Fehlende Werte in klinischen Verlaufsstudien - der Umgang mit Studienabbrechern

Abstract

Missing values are ubiquitous in clinical research. Especially in case of a longitudinal study, the complete acquisition of all study relevant data is complex and can therefore hardly be realized. In the course of an analysis, the parameters that should be estimated then may be biased. This leads to a restriction and falsification of study results. There are different approaches to figure the problems of missing data. In the present manuscript, imputation methods for the handling of missing data in longitudinal studies are investigated. It is discussed which preconditions has to be fulfilled for their appropriate usage and on which statistical properties they are based on. Their performance is mutually compared by the conduction of a simulation study. A completely observed longitudinal data set is used to simulate different missing data szenarios artificially. The results suggest the application of the so called Markov Chain Monte Carlo approach as imputation method of choice to treat missing data. Only this method showed a good imputation quality in a sense of acceptable validity and precision even under large amounts of missing data. However, the topic of missing data in clinical trials should still be a part of research to improve the existing imputation strategies and methods.