| Sumario: | Neural networks are presented as a complementary methodological tool to common statistical methods for certain types of classification problems. Prior research has suggested that neural networks can classify cases more accurately than many commonly used statistical techniques in situations where a data set does not fully meet the required assumptions, there are missing data, or there is a large amount of variance in several of the variables. There is also evidence to suggest that neural networks are a useful interpretive tool in these situations. Several neural networks were applied to a classification problem involving a problematic data set. The analyses were compared to a series of logistic regressions. Initially the logistic regression models provided greater predictive accuracy than the neural network models. However, as the data sets became more problematic, the accuracy of the neural network models surpassed that of the logistic regression models. It was concluded that neural networks are most useful as classification tools when insight and explanation are of less interest than predictive accuracy in problematic data sets. Reprinted by permission of the publisher.
|
|---|