Abstract:
|
Missing data due to nonresponse is a recurring problem in survey statistics, and can lead to biased inferences. Long questionnaires are part of the problem, as they can induce both, unit nonresponse if the respondent has information on the expected completion time, and item nonresponse due to interview break-off or question skipping. Split questionnaire survey (SQS) designs provide a mean to reduce this uncontrolled missingness. A crucial aspect of the selected designs is the avoidance of identification problems, i.e. everything to be analyzed jointly remains jointly observed for a subset of the original sample. The reduced questionnaire decreases response burden, but the data is now partially missing-by-design. However, the missing data can be assumed to be missing completely at random, and Multiple Imputation can provide the basis for unbiased complete-data analysis. Based on the work of Raghunathan and Grizzle (1995), we introduce a very flexible approach to identify SQS designs which preserve the information of the data as good as possible. Our proposed method is based on genetic algorithms allowing for flexible constraints, such as the definition of inseparable items.
|