Graduate School of Social Sciences
MSc Sociology, track Gender, Sexuality and Society
Thesis
STEM-WOMEN AND EHW-MEN: GENDER ROLE EXPECTATIONS AND ATTITUDES IN EXPLAINING SEX DIFFERENCES IN ACADEMIC PERFORMANCE
Name student: Lyydia Alajääskö Student ID: 12113263
Supervisor: Dr Agnieszka Kanas Second reader: Dr Herman van de Werfhorst
Date submitted: 01/07/2019 in Amsterdam, the Netherlands
Abstract
This paper reviews explanations for sex differences in academic performance in lower education. More specifically, it aims to answer the following research questions: “To what extent do gender role expectations mediate the effect of gender on academic performance?” (RQ1) and “To what extent do the prevalent gender role attitudes in the country affect the influences of gender and gender role expectations on academic performance?” (RQ2). Using data from the latest cycle of OECD’s Programme for International Student Assessment, four multi-level models are built. From these, it is found that gender role expectations do not mediate the effect of gender on academic performance to a meaningful extent. Furthermore, the results indicate that as country-level gender role attitudes become more egalitarian, sex differences in mathematics decrease, but sex differences in reading increase. Furthermore, gender role attitudes are not significantly related to science performance. In all subject fields, it is found that gender role attitudes have a contextual effect on gender role expectations (operationalised as expected occupation). Namely, as gender role attitudes become more traditional, going against gender role expectations (i.e. expecting an occupation traditionally associated with the opposite sex) is related to lower performance in the subject field related to the occupation. Importantly, the gender stratification hypothesis, according to which sex differences in academic performance emerge from gender inequalities in educational opportunities, fails.
Contents
Abstract ... 2
Contents ... 3
1. Introduction ... 5
2. Theoretical framework ... 7
2.1 Explanations: nature versus nurture ... 7
2.2 Gender inequalities in education ... 8
2.2.1 Gender equality indicators ... 9
2.3 Gender role attitudes ... 10
2.4 Gender role expectations ... 11
2.5 Gender role attitudes and gender role expectations ... 13
2.6 Theoretical model and hypotheses ... 14
3. Data ... 15
3.1 Program for International Student Assessment (PISA) ... 15
3.1.1 Variables of interest ... 15
3.2 Gender role attitudes (GRA) ... 16
3.3 Controls ... 17 3.3.1 PISA ... 17 3.3.2 Institutions ... 19 3.4 Sample selection ... 19 3.5 Descriptive statistics ... 20 3.5.1 Variables of interest ... 20 4. Methodology ... 21 4.1 PISA data ... 21 4.2 Regression equations... 22 5. Analysis ... 23
5.2 Regression results ... 24
5.2.1 Models one to three ... 24
5.2.2 Model four... 29 6. Conclusion ... 31 6.1 Discussion ... 31 6.1.1 Hypothesis one ... 31 6.1.2 Hypothesis two ... 32 6.2.3 Hypothesis three ... 32 6.2.4 Hypothesis four ... 33 6.2 Critical evaluation ... 33
6.2.1 First research question ... 33
6.2.2 Second research question ... 34
References ... 36
Appendix A ... 39
1. Introduction
Women’s labour force participation rates in developed nations have increased over the last decades (Buchmann, DiPrete & McDaniel, 2008; Stoet & Geary, 2013). In many European countries the rates of female participation are almost equal with male participation rates (European Commission, 2016). Furthermore, even though occupational sex segregation declined heavily in developed nations between the 70s and the 90s (Buchmann et al., 2008), occupational segregation still persists. Researchers today note an underrepresentation of women in the sectors of science, technology, engineering, and mathematics (STEM) (Stoet & Geary, 2013; Reilly, 2012), and an overrepresentation of women in sectors of education, human health and welfare (EHW) activities (EIGE, 2017).
Occupational segregation has been linked to educational segregation. Firstly, there is an equivalent separation in higher education with more women choosing humanistic majors in college and more men opting to major in physical science, technology, engineering, and mathematics (Stoet & Geary, 2015). Secondly, there are comparable sex differences in academic performance in lower education. On average, girls over perform in reading but underperform in mathematics and science (relative to boys) (OECD, 2018a). It is suggested that girls’ relatively higher language-related competencies lead them to choose humanities fields in college, whereas boys’ outperformance in mathematics guides them to STEM fields (Stoet & Geary, 2015).
There is considerable international variation in terms of magnitude as well as direction of sex differences in lower education (Buchmann et al., 2008). For this reason, many researchers have looked at macro-level variables to explain cross-country differences (Stoet, Bailey, Moore & Geary, 2016; Stoet & Geary, 2013, 2015; Reilly, 2012). Naturally, country-level gender equality has been an area of concentration in research into sex differences. More specifically, it has been argued that the level of gender equality in the country of residence is linearly related to sex differences in secondary education (Stoet & Geary, 2015).
In using gender equality indicators, researchers are assessing the effect of differences in social, economic and political opportunities for boys and girls (Stoet & Geary, 2015). However, it could be argued that these indicators tend to capture the effect of institutional inequality (e.g. restricted physical integrity for females and discriminatory family code (OECD, n.d.). Thus, using gender equality indicators to explain sex differences in more developed countries, where institutional barriers are mostly eliminated (Buchmann et al., 2008), can pose a problem.
Less prominent in research is a specific focus on both macro-level gender role attitudes as well as related micro-level gender role expectations. On the individual level, students’ perception of their own competencies is influenced by surrounding gender role attitudes (Buchmann et al., 2008).
This happens through a process of socialisation where students internalise gender role expectations (van der Vleuten, Jaspers, Maas & van der Lippe, 2016). For example, boys with more traditional gender role attitudes, and thus, traditional gender role expectations, choose traditionally masculine educational tracks in secondary education (Ibid.). As such, gender role expectations may explain some part of sex differences in academic performance in lower education.
With respect to the macro-level, it could be argued that by using gender equality indicators researchers are implicitly testing the effect of gender role attitudes. Rodríguez-Planas and Nollenberger (2018) also explicitly address the concept of gender role attitudes, but do not use them in their research. While gender role attitudes and gender equality indicators are evidently related, they are not necessarily equal. Gender role attitudes are reflected in institutional inequalities, however, they also exist outside of institutional inequalities.
An example of this is “the good housewife and mother” norm, which is still prevalent in contemporary Europe (EIGE, 2013, p. 44). For instance, in Greece women have the same legal rights and decsion-making abilities as men (SIGI, 2019). However, in religious marriages, women formally ackonwledge their role of taking care of the house, children, and more distant family (Ibid.). As such, there is a deviation between institutional practices and gender role attitudes within the country.
Considering this together with the suggested link between sex differences in lower education and higher education, it is noted that prevalent gender role attitudes within a country may affect students’ gender role expectations. This on the other hand has implications on academic performance in that students are more likely to concentrate on subjects that are associated with their sex.
