A Split Questionnaire Survey Design applied to German Media and Consumer Surveys

On the basis of real data sets it is shown that splitting a questionnaire surveyaccording to technical rather than qualitative criteria can reduce costs andrespondent burden remarkably. Household interview surveys about media andconsuming behavior are analyzed and splitted into components. Following the matrixsampling approach, respondents are asked only the varying subsets of thecomponents inducing missing data by design. These missing data are imputedafterwards to create a complete data set. In an iterative algorithm every variable withmissing values is regressed on all other variables which either are originally completeor contain actual imputations. The imputation procedure itself is based on the socalledpredictive mean matching. In this contribution the validity of split andimputation is discussed based on the preservation of empirical distributions, bivariateassociations, conditional associations and on regression inference. Finally, we findthat many empirical distributions of the complete data are reproduced well in theimputed data sets. Concerning these long media and consumer questionnaires welike to conclude that nearly the same inference can be achieved by means of such asplit design with reduced costs and minor respondent burden.