Ergänzung fehlender Daten in Umfragen

In multivariate datasets missing values due to item nonresponse may occur on any or all variables. Since nonrespondents often differ systematically from respondents deletion of incomplete cases would lead to substantial bias as far as inference is intended to the population of all cases rather than the population of cases with no missing data. Therefore a variety of techniques to fill in missing data with plausible values have been developed and are offered to a broad audience in statistical software packages. An overview of the most common ad hoc and modern modelbased imputation techniques is given herein. Since these techniques rely at least implicitly on the assumption of the so-called ignorability of the missing-data mechanism a simulation study is performed to investigate the power of several imputation techniques especially when the data are missing not at random. Bivariate normal datasets are used to discuss possible biases of common estimates of means, variances and correlations based on the imputed datasets. Following the results of the simulation study the imputation techniques often seem to produce substantially biased estimates, when the missing-data mechanism is nonignorable. At least the modelbased techniques perform somewhat better than the ad hoc ones.