they are determined before the course begins.
Other teacher characteristics are used as controls in some specifications or robustness checks, such as their age and how many times they have held the position of either seminar or exercise teacher. The latter could also be seen as a control for teacher quality. In the main model, exercise teachers’ quality is proxied by the grade point average from their first semester of the master’s program, a time during which most teachers are recruited for the job. Seminar teachers’ quality is proxied by a dummy variable equal to one if the teacher was a PhD student at the time of teaching rather than a master’s student. Female exercise teachers have a slightly higher grade point average whilst male seminar teachers are more likely to be PhD students.
4.6
Afternoon sessions
All exercise sessions and seminars are not held at the same time. Instead, a number of groups have their exercise sessions and seminars during the late afternoon instead of directly after lunch as the other groups do. When students choose what group they want to belong to, they might take this into consideration. From my own experience, both as a student attending this course and as a teacher for the exercise sessions, there is significant sorting into the groups based on this, where low-performing students more often opt for sessions held later in the day. For this reason, it is controlled for and defined as the share of the total number of both seminar and exercise sessions that are held in the afternoon.
Table 1: Summary Statistics: Student characteristics
All students Female students Male students
Total N Mean SD N Mean SD N Mean SD
Share of female students 2489 0,49 0,50
Age 2489 22,70 2,89 1213 22,42 2,73 1276 22,96 3,02
Share of students with 2489 0,52 0,50 1213 0,52 0,50 1276 0,52 0,50
female exercise teacher
Share of students with 2489 0,37 0,48 1213 0,40 0,49 1276 0,35 0,48
female seminar teacher
Share of female teachers 2489 0,45 0,36 1213 0,46 0,37 1276 0,44 0,36
Exam score 2163 25,64 7,37 1077 26,38 7,11 1086 24,90 7,55
Share of students that 2489 0,13 0,34 1213 0,11 0,32 1276 0,15 0,36
did not attempt exam
Attended NEK B within 2489 0,16 0,37 1213 0,13 0,34 1276 0,19 0,40
one year
Afternoon sessions 2489 0,32 0,45 1213 0,24 0,41 1276 0,40 0,47
Table 2: Summary Statistics: Program frequencies
All students Female students Male students
N Share N Share N Share
Politics 489 0.20 316 0.26 173 0.14 Social science 281 0.11 149 0.12 132 0.10 Business 1317 0.53 582 0.48 735 0.58 Historian 1 0.00 0 0.00 1 0.00 Planning 20 0.01 10 0.01 10 0.01 Mathematics 7 0.00 1 0.00 6 0.00
Peace & Conflict 6 0.00 6 0.00 0 0.00
Table 3: Summary Statistics: Teacher characteristics
Exercise teachers Female Male Seminar teachers Female Male
Quality Mean 12,83 12,31 Quality Mean 0,74 0,86
N 30 26 N 27 37
Age Mean 25,84 25,34 Age Mean 27,70 28,47
N 31 29 N 27 36
5
Empirical approach
Following the example of Dee (2007) and Hoffman & Oreopoulos (2009), the effects of a female (same-sex) teacher is first estimated for male and female students separately. By doing this, variables are not restricted to have the same effect for male and female students as they would be in a pooled regression. The gender separate specification is as follows:
yis = α0 + β1f emteacher(e)i+ β2f emteacher(s)i+ γs+ ωZ + (7)
Where yis is the outcome, either exam results or attending the
subse-quent course, of student i during semester s. The main variables of interest
are f emteacher(e)i, a dummy variable equal to 1 if the student’s exercise
teacher was female, and f emteacher(s)i, a dummy variable equal to 1 if the
student’s seminar teacher was female. This specification includes semester
fixed effects, γs, as the pool of students may differ significantly from semester
to semester and the difficulty of the exam is not constant. It also includes a vector of controls, Z, such as the age of the student, if exercise sessions and seminars are held primarily in the afternoon and the share of female students in each group. Different proxies for teacher quality are also included in this vector of controls. Interacting the two dummies for teacher types, is also considered as they may have either a dulling or a multiplicative effect on the
outcomes of students.