Following the above, the research questions are formulated as follows:
1. To what extent do gender role expectations mediate the effect of gender on academic performance?
2. To what extent do the prevalent gender role attitudes in the country affect the influences of gender and gender role expectations on academic performance?
Given the gendered reality of not only the current labour market, but also of higher education, studying sex differences in academic performance on earlier stages is socially relevant. In terms of educational policy-making as well as adopting successful policies to combat gender inequalities at later stages, understanding the determinants of sex differences in earlier education is of vital importance.
OECD’s Programme for International Student Assessment (PISA) datasets have been used extensively to analyse the sex differences. PISA is a test for 15-year-olds organised triennially to
assess educational systems worldwide (OECD, 2018a). This is valuable in assessing whether country-level socio-cultural determinants are at play. Research concentrating into gender role attitudes, specifically, is scientifically valuable to address a gap in the literature that has previously concentrated on more institutional factors. However, as institutional factors are not fully eliminated even in the most developed countries, this paper will make use of them as controls in the regressions. Furthermore, assessing potential mediation effects is beneficial in the aim of deciding successful policies.
Moreover, most research into gender inequality has concentrated mostly on mathematics (and reading) scores. As such, extending the analysis onto science scores provides a relatively unexplored avenue for research. Thus, because this paper also looks at students’ performance in science, the scientific relevance is amplified.
2. Theoretical framework 2.1 Explanations: nature versus nurture
Sex differences in lower secondary academic performance have been widely documented and researchers generally agree that they exist. There is considerable international variation in the observed directions of the differences (Buchmann et al., 2008). However, in most countries it is found that on average girls outperform boys in reading tests, whereas boys outperform girls in mathematics (Ibid.). Sex differences in science are less studied, perhaps because the differences are smaller (OECD, 2018a). Nevertheless, in 2015, 33 countries out of 72 had more boys than girls among the top performers in science (Ibid.). Nevertheless, across countries the magnitude of the reading difference is consistently observed to be the largest of the three (Stoet & Geary, 2013).
Thus, the first hypothesis is:
H1a: On average girls outperform boys in reading scores.
H1b: On average boys outperform girls in mathematics and science scores.
H1c: Sex differences in reading are largest, whereas sex differences in science are smallest.
Moreover, results are inconclusive with regards to whether sex differences in lower secondary education are declining over time (Buchmann et al., 2008). In other words, it has not been established whether over the years, boys and girls in the age cohort of lower education are becoming better at reading and mathematics & science, respectively. A related question is about the determinants of sex differences. Namely, researchers reporting persisting large sex differences explain their results with reference to sex differences in the cognitive ability of girls and boys (Stoet & Geary, 2015). On the other hand, researchers arguing that sex differences are narrowing refer to environmental factors as an explanation (Ibid.).
This debate about the determinants of sex differences in academic performance lies within the long-standing nature versus nurture debate. Most sociological research into sex differences concentrates on socio-cultural variables as explanation, rather than biological ones (Buchmann et al., 2008). However, a justification for this may be that cognitive differences between the sexes are minor as compared to other biological differences (Ibid.). More specifically, a meta-analysis finds that the sex differences on most psychological variables (including mathematics performance and verbal ability) are small or close to zero for all ages (Hyde, 2005). Hyde notes only three areas in which differences between females and males are moderate to large: motor performance, sexuality, and aggression.
As such, Hyde (2005) finds support for the gender similarities hypothesis, namely, that females and males are psychologically more similar than they are different. What this means is that both sexes should have an equivalent aptitude for learning the mathematical and verbal subjects. From this it follows that biological differences may not be the driving force behind sex differences in academic performance. Instead, sex differences appear to be due to sociocultural-economic variables. This is supported by findings that boys and girls start their educational career by performing similarly in reading and mathematics but develop increasing differences as they progress through grades in school (Buchmann et al., 2008).
2.2 Gender inequalities in education
Based on the gender similarities hypothesis, the gender stratification hypothesis is a prominent version of “nurture” based arguments about educational differences between boys and girls (Stoet & Geary, 2013, 2015; Reilly, 2012). This hypothesis originates from sociology and carries ties with the historically dominant position of men in society (Stoet & Geary, 2015). Much like men’s dominance in history is based on superior income and power, the sex differences in academic performance are a result of gender inequalities in economic and educational opportunities (Ibid.).
An implication of this hypothesis is that implementing gender equality policies should inhibit and ultimately eliminate educational sex differences (Stoet & Geary, 2015). As such, proponents of the hypothesis argue that the cross-country variation in sex differences among students in lower secondary education is directly linked to country-level variation in gender equality (Stoet & Geary, 2013). This idea has been tested particularly in the context of mathematics performance (Ibid.). More recently, research into levels of gender equality as explanation to academic sex differences has spread to reading and science performance as well (Stoet & Geary, 2015).
2.2.1 Gender equality indicators
Most research using PISA data and studying the effect of gender equality indicators work with multiple indicators at once. For instance, using the 2009 cycle, González de San Róman and de la Rica Goiricelaya (GSR & RG) (2012) test the gender stratification hypothesis with the following gender equality indicators: Gender Gap Index (GGI) and its sub index political empowerment, Female Labour Force Participation rate (FLFP), gender housework ratio, as well as an average World Values Survey indicator (WVS). They find that as gender equality increases, girls’ performance on the country-level increases in both reading and mathematics. This is also the case on the individual level while controlling for individual-, school- and country-level variables.
By using variables such as the WVS indicator and gender housework ratio, the authors are as a matter of fact assessing the effect of gender role attitudes. However, their analysis falls short because the WVS indicator is obtained by simply taking an average of the survey items. Using more complex statistical techniques, such as the principal component analysis, should yield a more accurate indicator. Furthermore, the gender housework ratio is only available for 2003, meaning that there is a six-year gap between the independent and dependent variables. More up to date variables are desirable.
Rodríguez-Planas and Nollenberger (2018) also make use of the GGI and all its subindexes. Much like GSR & RC (2012), they use a multi-level model with individual test scores regressed on the interaction between gender and country-level gender equality. In this model, individual, and school variables are controlled for. The authors find that higher gender equality has a positive effect on girls’ test scores relative to boys’ in reading, mathematics and science.
However, the results discussed above are contested by Stoet and Geary (2013; 2015) who do not find a relation between academic achievement and gender equality indicators. The authors make use of GGI as well as a multitude of other gender equality indicators. They additionally include gender inequality index, gender development index, & gender empowerment index in 2013, and gender empowerment index, women in research (WIR), & FLFP rates in 2015. In both studies, the PISA cycles 2000, 2003, 2006 and 2009 are pooled. This means that the data set differs by two cycles from the data set of Rodríguez-Planas and Nollenberger (2018). Furthemore, instead of a multilevel analysis, Stoet and Geary analyse sex differences only on the country-level. Finally, their model does not include any control variables.
