In this dissertation, several new statistical procedures in nonparametric and semiparametric models are proposed. The concerns of the research are efficiency, robustness and sparsity.

In Chapter 3, we propose complete composite quantile regression (CQR) procedures for estimating both the regression function and its derivatives in fully nonparametric regression models by using local smoothing techniques. The CQR estimator was recently proposed by Zou and Yuan (2008) for estimating the regression coefficients in the classical linear regression model. The asymptotic theory of the proposed estimator was established. We show that, compared with the classical local linear least squares estimator, the new method can significantly improve the estimation efficiency of the local linear least squares estimator for commonly used non-normal error distributions, and at the same time, the loss in efficiency is at most 8.01% in the worst case scenario.

In Chapter 4, we further consider semiparametric models. The complexity of semiparametric models poses new challenges to parametric inferences and model selection that frequently arise from real applications. We propose new robust inference procedures for the semiparametric varying-coefficient partially linear model. We first study a quantile regression estimate for the nonparametric varying-coefficient functions and the parametric regression coefficients. To improve efficiency, we further develop a composite quantile regression procedure for both parametric and nonparametric components. To achieve sparsity, we develop a variable selection procedure for this model to select significant variables. We study the sampling properties of the resulting quantile regression estimate and composite quantile regression estimate. With proper choices of penalty functions and regularization parameters, we show the proposed variable selection procedure possesses the oracle property in the terminology of Fan and Li (2001).

In Chapter 5, we propose a novel estimation procedure for varying coefficient models based on local ranks. By allowing the regression coefficients to change with certain covariates, the class of varying coefficient models offers a flexible semiparametric approach to modeling nonlinearity and interactions between covariates. Varying coefficient models are useful nonparametric regression models and have been well studied in the literature. However, the performance of existing procedures can be adversely influenced by outliers. The new procedure provides a highly efficient and robust alternative to the local linear least squares method and can be conveniently implemented using existing R software packages. We study the sample properties of the proposed procedure and establish the asymptotic normality of the resulting estimate. We also derive the asymptotic relative efficiency of the proposed local rank estimate to the local linear estimate for the varying coefficient model. The gain of the local rank regression estimate over the local linear regression estimate can be substantial. We further develop nonparametric inferences for the rank-based method. Monte Carlo simulations are conducted to access the finite sample performance of the proposed estimation procedure. The simulation results are promising and consistent with our theoretical findings.

All the proposed procedures are supported by intensive finite sample simulation studies and most are illustrated with real data examples.