A Model-Based Approach to Predicting Graduate-Level Performance Using Indicators of Undergraduate-Level Performance

Abstract

The graduate admissions process is crucial for controlling the quality of higher education, yet, rules-ofthumb and domain-specific experiences often dominate evidence-based approaches. The goal of the present study is to dissect the predictive power of undergraduate performance indicators and their aggregates. We analyze 81 variables in 171 student records from a Bachelor’s and a Master’s program in Computer Science and employ state-of-the-art methods suitable for high-dimensional data-settings. We consider regression models in combination with variable selection and variable aggregation embedded in a double-layered cross-validation loop. Moreover, bootstrapping is employed to identify the importance of explanatory variables. Critically, the data is not confounded by an admission-induced selection bias, which allows us to obtain an unbiased estimate of the predictive value of undergraduatelevel indicators for subsequent performance at the graduate level. Our results show that undergraduatelevel performance can explain 54% of the variance in graduate-level performance. Significantly, we unexpectedly identified the third-year grade point average as the most significant explanatory variable, whose influence exceeds the one of grades earned in challenging first-year courses. Analyzing the structure of the undergraduate program shows that it primarily assesses a single set of student abilities. Finally, our results provide a methodological basis for deriving principled guidelines for admissions committees.