Much like Stoet and Geary (2013, 2015), Reilly (2012) analyses country-level sex differences in reading, mathematics and science. Similarly to GSR & RG (2012), he only employs the 2009 cycle. Using the indicators global Gender Gap Index (GGI), relative status of women, and WIR, the author does not find support for the gender stratification hypothesis. However, he
emphasises that this is not a refutation of the hypothesis, but rather an indicator that the instruments generally used to test the hypothesis should be reconsidered.
An important difference between papers that do and do not find support for the gender stratification hypothesis is found in the methodology. Using multi-level analyses researchers find that as gender equality increases girls’ academic performance increases. However, country-level analyses do not find any relation. In an analysis conducted on the country level there is no possibility to observe or control for variance on other levels, such as school level. Furthermore, country-level analysis reduces the number of observations and, as such, the robustness of the results is questionable (GSR & RG, 2012). By looking at other levels, research is more informative. Due to this, this paper is more in line with the methodology of GSR & RG (2012) and Rodríguez-Planas & Nollenberger (2018).
2.3 Gender role attitudes
There are limitations in using gender equality measures, such as GGI, as indicators of gender inequalities in economic and educational opportunities. While these do measure institutional inequalities, and as such, relate to insititutional barriers to opportunity, the problem is that institutional barriers have mostly been addressed in industrial countries (Buchmann et al, 2008). This raises questions about the relevance of the gender stratification hypothesis in more developed socities. If most institutional barriers are accounted for, why would there be inequalities in economic and educational opportunities for boys and girls?
An important distinction is that of the private and the public sphere. Institutions and institutional barriers are situated in the public sphere. But gender inequalities exist in the private sphere too. Gender role attitudes are evidently related to the public sphere in that the aforementioned institutionals barriers reflect gender roles.
However, the distinction made is important because while institutional barriers are mostly eliminated, gender roles in the private remain. This is because they are slower to change (Rodríguez-Planas & Nollenberger, 2018). England (2010) suggests that this is because unlike in the public sphere, in the private sphere there is a lack of economic benefit for transgressing gender roles(Ibid.). As such, prevalent gender role attitudes within a country relating to family life and the relationship between man and woman, can strongly define an individual’s life course (EIGE, 2013).
In a qualitative study on gender perceptions in the EU the following gender roles were found to significantly affect the lives of the citizens (EIGE, 2013). Firstly, it is found that for women there is a pressure to be “the good housewife and mother” (EIGE, 2013, p. 44). This includes prioritising the well-being of the family at the cost of personal sacrifices, such as staying at home instead of working. Furthermore, housework is believed to be women’s rather than men’s work (Ibid.). Men,
on the other hand, are expected to work outside of the house and to be “the (stoic) provider”. Moreover, men are to have complete authority within the family and over the woman, who “should accept his decision” (Ibid, p. 44).
From the previous, four broad gender role categories are deduced: good housewife and mother, housework as women’s work, breadwinner husband, and man as decision maker and head of family. General belief in these gender roles can act as an ideological barrier for adolescents. Through the process of socialisation boys and girls internalise them from their social environment into their own gender role attitudes, and learn to behave accordingly (England, 2010; van der Vleuten et al., 2016). Through a system of accountability individuals are then judged on whether they are performing their gender roles appropriately (West & Zimmerman, 1987).
While rebelling against gender roles is possible, individuals who fail to meet the gendered standards are held accountable by their social surrounding (West & Zimmerman, 1987). In other words, non-compliant gender behaviour is punished and minimised by the social surrounding. Because teenagers are still forming their identity, it is highly probable that they will conform to gender roles (van der Vleuten et al., 2016). Consequently, ideological barriers are those that limit the options that individuals perceive themselves to have.
Because there are large practical consequences to these ideological barriers, the gender stratification hypothesis about unequal opportunities still applies. Just as men are seen as the sole income-provider, they are also associated with rationality and mathematical capability (van der Vleuten et al., 2016). Women on the other hand as caregivers and homemakers are seen to have more nurturing and verbal abilities (Ibid.).
Accordingly, existing gender roles not only prescribe appropriate behaviour, but also appropriate fields for boys and girls. These fields can be understood as educational fields (i.e. mathematics and science for boys, reading for girls), but the argumentation can also be extended to fields in the labour market (i.e. STEM fields for men, EHW fields for women). As it follows, there are ideological barriers for individuals entering a field that is opposite of gender role expectations. 2.4 Gender role expectations
On the individual level, gender inequalities in education are thought to emerge through interactions within family and school. For instance, there is some empirical evidence that boys and girls react differently to female and male teachers (Buchmann et al., 2008). While highly debated, this may suggest that teachers favour one gender over another (Ibid.). Furthermore, research indicates that the performance of girls and boys varies with classroom environments (Ibid.). More specifically, the stereotype threat (i.e. “the fear of conforming to a stereotype about a subgroup to which one belongs” (Ibid., p. 323)) is thought to affect students.
Stoet et al. (2016) explore the stereotype threat in the form of mathematics anxiety in the 2003 and 2012 PISA cycles. Using the GGI, they find that as gender equality increases, sex differences in mathematics anxiety also increase. This finding opposes the gender stratification hypothesis. Moreover, they also explore the role of parental involvement as a mediator variable. They find that parents value boys’ performance in mathematics more than girls’ even as gender equality increases. Finally, they do not find a relation between the proportion of mothers working in STEM and mathematics performance. Their overall findings point towards the rejection of the gender stratification hypothesis. However, much like Reilly (2012) and Stoet & Geary (2013; 2015), they only use a country-level model.
Van der Vleuten et al. (2016), on the other hand, study the effect of individual-level gender role attitudes on educational track choice. Competence beliefs, occupational values and subject preferences are suggested mechanisms through which gender role attitudes have an effect (Ibid.). The authors find that traditional gender role attitudes are associated with traditional competence beliefs, traditional occupational values as well as traditional subject preferences for both boys and girls.
Van der Vleuten et al.’s (2016) study is conducted within the context of the Netherlands, in which studying educational track choice in secondary education is relevant. However, this is not the case for all countries. In some countries no formal differentiation between students is made before finishing secondary education (OECD, 2005). Moreover, the level of stratification in education varies between countries (Han, 2016).
Another variable that may reflect gender role expectations is expected occupation. Expected occupation is also related to the mechanisms suggested by van der Vleuten et al. (2016) in the following ways. Firstly, high beliefs about one’s competence within a field increase the likelihood of choosing to work in that field (van der Vleuten et al., 2016). Secondly, what people value and want from the future can quite powerfully predict their chosen field (Ibid.). Thirdly, and intuitively, people are more likely to follow a track which they like (Ibid.). Furthermore, assuming that parents act as role models for their children, expected occupation may also reflect the information contained in the variable proportion of mothers working in STEM, used by Stoet et al. (2016).
