Closing the gender gap in science: new evidence from urban China

ABSTRACT

In this paper, we analyze recently collected data that conducts a unique assessment of high school student performance for over two thousand students from five Chinese provinces. Across three domains of scientific intelligence tested, we document heterogeneous gender gaps in academic performance. These differences generally arise due to differential productivity of inputs to the education production process and not differential levels of inputs. At many quantiles of the achievement distribution, girls perform better than boys when identifying scientific issues, whereas the converse holds on the portion of the assessment that measures whether one can apply scientific evidence. These differences may partially explain the subsequent gap in decision to major in specific STEM disciplines in college. Further, our results imply caution from using a single summative gender achievement gap measure when gender gaps in subject knowledge are not constant across each domain of intelligence examined within the test.

KEYWORDS:

• Gender gaps in STEM
• quantile regression
• education production function
• China
• decomposition analysis
• science education

JEL CLASSIFICATION CODES:

• J16
• C21
• I24

Acknowledgments

We would like to thank Steven Lehrer for many helpful comments and suggestions on earlier drafts. We thank two anonymous referees for valuable comments. We also wish to thank conference participants at the Canadian Economics Association Annual Meeting for additional comments.

Data availability statement

Data used in this article are available from the authors and is subject to permission from the National Innovation Center for Basic Education Quality Monitoring, P. R. China.

Disclosure statement

No potential competing interest was reported by the authors. We are responsible for all errors.

Notes

1 This is akin to mathematical competency being crucial to every citizen. After all, individuals make decisions that require some forms of mathematical intelligence such as how to save for retirement or involving a distinction between fixed versus adjustable rate mortgages. Individuals with more advanced understanding may possess skills essential for certain careers including becoming an accountant, economist or actuarian.

2 Further, if alternative variants of a test are used to measure gender gaps, a distorted comparison could arise if alternative versions of the test weight the domains differently in their overall score.

3 In addition, parents in urban China when choosing how to invest in their child may appear more gender neutral since they additionally consider recent transformations in the labor market such as the rapidly increasing returns to schooling beyond secondary school that exceed most Western nations and the fact that by and large women in China now have the same educational opportunities as men.

4 This latter finding also holds with the 2009 PISA math score, on which the OECD reports boys outperformed girls in 54 out of the 65 participating countries. Results from China, where the 2009 PISA was only completed by teenagers in Shanghai, find that it is 1 of only 11 countries where on average, girls scored slightly higher. Thus, our data collection provides information on adolescents that were previously not examined in the literature.

5 Within the economics of education literature test score gaps are decomposed between countries (Ammermuller,2007), schools (Krieg and Storer,2006) and ethnic groups (McEwan,2004); among many others.

6 The distributional decomposition considers two components and follows the steps in Hospido and Moral-Benito (2016) except for using a cross-sectional quantile regression estimator rather than the fixed effects quantile regression estimator of Canay (2011). The former estimator cannot remove the school indicators using standard demeaning (or differencing techniques. This substitution implicitly imposes a much stronger overlapping support assumption on the full set of observable inputs including school indicators to hold at each quantile.

7 These findings are interpreted as being consistent with evidence (see e.g. Bertrand,2010; Niederle and Vesterlund,2007) that women systematically underperform relative to men in competitive settings, but the possibility remains that women outperform men when competitive pressures are lower (e.g. Ors, Palomino, and Peyrache,2013).

8 Cimpian, Kim, and McDermott (2020) use U.S. data and find that the gender gap is similar across science subject areas suggesting the use of a single score. In contrast, by finding gender differences in competencies, we point out a challenge from using a single aggregate summative score even within a subject area.

9 In 2010, data from China National Bureau of Statistics states the per capita GDP of these five provinces ranked 9th, 10th, 13th, 23rd, 28th among the 31 administration regions in China. There is also an uneven distribution of resources for science and technology education across the 31 administrative regions in China as well as within individual provinces.

10 Chinese law mandates nine years of free education, from primary school to the end of middle school and the largest drop in secondary education enrollment occurs prior to senior high school. For example, Yi et al. (2015) report that more than 51 percent of junior high school students in poor rural areas do not go on to high school. Further, Loyalka et al. (2017) report across China that between 4.2% and 7.4% of students who enrolled in academic senior high school dropped out before graduating as well as small gender disparities that likely arise since boys can more easily find higher wage employment. The evidence is suggestive that any gender differences in dropout behavior between grades 10 and 11 would not be concentrated in specific locations of the distribution. Further, since we do not have information on the same cohort in the prior academic year. That said, we cannot directly investigate differential gender disparity in dropout behavior between grades 10 and 11. However, multivariate regression do not find evidence supporting the notion of gender differences across grades, conditional on school fixed effects in our sample.

