ABSTRACT
Neural networks (Venables and Ripley, 2002) are a class of predictive models that complement the previous models studied. The main advantage of neural networks is their structural flexibility. They can capture an unlimited variety of functional relations between predictor and target variables simply by including more nodes, or neurons, in the model. Their main disadvantage is that a calibrated model, while useful for prediction, is difficult to interpret. The estimated coefficients in regression models and split points in decision trees provide useful information to the analyst even before any attempt at prediction is made. The internal structure of a calibrated neural network model, however, is very complex, and thus not amenable to interpretation. We will first briefly describe the biological inspiration for neural network models, which were initially developed in the field of artificial intelligence (AI), and then their interpretation as a highly flexible statistical model.
