Predictive models for undergraduate admissions decision support systems: A test and comparison of statistical and artificial intelligence techniques in predicting first-year persistenc

This study focuses on comparing several predictive techniques that could be used in an undergraduate admission decision support system. The goal is to investigate the relative predictive performance of seven statistical and artificial intelligence models and to explore the potential of each technique for increasing the accuracy of current admission classification techniques. The mean accuracies of several techniques are compared. The statistical techniques selected are ordinary least squares regression, logistic regression, discriminant analysis and recursive partitioning. The techniques related to artificial intelligence are neural networks, genetic algorithms and interactive dichotomization.;Data on the background characteristics of students were obtained from the Cooperative Interinstitutional Research Program (CIRP) survey. The sample consisted of 1868 students of which 308 re-enrolled from the second to the third semester. The sample was drawn from the 1988 freshman cohort at a large northeastern research university. Independent random subsamples are used for cross-validation purposes. The criterion variable, persistence, is defined as re-enrollment from the second to the third semester. Each sample consisted of 50% persisters and 50% non-persisters. Each technique was applied to the three samples and mean accuracy calculated. Tests of hypotheses about the equality of mean accuracy were then performed. Conclusions regarding the tests and comparison are presented.;It was found that neural networks display higher predictive accuracy than the other techniques used in this research. However, the lack of explanatory power of the neural network model limits its use in admission decisions support systems.