As such, the second and third hypotheses are:
H2: Sex differences are (partly) explained by occupational expectations of girls and boys H3a: Sex differences are smaller in gender egalitarian countries.
2.5 Gender role attitudes and gender role expectations
Van der Vleuten et al. (2016) also find some evidence for the effect of gender role attitudes on educational track choice for boys. The authors measure the gender role attitudes variable earlier than they measure the educational track choice of students. This lends support for the causal claim that gender role attitudes influence educational track choice (Ibid.). This also implies that it is sensible to assume that gender role attitudes influence expected occupation.
Interestingly, the authors do not find a significant effect of gender role attitudes on girls’ educational track choice (van der Vleuten et al., 2016). As an explanation they suggest that gender role expectations are stricter for boys. Linking to the notion of accountability, the previous implies that boys are held accountable to a higher degree than girls for not following gender role expectations. As such, boys are expected to face higher ideological barriers, and thus, underperform by a larger magnitude in reading than girls underperform in mathematics and science.
England’s (2010) theorising on asymmetric desegregation of the labour market relates to the above. She observes that the egalitarian gender role attitudes in developed nations are influenced by the positive valuation of upward mobility. Her argument is as follows. The emphasis on upward mobility means that individuals are incentivised to aim towards higher earnings or towards a higher social status than their preceding generations (Ibid.). Combined with the fact that activities and occupations typical for women are devalued and poorly rewarded, there is an incentive for both men and women to seek “male” occupations. It is also due to this devaluation that boys are held more accountable for displaying traditionally female gender roles. Their choice is deemed to make them worse off.
Additionally, according to the accountability system (West & Zimmerman, 1987), within contexts that have more traditional gender roles, displaying inappropriate gender roles should be punished more heavily. This implies that ideological barriers are larger in countries with more traditional gender roles. Gender stratification hypothesis, on the other hand, stipulates that as inequalities in opportunities increase, sex differences in education increase. Given that ideological barriers constitute an inequality, the following relationship is posited: in countries with more traditional gender roles, choosing a labour field inappropriate for one’s gender is associated with underperformance in the related educational field. This underperformance is relative to peers who choose a labour field appropriate for one’s gender.
Following the above, hypothesis 4 is:
H4a: Boys who expect an EHW occupation in a gender egalitarian country perform better in reading than boys who expect an EHW occupation in a less egalitarian country.
H4b: Girls who expect a STEM occupation in a gender egalitarian country perform better in mathematics and science than girls who expect a STEM occupation in a less egalitarian country. 2.6 Theoretical model and hypotheses
The theoretical model is presented in figure 1. Firstly, it is expected that there is an effect of gender on academic performance (H1). Academic performance includes variables for reading, mathematics, and science performance. Secondly, it is expected that part of this effect is indirect and mediated by expected occupation (H2). Thirdly, it is expected that sex differences are smaller in countries with more gender egalitarian attitudes. This is presented as an interaction between gender and gender role attitudes (H3). Finally, it is expected that the effect of expected occupation is also conditional on gender role attitudes (H4). Table 1 summarises the hypotheses.
Figure 1. Theoretical model Table 1
Hypotheses
H1a : On average girls outperform boys in reading scores.
H1b : On average boys outperform girls in mathematics and science scores.
H1c : Sex differences in reading are largest, whereas sex differences in science are smallest. H2 : Sex differences are (partly) explained by occupational expectations of girl and boys H3a : Sex differences are smaller in gender egalitarian countries.
H3b : Sex differences in reading remain largest, whereas sex differences in science remain smallest.
H4a : Boys who expect an EHW occupation in a gender egalitarian country perform better in reading than boys who expect an EHW occupation in a less egalitarian country.
H4b : Girls who expect a STEM occupation in a gender egalitarian country perform better in mathematics and science than girls who expect a STEM occupation in a less egalitarian country.
3. Data
3.1 Program for International Student Assessment (PISA)
All variables at the individual- and school-levels that are used in the analysis are acquired from OECD’s Program for International Student Assessment (PISA) dataset. More specifically, this paper makes use of the dataset collected in 2015.
The PISA is a comprehensive international test conducted triennially. It aims to evaluate schooling systems worldwide, by testing the knowledge and ability of 15-year-olds in the participating countries. 15-year-olds are chosen as the subjects of the test, because in most countries this age marks the end of compulsory schooling (Stoet & Geary, 2015). Fields of knowledge that are tested include mathematics, science and reading as well as collaborative problem solving and financial literacy (OECD, 2018a).
A two-fold stratification technique is used in obtaining the final sample (OECD, 2017b). First, within each country a representative sample of schools is selected (Ibid.). Second, schools provide a list of eligible students from which again a representative sample is selected (Ibid.). In 2015, about 540,000 students in 72 countries and economies participated in taking the test (OECD, 2018a).
3.1.1 Variables of interest
Reading, mathematics and science performance
Reading, mathematics and science performance are dependent variables of three separate and distinct sets of regressions. The dependent variables are measured by observing students’ scores in PISA test questions pertaining to reading, mathematics and science, respectively. To maintain a reasonable timeframe for completing the test, not all students complete all test items (OECD, 2018a). As such, instead of one single score PISA provides several plausible values (PV) for test scores (OECD, 2009).
Plausible values represent the possible test scores a student would have obtained had they completed the whole test (OECD, 2009). In the 2015 cycle, 10 PVs were calculated per student. These scores are standardised so that they average at 500 points and have a standard deviation of 100 points (Ibid.). The dependent variables read, math, and scie, that are constituted from these scores, are continuous variables. Section 4.2 explains the methodology regarding PVs.
Female
The variable female (F) is a dummy that equals 1 when the respondent is female, and 0 when the respondent is male. It is used as an independent variable as well as in interactions with other variables of interest.
Gender role expectation: Expected occupation
Gender role expectations are operationalised as expected occupation. More specifically, the variable consists of two dummy variables STEM and EHW, where each equal 1 when the respondent expects an occupation in the respective field and 0 otherwise. The variables are obtained by recoding dummies from expected occupations expressed in ISCO-08 codes. STEM fields fit into three occupational categories: science and engineering professionals (ISCO 21), information and technology professionals (ISCO 25), and science and engineering associate professionals (ISCO 31) (European Commission, 2015). EHW fields, on the other hand, cover the following ISCO categories: health professionals (ISCO 22), teaching professionals (ISCO 23), health associate professionals (ISCO 32), and personal care workers (ISCO 53) (EIGE, 2018).
3.2 Gender role attitudes (GRA)
Macro-level gender role attitudes (GRA) are used as explanatory variables.They are constructed from survey data regarding gender values. This is obtained from the European Institute for Gender Equality (EIGE), but was initially collected by the Eurobarometer. EIGE is an independent European Union institute that advocates for gender equality within EU policy (at the EU level as well as national levels) (EIGE, n.d.). Furthermore, it aims to spread information amongst European citizens about gender equality (Ibid.). As such, they have a mission to provide high quality data (Ibid.).
