Robust and efficient multiple imputation of complex data sets

点击翻译按钮获取中文摘要

Missing data occur even in carefully conducted scientific surveys. However, valid inferences based on incompletely observed data sets are only possible if the missing data problem is handled properly. One increasingly accepted method supported by data base producers to compensate for missing data is the method of multiple imputation. Available model-based techniques of generating multiple imputations are restricted to fully parametric models, which, if misspecified may produce unnecessarily imprecise or even biased inferences. Furthermore, most of the available software is not designed to efficiently handle large complex clustered or panel data sets. In this project, multiple imputation procedures will be extended to enable efficient and robust imputation of complex data sets based on an approximate Bayesian and a Bayesian approach, thus allowing valid and more precise inferences. Guidelines for the use of the multiple imputation method, based on currently available software and on functions and modules to be developed (callable in R), will be published, particularly with regard to possible limitations discussed in the literature. The need of the extensions developed will be illustrated based on substantive applications and through analyses of real data sets. The imputation programs will be made available to the scientific community.