9730688 Ichimura Much recent work in econometrics has been directed towards developing non-parametric and semi-parametric estimation approaches. These methods improve on traditional methods by allowing researchers to better uncover relationships in the data without having to make as many a priori modeling assumptions. They can also be used to test for the validity of parametric restrictions. Yet despite the potential broad applicability of these new estimation methods in empirical economic research, they are not widely used. This project addresses the reasons these methods are not used by: (1) resolving arbitrariness left in many semi-parametric procedures; (2) deriving general approximation results for data-dependent smoothing and trimming methods; (3) developing methods and software that overcome difficulties in computation with large data sets and (4) demonstrating how semi-parametric methods are useful for empirical work. The arbitrariness left in many semi-parametric procedures, which takes the form of parameters left unspecified, is addressed by developing effective, data-driven methods for choosing unspecified parameters. A preliminary Monte Carlo study comparing alternative methods demonstrates the effectiveness of the new methods. New asymptotic approximation results, needed to justify estimators that use data-dependent smoothing and trimming methods, are developed. Computational problems are resolved by adapting highly effective approximation techniques for use in estimating semi-parametric models. Two classes of empirical problems are shown to be promising areas to apply semi-parametric models: regression comparisons of two or more groups and the binary choice model. ??
