Predicting behavior with an artificial neural network: A comparison with linear models of *prediction
The purpose of this study was to empirically validate a set of artificial neural networks (ANN'S) as viable methods of predictive classification in a clinically relevant context. Classification performance of backpropagation ANN models was compared with discriminant analysis and logistic regression analysis. Because of their inherent non-linear properties, it was held that the ANN models could implement more complex functions in solving these prediction tasks, and their performance accuracy might therefore exceed the performance of the generalized linear statistical prediction methods mentioned above. Performance of all classification methods in the prediction of recent alcohol binging, marijuana use in the prior year, and lifetime use of speed, amphetamines, and/or “uppers” was compared and critically evaluated. The data in this study were ascertained from a national survey of substance use and related behaviors among American college students in 1993, by Wechsler, Davenport, Dowdall, Moeykens, and Castillo (1994). Predictor variables or inputs included demographics, current academic performance, use of other substances, and opinions about alcohol use and risk taking behaviors. The sample of the survey was divided by random assignment into a training set (n = 9,950), and two test sets of differing size (n = 510 and n = 7,042). Fifty-four ANN models were trained and tested using the three dependent variables, two test sets, three different hidden unit architectures, and three different random starting points. Incidental cluster analyses and examination of weight configurations were done with four ANN models, and 48 supplemental ANN models were run to investigate performance bias. Results indicated a pervasive negative classification bias in the ANN models predicting marijuana and speed usage. ANN models were highly sensitive to disparate ratios of student responses in predicted substance abuse behaviors. Statistical models showed less bias. ANN models exceeded discriminant analysis performance on only one prediction task, and did not exceed logistic regression analysis performance on any of the classification tasks. Clinical and research implications of the findings are discussed, as are the limitations of the present study. Suggestions for future related research are considered as well.
Maucieri, Lawrence P., "Predicting behavior with an artificial neural network: A comparison with linear models of *prediction" (2003). ETD Collection for Fordham University. AAI3098134.