College Course Map: Applications

Students' postsecondary course-taking is of interest to researchers, yet has been difficult to study at large scale because administrative transcript data are rarely standardized across institutions or state systems. This paper uses machine learning and natural language processing to standardize college transcripts at scale. We demonstrate the approach's utility by showing how the disciplinary orientation of students' courses and majors align and diverge at 18 diverse four-year institutions in the College and Beyond II dataset. Our findings complicate narratives that student participation in the liberal arts is in great decline. Both professional and liberal arts majors enroll in a large amount of liberal arts coursework, and in three of the four core liberal arts disciplines, the share of course-taking in those fields is meaningfully higher than the share of majors in those fields.

Suggested citation: Paulson, Annaliese, Kevin Stange, and Allyson Flaster. (2024). Classifying Courses at Scale: a Text as Data Approach to Characterizing Student Course-Taking Trends with Administrative Transcripts. (EdWorkingPaper: 24-1042). Retrieved from Annenberg Institute at Brown University: https://doi.org/10.26300/7fpa-s433

Identifying Curricular Deserts

Drawing on our models, we study the geographic availability of postsecondary curriculum. Prior work has described the geographic accessibility (or inaccessibility) of postsecondary opportunities, primarily focused on institutional proximity. We instead examine proximity to the specific types of courses offered by institutions. Access to a nearby college may not be as valuable if the college doesn't offer the courses you want.

We use our finetuned model to classify all courses offered at public institutions in Texas between the 2015-16 and 2023-24 academic year at the six-digit level, then calculate the distance from each census block to the nearest offering of each course by year. Here we present the number of CCM codes available within 50 miles of each census block during the 2023-2024 academic year. We see course availability is concentrated around major metropolitan areas in the east of the state but many census blocks do not have easily accessible coursework.

In the figure below, we plot a six-digit CCM code's logged statewide enrollments against the average distance from high schools in the state to the nearest offering of that CCM code and the number of institutions that ever offer that CCM code in our sample. We observe a strong positive correlation between log enrollments and the number of institutions offering a course of 0.782 and a strong negative correlation of -0.802 between average distance and logged enrollments. The availability of courses has a strong log-linear relationship with enrollments.

College vs. Curricular Deserts

The figures suggest the presence of an institution is not sufficient to ensure all students have geographic access to courses. We more formally test this hypothesis by regressing the minimum number of miles a student must travel to enroll in a given course on the minimum miles a student must travel to enroll in the nearest institution. We regress on distance to the nearest six-digit CCM offering on measures of distance to the nearest institutions, controlling for six-digit CCM by academic year fixed effects. The table below shows that accounting for distance to the nearest institution and statewide trends in course availability leaves more than half of the variation in high school distance to course offerings unexplained. The presence of an institution does not ensure access to all higher education curricula.

In a new project, in support of the State of Michigan's effort to increase success for older learners, our team is using classified course data to identify the courses that give students the most difficulty. Which courses tend to stand in the way of progress for older learners? Our classifier is being used to make courses comparable across institutions.