Standard errors are clustered at the exercise group level, where I allow for correlation between students who are treated together as the exercise grouping determines the assignment of teachers, and thus a large portion of the treatment. As treatment in this scenario is administered at two stages, firstly by the smaller seminar groups and secondly through the larger ex-ercise groups, two levels of clustering are potentially viable. The larger is chosen as they spend more time together in the larger groups than in the smaller. Comparisons between the two standard errors are included in the appendix in tables A3 through A8. All in all, they for the most part render similar standard errors, with no clear pattern regarding size, and identical conclusions.
An alternative specification is also used where exposure to female teach-ers is simply defined as the share of female teachteach-ers among the total share of teacher types that vary in gender. This variable, f emteacher(%), can then take on the values 0, 0.5 and 1.
yis = α0+ β1f emteacher(%)i+ γs+ ωZ + (8)
The coefficient β1 would then give the average effect of increasing the
share of female teachers from 0 to 1 for female (male) students.
The next specification pools all students and the effect of a female stu-dent having a female teacher is given by the interaction terms.
yis = α0+ β1f emstudenti + β2f emteacher(e)i+ β3f emteacher(s)i+
β4interaction(e)i+ β5interaction(s)i+ γs+ ωZ +
(9)
This specification includes dummies for being a female student, having a female seminar and/or exercise teacher as well as interactions between
female student and female teacher for each type of session. The value added of having a female teacher as a female student compared to male students
with male teachers is given by β4 and β5 respectively.
The alternative definition of exposure to female teachers is also applied to the pooled specification, where the effect of interest is given by the inter-action between being a female student and share of female teachers.
yis = α0+β1f emstudenti+β2f emteacher(%)i+β3interaction(%)i+γs+ωZ+
(10) The value added for female students having only female teachers
com-pared to male students with male teachers is given by β3.
5.1
Identifying assumptions
The ideal scenario for identifying a causal effect of having a same-sex teacher on a given outcome is for all students to be randomly assigned a teacher. In other words, the gender of the student should not correlate with the gender of the teacher, nor should predetermined outcomes determine the gender of the teacher to which you are assigned. Adding to this, because the gender com-position of the classroom is of interest, students should as well be randomly assigned to groups. Other studies have either relied on the random nature of teacher and class assignment (Carrell et al, 2010) or within-student-subject variation to remove the possible threat of female teachers being assigned groups of relatively high (or low) performing female students (Holmlund & Sund, 2008).
At the department for Economics at Uppsala University, the group to which you are assigned determines which teacher you are assigned, and thus the gender of your teacher. Sorting into these groups is however not random
as students are required to choose which group they want to be in. On the other hand, the teacher assigned to each group is not based on any known student characteristics of the group and the teacher is unknown to the student when they choose their group. This means that the sorting into groups is self-selected, but the gender of your teacher is not. Sorting into groups is not based on your prior outcomes but could correlate due to homosocial reproduction (equally able students grouping together) when making sure
you are assigned to the same group as your friends. The assignment of
teachers to each group is not based on students’ prior achievements nor the gender composition of the group. Some steps are taken to control for non-random group assignments, such as controlling for gender composition and the program to which students are enrolled.
Economics A is a mandatory course for several large and popular pro-grams such as the Business and Political Science bachelors. This creates a context where the level of self-selection into the course is relatively low compared to stand alone courses that are taken independently of a bachelors program. The gender composition in Economics A is also even, which points to a relatively low gender bias in the extensive margin. Hoffman & Oreopou-los (2009) also point to this as a strength in their study of first year students, as they are less familiar with the faculty for any given program.
