Abstract

Longitudinal surveys face many challenges, including high non-response rates and increasing data collection costs, which threaten the quality and utility of the collected data. The planned research will focus on two strategies for overcoming these challenges: developing adjustment methods for potential biases from non-consent if survey data are linked with other data sources and measuring and accounting for nonresponse bias in longitudinal studies. All research will make extensive use of data from the National Educational Panel Survey (NEPS) with the aim of developing guidelines how to address these challenges for the NEPS.
Many surveys, including the NEPS, link their data to large-scale administrative databases in order to minimize data collection costs and enhance data utility. The major concern regarding the linkage is that linkage-consent which needs to be obtained before the linkage is selective and introduces bias in linked-data analyses. Our planned research addresses this issue by evaluating alternative bias correction methods. The first method is based on the idea of vertically partitioned data, which makes it possible to analyze variables from two different files without actually linking them; thus, avoiding the need for linkage consent. The second method builds on the idea of matching non-consenting survey units to statistically similar units in the administrative data. We propose an innovative matching procedure that enhances the pool of matching variables by imputing some of the missing information prior to matching. Furthermore, we will develop methods of softening the conditional independence assumption that is required in most statistical matching applications.
To address the issue of nonresponse, we develop methods for assessing and adjusting for unit nonresponse and propose imputation strategies for item nonresponse that specifically account for the multilevel longitudinal design of the NEPS. Using linked administrative data from a forerunner survey (the ALWA survey) to the NEPS adult cohort, we will study the negative impacts of panel attrition. Information from the administrative data is available in consecutive years even if the ALWA respondent refused to participate in the NEPS survey. We can use this information to identify potential factors of panel attrition and evaluate the extent to which panel attrition bias exists in the NEPS survey. Furthermore, we utilize the linked administrative data to enhance nonresponse bias adjustment procedures (e.g. weighting) and compare the proposed approach to current nonresponse adjustment methods. Finally, we address the issue of item nonresponse in longitudinal surveys with hierarchical data structures by developing new imputation strategies for panel data that account for multiple sources of clustering (e.g. repeated measurements, students within schools). We will also compare these methods with previously proposed strategies for imputing missing values in longitudinal contexts.