Bayesian Survival Modelling of University Outcomes

Catalina A. Vallejos, Mark F. J. Steel

Abstract

This paper presents a Bayesian framework for modeling the duration of university registration and the resulting academic outcomes for undergraduate students at the Pontificia Universidad Católica de Chile (PUC). By defining survival time as the period until enrollment ends—whether through graduation or dropout—the study employs a competing risks model. To address statistical challenges like "separation" (structural zeros in outcomes) and model uncertainty, the authors utilize Bayesian inference with Cauchy priors and Bayesian Model Averaging (BMA). The methodology is applied to specific degree programs characterized by high dropout and late graduation rates, identifying key determinants of student success.

Introduction: The Challenge of Educational Survival

Predicting student outcomes is a fundamental task for university administration, yet traditional binary models often overlook the temporal dimension of student trajectories. Standard survival models, while accounting for time, frequently treat graduation as "censoring," effectively ignoring it as a terminal event. This paper argues that graduation and dropout are competing risks: the occurrence of one precludes the other.

Furthermore, educational datasets often exhibit "separation." For instance, graduation is impossible in the first few semesters. This structural zero causes Maximum Likelihood Estimation (MLE) to fail, as it attempts to assign infinite negative values to these impossible events.

Methodology: A Bayesian Approach

The authors propose a Discrete-Time Competing Risks Model using a Proportional Odds (PO) specification. To overcome the limitations of standard frequentist methods, they implement:

  1. Bayesian Inference for Separation: Weakly informative Cauchy priors are used to estimate low but finite hazards for periods where certain events (like early graduation) are impossible.
  2. Bayesian Model Averaging (BMA): Instead of selecting a single "best" model, BMA weights all possible covariate combinations by their posterior probability, formally addressing model uncertainty.
  3. Efficient Sampling: The model is fitted via MCMC using Polya-Gamma latent variables, enhancing the efficiency of sampling from complex logistic-style structures.

Data and Program Heterogeneity

The study analyzes data from 27,185 students at PUC Chile enrolled between 2000 and 2011. The dataset includes demographics, socioeconomic status, and admission factors. A key finding is the significant heterogeneity across different academic programs.

Figure 1: Proportion of students by academic outcome across various degree programs at PUC Chile.

As shown in Figure 1, the distribution of outcomes varies significantly between disciplines. Programs like Medicine show very high graduation rates, while science-heavy programs like Chemistry and Physics face higher rates of dropout and delayed graduation.

Key Insights and Results

The analysis reveals several critical predictors of student success:

  • The "Gap Year" Effect: Students who took a break between high school and university were found to graduate more slowly and faced a higher risk of expulsion.
  • Motivation and Preference: Students enrolled in their first-choice program were significantly more likely to graduate on time and showed lower rates of voluntary dropout.
  • Temporal Hazard Spikes: The risk of different outcomes is not constant over time.
Figure 4: Empirical hazard rate over semesters, showing a sharp peak in risk early in the academic career.

In programs like Chemistry, the risk of involuntary dropout (expulsion) peaks sharply at the end of the second semester, as illustrated by the early hazard spike in Figure 4. Voluntary dropout tends to spike at the end of academic years (even semesters), while graduation hazards only emerge after the nominal program duration.

Conclusion

By treating graduation and dropout as competing risks within a Bayesian framework, this research provides a more nuanced understanding of student trajectories. The use of BMA and Cauchy priors allows for robust estimation even in the presence of structural zeros and model uncertainty. These insights suggest that university interventions should be program-specific and timed to coincide with identified hazard peaks to effectively support student retention and timely graduation.

Summary updated

Abstract

The aim of this paper is to model the length of registration at university and its associated academic outcome for undergraduate students at the Pontificia Universidad Católica de Chile. Survival time is defined as the time until the end of the enrollment period, which can relate to different reasons - graduation or two types of dropout - that are driven by different processes. Hence, a competing risks model is employed for the analysis. The issue of separation of the outcomes (which precludes maximum likelihood estimation) is handled through the use of Bayesian inference with an appropriately chosen prior. We are interested in identifying important determinants of university outcomes and the associ- ated model uncertainty is formally addressed through Bayesian model averaging. The methodology introduced for modelling university outcomes is applied to three selected degree programmes, which are particularly affected by dropout and late graduation.
BibTeX
@article{vallejos2016bayesian,
  author = {Catalina A. Vallejos and Mark F. J. Steel},
  title = {{Bayesian Survival Modelling of University Outcomes}},
  journal = {Journal of the Royal Statistical Society Series A (Statistics in Society)},
  year = {2016},
  doi = {10.1111/rssa.12211},
}
Figure