One issue with the identification in this setting stems from the optional nature of the exercise sessions where each combined group is assigned a teacher, but this does not guarantee that the students attend these sessions nor that they attend the sessions held by the teacher that they are assigned. Because attendance is not taken, the students are in theory free to attend whichever session, held by any teacher, as they please. The assignment of teachers could then only be considered a proxy variable for the teacher that the student is exposed to and the estimated effect only the intention-to-treat-effect. On the other hand, this is preferred over attendance records as decisions made regarding exercise teachers after the initial allocation could be endogenous to the model. Seminars are mandatory and must be attended
to complete the course.
An implicit test of the identifying assumption is performed by regress-ing pre-determined class and teacher characteristics on the variables and interactions of interest, that being the gender of teachers and students and the interaction between the two. The pooled regressions, including semester fixed effects are used but all pre-determined background characteristics are removed and are instead used, one by one, as the dependent variable. Table 4 presents the results from these regressions where the two teacher types are included separately. Column 3 shows that the interaction between being a female student and having female seminar teacher is somehow significantly correlated with the quality of the exercise teacher. This significant correla-tion is however not robust to including the quality of the seminar teacher and the share of afternoon sessions, motivating their use as control variables as to control for this spurious correlation between the gender interaction of one type of teacher and the quality of the other. Columns 5 and 6 show that female students are less likely to attend the afternoon sessions and are more likely to be in exercise groups with a larger share of women and should therefore be controlled for. Positive for the identification in this study is that the interactions that serve as the main variables of interest are, for the most part, not significantly correlated with pre-determined characteristics.
Table 5 show the results from the same analysis as Table 4 but with exposure to female teachers defined as the share of female teachers. Column 2 shows that the interaction between being a female student and the share of female teachers is significantly correlated with the age of seminar teacher but at economically insignificant levels (half a year). This significant correlation disappears when adding the other controls. Column 5 shows, just as in Table 4, that female students are less likely to attend the afternoon sessions but also that the interaction is negatively correlated with the share of afternoon
sessions. This significant correlation disappears when controlling for the
T able 4: Balancing test: Separate teac hers (1) (2) (3) (4) (5) (6) V ARIABLES Age of exercise teac her Age of seminar teac her Qualit y of exerci se teac her Qu alit y of seminar teac her Share of afterno o n sessions Share of w omen in group F emale studen t -0.0490 -0.271 -0.0101 0.0107 -0.104** 0.0854** * (0.128) (0.141 ) (0.107) (0.0318) (0.0359) (0.0141) F emale exercise teac her 0.730* -1.470 0.335 -0.0387 -0.00215 -0.000984 (0.330) (1.134 ) (0.350) (0.104) (0.147) (0.0346) F emale studen t x -0.0859 0.376 0.294 -0.0233 -0.0525 -0.0121 F emale exercise teac her (0.221) (0.327 ) (0.285) (0.0544) (0.0471) (0.0150) F emale seminar teac her -0.7 92 -0.560 0.274 -0.118 -0.210 0.0303 (0.457) (1.093 ) (0.407) (0.0956) (0.152) (0.0386) F emale studen t x -0.159 0.257 0.196** 0.0640 -0.0219 0.0295 F emale seminar teac her (0.233) (0.335 ) (0.0724) (0.0482) (0.0644) (0.0207) Constan t 24.66*** 27.92*** 12.84*** 1.049*** 0.180* 0.427*** (0.671) (1.318 ) (0.842) (0.0448) (0.0908) (0.0489) Observ ations 2,391 2,308 2,201 2,348 2,489 2,489 R-squared 0.30 9 0.174 0.207 0.204 0.205 0.213 Standard errors clustered at exercise group lev el. Significance lev els: ∗p < 0.1; ∗∗ p < 0.05; ∗∗∗ p < 0.01. T able 5: Balancing test: Share of female teac hers (1) (2) (3) (4) (5) (6) V ARIABLES A g e of exercise teac her Age of seminar teac her Qualit y of exercise teac her Qualit y of seminar teac her Share of afterno on sessions Share of w omen in group F emale studen t -0.0620 -0.272 -0.00308 0.00665 -0.109*** 0.0836*** (0.113) (0.150 ) (0.0985) (0.0287) (0.0273) (0.0127) Share of T As that are female -0.118 -1.982 0.605 -0.156 -0.218 0.0320 (0.517) (2.154 ) (0.530) (0.165) (0.191) (0.0488) F emale studen t x -0.274 0.656* 0.49 0 0.0387 -0.0782** 0.0170 Share of female teac hers (0.246) (0.270 ) (0.295) (0.0875) (0.0220) (0.0126) Constan t 24.57*** 27.96*** 12.8 4*** 1.048*** 0.171* 0.430*** (0.673) (1.312 ) (0.843) (0.0469) (0.0858) (0.0488) Observ ations 2,391 2,308 2,20 1 2,348 2,48 9 2,489 R-squared 0.244 0.168 0.207 0.201 0.187 0.201 Standard errors clustered at exercise group lev el. Significance lev els: ∗p < 0.1; ∗∗ p < 0.05; ∗∗∗ p < 0.01.
