Multiple Imputation for Complex Data Sets

Abstract

Data analysis, common to all empirical sciences, often requires complete data sets, but real-world data collection will usually result in some values being not observed. Many methods of compensation with varying degrees of complexity have been proposed to perform statistical inference when the data set is incomplete, ranging from simple ad hoc methods to approaches with refined mathematical foundation. Given the variety of techniques, the question in practical research is which one to apply. This dissertation serves to expand on a previous proposal of an imputation method based on Generalized Additive Models for Location, Scale, and Shape. The first chapters of the current contribution will present the basic definitions required to understand the Multiple Imputation field. Then the work discusses the advances and modifications made to the initial work on GAMLSS imputation. A quick guide to a software package that was published to make available the results is also included. An extensive simulation study was designed and executed expanding the scope of the latest published results concerning GAMLSS imputation. The simulation study incorporates a comprehensive comparison of multiple imputation methods.