Gender differences in the intention to study math increase with math performance

The study investigates why females, despite outnumbering males in higher education, remain underrepresented in math-intensive fields (math, computer science, physical sciences, geosciences, engineering, economics). This underrepresentation matters because math-intensive jobs pay more and exhibit smaller gender wage gaps, and because rising demand for math skills risks shortages—especially if high-ability females avoid math paths. Prior work documents gender gaps in STEM choices but rarely examines how these gaps vary along the math ability distribution. Understanding the ability gradient is crucial for assessing potential impacts on talent allocation, performance gaps among math-track students, and for targeting policy interventions at the most relevant parts of the distribution. The research question is whether the gender gap in intentions to pursue math-related studies or careers increases with math performance among 15-year-olds, and what implications this has for subsequent enrollment and labor market outcomes.

A large literature examines gender differences in STEM participation and mechanisms such as comparative academic advantages, self-concepts, interests, and stereotypes. However, few studies analyze how the gender gap varies with math ability. Cimpian et al. (2020) studied ~6000 U.S. 4-year college students and found an overrepresentation of lower-performing males in physics, engineering, and computer science relative to females, suggesting complex dynamics by ability at the postsecondary level. Other longitudinal research indicates that high school course-taking strongly shapes later STEM enrollment, with many high-achieving females opting out of advanced math/science courses, thereby affecting STEM readiness. The current study extends this literature by analyzing intention-to-pursue-math as a function of math ability across 61 countries, and by testing explanations related to comparative advantage, self-concept, interests, and socioeconomic status (SES).

Data come primarily from PISA 2012, covering 15-year-old students in 61 countries (32 OECD, 29 non-OECD). The core outcome is students’ intention to pursue math-related studies/careers, measured via a five-item forced-choice battery (ST48) contrasting math with reading/language or science. The main measure is a binary indicator (“strong intentions”) equal to 1 if a student chooses math on all five items. Due to questionnaire rotation and item nonresponse, this variable is available for 251,120 students. Math performance is standardized within each country to mean 0, SD 1; analyses use plausible values and student weights, with replication weights for correct standard errors. Intentions are examined across ventiles of standardized math performance and via linear probability models: Math intention_i = α·Girl_i + β·Math_ability_i + γ·(Math_ability_i × Girl_i) + δX_i + ε_i, where X_i includes various controls (country fixed effects, reading/science ability, self-concept, interest, SES, and interactions). Robustness includes alternative performance standardizations, continuous indices of intention, alternative cutoffs, and non-linear models (logit/probit). External validation assesses extension to actual enrollment using French administrative/survey data (∼12,000 10th graders and 5,500 12th graders in 2016–2018), and U.S. HSLS:09 data on course-taking in engineering/computer science/math. Labor market extensions use PIAAC adult data (∼150,000 adults across 31 countries) to study math/STEM occupations and numeracy usage at work as functions of numeracy ability. Additional analyses reweight gender-specific ability distributions to isolate the role of differential intentions by ability and explore theoretical implications when boys and girls share identical ability distributions but differ in intention gradients.

