Parametric and non-parametric methods to enhance prediction performance in the presence of missing data

Abstract

Most missing data analysis techniques have focused on using model parameter estimation which depends on modern statistical data analysis methods such as maximum likelihood and multiple imputation. In fact, these modern methods are better than traditional methods (for example, complete data analysis and mean imputation approaches), and in many particular applications can give unbiased parametric estimation. Because these modern approaches depend on linear parametric regression, they do not give good results, especially if the data distribution has highly nonlinear behaviour. This paper explains parametric estimation in cases of missing data, including an overview of parametric estimation with missing data, and provides accessible descriptions of nonlinear parametric and nonparametric estimation with missing data. In particular, this paper focuses on the effect of model selection methods on nonlinear parametric and nonparametric estimation in the presence of missing data. We also present analysis of an example to illustrate the performance of the two methods.