5.2
Identifying mechanisms at play
Paredes (2014) models two scenarios where the teacher bias could possibly be distinguished from role model effects, and in a less precise sense, separating student effects from teacher effects.
With the data available, I will attempt a similar identification by ex-ploiting interactions between female teachers, female students and the gender composition of the group that they belong. This follows from the assumption that the effect of being assigned a same-sex teacher, arising from the actions and decisions of biased teachers, will have a smaller effect on the individ-ual if the share of students belonging to the favored group is larger. Even though a biased teacher will allocate more time and engagement (in Pare-des (2014) referred to as hours teaching) towards the group they favor, the time per favored student will be lower. What we should see then is that the female student-female teacher interaction effect is larger for female students belonging to groups with fewer women.
To attempt to quantify the effect of teacher bias, the following regression is performed:
yis = α0+ β1f emstudenti+ β2f emteacher(e)i+ β3f emteacher(s)i+
β4interaction(e)i+ β5interaction(s)i+ β6sharewomen(e)i+
β7interaction × majoritymen(s)i + γs+ ωZ +
(11)
This specification mimics equation (3) but with an added variable that distinguishes female students with female seminar teachers in groups with a majority of female peers from those with a majority of male peers. Holding the effect of being a female student, having a female exercise teacher and the share of female students in your exercise group constant, the effect of interest
comes from β7, that should be positive if female students are benefited from
teach-ers having a bias towards students of their own gender. For comparison, I will also present the results from this specification but with share of female teachers as the variable indicating exposure to female teachers.
As for the role model effect, Paredes presents a definition of the role model effect as a function of the gender of the teacher, pre-existing role models and society. I will venture to proxy pre-existing role models with the program that students attend, as many students are enrolled into a bachelor’s program where Economics A is a mandatory course but seldomly the first one that they take. Instead, many students have already attended a semester or two of other courses depending on which program that they are enrolled to and have thus been exposed to different shares of female teachers. By interacting the program to which the student was enrolled at the time of attending Economics A with the female teacher-female student interaction we can potentially detect the presence of a role model effect, keeping teacher bias constant as there is within group variation in the programs to which the students are enrolled.
84% of all students in this study attended one of the three most common bachelor programs, political science, social science and business, all of which could entail majoring in Economics if the student chooses to do so. I have limited this part to only include students from these three programs as to avoid very small (1-7 observations) groups in the analysis.
The specification used for this analysis is as follows:
yips= α0+ β1f emstudenti+ β2f emteacher(e)i+ β3f emteacher(s)i+
β4interaction(e)i+ β5interaction(s)i+
β6interaction(e)i × program + β7interaction(s)i× program + δp+ γs+ ωZ +
(12)
Where δp is a series of dummy variables equal to one depending on which
interest lie in the interactions between female student, female teacher (either type) and the type of bachelor’s program. According to theory, female stu-dents attending either the political science program or social science program should find smaller effects of having a same-sex teacher than those attend-ing the business program, as they have previously been exposed to a larger number of female academic role models.
