The purpose of this study is primarily twofold. First, to demonstrate that social sciences and more specifically social gerontology might be an important new application area for neural networks research. Second, to propose several simple feature selection procedures and investigate their efficiency in improving the generalization performance of feedforward neural networks of the Multilayer Perceptron (MLP) type when they are applied to classification tasks, using a specific social gerontology mapping problem as a real world benchmark The suggested feature selection methods are based on statistical concepts and techniques and more specifically on the X² test of independence for qualitative random variables, principal component analysis and stepwise discriminant analysis. Both study?s objectives are novel and the associated analysis is conducted through using cross-validation methodology. In addition to the above stated objectives, the final major goal of this research effort is to compare the generalization performance of MLPs employing different feature selection techniques with that of conventional neural network models and statistical pattern recognition techniques in a multidimensional classification real world task. The inputs to the methods involved come from a relatively small database of retirees containing 206-item structured interviews based on a schedule suitable for aged people. It is attempted to classify mental health as well as mental disorder in the sample comprising repatriated-retired as well as non-repatriated retired subjects. This study clearly shows the feasibility of successfully employing efficient artificial neural network (ANN) models in interview based social sciences investigations, despite the very large dimensionality of the pattern space.