Predicting STEM Career Interest in High School: A Machine Learning Study on In-School and Out-of-School-Time Experiences

• In Event: AERA Roundtable Session Friday 9:50 am Four Seasons Ballroom 4
  In Roundtable Session: Machine Learning Techniques I: Advancing Equity, Prediction, and Measurement (Table 4)

Abstract

Grounded in theoretical and empirical literature, this study utilized supervised machine learning to address four key questions on predicting STEM career interest at the end of high school. Using the Outreach Programs and Science Career Intentions dataset (N = 15,847), we identified 75 predictors of STEM career interest across various domains. ML models were trained using lasso regression, decision tree, random forest, extreme gradient boosting, and support vector machines. Key findings include: (1) XGBoost achieved the highest prediction accuracy of 80.76%, (2) Fifteen key predictors were identified, (3) Nonlinear patterns and interactions, such as gender and course likeness, were highlighted, and (4) ML revealed more variables and complex effects than traditional logistic regression. It advances theoretical understanding and complements traditional methods.

Authors

• Rongxiu Wu, Harvard University

• Tingting Reid, Harvard University

• Sue Sunbury, Harvard University

• Philip M. Sadler, Harvard University

• Gerhard Sonnert, Harvard University