| Much research and practical effort has gone into the classification of various natural phenomena. Classification is, after all, a requisite step in the process of explaining the complex relationships among such phenomena. A complete understanding of the efficacy of each of the myriad statistical and non-statistical classification models across all possible problem domains does not currently exist.;The current research uses a live transportation planning data set from Dallas/Fort Worth and five simulated data sets to compare the classificatory performance of six classification models: logit, linear discriminant analysis, quadratic discriminant analysis, backpropagation neural network, modular neural network, and radial basis function neural network. The five simulated data sets are all two-group problems with two independent variables. The two groups range from both being bivariate normal and linearly separable to being contaminated such that they are neither linearly nor quadratically separable.;In addition, a new method for potentially enhancing the performance of the models is introduced. This approach entails using outputs from matched statistical and neural network model pairs in an iterative fashion to increase the classificatory performance of the combined model pair over the performance of the separate models.;The study also examines the impact of proportional mix of observations on the relative classificatory performance of the six models.;The results indicate that the included classification models can be improved upon absolutely, and often significantly, by incorporating the results of a complementary model in the model estimation process. The results also indicate that among the statistical and neural network models included, the quadratic discriminant analysis and modular neural network models, respectively, tend to be the most efficacious. This is true whether they are used alone or are paired with complementary models using the iterative procedure. Lastly, the neural network models appear to be less sensitive than the statistical models to changes in the proportional mix of observations in each group. |
