Putting the Pieces of the Puzzle Together in Modeling Gendered Educational Choices

Background: The underrepresentation of women in science, technology, engineering, and mathematics (STEM) remains a persistent international challenge with complex causes. This study draws on five complementary theoretical perspectives–theory of circumscription and compromise, situated expectancy-value theory, the internal/external frame of reference model, the differential effects model, and the goal congruity model–to identify the strongest predictors of women's entry into STEM fields in higher education from 21 variables and their up to 441 nonlinear pairwise interactions. No previous study has included such a comprehensive set of variables and interactions. To address this gap, we applied Prediction Rule Ensembles, a machine learning technique that balances interpretability and predictive accuracy, capturing complex, nonlinear interactions between variables. We used a nationally representative longitudinal German NEPS sample of 4,383 academic-track students (53% female, 25% second-generation immigrants).Results: Our analyses showed that high school students' career aspirations with a higher proportion of female than male incumbents (gendered career aspirations), along with their self-concept of math and verbal ability, were the strongest predictors of STEM major choice. Gendered career aspirations accounted for between 71% and 77% of the gender gap in STEM major choice, while students’ achievement variables had a much lower importance for the STEM major choice prediction. Notably, the association between math self-concept and STEM major choice was stronger for students with higher male-dominated career aspirations, highlighting the interplay between gendered career aspirations and domain-specific self-concepts. The application of prediction rule ensembles allowed us to uncover nonlinear interactions between predictors that had not previously been identified, demonstrating the potential of machine learning approaches in educational research. Conclusion: Our findings highlight the importance of considering complex, nonlinear interactions between theory-driven predictors to better understand the gender gap in STEM. These findings have significant implications for both theory and practice, suggesting that interventions targeting the perceived gender typicality of STEM careers and students' math self-concepts may be effective in increasing women's participation in STEM. Furthermore, the methodological approach adopted in this study offers a promising framework for future research, paving the way for more nuanced investigations of the multiple factors that influence educational and career choices in STEM.