Publication type
Research Paper
Series Number
736
Series
DIW Discussion Papers
Authors
Publication date
June 1, 2007
Abstract:
Using data on annual individual labor income from three representative panel datasets (German SOEP, British BHPS, Australian HILDA) we investigate a) the selectivity of item non-response (INR) and b) the impact of imputation as a prominent post-survey means to cope with this type of measurement error on prototypical analyses (earnings inequality, mobility and wage regressions) in a cross-national setting. Given the considerable variation of INR across surveys as well as the varying degree of selectivity build into the missing process, there is substantive and methodological interest in an improved harmonization of (income) data production as well as of imputation strategies across surveys. All three panels make use of longitudinal information in their respective imputation procedures, however, there are marked differences in the implementation. Firstly, although the probability of INR is quantitatively similar across countries, our empirical investigation identifies cross-country differences with respect to the factors driving INR: survey-related aspects as well as indicators accounting for variability and complexity of labor income composition appear to be relevant. Secondly, longitudinal analyses yield a positive correlation of INR on labor income data over time and provide evidence of INR being a pre-dictor of subsequent unit-non-response, thus supporting the 'cooperation continuum' hy-pothesis in all three panels. Thirdly, applying various mobility indicators there is a robust picture about earnings mobility being significantly understated using information from completely observed cases only. Finally, regression results for wage equations based on observed ('complete case analysis') vs. all cases and controlling for imputation status, indicate that individuals with imputed incomes, ceteris paribus, earn significantly above average in SOEP and HILDA, while this relationship is negative using BHPS data. However, once applying the very same imputation procedure used for HILDA and SOEP, namely the 'row-and-column-imputation' approach suggested by Little & Su (1989), also to BHPS-data, this result is reversed, i.e., individuals in the BHPS whose income has been imputed earn above average as well. In our view, the reduction in cross-national variation resulting from sensitivity to the choice of imputation approaches underscores the importance of investing more in the improved cross-national harmonization of imputation techniques.
Subject
Link
- http://EconPapers.repec.org/RePEc:diw:diwwpp:dp736
Notes
Econpapers search
working paper
#512018