Abstract:

High dropout rates and poor academic performance among university students remain critical challenges for higher education institutions. Accurate prediction of student outcomes is essential for strategic resource allocation, improving institutional reputation, and enhancing educational interventions. This study compares the efficacy of Multinomial Logistic Regression (MLR) and Linear Discriminant Analysis (LDA) in predicting first-year academic performance, measured by three categories: Clear Pass (CP), Proceed and Repeat (PR), and Part-Time (PT).

Using institutional data from Copperbelt University (2015–2020), including Grade 12 subject scores and enrollment records across five schools i.e. School of Mathematics & Natural Sciences (SMNS), School of Mines & Mineral Sciences(SMMS) , School of Business (SB), School of Built Environment (SBE), and School of Engineering (SE), we evaluated both models’ classification accuracy, precision, and sensitivity via confusion matrices. Key findings indicate:

MLR achieved higher overall accuracy (82.5%) compared to LDA (71.5 %). MLR excelled in predicting CP (94.7% precision) and PR (86.3% precision), while LDA showed balanced performance across all classes, particularly for PT (56.3% recall).

Significant predictors included Mathematics, Geography/History, and Principles of Accounts, with school affiliation and gender also influencing outcomes.

These results demonstrate that predictive modeling can effectively identify at-risk students early, enabling targeted academic support. While MLR offers superior accuracy for high-performing students, LDA provides subtle insights for borderline cases. The study contributes to educational data mining by validating these methods in a Zambian context and recommends their integration into institutional early-warning systems.

Keywords: Academic performance prediction, Multinomial Logistic Regression, Linear Discriminant Analysis, Confusion matrix.