One of the main findings of this study is that, paradoxically, countries with lower levels of gender equality had relatively more women among STEM graduates than did more gender-equal countries. This is a paradox, because gender-equal countries are those that give girls and women more educational and empowerment opportunities and that generally promote girls’ and women’s engagement in STEM fields (e.g., Williams & Ceci, 2015).

In our explanation of this paradox, we focused on decisions that individual students may make and decisions and attitudes that are likely influenced by broader socioeconomic considerations. On the basis of expectancy-value theory (Eccles, 1983; Wang & Degol, 2013), we reasoned that students should at least, in part, base educational decisions on their academic strengths. Independently of absolute levels of performance, boys on average had personal academic strengths in science and mathematics, and girls had strengths in reading comprehension. Thus, even when girls’ absolute science scores were higher than those of boys, as in Finland, boys were often better in science relative to their overall academic average. Similarly, girls might have scored higher than boys in science, but they were often even better in reading. Critically, the magnitude of these sex differences in personal academic strengths and weaknesses was strongly related to national gender equality, with larger differences in more gender-equal nations. These intraindividual differences in turn may contribute, for instance, to parental beliefs that boys are better at science and mathematics than girls (Eccles & Jacobs, 1986; Gunderson, Ramirez, Levine, & Beilock, 2012).

We also found that boys often expressed higher self-efficacy, more joy in science, and a broader interest in science than did girls. These differences were also larger in more gender-equal countries and were related to the students’ personal academic strength.

. . . .

We propose that when boys are relatively better in science and mathematics while girls are relatively better at reading than other academic areas, there is the potential for substantive sex differences to emerge in STEM-related educational pathways. The differences are expected on the basis of expectancy-value theory and are consistent with prior research (Eccles, 1983; Wang & Degol, 2013). The differences emerge from a seemingly rational choice to pursue academic paths that are a personal strength, which also seems to be common academic advice given to students, at least in the United Kingdom (e.g., Gardner, 2016; Universities and Colleges Admissions Service, 2015).

The greater realization of these potential sex differences in gender-equal nations is the opposite of what some scholars might expect intuitively, but it is consistent with findings for some other cognitive and social sex differences (e.g., Lippa, Collaer, & Peters, 2010; Pinker, 2008; Schmitt, 2015). One possibility is that the liberal mores in these cultures, combined with smaller financial costs of foregoing a STEM path (see below), amplify the influence of intraindividual academic strengths. The result would be the differentiation of the academic foci of girls and boys during secondary education and later in college, and across time, increasing sex differences in science as an academic strength and in graduation with STEM degrees.

Whatever the processes that exaggerate these sex differences, they are abated or overridden in less gender-equal countries. One potential reason is that a well-paying STEM career may appear to be an investment in a more secure future. In line with this, our mediation analysis suggests that OLS partially explains the relation between gender equality and the STEM graduation gap. Some caution when interpreting this result is needed, though. Mediation analysis depends on a number of assumptions, some of which can be tested using a sensitivity analysis, which we conducted (Imai, Keele, & Yamamoto, 2010). The sensitivity analysis gives an indication of the correlation between the statistical error component in the equations used for predicting the mediator (OLS) and the outcome (STEM graduation gap); this includes the effect of unobserved confounders. Given the range of ρ values in the sensitivity analysis (Fig. S1), it is possible that a third variable could be associated with OLS and the STEM graduation gap. A related limitation is that the sensitivity analysis does not explore confounders that may be related to the predictor variable (i.e., GGGI). Future research that includes more potential confounders is needed, but such data are currently unavailable for many of the countries included in our analysis.