Predicting College Dropouts by Combining Automatic Interaction Detector and Discriminant Analysis

摘要

In this thesis a two group discriminant analysis was performed on two large samples from the Air Force Academy to predict college success and failure. The efficiency of the model was estimated by the hit rate (i.e. the proportion of correctly classified subjects) and by a cross-validation process in which difference in hit rates (shrinkage) was calculated. A new procedure, MAIDDA, was developed which combines a modified automatic interaction detector (MAID) with two group discriminant analysis. The MAID procedure does not require the conversion of continuous variables to categorical variables. In addition, MAID is easily performed with existing statistical software packages. The unique contribution of MAID to the prediction process was estimated by differences in hit rates and shrinkage for the two group discriminant analysis and MAIDDA when applied to the same sample data. The results from the samples described above indicate a substantial improvement in prediction when MADDA is used. It is postulated that MAIDDA will provide prediction improvement for most samples where N > or = 1000, numerous predictors are available and interaction of predictor variables exists. Further tests of MAIDDA are needed to ascertain its full potential.