Using key search words1, statements and questions in the EIGE/Eurobarometer database relating to gender role attitudes are collected. The following items are included as potential factors in a principal component analysis.
for the gender role good housewife and mother e2 family life suffers with a working mother e4 men should work in childcare sectors for the gender role of housework as women’s work
e3 women care less about career than men
e5 men are less competent than women in housekeeping for the gender role breadwinner husband
e6 fathers should prioritise career over childcare
e8 gender equality enhances women’s economic independence e9 more women in labour market improves the economy for the gender role of man as decision maker and head of family
1 The search words used are: ”gender”, ”woman”, ”man”, ”wife”, ”husband”, ”women”, ”men”, ”mother”, ”father”,
e1 tackling gender inequality as a priority e7 gender equality as a fundamental right e10 gender equality is needed for fairer society
Table A1 in the appendix presents the exact wordings of the items. Furthermore, table A2 presents the pairwise correlation between the items.
The items are constructed such that the answer indicates the level of agreement or disagreement with the statement. Thus, the first step in composing a gender role attitudes index is to recode all the answers in a way so that a higher value indicates more egalitarian gender role attitudes. More specifcally, the value 1 is attached to answers representing the most traditional gender role attitudes, whereas the value 5 is attached to the most egalitarian gender role attitudes.
Table A3 in appendix displays the results from a simple PCA. Based on the adjusted version of Kaiser’s criterion and a visual estimation of the scree plot (Jolliffe, 2002) (see figure A1), it is decided to include three components (see table A4). Promax and varimax rotations (see table A5) are used to determine simplifications as well as enhance the validity of the created index (Ibid.). The Kaiser-Meyer-Olkin (KMO)2 postestimation (see table A6) is also used to ensure the validity of the index (StataCorp, 2017). Furthermore, the Cronbach’s alpha3 measure (see table A7) is obtained to guarantee that the summative rating is reliable (Ibid.).
Three components are formed as follows. The first component (GRA1) includes variables e1, e3, e6, e7 and e10. As such, it covers the gender role of man as decisionmaker and head of family. Furthermore, it includes related aspects from the roles of housework as women’s work and breadwinner husband. The second component (GRA2) consists of variables e2, e4, e5. It covers the gender role good housewife and mother, as well as incorporates a part of the belief that housework is women’s responsibility. Finally, in component three (GRA3) there are variables e8 and e9, these reflect the gender role breadwinner husband.
3.3 Controls 3.3.1 PISA
A unique advantage of the PISA test is that both students and schools are surveyed for background information (OECD, 2017a). From is information, control variables can be operationalized. Following Rodríguez-Planas & Nollenberger (2018) and González de San Román & De La Rica (2016), several variables are controlled for, to arrive at the effect of gender role attitudes on academic performance. These are listed below.
2 Small KMO values (below 0.5) indicate that the variables do not have enough in common to justify PCA (StataCorp,
2017). When PCA is used mainly to reduce data, the assumptions are less important (Universiteit van Amsterdam, n.d.)
3
Economic and socio-cultural status (ESCS)
ESCS is an index of economic, social and cultural status. It is a continuous variable that is constructed, via a principal component analysis, using the following indicators: the level of parental education, highest level of parental occupation, and home possessions (OECD, 2017b). It is included as a control because previous literature suggests that sex differences in educational performance are more pronounced among children from economically disadvantaged families (Buchmann et al., 2008). The ESCS scores are standardised so that the score of an average student is zero, and the standard deviation is one (OECD, 2017b).
Grade
Delayed entry into kindergarten in order to allow the student extra time for socioemotional or intellectual development is more likely for boys than girls (Buchmann et al., 2008). It is also more common in elementary school that boys experience grade retention than girls (Ibid.). This means that among 15-year-olds, on average, girls will have attained a higher grade. Attaining a higher grade implies having been taught more. As such, it is important to control for the effect of grade on academic performance.
Grade is a categorical variable which gives the grade of the student as compared to the modal grade in the respondent’s country. For instance, it gives a value 0 for the modal grade and values -1 and 1 for one grade lower and higher, respectively.
Immigration status
For immigrant children there may be more than one model of gender roles: the one from the country of origin and the one from the country of residence (Rodríguez-Planas & Nollenberger, 2018). Furthermore, these two models may differ significantly. As such, the effect of gender roles in the country of residence on academic performance can be different for immigrant children than their native peers. For example, even if a student lives in a country with non-traditional gender roles, they may still be influenced by traditional gender roles in their homes. Most importantly, students of immigration background tend to perform worse than their native peers (OECD, 2018b). Therefore, it is important to control for students’ immigration background.
Immig is a categorical variable for the respondent’s immigration status. It takes on the value 1 for a native respondent, the value 2 for a second-generation immigrant, and the value 3 for a first-generation immigrant.
School ID
Assuming that parents choose in which schools their children enrol and that their choice is influenced by prevalent cultural norms, school-level variables can be understood as endogenous (Rodríguez-Planas & Nollenberger, 2018). Examples of such variables include school type (whether
a school is private, public or something in the middle) and student-teacher ratio. As these have clear effects on academic performance, it is important to control for their effect.
Alternatively, it is argued that more social interaction between the two sexes can lead to heightened between-sex comparisons, which in turn can lead to larger sex differences in performance (Stoet et al., 2016). This is because as competition increases, sex differences in subject-related anxiety increase facilitating sex differences in performance (Ibid.). The extent of social interaction between boys and girls depends on percentage of girls versus boys in the school.
Given the above, school ID is included as a control variable. It is a categorical variable, which has a different value for each school. It takes into account all the variation between schools. In other words, it includes the effect of school type, student-teacher ratio, as well as girl percentange. The advantage of controlling for school ID only – rather than three separate variables – is that this allows for a more parsimonious model. Furthermore, there is a large number of missing observations in the school specific variables.
3.3.2 Institutions
Finally, to arrive at the effect of gender role attitudes only, the following institutional factors are controlled for. For direct comparability, this data is obtained for the same time period as the gender role attitudes.
Gender Gap Index (GGI)
As discussed throughout this paper, it has been argued that gender equality indicators are linked to sex differences in academic performance. Namely, it is suggested that a higher level of equality is related to smaller sex differences. As mentioned in the theoretical framework, the most commonly used indicator is the Gender Gap Index (GGI) by the World Economic Forum. As such, this paper makes use of the GGI value in 2014 as a control for institutional gender equality. This variable takes on values between 0 and 1, with larger values indicating higher gender equality.
Log of GDP per capita
Gender inequalities in education vary largely from less developed societies to more developed societies (Buchmann et al., 2008). Thus, this paper controls for this effect by including the log of GDP (log GDP) per capita as a control variable. This is obtained for each country in 2014 from the World Bank open data base. It is measured in thousands of current U.S. dollars.