6
Results Part One – The effect of a female
teacher
In this section I will present the results from specifications (7) through (10). First, the results from the regressions performed on each gender separately and then the results from the pooled regressions. Furthermore, I will present the results from the specifications aimed at quantifying the discussed mech-anisms.
6.1
Separate by gender
Table 6 presents the results from regressions performed only on female stu-dents. The first four columns present the results from regressions with exam points as the dependent variable and the last four columns present the results from regressions with attending the subsequent course in Economics as the dependent variable.
Columns 1 & 2 presents the results from analysing the effect of the two teacher types separately, with and without controls for share of afternoon ses-sions and quality of teachers for example. Having a female teacher, of either sort, does not have any statistically significant effect on female students’ per-formance during the first ordinary exam, compared to female students with male teachers. When controlling for background variables, the estimates
become smaller in absolute size and approaches zero. Interacting the two teacher types does not change this conclusion as can be seen in Table A9 in the Appendix. No included factors, apart from certain semester and program dummies, significantly affects the performance on the exam.
Columns 3 & 4 redefines exposure to female teachers as the share of teachers (that vary in gender) that the student encounters during the course.
This analysis renders insignificant results as well. Including controls
de-creases the absolute size of the estimates, mimicking the tendencies in columns 1 and 2.
At this stage, there does not seem to be any evidence that female stu-dents perform better on the exam if they are exposed to more female teachers. If anything, having a larger share of female teachers, specifically a female ex-ercise teacher, seems to have a negative, although insignificant, effect on their performance.
Columns 5 through 8 mirror the analysis from columns 1 through 4 but this time on the likelihood that the student attends the subsequent course in Economics (rather than choosing another major or ending their schooling altogether).
Columns 5 & 6, where teacher types are included separately show that having a female exercise teacher and/or a female seminar teacher has a pos-itive effect on the likelihood that a female student chooses to attend Eco-nomics B. The former is only significant when controlling for covariates such as the age of the student, share of afternoon sessions and program. Inter-acting the two teacher type dummies reveals no further conclusions as can be seen in Table A9 in the appendix. When adding controls, the size of the estimate increases and does not appear to approach zero as could be seen for the estimates regarding the performance of the exam. The effect of having a female seminar teacher as a female student is particularly stable.
increases the likelihood that female students opt to attend the subsequent course. Increasing the share from 0 to 1, in this case increasing number of female teachers from 0 to 2, increases the likelihood with approximately 10 percent compared to female students with no female teachers.
Columns 6 & 8 also show that the likelihood of attending the subsequent course is negatively affected by the age of the seminar teacher, however at weak levels of statistical significance. In conclusion, although female teachers do not seem to improve upon female students’ performance, they do inspire them to continue their studies within economics.
Table 7 shows the exact same analysis as the one previously described, but this time on the male student population. Column 1, that regresses the exam points on the two teacher types separately but with no controls other than semester fixed effects, show that being assigned a female exercise teacher and/or a female seminar teacher has no apparent effect on male stu-dents’ performance. However, in column 2, when controlling for the age and quality of teachers and controlling for the share of afternoon sessions, it ap-pears that having a female exercise teacher affects male students negatively. Male students with a female exercise teacher score on average 4.5 percentile ranks lower compared to male students with male exercise teachers, all else equal.
Columns 3 & 4, when we instead look at the share of female teachers, show no significant results regarding exposure to female teachers. Adding controls slightly increases the size of the estimate whilst decreasing the size of the standard errors which points to somewhat robust negative effects of female teachers for male students, but both statistically and economically negligible.
Male students that attend the afternoon sessions perform on average worse on the exam, by close to 8 percentile rank points, compared to male students attending the morning or early afternoon sessions. We also find that male students are positively affected by the quality of the seminar teacher
and negatively by their age in the sense that male students with older seminar teachers perform worse on the exam.
