A comparative case study of neural network analysis and statistical discriminant function analysis for predicting law students passing the bar examination

The purpose of this study was to compare the predictive accuracy of the discriminant function analysis to the predictive accuracy of a neural network analysis using data from graduates of a law school in the Northwest.There were four research questions explored by this study. The first question explored if there was a significant difference between the discriminant function analysis method and the neural network analysis method in rate of accuracy in predicting graduates' success in passing a bar examination. Neural network analysis was more accurate than discriminant analysis. The difference between the two accuracy rates appears not to have occurred by chance.The second research question analyzed if there was a significant difference between the discriminate analysis and the neural network analysis method in rate of accuracy in predicting graduates' success in passing a bar examination. Neural network analysis was more accurate than discriminant analysis. The difference between the two accuracy rates appears not to have occurred by chance.The second research question analyzed if there was a significant difference between the discriminate analysis and the neural network analysis in the rate of accuracy of predicting graduates' success in passing different state bar examinations from three geographic locations. Neural network analysis was more accurate than discriminant analysis for the three groups however the test of significant difference indicated that the difference could have happened by chance.The third research question explored if there was a significant difference between the two methods in predicting minority graduates' success in passing a bar examination. Discriminant analysis was more accurate than neural network analysis in predicting minority graduates passing a bar, but the test for significant difference in the proportion revealed that the difference could have occurred by chance.The final research question investigated the assumptions of neural network analysis dealing with the concepts that: (1) longer training provided more accurate predictions (2) a representative test set trained a more accurate network (3) and a retrained network was consistent in its accuracy rate but was based on different connections. The results of this study support that longer training does not result in better predictions, a random training test set provides similar results as a representative training test set if the sample set is large, and a retrained network produces consistent results while based on different connections. (Abstract shortened by UMI.)...