NONRESPONSE - Nonresponse analysis in the German labor force using individual social security administration data
Nonresponse causes concern in all (but not only) social scientific surveys. Nonresponse denotes the problem if a unit sampled to take part in a survey fails for different reasons in participating in the survey.
Thus, sample sizes of successfully interviewed persons, households and so on are reduced, thereby increasing uncertanity in estimation based on the reduced sample compared to the full samples without nonresponse.
However, things are complicated by the fact that nonparticipants in the survey may differ from participants. Things become even worse if they differ systematically from participants in aspects relevant to the topic of the survey.
Thereby nonresponse not only increases uncertanity in estimates, even seriously biased estimates may result from systematically missing nonparticipants. Furthermore, nonparticipants may fail to participate in the survey for different reasons, introducing different biases.
Several different techniques for correction of nonresponse are discussed and implemented, e.g. different weighting and imputation techniques. All of these techniques make specific assumptions about the mechanism that leads to nonresponse (Missing completely at random, missing at random, ...). If these asssumptions don't hold nonresponse corrections fail and biases in estimates remain.
To check if assumptions hold and to correct for nonresponse one needs information on participants and nonparticipants. This information on nonparticipants is typically sparse, typically comes from the sampling frame. Whether this sparse information allows to check for differences between participants and nonparticipants and to correct for nonresponse isn't taken for granted.
This project however uses detailed individual data on respondents and nonrespondents from administrative data of the german social security system, namely the German Employment Agency (Bundesagentur für Arbeit). Thereby it's possible to analyse differences between participants and nonparticipants on many different variables, in order to identify characteristics of (non)respondents and to compare different correction techniques using complete information for participants and nonparticipants from administrative data.
Survey data cover three different data sets which all come from the Institute for Employment Research (IAB), Nuremberg:
- Panelsurvey WeLL (Further Training as a part of lifelong learning)
WeLL is a linked employer-employee dataset of about 100 firms providing further training and about 21000 employees within these firms to study their characteristics related to further training.
see the WeLL homepage for further details.
- Crosssectional Survey LSS2005 (Life Situation and Social Security 2005)
LSS is a sample of about 50000 persons aged 15 till 65 in households drawing social security benefits in December 2004 resp. January 2005.
see the LSS homepage for further details.
- and the recently started Panelsurvey PASS (Labour Market and Social Security)
the yearly conducted PASS Panel consists of two independent subsamples (Sample 1: households drawing social security benefits. Sample 2: general population, households with low household income oversampled) each with about 6000 households sampled with all persons aged 15 and older to be interviewed.
see the PASS homepage for further details.
For all of the survey data sets individual data can be linked to from administrative data from the German Employment Agency.
The project is in cooperation with the Research Data Center (FDZ) of the German Employment Agency (Bundesagentur für Arbeit, Nuremberg) at the Institute for Employment Research (IAB, Nuremberg)
Schnell,R./Gramlich,T./Mosthaf,A./Bender,S.: Using complete administration data for nonresponse analysis: The PASS survey of low-income households in Germany; in: Proceedings of Statistics Canada Symposium 2010. Social Statistics: The Interplay among Censuses, Surveys and Administrative Data (forthcoming).
Staff and Contact:
Dipl.-Sozw. Alexander Mosthaf, University of Duisburg-Essen, IAB Nürnberg. eMail
Dipl.-Ver.Wiss Tobias Gramlich, University of Duisburg-Essen. eMail
The project is part of the DFG Priority Programme 1292 on Survey Methodology.