Model-Based Methods

Abstract

Model-based methods are often used to impute missing values. If the model assumptions are satisfied, these types of methods are often superior to model-free methods. In this chapter, linear models are discussed, while the following chapters focus on nonlinear methods.

First, we introduce linear regression based on (classical) ordinary least squares (OLS) at a very basic level. OLS has some nice mathematical properties, but this type of method is strongly influenced by outliers. Robust methods give roughly the same results in the case of a multivariate normal distribution, but give reliable results when the data contain artifacts or/and outliers.

Therefore, after an introduction to common concepts and implementations based on mice, we focus on robust imputation available in the R package VIM, as it gives roughly the same results in the case of elliptically normally distributed data, but is better in practice when obvious or masked outliers are present.