The unique contributions of math and science motivation to STEM outcomes for high school students: A model comparison study

Abstract

The most recent U.S. National Science Foundation report on women, minorities, and persons with disabilities in science and engineering (2015) shows that the well-known gender gap in STEM continues to persist and yet the urgent need for a STEM-educated workforce grows. Thus, research in the area of STEM achievement motivation is one avenue for potentially understanding and intervening in the gender gap. The present study contributes to this literature by: 1) investigating the relationships between grade 9 motivational subconstructs (e.g., identity with the subject matter, and interest in the subject matter) in relation to the larger construct of motivation within the specific domains of math and science; 2) testing whether those motivational variables are related to end of high school STEM outcomes, including STEM career aspirations, STEM course credits earned, and STEM GPA; and 3) testing whether the relationships between STEM motivation variables and STEM outcomes are moderated by gender (in other words, are relationships stronger or weaker for females compared to males?). Specifically, the High School Longitudinal Study 2009 (HSLS:09) waves 1, 2, and 3 were employed, which included a nationally representative sample of N = 21,444 high school students who were followed from grade 9 in fall 2009 to grade 12 in spring 2013. In addition to investigating these substantive research questions, the present study also compared model estimates from Maximum Likelihood (ML), Weighted Least Squares Minimum Variance (WLSMV), and Bayesian estimators of each of the structural equation models, with and without use of complex sample survey weights. The results indicated that STEM motivation subconstructs are measuring similar but distinct domain-specific constructs, and that these constructs generally do predict STEM outcomes. Importantly, these relationships were not moderated by gender. Results of the model comparisons showed that ML estimation was the most flexible for analyzing these complex survey data, with WLSMV estimation the second most flexible method, and Bayesian the least flexible method. Discussion of substantive and statistical results and recommendations are provided.