3.4 Sample selection
As both EIGE and Eurobarometer operate within the EU exclusively, the initial PISA sample is restricted to the countries that are also included in the survey by Eurobarometer. This consists of the countries in the EU in 2014, excluding the United Kingdom. This limits the sample to 26 countries with 162,386 respondents. Furthermore, there are a number of missing observations for the
independent variables STEM and EHW, as well as for the following control variables: grade, ESCS, and immig. The missing observations make up only 4% of the entire sample. These are deleted in order to obtain the same number of observations in each model. This is important for the comparability of the models. The final sample consists of 155,337 respondents.
3.5 Descriptive statistics
Descriptive statistics for the variables of interest and control variables are displayed in appendix A. Table A8 contains information about test scores, whereas table A9 about independent variables. Control variables are displayed in table A10.
3.5.1 Variables of interest Test scores
Firstly, in table A8, countries and the number of observations per country are included. Italian respondents make up 11,103 and represent the largest country group. The smallest number of respondents (4,404) is found in Poland. The difference between the largest and smallest country groups is, thus, 6,699. Such differences are accounted for by using the weighting methodology4.
Secondly, average test scores are calculated by country and a test of difference is conducted on the means of girls and boys. It can be observed that the overall average subject scores do not differ largely. More specifically, the differences between subject scores within countries do not tend to exceed the threshold of 15 points. Only five countries have differences larger than this: Finland, Croatia, Ireland, Malta, and Slovakia. The countries with the smallest differences between subject scores are Czech Republic, Luxembourg and Poland. They all have differences of less than 5 points. The subject for which the highest average score is obtained differs per country and no clear pattern can be observed.
Within the test of gender differences, however, patterns can be observed. Importantly, the gender difference in reading score is significantly positive for every country at a 1% level. This means that without a single exception, on average girls are performing better than boys in reading throughout the EU. Furthermore, not only is the gender difference statistically significant, it is also practically non-negligible. Namely, for all countries gender differences in reading are larger than 10 standardised test points. This lends some support in the name of hypothesis 1A.
For the STEM fields (mathematics and science) most gender differences are negative or insignificant. This is especially the case in mathematics, for which only one country has a significantly positive value for the mathematics gender difference – Finland. Finland is an interesting country, as it is the only one where on average girls outperform boys in all three subjects
4
significantly (at a 1% level). Within the science scores, there are 14 countries with negative gender differences and eight with positive. The rest have insignificant differences. Interestingly, mathematics – the subject that has been at the centre of research into educational gender inequalities – has the highest number of countries (9) with insignificant sex differences.
Independent variables
In table A9, information on the percentage of respondents expecting an occupation in the fields of STEM and EHW are calculated. Furthermore, a test of gender difference is executed to see whether there are more boys or more girls expecting a certain field. Finally, the values the three gender role attitude variables are listed in the last three columns.
In most countries the share of students expecting to work in STEM or EHW fields ranges around 10%, with some countries having higher shares at around 20%. In 22 countries more students are expecting an EHW occupation than STEM occupation. The four countries in which the share of STEM is higher than EHW are Bulgaria, Estonia, Hungary, and Latvia.
Almost all tests of gender difference are significant at a 1% level and confirm the observed gender segregation in occupation fields. More girls are expecting to work in the fields of education, health and welfare, whereas more boys are expecting to work in the fields of science, technology, engineering and mathematics. The only insignificant gender difference value is observed in Denmark for EHW.
With regards to the gender role attitude variables, most countries (16) are attached with both positive and negative values. This means that they exhibit both traditional and non-traditional gender role attitudes, respectively. From the remaining ten countries, seven obtain negative values for all three variables. These countries are Austria, Czech Republic, Hungary, Lithuania, Latvia, Poland, and Slovakia. The countries with positive values only and, thus, egalitarian gender role attitudes, are Denmark, Finland and Sweden.
4. Methodology 4.1 PISA data
Following the instructions in the PISA analysis manual (2009) all regressions are performed separately on each PV and the final model is obtained by aggregating the results. Furthermore, as suggested, variables are weighted throughout the data analysis (OECD, 2009). This ensures that estimates are not biased due to different selection probabilities in the two-stage sampling process (Ibid.). For the 2015 PISA cycle, there is only one usable STATA routine to estimate the regressions. This is the command REPEST. It is a user written command for complex survey setups. As such, it accommodates weighting procedures and has the ability to use plausible values in the estimation process (Keslair, 2017).
However, unlike the command PISATOOLS for previous cycles, it does not accommodate post estimation (Ibid.). This is a disadvantage because it means that heteroscedasticity or multicollinearity cannot be formally tested for in the models. An alternative for formal test of multicollinearity is a pairwise correlation table. Correlations of above 0.8 are thought to indicate severe multicollinearity (Berry & Feldman, 1985). In large samples the cut-off point may even be 85% (Ibid.). Moreover, large standard deviations or confidence intervals and sensitiveness to small changes in specification are indicators of multicollinearity (Stock & Watson, 2014). However, these checks are not necessarily reliable (Alin, 2010).
Furthermore, without the availability of post estimation commands, information criteria needed for the comparison of models cannot be obtained. Thus, the best model is chosen according to adjusted R-square, log-likelihood (ll), F-statistics, and log-likelihood ratio tests (LRT). For the LRT test-statistic the following formula is used:
−2 ∗ (𝑙𝑙𝑓𝑢𝑙𝑙 𝑚𝑜𝑑𝑒𝑙− 𝑙𝑙𝑛𝑢𝑙𝑙 𝑚𝑜𝑑𝑒𝑙)
The significance of the test statistic is assessed via comparison to the chi-square critical value with (𝑘𝑓𝑢𝑙𝑙 𝑚𝑜𝑑𝑒𝑙− 𝑘𝑛𝑢𝑙𝑙 𝑚𝑜𝑑𝑒𝑙)5 degrees of freedom at a 1% significance level.
4.2 Regression equations
Given that the dependent variable is at the individual level, whereas one of the explanatory variables is an index on the country level, this paper implements a multilevel regression analysis. More specifically, four different sets of models are constructed. Each set of model has three regressions with a different subject as dependent variable. The baseline regressions are provided below. Although controls are already included in the equations, the regressions are performed with and without the control variables for robustness checks. The subscripts i, s and k indicate whether a variable is on the individual-, school- or country-level, respectively.
Firstly, to test hypothesis 1a-b, the dependent variables are regressed on the female dummy (F) (Models 1). Furthermore, to test hypothesis 1c, a t-test with pooled variance6 is performed on female coefficients in each model.
𝑟𝑒𝑎𝑑𝑖𝑠𝑘 = 𝐹𝑖+ 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑖𝑠𝑘
𝑚𝑎𝑡ℎ𝑖𝑠𝑘 = 𝐹𝑖+ 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑖𝑠𝑘 𝑠𝑐𝑖𝑒𝑖𝑠𝑘 = 𝐹𝑖 + 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑖𝑠𝑘
5 k = number of parameters in the model.