• Intentions increase with ability for both sexes but more steeply for boys: A one SD increase in math performance raises the probability of strong intentions by 5.4 percentage points for boys (z=15.86, p<0.001, 95% CI [0.047, 0.061]) and 3.3 percentage points for girls (difference in slopes γ=−0.021; z=−4.84, p<0.001, 95% CI [−0.029, −0.012]). The relationship is approximately linear across the distribution, with slightly higher growth for girls only at the extreme top (~top 5%). R-squared from intent-on-performance is more than twice as large for boys (0.015 vs 0.006). - Gender gap widens with ability: The boy–girl difference in strong intentions is ~1.8 percentage points and not significant at the bottom, increasing steadily to ~8.0 percentage points at the top. The boys-to-girls ratio in strong intentions rises from ~1 among low performers to ~1.5 among high performers; it increases from 1.18 in the bottom quartile to 1.40 in the top quartile. - Cross-country prevalence: In countries where intentions significantly rise with ability, the slope is larger for boys in most OECD (24/29) and non-OECD (20/23) countries; the difference is statistically significant in 14 OECD and 7 non-OECD countries. The magnitude of the gender-differential slope is not significantly correlated with country GDP, HDI, or Gender Gap Index. - Implications for performance among math-intenders: The global gender gap in math performance (girls − boys) is larger among students with strong intentions than in the overall population (−0.190 vs −0.137 SD; ratio 1.389). The girls-to-boys ratio in the top decile of performers declines among intenders relative to all students (global: 0.536 vs 0.743; OECD: 0.482 vs 0.671; non-OECD: 0.596 vs 0.832). Reweighting to equalize gender ability distributions confirms that differential intention gradients alone increase the performance gap and reduce girls’ representation at the top among intenders. - Actual enrollment evidence: In France, intentions correlate strongly with subsequent actual choices (r=0.78 for 10th graders, r=0.74 for 12th graders; both p<0.001). Intentions and actual enrollment in math-related majors rise with performance but more slowly for females; for 12th graders, the slope for females is about half that for males. U.S. HSLS:09 shows enrollment in engineering/computer science/math courses rises more with ability for males (from 7% to 25%) than females (from ~7% to ~10%); the male–female gap grows from ~1 pp in the bottom decile to ~17.9 pp in the top decile; the ratio rises from 1.14 to 2.91. Among all high schoolers, the math performance sex gap is near zero, but among those choosing these courses it is ~0.25 SD to girls’ detriment. - Labor market patterns: In PIAAC, the probability of working in a math-related or STEM job and the intensity of numeracy usage at work increase with numeracy ability more for men than women; the male gradient is 86% larger for math-related jobs. The sex gap in math performance is 15–20% larger among math/STEM workers than in the general adult population. Females are underrepresented among top performers in math-intensive jobs (top-decile female share drops from 38% overall to 20% among math practitioners). - Mechanism tests: The widening intent gap with ability is not explained by comparative advantage (math vs reading), self-concept, or declared interest. Including reading/science controls magnifies the differential slope. Math self-concept and interest gaps narrow with ability, opposite the intention pattern. SES explains about half of the gender-differential slope; controlling for SES reduces γ from −0.0210 to −0.0116.

The findings show that while intentions to pursue math increase with ability for both sexes, the relationship is stronger for boys, producing a widening gender gap with rising performance. This pattern helps explain why, among students intending to study math, girls exhibit lower average math performance relative to boys than in the general population and are increasingly underrepresented among top performers. Extensive robustness and cross-country analyses indicate generalizability, and extensions to actual enrollment (France, HSLS:09) and adult labor markets (PIAAC) suggest persistence into educational and occupational outcomes. Tests of potential mechanisms indicate that comparative academic advantages (math vs reading), math self-concept, and interest do not account for the observed pattern; SES accounts for about half. The remaining gradient may plausibly reflect gender stereotypes about math excellence being male-typed and requiring innate brilliance, potentially more salient among high performers. Policy implications include targeting high-performing girls with interventions (e.g., role models, stereotype-challenging environments) and evaluating policies across the full ability distribution to reduce widening gaps at the top.

This study documents that the gender gap in intentions to pursue math increases with math performance among 15-year-olds across 61 countries. Because girls’ intentions are less sensitive to their math performance than boys’, girls become relatively less represented among high-ability math intenders, contributing to larger performance gaps within the math-intending subgroup and potentially perpetuating stereotypes and long-term underrepresentation in math-intensive fields. The results are robust across specifications and datasets and extend to actual enrollment and labor market patterns. Future research should identify causal mechanisms behind the differential gradients—particularly the roles of SES, parental and school environments, and stereotypes—and rigorously test interventions aimed at encouraging high-performing girls to pursue math-related studies and careers.

• Mechanism indeterminacy: While SES explains about half of the differential gradient, the precise causal mechanisms remain unclear; stereotypes are posited but not directly tested causally here. - Intentions vs actual behavior: The main outcome is intentions, not enrollment; although validated in France and the U.S., generalization to all countries cannot be fully confirmed. - Adult analyses: In PIAAC, reverse causality or on-the-job learning could shape measured ability, complicating causal interpretation of labor market gradients. - Country heterogeneity and power: In some countries, intentions do not significantly increase with ability, limiting within-country gender-slope comparisons. - Measurement and sample design: The intention items are in a rotated questionnaire answered by two-thirds of students, and item nonresponse reduces the analytic sample; despite weights and plausible value methods, residual measurement error may remain.