Source-linked AI summary
Predicting Student Dropout in Higher Education
Lovenoor Aulck, Nishant Velagapudi, Joshua Blumenstock, Jevin West
TL;DR
Student dropout is a major concern, but quantitative evidence on its predictors and remedies has been limited. The paper models dropout using demographics and transcript records from over 32,500 students at a large public university, finding that first-term data provide predictive signal while timing dropout remains difficult to predict.
Problem
Student attrition remains costly and important, while quantitative research on its causes and possible remedies has been limited.
Method
The paper uses registrar data, demographic and transcript-derived features, and three machine-learning models to predict dropout and analyze individual predictors and dropout timing.
Results
AUC ranged from 0.66 to 0.73 across dropout models, with regularized logistic regression strongest; timing prediction yielded RMSE of about 5 quarters.
Takeaways & Limitations
First-term transcript and demographic data contain useful early indicators of student attrition, supporting further machine-learning work on retention and success.
Abstract
from arXiv · showhide
Each year, roughly 30% of first-year students at US baccalaureate institutions do not return for their second year and over $9 billion is spent educating these students. Yet, little quantitative research has analyzed the causes and possible remedies for student attrition. Here, we describe initial efforts to model student dropout using the largest known dataset on higher education attrition, which tracks over 32,500 students' demographics and transcript records at one of the nation's largest public universities. Our results highlight several early indicators of student attrition and show that dropout can be accurately predicted even when predictions are based on a single term of academic transcript data. These results highlight the potential for machine learning to have an impact on student retention and success while pointing to several promising directions for future work.
1. Introduction
Student dropout remains a major education and policy concern despite longstanding attrition theories. This study models dropout in a large, heterogeneous university population using demographics and first-term transcript records.
- Motivation: About 40% of bachelor’s-seeking students do not complete their degree within 6 years, while first-year attrition involves 30% not returning for a second year.Governments spent over $9 billion educating these first-year students from 2003–2008.
- Motivation: Longstanding theories of postsecondary attrition have not eliminated dropout as a major concern for students, institutions, and policymakers.Attriting students lose time and effort, while institutions cannot recoup scarce educational resources.
- Study objective: The study uses registrar data from a large, publicly funded, four-year US university and describes this as the largest dataset used to study student attrition at scale.The initial analysis uses demographic information and first-term transcript records to predict dropout.
- Related work: Prior studies generally examined small, homogeneous populations, whereas this approach considers an extremely heterogeneous population at one of the nation’s largest public universities.Earlier examples included cohorts of 453, 528, 648, and 1,508 students, or focused on second-term enrollment rather than eventual graduation.
2. Methods
The study defines dropout as failure to complete an undergraduate degree within six years and predicts this binary outcome from demographic, pre-college, and transcript-derived features using three machine-learning models.
- Data: The dataset contains demographics, pre-college entry information, and complete transcript records for main-campus undergraduates first enrolled between 1998 and 2006.The data came from de-identified University of Washington registrar databases.
- Outcome definition: Students are classified as non-completions if they do not complete at least one undergraduate degree within 6 calendar years of first enrollment.In the quarter-based system, this corresponds to 24 calendar quarters after the initial enrollment quarter.
- Feature construction: Transcript features encode departments, credits, classes, and departmental GPA, producing 784 additional features alongside demographic and entry variables.The feature construction also included dummy variables for frequently declared majors and gatekeeper-course analyses.
- Experiments: Regularized logistic regression, k-nearest neighbors, and random forests predict the binary dropout outcome.Performance uses a held-out 30% test sample, while 10-fold cross-validation on the remaining 70% tunes model parameters.
- Additional analyses: Separate feature-level logistic regressions assess individual predictors, while regularized linear regression predicts enrollment duration among non-completions.The duration model uses 70% of non-completions for tuning and reports performance on the remaining 30%.
3. Results
Dropout was predicted with a strong signal from first-term transcript and demographic data, while individual predictors were limited and timing-of-dropout prediction was only marginally successful.
- 3.1. Predicting dropout: AUC ranged from 0.66 to 0.73 across models, with modest gains for regularized logistic regression.The authors characterize these results as naive baselines despite the strong signal in transcript data.
- 3.2. Correlates of Dropout: GPA in math, English, chemistry, and psychology courses ranked among the strongest individual predictors of attrition.Indicators of previous schooling, birth year, enrollment year, and first enrollment quarter were also strong predictors.
- 3.2. Correlates of Dropout: 27.6% of students in the 1998 entering class and 20.2% in the 2006 entering class experienced attrition.The declining trend across enrollment years was reflected in the predictive strength of birth year and enrollment year.
- 3.2. Correlates of Dropout: No single feature achieved predictive accuracy higher than 54%.The seven features with the highest isolated predictive performance are reported in Table 3.
- 3.3. Timing of Dropout: 5.03 quarters was the RMSE for predicting when all non-completers stopped pursuing their degree.After excluding the least accurate 5% and 10% of predictions, RMSE fell to 4.14 and 3.74 quarters; non-completers enrolled for 7.35 quarters on average.
4. Future Directions
Future work will extend the analysis across the full dataset and other universities, broaden transcript coverage, and improve prediction models and feature representations.
- Data and scope: The researchers plan to analyze the full dataset and other universities while addressing class-imbalance issues.They also intend to move beyond first-term transcript data to study attrition processes more comprehensively.
- Model development: Planned model improvements include feature engineering and convolutional or recurrent neural networks.These neural approaches are intended to reduce reliance on hand-engineered features.
5. Conclusions
The study finds promising results for predicting eventual student attrition in a large, heterogeneous dataset, with regularized logistic regression performing strongest. Academic performance, enrollment timing, and birth year were among the strongest predictors, while predicting time to dropout remained difficult.
- Over 32,500 students formed the balanced dataset used to predict eventual attrition.
- Regularized logistic regression provided the strongest predictions of eventual student attrition.
- GPA in math, English, chemistry, and psychology courses, along with enrollment year and birth year, were among the strongest individual predictors.
- An RMSE of about 5 quarters indicated marginal performance when predicting how many quarters students took before dropping out.
- Next steps include working with university administrators, improving predictive models, and potentially expanding the dataset to other institutions.