6 Calculated based on standard errors obtained by estimating seemingly unrelated regressions. This allows correlation of
error terms across the regressions, which is important because the patterns in academic performance across fields may have common characteristics.
Secondly, to test hypothesis 2, occupational expectation dummies STEM and EHW are added to the regressions above (Models 2). In the regression on read only the variable EHW is of interest, whereas in regressions on math and scie it is the variable STEM. A test of mediation is conducted to assess the statistical significance of the mediation effect. More specifically, the KHB methodology is followed (Karlson, Holm & Breen, 2012).
𝑟𝑒𝑎𝑑𝑖𝑠𝑘 = 𝐹𝑖 + 𝐸𝐻𝑊𝑖+ 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑖𝑠𝑘 𝑚𝑎𝑡ℎ𝑖𝑠𝑘 = 𝐹𝑖 + 𝑆𝑇𝐸𝑀𝑖 + 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑖𝑠𝑘
𝑠𝑐𝑖𝑒𝑖𝑠𝑘 = 𝐹𝑖 + 𝑆𝑇𝐸𝑀𝑖 + 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑖𝑠𝑘
Thirdly, to test hypothesis 3, the dependent variables are regressed on all three explanatory variables: gender, occupational expectation and the interaction effect of gender role attitudes (GRA) and gender (F) (Models 3). The GRA variables must be included although there is no expectation on their coefficients. Again, a t-test with pooled variance6 is performed on female coefficients in each model to test their equality.
𝑟𝑒𝑎𝑑𝑖𝑠𝑘 = 𝐹𝑖+ 𝐸𝐻𝑊𝑖+ 𝐺𝑅𝐴𝑘𝐹𝑖 + 𝐺𝑅𝐴𝑘+ 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑖𝑠𝑘
𝑚𝑎𝑡ℎ𝑖𝑠𝑘 = 𝐹𝑖 + 𝑆𝑇𝐸𝑀𝑖 + 𝐺𝑅𝐴𝑘𝐹𝑖 + 𝐺𝑅𝐴𝑘+ 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑖𝑠𝑘 𝑠𝑐𝑖𝑒𝑖𝑠𝑘 = 𝐹𝑖+ 𝑆𝑇𝐸𝑀𝑖+ 𝐺𝑅𝐴𝑘𝐹𝑖+ 𝐺𝑅𝐴𝑘+ 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑖𝑠𝑘
Finally, to test hypothesis 4, an interaction effect of occupational expectations and gender role attitudes is included (Models 4). Ideally the hypothesis would be tested using the moderated mediation methodology (Preacher, Rucker & Hayes, 2007). However, neither the normal based estimation nor bootstrapping used for obtaining standard errors for direct and indirect effects, allow for multiple moderator variables.
Alternatively, hypothesis 4 is tested by running the regressions separately for girls and boys, hence why variable F is not included. Again, in the regression on read only the variable EHW is of interest, whereas in regressions on math and scie it is the variable STEM.
𝑟𝑒𝑎𝑑𝑖𝑠𝑘 = 𝐸𝐻𝑊𝑖 + 𝐺𝑅𝐴𝑘𝐸𝐻𝑊𝑖 + 𝐺𝑅𝐴𝑘+ 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑖𝑠𝑘 𝑚𝑎𝑡ℎ𝑖𝑠𝑘 = 𝑆𝑇𝐸𝑀𝑖 + 𝐺𝑅𝐴𝑘𝑆𝑇𝐸𝑀𝑖+ 𝐺𝑅𝐴𝑘+ 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑖𝑠𝑘
𝑠𝑐𝑖𝑒𝑖𝑠𝑘 = 𝑆𝑇𝐸𝑀𝑖 + 𝐺𝑅𝐴𝑘𝑆𝑇𝐸𝑀𝑖 + 𝐺𝑅𝐴𝑘+ 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑖𝑠𝑘 5. Analysis
5.1 Pairwise correlation and multicollinearity
A pairwise correlation table including all the variables used in the analyses is displayed in appendix B (table B1). Correlations between the plausible values are excluded from the table because these variables never appear in the same regressions. Among the remaining variables, the highest
correlation is found between GGI and GDP (71.8%). As this is still below the cut-off point, it is not expected for multicollinearity to pose problems.
Nevertheless, the sensitivity of coefficients to small changes in model specification is still observed. This is in order to ensure the legitimacy of the conclusion that multicollinearity is not a problem in the data set.
5.2 Regression results
Firstly, the overall relevance of the control variables is tested using the LRT. It is found that for all dependent variables in all four models (1-4), the full model is significantly different from the null model. In this case, the full model refers to the regression which includes control variables, whereas the null model refers to the regression in which only the variables of interest are included. This means that the models with controls fit the data better. Table B2 in appendix B presents the likelihood ratio test statistics and their significances.
5.2.1 Models one to three
Table 4 displays the results for models 1-3 with control variables. Similarly, table B3 in appendix B displays the results for models 1-3 without control variables.
Direction and magnitude of sex differences
Hypothesis 1a stipulates that girls perform better than boys in reading. In table 4 the direction of the female coefficient in model 1 is as expected: significantly positive. On average girls obtain about 21 points more than boys in reading related test questions. Since on average boys score 485.695 points in reading, girls outperform boys by 4.24%.
According to hypothesis 1b girls underperform in mathematics and science relative to boys. As can be seen from model 1 for mathematics, the coefficient of female is significant and negative. On average girls obtain about 15 points less in mathematics than boys. In model 1 for science performance, the coefficient of female is also significantly negative. On average girls underperform by about 10 points in science. Average scores for boys in mathematics and science are 501.523 and 499.434, respectively. As such; girls’ scores are on average 2.97% and 2.00% lower than boys’.
Furthermore, these values lend some support in favour of hypothesis 1c, i.e. that sex differences in reading scores are largest and sex differences in science scores are lowest. To ascertain that reading scores are in fact largest and science scores lowest, a t-test on the coefficients in performed. This is displayed in table B4 in appendix B. It can be observed that with low standard errors, all differences are highly statistically significant. As such, it is concluded that the female coefficients across the models are statistically different.
Mediation effect of expected occupation
In H2 it is hypothesized that occupational expectations of girls and boys act as a mediator for the effect of gender. As a first step it is confirmed that adding expected occupation as an explanatory variable increases the fit of the model for all dependent variables. The first row of table B5 in appendix B presents the results of this LRT. All test statistics are highly significant, implying that expected occupation is a relevant variable in explaining academic performance.
For reading performance, it is predicted that model 2 has a positive coefficient for EHW. As can be seen in table 4 column 2, this prediction is confirmed. The significant coefficient of EHW implies that the average reading score for student who expects to work in the EHW fields is 10 points higher than the score of students who do not. Given that the average reading score for students not expecting an EHW occupation is 494.105, students expecting EHW occupations outperform those who do not by 2.03%.