Column 5, where we now turn our attention to the likelihood of attending Economics B, shows a negative effect of being assigned a female seminar teacher, but this result is not robust to including further controls, as can be seen in column 6. When adding controls, the estimates approach zero and the standard errors increase in size. Columns 6, 7 and 8 show that male students that are exposed to female teachers are not significantly less or more likely to attend Economics B compared to male students with male teachers.
T able 6: Regression results: F emal e studen ts only Y = Exam p oin ts, p ercen tile ranking Y = attending subsequen t course (1) (2) (3) (4) (5) (6) (7) (8) V ARIABLES F emale exercise teac her -3.778 -1.7 18 0.0280 0.0520 ** (1.999) (2.152) (0.0177) (0.0208) F emale seminar teac her 1.313 0.528 0.0538* 0.0546 ** (2.203) (1.541) (0.0268) (0.0218) Share of T As that are female -2.166 -0.897 0.0833 * 0.107** (1.288) (2.729) (0.0387) (0.0406) Share of afterno on sessions -4.381 -4.327 0.0212 0.0212 (3.072) (2.816) (0.0315) (0.0316) F emale quota/gro up2 16.69 17.80 -0.0149 -0.0138 (10.90) (9.718) (0.0515) (0.0536) Age of exercise teac her 0.660 0.590 3.22e-05 -5.02e-0 5 (0.414) (0.461) (0.00323) (0.00343) Age of seminar teac her -0.1 78 -0.167 -0.00477* -0.0047 6* (0.306) (0.304) (0.00211) (0.00207) Qualit y of exercise teac her -0.451 -0.583 -0.00437 -0.00450 (0.640) (0.575) (0.00665) (0.00634) Qualit y of seminar teac her 1.8 29 1.631 0.0210 0.0208 (3.815) (3.459) (0.0272) (0.0270) Constan t 61 .84*** 42.04* 61.98*** 44.84* 0.08 40 0.330** 0.0848 0.333** (2.910) (19.70) (2.746) (21.41) (0.0441) (0.131) (0.0441) (0.128) Semester FE YES YES YES YES YES YES YES YES Con trols NO YES NO YES NO YES NO YES Observ a tions 1,077 855 1,077 855 1,213 969 1,213 969 R-squared 0.017 0.155 0.014 0.154 0.022 0.056 0.022 0.056 Standard errors clustered at exercise group lev el. Significance lev els: ∗p < 0.1; ∗∗ p < 0.05; ∗∗∗ p < 0.01.
T able 7: Regression results: Male studen ts only Y = Exam p oin ts, p ercen tile ranking Y = attending subsequen t course (1) (2) (3) (4) (5) (6) (7) (8) V ARIABLES F emale exercise teac her -3.753 -4.5 82** 0.0338 0.0313 (2.754) (1.446) (0.0265) (0.0216) F emale seminar teac her 0.0152 0.466 -0.0487*** -0.0278 (1.193) (1.870) (0.0116) (0.0222) Share of T As that are female -3.644 -3.88 3 -0.0176 -0.000219 (3.265) (2.566) (0.0298) (0.0312) Share of afterno on sessions -7.835*** -7.971** 0.00708 0.00725 (2.016) (2.409) (0.0202) (0.0268) F emale quota/gro up2 -3.828 -3.349 -0.0 596 -0.0622 (10.13) (10.84) (0.0562) (0.0607) Age of exercise teac her -0.160 -0.310 -0.0 0645 -0.00476 (0.651) (0.644) (0.00724) (0.00673) Age of seminar teac her -0.6 04** -0.560* 0.00 425 0.00363 (0.214) (0.232) (0.00330) (0.00340) Qualit y of exercise teac her 0.796 0.568 0.00102 0.00361 (0.888) (0.913) (0.00720) (0.00709) Qualit y of seminar teac her 5.7 92* 5.034* -0.0647 -0.0556 (2.553) (2.201) (0.0422) (0.0393) Constan t 40 .24*** 51.25 40.55*** 56.44* 0.261*** 0.424 0.25 4*** 0.364 (4.761) (28.24) (4.863) (26.47) (0.0411) (0.301) (0.0460) (0.304) Semester FE YES YES YES YES YES YES YES YES Con trols NO YES NO YES NO YES NO YES Observ a tions 1,086 869 1,086 869 1,276 1,03 0 1,276 1,030 R-squared 0.018 0.101 0.016 0.098 0.018 0.05 1 0.013 0.049 Standard errors clustered at exercise group lev el. Significance lev els: ∗p < 0.1; ∗∗ p < 0.05; ∗∗∗ p < 0.01.
