This study investigates the effects of project-based learning (PBL) courses on higher-education STEM students by examining changes in students’ non-cognitive variables. The variables examined include learning satisfaction, grit, growth mindset, fixed mindset, and self-efficacy. To reduce response shift bias and missing data while providing a more accurate baseline, this study adopts a retrospective pretest–posttest (RPP) design to assess changes in these variables from the beginning to the end of the semester following students’ participation in PBL courses. To explore the latent structures underlying these changes, the difference scores between the retrospective pretest and posttest are used as dependent variables. Multidimensional scaling (MDS) is employed to conduct variable reduction and structural analysis. The participants consist of students from four departments at a university in Taiwan, representing the four STEM domains: science, technology, engineering, and mathematics. The MDS results reveal that the latent structure of the non-cognitive variables can be represented by three dimensions. Dimension 1 reflects overall improvements across all non-cognitive variables. Dimension 2 represents improvements in all non-cognitive variables except fixed mindset. Dimension 3 indicates improvements in learning satisfaction accompanied by a decrease in growth mindset. These findings provide new insights into the latent structure of non-cognitive development among STEM students engaged in project-based learning. Furthermore, the methodological approach combining the RPP design with multidimensional scaling offers a novel analytical framework for examining changes in non-cognitive variables in educational research.