Prediction and classification with neural network models

This article compares neural network models to the legit and probit models, the most widely used choice models in current empirical research, and explores the application of neural network models to social science choice/classification problems. Social and political relationships are generally characterized by nonlinearity and complexity and are usually of unknown functional forms. The legit and probit models assume exact and in general, linear functional forms for the utility functions underlying the observed categorical darn. Neural network models, on the other hand, are capable of approximating arbitrary functional forms under general conditions and can handle rich patterns of nonlinearity in the utility functions. They are therefore potentially better suited to typical social science data than the legit and probit models, which are shown to be special cases of the neural network class. Simulation results show that the neural network models perform significantly better than the legit models and are indistinguishable from the "true" models.