11 In Wuhan, the project team collected data from 3 schools and as a result six classrooms.

12 That said, what is a scientific capability and how to evaluate it has long been questioned and explored by educational researchers. As an influential and well-recognized international assessment program, PISA proposed the concept of scientific literacy and successfully elaborated this concept from four aspects -- context, knowledge, attitude and competencies (OECD,2007). Scientific competencies are often portrayed as being the most important component of the scientific literacy, and it mainly refers to the three measurable cognitive capabilities that we explore in our assessment.

13 That said, analysis of our surveys conducted in the middle schools reveal that within cities, there is substantial heterogeneity in the instruments and equipment available in the physics, chemistry and biology labs. In many regular middle schools these materials are insufficiently provided, and our analysis further reveals that there are much fewer popular science magazines and books in the school libraries.

14 Financial releases from Tencent the company that manufactures QQ indicate that there were 829 million active QQ accounts in late 2014. We should note that in the survey, the subjects not only provided details on their QQ account but also provided information on whether they or their parents had an email account and included the email address.

15 Specifically, we conducted tests of differences in the proportion of families where the subject is the only child by gender across regions. In eastern China (Fujian and Liaoning) and central China (Hubei) two-sided tests are unable to find significant differences (z = 0.93 Pr($z>z\ast =0.35$), and z = −0.50 Pr($z>z\ast =0.62$) respectively. However, in the Western provinces (gansu and Sichuan) 58% of the boys are the only child in their family compared to 48.9% of the girls, which is statistically different z = 3.34 Pr($z>z\ast =0.001$) at the 1% level.

16 This hypothesis generates controversy since it has been used to explain dimensions of occupational segregation by suggesting that fewer women have the requisite ability for certain status jobs. O'Dea et al. (2018) recently compared gender differences in academic grades from over 1.6 million students and concluded that gender differences in both the mean and variance of grades are smaller in STEM than non-STEM subjects, suggesting that greater variability in performance is insufficient to explain male over-representation in STEM careers.

17 An additional advantage of using this methodology and moving beyond carrying out a decomposition at the mean is that we estimate a more flexible model for heterogeneous data at various points of the conditional achievement distribution. Last we note that this decomposition builds off results in Melly (2005) and it is also considered to be more efficient than the semiparametric decomposition developed in Machado and Mata (2005).

18 Ex ante we do not see any reason why this assumption would not hold between genders in China or elsewhere. Note that, whereas OLS estimates have the property that the means of both the dependent variable and explanatory variables lie on the regression line, thereby making the Blinder-Oaxaca decomposition of the dependent variable fairly straightforward; quantile regression estimators do not share this property.

19 Rank preservation is not needed with this strategy so that girls do not need to have the same rank in the counterfactual distribution as the actual distribution. That said, a large literature focusing on unconditional statistics reports gender differences in education inputs in both the developed world (e.g. see Lundberg and Rose,2002; Baker and Milligan, 2016, among others) as well as in China (see Knight, Shi, and Quheng, 2010; Ding and Zhang, 2014, among others).

20 This procedure involves exchangeable bootstrap distributions and is time intensive with larger datasets than we utilize in this paper. In our implementation we consider 200 repetitions. With larger data, Melly (2005) provides formulas to compute the analytic standard errors. Last, to consider multiple testing issues, Lehrer, Pohl, and Song (2021) proposed an alternative bootstrap-based testing procedure for quantile treatment effect heterogeneity under the assumption of selection on observables, and additionally shows its asymptotic validity.

21 Ding and Lehrer (2007) and Hoekstra, Mouganie, and Wang (2018) present evidence of the various dimensions that school quality differs in urban China and find that teacher quality is quite important in boosting student achievement. If we replace key school with teacher characteristics, we reach a similar conclusion but the individual conditional quantile regressions have a lower pseudo R-square.

Additional information

Funding

Ding wishes to thank funds for the Project 985 Visiting Scholar, Institute of the Economics of Education, Beijing Normal University for research support that facilitated conducting this project via research visits to Beijing Normal University.