Prediction and Classiication with Neural Network Models Prepared for Delivery at the American Political Science Association Annual Meeting

This paper compares neural network models with the standard logit and probit models, the most widely used choice/classi cation models in current empirical research, and explores the application of neural network models in analyzing political choice/classi cation problems. Political relationships are usually nonlinear and of unknown functional forms, and political data are likely noisy. The logit/probit models assume exact and in general linear functional forms for the utility/classi cation functions underlying the observed categorical data, and are sensitive to noise. Neural network models, on the other hand, can be nonlinear, relatively robust to data noise, and capable of approximating arbitrary functional forms under general conditions. The latter are therefore potentially better suited to typical political data than the former. I rst compare the models for this more likely case of nonlinear unknown generating function by Monte Carlo simulations, in which a "true" bench mark model is available and the noise level can be controlled to study its e ect on model performance. The simulation results show that the neural network models perform signi cantly better than the logit, and indistinguishable from the \true" model, at all noise levels. They also show that as the noise level increases, the performance of the network and the true models gets closer to that of the logit. The paper then compares the models on real political data used in several existing studies. Re-analysis of the real data also indicates better performance of the network models compared with the logit/probit models, although for some cases the di erences are not statistically signi cant, due to low variance in the dependent variables as well as data noise. Issues of inference is then discussed and sensitivity analysis is conducted on a network model that signi cantly outperforms the logit model in predicting the voting behavior of voters who favored Perot most in the 1992 election, revealing more convincing substantive relationships than the existing logit model suggests.