Predicting Who Will Graduate

Predicting who will graduate from a university is a difficult challenge, especially for US public universities whose missions serve diverse populations under relaxed admission criteria. Building predictive models for entering freshmen poses many problems: some students receive financial aid, others do not; some enter with SAT scores, others with ACT scores; some students stop out and then return. And, with the advent of the modern data warehouse, a dizzying array of data exists, which might, or might not, help build predictive models. This doctoral study examines the work required to build four predictive models for entering freshmen: logistic regression, automatic cluster detection, neural network, and decision tree. Practical problems are addressed squarely: Cleaning institutional data, dealing with missing data, adjusting model parameters, recognizing model drift, grouping students into prediction bands, and evaluating disparate model types are just some of the practical solutions shared in this work.