Regarding the coefficients in model 2 for mathematics and science performance, it is hypothesized that expecting a STEM occupation has a significantly positive effect on performance. For both subject fields, the expectation is fulfilled. The STEM coefficients for mathematics and science are about 29 and 32, respectively. These reflect the over performance of students who expected a STEM occupation in relation to students who do not. Since average scores in mathematics and science are 489.947 and 490.056, respectively, the over performance in percentages is 5.87% and 6.50%.
Noticeably, the effects of the STEM variables are larger than the effect of EHW. A reason for this could be that STEM fields are viewed as more prestigious and, thus, are highly valued. As such, anticipating a STEM job incentivizes pupils to work hard and to perform better. EHW fields, however, are not necessarily valued as highly – partly due to them being female dominated (England, 2010). This would explain why incentive to work hard is smaller and why the effect of EHW on reading is smaller in the magnitude.
In table 2, the results from the KHB estimation are displayed. From the table, the statistical significance and the magnitude of the mediation can be assessed. The row total represents the effect of gender on academic performance when only female and control variables are included in the regression. This effect is then decomposed into direct and indirect effects, where the indirect effect is the effect transmitted through expected occupation.
It is expected that the indirect effect is statistically significant and acts in the same direction as gender. These expectations are confirmed for all dependent variables. For reading performance the indirect effect represents 7.49% of the full effect, whereas for mathematics and science the indirect effect represents 18.63% and 30.85%, respectively. This implies that the mediation effect of
EHW is smaller than the mediation effect of STEM. Furthermore, this provides further support for the speculation above about the difference in valuation of EHW and STEM fields.
However, in absolute terms, the coefficients for the indirect effect are not substantial. About two test points in reading performance are transmitted through expected occupation. For mathematics and science performance about three test points are transmitted through expected occupation.
Table 2
Gender role attitudes and sex differences
Hypothesis 3 deals with the interaction effect of gender and country-level gender role attitudes. Firstly, it is stipulated that sex differences are smaller in gender egalitarian countries (H3a). Much like for model 2, the LRT is used to confirm that adding the GRA*F and GRA variables increases the fit of the model. These results are found in table B5, row two. As the test statistics are highly significant, gender role attitudes are relevant factors in explaining sex differences in academic performance.
Regarding reading performance, it is expected that the GRA*F interaction terms are negative. In this way, girls in gender egalitarian countries on average have a smaller advantage than girls in more traditional countries. As can be observed from column 3 under reading performance, only the interaction with component 2 is statistically significant at a 5% level. Furthermore, it has a positive coefficient which is not in line with hypothesis 3a. In fact, girls in more egalitarian countries have a wider advantage than girls in less egalitarian countries.
With regards to the direction of the GRA*F coefficients in the regressions on mathematics and science, it is expected that the interaction term between gender role attitudes and gender is positive. In this way, the disadvantage of girls relative to boys is smaller in gender egalitarian countries. For mathematics performance, only the interaction with component 2 has a statistically significant positive coefficient – much like for reading performance. In this regression, however, the significance level is 1% and the direction of the coefficient is as expected. For science performance, all interaction terms are statistically insignificant, implying that gender role attitudes do not influence sex differences in science performance.
Effect Reading performance Mathematics performance Science performance
Total 20.580*** (-14.914)*** (-9.974)***
Direct 19.039*** (-12.135)*** (-6.897)***
Indirect 1.541*** (-2.779)*** (-3.077)***
† Expected occupation refers to variable EHW in the regression on reading performance and to variable STEM in the regressions on mathematics and science performance.
10%, 5%, 1% significance levels are denoted by *, **, ***, respectively.
Component 2 pertains to the gender role good housewife and mother, and it also incorporates a part of the belief that housework is women’s responsibility. This means that, on average, as less people in the country agree with these gender roles, girls perform better in reading as well as mathematics. To put the results into context, the effects of the GRA2*F variable on reading and mathematics performance are calculated for a more gender egalitarian country (Finland) and a more traditional country (Lithuania). As science has insignificant coefficients only, the effect is not calculated.
In Finland the value for the second gender role attitude component is 2.958, thus, the effect of GRA2*F on reading is about 4 test points. This means that in the country where this gender role attitude is not prevalent, on average, girls’ advantage in reading is increased by four test points. In Lithuania, the value of GRA2 is (-1.015), and as such, the effect of the interaction term is about (-1) point in. In other words, in the country where the gender role is prevalent, on average, girls’ advantage is decreased by one test point.
On the other hand, using the information above, it is found that the effect of GRA2*F on mathematics is about 6 test points in Finland and (-2) points in Lithuania. This means that in the country where the specific gender role is not as prevalent, girls’ disadvantage in mathematics is reduced by six points, whereas in the country where it is prevalent it is increased by two points.
Furthermore, it is expected that after controlling for other effects, the sex differences in reading remain largest (H3b). This is tested by applying a t-test on the female coefficients across the regressions. The results of this test can be found in table B6 in appendix B. All test statistics are highly significant. This implies that the female coefficients across the regressions are statistically different. As such, it can be concluded that sex differences in reading performance remain largest and sex differences in science performance remain smallest.
Dependent variable Model 1 2 3 1 2 3 1 2 3 Female 20.580*** 19.039*** 18.915*** (-14.914)*** (-12.135)*** (-12.170)*** (-9.974)*** (-6.897)*** (-7.065)*** EOCC † 10.014*** 9.465*** 28.783*** 29.479*** 31.866*** 32.195*** GRA1*F -0.766 -0.550 0.279 GRA2*F 1.451** 2.114*** 0.795 GRA3*F -1.421 -0.651 -0.529 GRA1 1.846*** (-1.821)*** 0.235 GRA2 (-2.789)*** -0.587 (-2.612)*** GRA3 (-6.760)*** (-5.617)*** (-7.286)*** Constant 236.975*** 234.830*** 228.824*** 311.518*** 302.028*** 254.625*** 270.644*** 260.137*** 219.704***
Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes
Weighted Yes Yes Yes Yes Yes Yes Yes Yes Yes
N 155,337 155,337 155,337 155,337 155,337 155,337 155,337 155,337 155,337 Adjusted R2 0.237*** 0.239*** 0.245*** 0.237*** 0.248*** 0.255*** 0.232*** 0.245*** 0.251*** Log-likelihood (-914,386.4) *** (-914,225.2) *** (-913,605.2) *** (-901,523) *** (-900,359) *** (-899,680.7) *** (-909,926.8) *** (-908,645.6) *** (-908,023) *** F-statistic 6,891.239 *** 6,082.666 *** 3,592.440 *** 6,889.717 *** 6,412.679 *** 3,793.704 *** 6,707.047 *** 6,289.159 *** 3,711.883 ***
† EOCC (i.e. expected occupation) refers to variable EHW in the regression on reading performance and to variable STEM in the regressions on mathematics and science performance.
10%, 5%, 1% significance levels are denoted by *, **, ***, respectively.