6.2
Pooled regressions
Tables 8 and 9 present the results from the pooled regression, where male and female students are analysed simultaneously. Table 8 presents the results from regressions with exam points as the dependent variable and Table 9 presents the results from regressions with attending Economics B as the dependent variable.
Columns 1 & 2 from Table 8 present the results from the pooled regres-sion of exam points on the two teacher types separately and their respective interactions with and without controls. They show that female students per-form on average better than male students on the exam. With controls, the difference amounts to around 5 percentile points when including background characteristics. The gender of the teacher, both exercise and seminar, does not have an effect on the outcome of male students, nor female students as can be seen from the interactions. The effect of having a female exercise teacher is negative, as is the interaction between female student and female exercise teacher, but insignificant. The opposite holds for having a female seminar teacher where both the dummy variable and the interaction have positive estimates, but insignificant as well.
Columns 3 & 4 render similar non-effects where the share of female teachers does not have a significant effect on the performance of neither male nor female students.
In line with the results from regressions performed only on the male students population, attending the afternoon sessions is associated with on average lower exam points as is having an older seminar teacher.
Table 9 presents the results from the analogous analysis to the one scribed above but with the attendance of the subsequent course as the de-pendent variable. It is evident that female students attend the subsequent course at a significantly lesser extent than the male students, even when controlling for program and semester. Keep in mind that the results from
Table 8 showed that female students perform better on average and this dif-ference in attendance is unlikely due to female students not performing well enough.
Columns 1 & 2, where each teacher type is included separately, show that having a female exercise teacher has seemingly no effect on the likeli-hood of attending the subsequent course, both with and without controls. Having a female seminar teacher has negative effects on this likelihood but this estimate is not robust to including controls for teacher quality, age and afternoon sessions et.c. It does however show that having a female seminar teacher has positive effects for female students, almost to the capacity that it offsets the initial lower propensity of female students attending Economics B.
The results from columns 3 & 4, where exposure of female teachers is defined as the share of total teachers that are female, are in line with those from columns 1 & 2. A larger share of female teachers has an initial negative effect, albeit insignificant, but the interaction between being a female student and the share of female teachers show a positive effect on the likelihood of continuing on your studies in economics. The estimates are, just as in columns 1 & 2, large enough to offset the negative effect on the likelihood of attending Economics that is associated with being a female student. In other words, female students that are exposed only to female teachers (other than the head teachers that have always been men) attend Economics B at the same extent as male students, all else equal.
