Data modeling using constrained categorical regression

We apply a sparsely-connected neural network to the problem of recognizing statistical regularities and patterns in a National Labor and Alcohol Survey database. The network architecture, called Constrained Categorical Regression (CCR), is designed to identify valid statistical inferences even in the presence of a mis-specified model and offers fast training with guaranteed convergence. Each weight within the network can be tested for statistical significance and the overall network is interpretable for meaning and validity.