Identifizierung von Ausreissern in eindimensionalen gewichteten Umfragedaten

Outliers and influential observations are a frequent concern in all kind of statistics, data analysis and survey data. Especially, if data are asymmetrically distributed or heavy-tailed, outlier detection is not clear-cut. Even more, for periodic data and data with multiple subsets, in which the distributional characteristics of each data sample may differ, outlier detection is challenging as the method used may need to be adjusted each time. In this paper we examine various non-parametric outlier detection approaches for (size-) weighted growth rates from surveys and propose new respectively modified methods which can account better for non-normal data and particularly for altering levels of dispersion and asymmetry. As outlier detection (and treatment) involves in practice a lot of subjectivity, we pursue an approach in which as few as possible parameters need to be defined. We conduct a simulation study to compare these methods under various models.