Abstarct: Semiparametric models are often used in real data analysis for their flexibility and parsimony. Two well-known examples of semiparametric models are the partially linear additive model and the partially linear varying coefficient model. Statistical inference of these models is restricted to a condition that the parametric and nonparametric parts are known in advance. However, such prior information is usually unavailable, especially when the number of covariates is large. Therefore, it is of great interest to develop some efficient methods to distinguish parametric components from nonparametric ones. In this dissertation, we introduce a two-step procedure, in the context of ultra-high dimensional additive models, which aims to reduce the size of covariates vector and distinguish linear and nonlinear effects among nonzero components. Also, we propose a robust method for simultaneously variable selection and parametric component identification in varying coefficient models baxsed on modal regression, which is robust with respect to non-normal errors and outliers in the response. The performance of the two proposed methods is examined by simulation studies and real data analysis. Results of numerical studies demonstrate the superiority in comparison with the existing methods.