Table 8: Regression results: All students
Y = Exam points, percentile ranking (1) (2) (3) (4)
VARIABLES
Female student 7.558* 5.109** 7.337** 4.988**
(3.133) (2.000) (2.785) (1.689)
Female exercise teacher -2.690 -3.509*
(3.082) (1.773)
Female student x Female exercise teacher -2.036 -0.436
(4.445) (3.332)
Female seminar teacher -0.709 0.0258
(1.773) (1.437)
Female student x Female seminar teacher 2.647 -0.112
(3.054) (2.525)
Share of TAs that are female -3.165 -3.281
(3.927) (2.753)
Female student x Share of female teacher 0.576 -0.338
(4.159) (1.440)
Share of afternoon sessions -6.452** -6.434**
(2.191) (2.174)
Female quota/group2 5.788 6.990
(8.713) (8.408)
Age of exercise teacher 0.168 0.0487
(0.444) (0.450)
Age of seminar teacher -0.445** -0.423*
(0.160) (0.181)
Quality of exercise teacher 0.212 0.0112
(0.684) (0.712)
Quality of seminar teacher 4.037 3.606
(2.965) (2.627)
Constant 47.33*** 48.52* 47.62*** 53.18*
(4.309) (21.47) (4.163) (22.18)
Semester FE YES YES YES YES
Controls NO YES NO YES
Observations 2,163 1,724 2,163 1,724
R-squared 0.021 0.115 0.018 0.113
Table 9: Regression results: All students
Y = Attending subsequent course (1) (2) (3) (4)
VARIABLES
Female student -0.115*** -0.146*** -0.121*** -0.149***
(0.0223) (0.0246) (0.0226) (0.0255)
Female exercise teacher 0.0211 0.0156
(0.0221) (0.0245)
Female student x Female exercise teacher 0.0158 0.0559
(0.0343) (0.0323)
Female seminar teacher -0.0498*** -0.0336
(0.0121) (0.0222)
Female student x Female seminar teacher 0.108** 0.100***
(0.0334) (0.0262)
Share of TAs that are female -0.0274 -0.0195
(0.0306) (0.0416)
Female student x Share of female teacher 0.121* 0.155**
(0.0609) (0.0574)
Share of afternoon sessions 0.00679 0.00720
(0.0198) (0.0223)
Female quota/group2 -0.0589* -0.0604*
(0.0247) (0.0263)
Age of exercise teacher -0.00281 -0.00198
(0.00490) (0.00456)
Age of seminar teacher -0.000114 -0.000395
(0.00190) (0.00190)
Quality of exercise teacher -0.00398 -0.00274
(0.00391) (0.00402)
Quality of seminar teacher -0.0181 -0.0137
(0.0171) (0.0185)
Constant 0.230*** 0.483** 0.231*** 0.455**
(0.0213) (0.155) (0.0238) (0.155)
Semester FE YES YES YES YES
Controls NO YES NO YES
Observations 2,489 1,999 2,489 1,999
R-squared 0.022 0.047 0.020 0.046
6.3
Robustness and further analysis
Several robustness checks have been performed to look at the rigidity of the estimated effects presented above. First off, the two different ways of clustering standard errors are compared but they do not call to question the results and/or conclusions. These comparisons are found in the appendix in tables A3 through A8.
Secondly, not all students who attend the course will attempt the exam at the first opportunity, or at all. If female teachers lower the propensity for female students to opt out of taking the exam, and these marginal students tend to come from the left tail of the points distribution, our OLS estimates for the interaction effect may be downward biased. This point is made by Hoffman & Oreopoulos (2009) who, following the example in Lee (2005) apply a two-step procedure to re-estimate an upper bound for the interaction effect when excluding the worst performing female students in accordance to the relative dropout rates between female and male students assigned a female teacher. Replacing the dependent variable with a dummy equal to one if the student did not attempt the exam at the first opportunity renders no significant effects, neither for the gender of the student nor the gender of the teacher or any interactions. It is thus unlikely that the estimates are downward biased due to this. Results from these regressions are included in the appendix in Table A10. An alternative approach to this is to include students that have not attempted the exam by recoding their missing values to hold the value 0. The results from this is included in the appendix in table A11 but they do not contribute to any alternative conclusion.
Furthermore, one could suspect that having both a female seminar teacher and a female exercise teacher could have either a multiplicative or dulling ef-fect on either exam points or the likelihood of attending Economics B for female students. However, the results from regressions where this interaction is included show no such tendencies, as can be seen in the appendix in Table A9.
